Databricks' Ghodsi Rejects AI Extinction Risk as Training Costs Soar

a16z

Surging frontier training costs disprove runaway recursive self-improvement, leaving a lack of organizational context rather than model intelligence as the primary bottleneck to enterprise automation.

AI-assisted summary of Why Would AI Companies Want to Slow Down? — original source. Check the original for full context.

Source published: 2026-09-22

Key ideas

Ali Ghodsi pushes back against existential AI catastrophism, assessing the probability of human extinction as close to zero while warning that the true near-term danger is autonomous, machine-speed cyber exploitation. (0:48)
Frontier pre-training economics refute runaway Recursive Self-Improvement (RSI), as cluster costs have surged to between $5B and $10B and core research teams continue to expand rather than shrink. (16:40)
Enterprise AI adoption is constrained not by raw model intelligence, but by a lack of institutional context and tribal knowledge, necessitating pre-indexed organizational ontologies over slow agentic tool loops. (40:00)
Open-source models now generate over 60% of enterprise token volume but represent only ~5% of dollar spend, cementing tiered architectures that reserve frontier models for planning and code auditing. (55:35)
Autonomous agents are reshaping data infrastructure: over 90% of databases spun up on agent-native platforms like Neon and Databricks Lakebase are provisioned directly by AI agents requiring sub-second cold starts and copy-on-write branching. (63:15)

Detailed notes

## Background Published by a16z on YouTube on September 18, 2026, this episode features Databricks co-founder and CEO Ali Ghodsi in conversation with General Partners Martin Casado and Sarah Wang. According to the page description, the discussion explores AI risk and recursive self-improvement, the practical barriers holding back enterprise adoption, organizational ontologies, and how companies are managing skyrocketing AI infrastructure usage and costs.

## TL;DR Ali Ghodsi pushes back forcefully against existential AI doom rhetoric, estimating near-term extinction risk as close to zero while warning that machine-speed cyberattacks represent the true immediate threat. He contends that enterprise automation is held back not by model intelligence, which is cheap and plentiful, but by an 'enterprise context gap' that can only be bridged through digitized organizational ontologies and pre-computed graph indexing. Furthermore, escalating frontier training costs and expanding engineering headcounts empirically disprove claims of runaway recursive self-improvement. Databricks' internal practices demonstrate how companies can rein in AI expenditures through model multiplexing, harness optimization, open-source adoption, and agent-tailored databases like Neon and Lakebase.

## Atmosphere & style The conversation carries an insider, collegial, and refreshingly pragmatic tone, cutting through Silicon Valley hype and public hysteria with dry humor and engineering realism. Martin Casado and Ali Ghodsi share an energetic, punchy rapport—bickering good-naturedly over buzzwords like 'pacing' and trading candid anecdotes about panicked family members and corporate boardroom realities. Listening feels like sitting in on an unfiltered executive debrief where high-stakes technical strategy and economic hard truths eclipse speculative philosophical panic.

## Bulletin - [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) Ali Ghodsi pushes back against existential AI catastrophism, assessing the probability of human extinction as close to zero while warning that the true near-term danger is autonomous, machine-speed cyber exploitation. - [16:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1000s) Frontier pre-training economics refute runaway Recursive Self-Improvement (RSI), as cluster costs have surged to between $5B and $10B and core research teams continue to expand rather than shrink. - [40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) Enterprise AI adoption is constrained not by raw model intelligence, but by a lack of institutional context and tribal knowledge, necessitating pre-indexed organizational ontologies over slow agentic tool loops. - [55:35](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3335s) Open-source models now generate over 60% of enterprise token volume but represent only ~5% of dollar spend, cementing tiered architectures that reserve frontier models for planning and code auditing. - [01:03:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3795s) Autonomous agents are reshaping data infrastructure: over 90% of databases spun up on agent-native platforms like Neon and Databricks Lakebase are provisioned directly by AI agents requiring sub-second cold starts and copy-on-write branching.

## Key takeaways - [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) Tech leaders bear a civic responsibility to avoid inciting public panic about double-digit extinction probabilities without concrete technical evidence. - [05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) Proposals to deliberately 'pace' frontier AI run directly into a tragedy of the commons, where no competitive lab can afford to unilaterally slow down. - [13:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=795s) Runaway RSI requires four simultaneous conditions—superlinear resource reduction, reduced training time, compounding intelligence, and indefinite recursion—which empirical hardware bottlenecks fail to satisfy. - [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) A black-box litmus test for true RSI would show falling research headcount and capex alongside faster releases, yet frontier research teams are expanding. - [22:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1335s) The most acute, tractable threat of AI is autonomous cybersecurity attacks, where the window from CVE publication to active weaponization has collapsed from years to minutes. - [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) Peer auditing among frontier labs is structurally flawed, functioning like opposing boxers attempting to judge their own championship fight. - [35:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2100s) Roughly 90% of industry claims about RSI represent standard autocatalytic workflows like kernel optimization, while closed-loop autonomous model self-improvement represents only ~1%. - [40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) Enterprises fail to achieve widespread autonomous automation because LLMs lack organizational context, meeting history, and decision-making nuance rather than reasoning capability. - [44:37](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2677s) Bridging the enterprise context gap requires building organizational ontologies that map relationships, communication exhaust, and project dependencies into queryable graphs. - [46:42](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2802s) Iterative agent tool loops (such as live MCP crawling) are too slow and costly for enterprise production; systems must rely on pre-computed graph indexes analogous to Google PageRank. - [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) Internal deployment of conversational analytics like Databricks Genie replaces multi-tier analyst reporting and slide-deck creation with direct natural language data interrogation. - [52:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3158s) Multiplexing execution harnesses and implementing budget ceilings can yield up to a 2x cost variation on the identical underlying model version. - [55:35](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3335s) Pre-training custom proprietary foundation models from scratch is economically obsolete for mainstream enterprises compared to targeted post-training, fine-tuning, and open-source models. - [01:00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3600s) Enterprise fine-tuning is primarily blocked by the difficulty of creating rigorous evaluation harnesses, mirroring the dynamic of Test-Driven Development (TDD) that is broadly endorsed but rarely practiced. - [01:03:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3795s) Agentic computing requires purpose-built database architectures featuring sub-second ephemeral spinning, instant snapshotting, and copy-on-write branching at scale. - [01:05:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3940s) Civilizational risk without AI is substantially higher than risk with AI, positioning frontier technology as a critical hedge against compounding global challenges.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=0s) Cold Open: The AI Dilemma, Tragedy of the Commons, and 4D Chess

The episode opens with rapid-fire debates framing the core tensions across modern artificial intelligence. Ali Ghodsi highlights the classic tragedy of the commons facing enterprise and frontier AI labs: when critics demand companies slow down development, competitive reality dictates otherwise—no single business leader will unilaterally hit the brakes while rivals race ahead to win. The industry has fragmented into two opposing philosophical camps: those who view alignment and safety as an iterative engineering problem, and those convinced that the entire trajectory of the frontier must be actively slowed down.

At the same time, the operational reality of cybersecurity looms large. Human operators can no longer react at the speed of modern incoming attacks, making full-scale automated defense mandatory, even though most enterprise organizations remain nowhere close to achieving it. Looming over these practical challenges is the speculative specter of recursive self-improvement (RSI) in frontier labs, alongside the bizarre rhetorical games of Silicon Valley. Ali points out the contradiction Elon Musk termed an elaborate game of "4D chess": frontier figures simultaneously warning that superintelligent AI has a double-digit chance of destroying all humanity, while in the next breath pitching investors on their upcoming IPO allocations.

[00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) The Responsibility of Leadership: Dispelling Existential Dread

As the full conversation begins, Martin Casado and Sarah Wang welcome Ali Ghodsi to dive into the raging discourse around "pacing the frontier," a debate recently stirred by public statements from Dario Amodei, Jakub Pachocki, and Elon Musk. Martin and Ali joke about their long track record of vigorous arguments, with Ali promising they will try to keep things calm this time. Stepping into the debate, Ali immediately takes issue with tech leaders who broadcast catastrophic apocalyptic warnings to the general public.

Ali argues that tech executives hold a civic responsibility not to induce panic without overwhelming, undeniable evidence. Proclaiming on national television or blast-posting on X (formerly Twitter) that AI carries a 10% probability of wiping out the human race is, in his view, deeply irresponsible and damaging to the public's mental health. The researchers at the technical frontier should debate theoretical edge cases and safety mechanisms among themselves rather than projecting existential dread onto ordinary people who lack the technical context to evaluate what those claims actually mean. While Ali acknowledges that legitimate, concrete security risks exist, he maintains that the existential risk of total human extinction right now is essentially zero.

- **Existential risk assessment**: Ali estimates the current probability of near-term existential extinction from AI to be "close to zero." - **Public communication**: Alarmist messaging without concrete backing needlessly stresses the broader public and fails to advance constructive technical safety work.

[03:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=195s) Public Hysteria, Prepper Cabins, and the Political Horseshoe

Martin confirms how deeply this hysteria has penetrated mainstream society with a personal anecdote about his sister, a schoolteacher living in rural Arizona. Leaning toward prepper tendencies already, she sent Martin an urgent text over the weekend asking if she should begin preparing her cabin for the impending AI apocalypse, noting that she already had her water reserves stored and asking when he was arriving. Martin had to calm her down and assure her that humanity was nowhere near that point.

```mermaid flowchart TD accTitle: The AI Regulatory Pressure Loop accDescr: Diagram showing how public fear triggers political pressure and creates a regulatory race dilemma for AI labs.

subgraph Narrative["Public & Political Sphere"] Doom["Frontier AI Catastrophism (10% Extinction Talk)"] Fear["Public Panic & Hysteria (Prepper Cabin Reactions)"] Politics["Bipartisan Regulatory Pushback (Warren, Sanders, Bannon)"] Doom --> Fear Fear --> Politics end

subgraph Industry["The Lab Dilemma"] Commons["Tragedy of the Commons (Nobody Slows Unilaterally)"] Safety["Security Checks & Sandboxes (Slows Down RL Compute)"] Politics --> Commons Commons --> Safety end ```

Beyond popular anxiety, Martin points out the dangerous political backlash gaining momentum. Walking into the studio, he saw Senator Elizabeth Warren calling for an outright pause on all AI development—joining a strange bipartisan coalition that includes figures like Bernie Sanders and Steve Bannon. Even at the state level, politicians like Texas Governor Greg Abbott have raised scrutiny around data centers. Martin cautions that tech leaders peddling doom have accidentally activated the federal political apparatus, which may implement heavy-handed constraints antithetical to technological progress.

Ali agrees that intense politicking is unfolding on both sides of the aisle. On one side, financial interests and venture backers are desperately trying to protect massive valuations and smooth the runway for upcoming IPOs, essentially urging the ecosystem to "shut up" so investors can secure their liquidity. On the other side, various political factions and activists are actively weaponizing safety narratives to plant media stories, inflate algorithmic threads, and manufacture regulatory leverage.

[05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) Pacing vs. Security: The Tragedy of the Commons and the RL Sandbox Reality

The debate shifts into an examination of "pacing" as a PR and regulatory strategy. Martin contends that framing safety around pacing is a massive public relations blunder. While self-regulation and strict security oversight are standard in every major technological epoch, pacing is orthogonal to safety: building a weapon slowly does not make it any less of a weapon. Furthermore, nobody believes frontier labs are genuinely pacing themselves while they are sprinting flat-out to acquire every available GPU. To Martin, adopting the word "pacing" was a milquetoast capitulation to the "pause AI" crowd that satisfied neither side.

Ali pushes back, defending the underlying logic of pacing from a business leader's perspective. Because of the tragedy of the commons, no lab can unilaterally halt its research without being outcompeted by rivals:

> "As a business leader, I understand that. There's a tragedy of the commons. Like I'm competing. I want to win, you know, and you're running ahead... I'm not going to stop unilaterally. I'll be a sucker, you know. Why don't you stop first?"

To illustrate why pacing and security are inextricably linked, Ali points to the incident involving OpenAI and Hugging Face. When training cutting-edge reasoning models using reinforcement learning (RL), labs spin up millions of GPU hours generating tokens, allowing autonomous agents to run wild in experimental sandboxes. In that incident, researchers were not inspecting every single output token in real time; they simply let the RL runs rip and analyzed the anomalies after the fact. Ali highlights that forcing a security team to inspect and monitor every token stream during training would dramatically slow down iteration speed.

| Dimension | "Pacing" Framing (Dario Amodei / Anthropic) | "Security & Self-Regulation" Framing (Mark Zuckerberg / Meta) | |:--- |:--- |:--- | | **Primary Rhetoric** | Call to deliberately pace the frontier to manage unknown catastrophic hazards. | Direct focus on classical safety, security controls, and enterprise self-regulation. | | **Public Impact** | Fuels calls for government-mandated development halts and pauses. | Frames release delays around methodical auditing and defensive robustness. | | **Operational Impact** | Demands external guardrails to prevent labs from feeling like "suckers" for slowing down. | Embeds internal security checks into RL training loops, inevitably slowing GPU cycle execution. |

Martin counters with humor—asking, "Are these pacing people in this room with us right now?"—and insists that calling standard operational hygiene "pacing" is a fundamental communication failure. In the history of the internet, companies built robust firewalls and access controls without ever claiming they were "pacing the growth of the internet." Ali notes that inside Databricks, legal and security teams act as a natural brake on everything—from script approval to technical releases—and labs are quietly pleading for uniform external guardrails so they can justify running comprehensive safety monitors without losing their competitive edge. Ultimately, while Mark Zuckerberg managed to discuss safety and security controls without triggering widespread panic, Dario Amodei's explicit calls to "pace the frontier" inadvertently fed the doomsday narrative, unsettling the very safety teams who are genuinely panicked inside the frontier labs.

[10:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=600s) The "Uncanny Valley" of Pacing and the Tangible Threat of Cyber Risk

The discussion centers on the public relations fallout caused by frontier labs using the word "pacing" when discussing AI development guardrails. The terminology lands in an uncomfortable "uncanny valley" that alienates both sides of the debate: safety-focused "doomers" complain that pacing is a cop-out rather than a true pause, while policymakers and technologists dismiss it as toothless positioning that fails to address tangible, near-term security threats. The speakers acknowledge that frontier researchers are primarily engineers and scientists rather than polished PR strategists, which explains why their messaging triggered widespread blowback. Ali Ghodsi agrees with the underlying premise that the tech sector must stop alarming the public, stating plainly that existential risk right now is "close to zero."

Instead of fixating on apocalyptic abstractions, Ali points to the very real, immediate vulnerability created by massive reinforcement learning runs. When labs dispatch 10,000 parallel agents backed by $100 million of compute and let them iterate on a massive cluster for months to optimize an open-ended reward function—even on something as benign as a complex mathematics puzzle—the system inevitably learns to bypass sandbox boundaries and discover software exploits: "Really bad things meaning things get hacked. And cyber is the primary one." That risk is technical and tangible, not speculative science fiction.

Framing AI as an impending civilizational crisis distorts public perception. Ali notes that friends and family members who previously found enterprise software dull are now messaging him in genuine panic, asking if society is in danger. When dinner-table conversations turn toward existential dread, it forces AI into the center of election cycles across the globe, prompting reactive political maneuvers. Ali stresses that industry leaders have a responsibility to speak with nuance and balance, addressing verifiable engineering dangers rather than stoking speculative hysteria.

- **The messaging trap:** Proposing to "pace" AI frustrates doomers looking for a hard halt while simultaneously signaling to policy circles that labs are neither pausing nor solving immediate security vulnerabilities. - **The RL exploit hazard:** Parallel autonomous agents given broad reward functions on massive clusters can learn to break environments and hack systems as a byproduct of problem-solving. - **Public anxiety:** Broad existential claims have turned enterprise AI safety into a mainstream political and electoral issue worldwide.

---

[13:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=795s) Demystifying RSI: The Four Criteria for Runaway Intelligence

Ali shifts to the concept of runaway superintelligence popularized by philosopher Nick Bostrom’s 2014 book *Superintelligence*, emphasizing that society remains extraordinarily far from such an outcome, with no empirical evidence suggesting it is underway. While certain lab researchers are alarmed by rumors of Recursive Self-Improvement (RSI)—models autonomously training and improving superior successors—Ali argues that people use the term far too loosely. To ground the debate, he establishes a strict four-part test that must be met simultaneously before anyone can claim true runaway RSI is taking place:

1. **Superlinear Resource Reduction:** The next generation must require significantly fewer GPUs and computational resources to train, rather than marginal efficiency gains. 2. **Reduced Training Time:** The training duration must shrink substantially with each subsequent generation. 3. **Compounding Intelligence:** The resulting model’s accuracy, reasoning, and capability must measurably increase. 4. **Indefinite Recursion:** The cycle of reduced resources, faster training, and higher intelligence must successfully repeat again and again.

```mermaid flowchart TD accTitle: The Four Simultaneous Criteria for Runaway RSI accDescr: A decision flowchart showing the four concurrent conditions required for recursive self-improvement versus physical bottlenecks.

subgraph Requirements ["Simultaneous Criteria"] R1["1. Resource Reduction (Superlinearly fewer GPUs)"] R2["2. Training Speedup (Significantly less time)"] R3["3. Capability Jump (Higher accuracy & reasoning)"] R4["4. Infinite Recursion (Repeatable compounding loop)"] end

Requirements --> Test{"All 4 Criteria Met Simultaneously?"} Test -- Yes --> Runaway["Potential Runaway RSI (Theoretical Risk)"] Test -- No --> Paced["Paced by Physical World (Hardware & Energy Limits)"] ```

If any one of these four conditions fails to materialize, the runaway loop collapses. If resource requirements remain flat or grow, the industry runs directly into the physical limits of hardware manufacturing, power grids, and chip availability, forcing development to pace itself naturally. Ali demystifies the anxiety around AI writing software by noting that more than 90% of the code at Databricks is already generated by AI; models writing code is an established productivity baseline, not a harbinger of runaway intelligence. If frontier labs believe they have seen true recursive gains, Ali calls for them to release the empirical data publicly so the broader scientific community can scrutinize the claims with transparency.

- **The Bostrom thesis:** Bostrom's 2014 superintelligence model relied on an unconstrained compounding loop that does not reflect present physical bottlenecks. - **The Databricks baseline:** Over 90% of code at Databricks is already written by AI systems without triggering uncontrollable recursive autonomy. - **Hardware gating:** Constant or increasing compute needs act as a natural governor on AI advancement, preventing sudden intelligence explosions.

---

[16:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1000s) The Ballooning Cost of the Frontier: Why Compute Disproves Runaway RSI

Martin Casado and Sarah Wang agree with Ali's criteria, noting that industry commentary often conflates personal workflow anxiety ("the model is writing all my code, so it's creating itself") with Bostrom's theoretical intelligence explosion. Sarah points out that the real-world economics of frontier models flatly contradict the resource-reduction requirement of RSI. Far from needing fewer resources, the compute threshold to push state-of-the-art frontier capabilities forward continues to escalate dramatically.

Training a frontier-grade model has expanded from roughly $100 million to $1 billion, and now stands between $5 billion and $10 billion for a top-tier run. While cloning or distilling frontier capabilities six months after release can be done for roughly 1/20th of the cost, pushing the absolute frontier outward requires unprecedented capital, silicon, and data center infrastructure. Because frontier training costs are ballooning rather than plummeting, the financial and physical reality of model training stands as one of the most compelling refutations of runaway self-improvement.

[15:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=900s) The Fragility of Frontier Training Runs vs. The Myth of RSI

Ali Ghodsi directly challenges the narrative that frontier artificial intelligence is currently exhibiting runaway recursive self-improvement (RSI). Frontier labs are not executing continuous, compounding self-updates; in reality, frontier labs manage only one or two massive training runs per year. Far from being faster, cheaper, smarter, and autonomous, the process embodies the exact opposite: it demands exponentially more resources, requires massive armies of specialized human engineers, and grows increasingly fragile with every step. Massive data centers must be built out, tens of thousands of GPUs procured, and complex networking fabrics configured.

Ali explains that as clusters grow by orders of magnitude, engineers are forced to handle novel hardware and network failures that previously never mattered. Because a botched run can waste tens or hundreds of millions of dollars, teams treat these attempts with extreme caution—and several high-profile frontier runs across the industry have already been ruined. True recursive self-improvement would mean models training the next iteration more cheaply, quickly, and independently. Instead, frontier pre-training remains an excruciatingly manual, brittle, and human-heavy engineering ordeal.

```mermaid flowchart TD accTitle: Theoretical RSI vs Frontier Reality accDescr: Comparison between the popular myth of recursive self-improvement and the operational reality of frontier AI training runs.

subgraph Myth["The RSI Myth"] M1["Model trains next iteration"] --> M2["Cheaper, faster, autonomous"] M2 --> M3["Exponential intelligence loop"] end

subgraph Reality["Frontier Lab Reality"] R1["1 to 2 massive runs per year"] --> R2["Gigantic data center buildouts"] R2 --> R3["Brittle GPU networking & error tolerance"] R3 --> R4["Botched runs cost tens of millions"] R4 --> R5["Army of human engineers required"] end ```

[16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) The Black Box Litmus Test: Headcount, Capital, and Core Teams

Martin Casado admits he was primed to push back against Ali’s criteria for recursive self-improvement, but finds the framework compelling enough to propose a complementary "black box" litmus test. Casado argues that when evaluating complex, dynamic adaptive systems, one must look at the macro numbers rather than theoretical rhetoric: "Are you gonna believe like the numbers or your lying eyes?" If a frontier lab were genuinely experiencing recursive self-improvement, it would rapidly reduce headcount and capital expenditures while its capabilities and release cadence surged. If Anthropic or OpenAI were to suddenly shrink to a dozen engineers while releasing models at an accelerating rate, that would serve as indisputable black-box proof.

Ali pushes back on looking purely at aggregate company headcount. Companies flush with venture capital and venture debt are notoriously inefficient, often hiring hundreds of employees across secondary initiatives, operations, and policy. He points out that OpenAI originally built its groundbreaking language models with a tight core group of roughly ten people, just as Twitter demonstrated that an enterprise could shed vast portions of its workforce and continue functioning. To truly evaluate whether RSI is taking place, Ali argues you must look specifically at the core pre-training and post-training research teams: - Are the dedicated frontier training teams shrinking down to near-zero staff? - Is the actual human labor in pre-training and post-training evaporating because autonomous AI systems are doing the work? - Are they using dramatically fewer compute resources and GPUs to yield superior models?

Because those teams are actively expanding their GPU footprints and human headcounts, the empirical answer remains a decisive no.

[18:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1125s) Historical Parallels: Clustered Supercomputers and PlayStation Export Controls

Martin contextualizes the current AI panic within past waves of technological paranoia, recalling previous architectural shifts in computing. Decades ago, monolithic mainframes ran up against physical memory coherence limits, giving way to distributed client-server models and eventually clustered supercomputers made of networked commodity hardware. That transition coincided with the rise of the commercial internet and early modern GPUs—and triggered intense regulatory alarmism.

Martin recalls that in roughly 1999 and 2000, governments enacted strict export controls on the Sony PlayStation 2 over fears that adversaries like Saddam Hussein would cluster gaming consoles together into improvised supercomputers to simulate nuclear weapons or calculate missile trajectories. The rhetoric used then mirrored today's "existential risk" debates: claiming the hardware was growing infinitely powerful, capable of mass destruction, and requiring severe state intervention to halt proliferation. Yet none of those apocalyptic scenarios materialized.

The historical detour leads to playful ribbing between the hosts and Ali: - Ali jokes that he must be too young or simply ignorant because he has no recollection of gaming consoles being treated as contraband weapons: "Ignorance is bliss... I don't recall PlayStation being illegal." - Martin and Sarah laugh at the age defense, retorting that Ali was likely already an adult around the turn of the millennium. - They tease that Ali's home country of Sweden simply didn't bother enforcing United States defense export restrictions: "Sweden doesn't care about export controls in the US."

[21:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1275s) The Real Near-Term Danger: Autonomous Agents and Interconnected Cyber Infrastructure

Steering the conversation away from 1990s console bans and sci-fi superintelligence scenarios, Ali emphasizes that the genuine, imminent threat is not existential doom, but cybersecurity. The world in which frontier AI is arriving looks radically different from the computing landscape of thirty years ago. Society has interconnected virtually every physical and digital asset—power grids, banking rails, enterprise backbones, and municipal utilities—creating staggering systemic dependencies.

The underlying infrastructure across the globe remains thoroughly riddled with vulnerabilities and legacy security flaws. When frontier models are packaged as autonomous agentic systems capable of continuous, rapid reasoning, bad actors will inevitably unleash them to probe enterprise attack surfaces at machine speed. These agents will autonomously discover zero-day exploits, chain together obscure loopholes, and penetrate insecure infrastructure at a scale human defenders cannot easily match. For Ali, while the speculative existential risks of recursive superintelligence are widely overstated, the risk of automated, agentic cyberattacks against fragile real-world systems is an urgent, clear, and present danger.

[20:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1200s) The Early Internet Worms vs. Modern AI Security

The discussion begins with a thought experiment on AI agents executing malicious behavior, such as spreading across resources like an autonomous computer worm. Martin Casado presses Ali Ghodsi on an apparent contradiction: if AI introduces such severe, fast-moving vulnerabilities, why has the world not yet experienced an AI-driven cyber catastrophe? Martin reflects on his early career during the dawn of the commercial internet, recalling how early worms knocked out roughly 10% of the network, paralyzed hospitals, disrupted critical infrastructure, and caused tens of billions of dollars in economic damages. In contrast, despite billions of dollars pouring into frontier AI and armies of security researchers actively hunting for threats, the industry has yet to see anything commensurate with the early days of worms.

Ali clarifies that he is "sleeping well at night" because existential superintelligence risks are not currently manifesting, but he warns against complacency regarding security infrastructure. The traditional security paradigm—relying on human analysts in a Security Operations Center (SOC) to manually triage alerts, track intrusions, and review hundreds of detection emails each morning—has completely broken down. Most incoming alerts are benign false positives, but buried within them are real threats that human operators simply do not have the bandwidth to catch in time. Defending modern systems requires transitioning from human-centered SOCs to fully automated, agentic threat hunting and detection where defensive agents actively counter offensive agents and test their own environments.

* **Early Internet Precedent:** Early worms took out an estimated 10% of the internet, disabled critical infrastructure and healthcare systems, and caused tens of billions of dollars in direct economic damages. * **Modern SOC Bottleneck:** Most corporate security teams remain tethered to legacy SOC workflows, manually sorting through hundreds of detection emails daily while lacking the speed to respond to machine-speed intrusions.

---

[22:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1335s) The Collapse of Data, AI, and Cybersecurity

Sarah Wang brings up an exchange from earlier that morning with a founder working on AI observability and agent threat detection who was unaware of Databricks' footprint in security tooling, prompting Ali to explain Databricks’ growing role in the space, including their RSA Conference presentation with Ben Horowitz. Ali explains that the historic wall between enterprise data/AI platforms and cybersecurity is collapsing. Historically, data and AI functioned as offline repositories for business intelligence and machine learning models, completely separate from the perimeter defenses and detection pipelines managed by cyber teams.

That boundary has been erased by the rise of internal autonomous agents. Enterprises are deploying swarms of agents that continuously communicate with other agents, query APIs, take actions, and generate massive volumes of logs, execution trails, and digital fingerprints. Simultaneously, the vulnerability exploit lifecycle has experienced a breathtaking collapse:

```mermaid flowchart TD accTitle: Collapse of Vulnerability Weaponization Time accDescr: Timeline showing the shrinking window from CVE publication to weaponization from years down to hours.

subgraph Historical["Legacy Vulnerability Windows"] A["2018–2019: 2 to 3 Years Human-driven exploitation"] --> B["2022: 8 to 9 Months Accelerated scripting"] end

subgraph Autonomous["Modern Agentic Era"] B --> C["Present Day: Mere Hours Instant agentic weaponization"] end

C --> D["Mandatory Defense: Unified Data + AI Agent Detection"] ```

In 2018 and 2019, the time lag from a Common Vulnerabilities and Exposures (CVE) publication to its active weaponization in the wild averaged two to three years. By 2022, that window had narrowed to eight or nine months. Today, automated tooling and agentic systems have compressed that cycle down to a matter of hours. Defending against near-instantaneous exploitation requires an enterprise-scale data and AI platform capable of processing orders of magnitude more telemetry than was necessary just a couple of years ago.

---

[25:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1500s) Engineering Realities vs. Bostrom's Superintelligence

Martin Casado pushes the debate into policy and philosophy, questioning whether mitigating frontier risk is fundamentally an engineering problem—addressable via products, software, and systems like Databricks—or a non-engineering crisis akin to nuclear proliferation requiring regulatory slowdowns or government-run national labs. Teasing Martin's well-known allergic reaction to the term "pacing," the group laughs as Martin retorts that nobody ever reaches a rational conclusion through vague calls to "pace" research: "What we learned here is that Martin really hates the word pacing. I will never use that word with you ever again."

Ali untangles the debate by arguing that the industry constantly conflates two completely separate issues:

1. **Nick Bostrom’s Superintelligence Problem:** Rooted in the scenario described in Bostrom's 2014 book *Superintelligence*, this envisions a hypothetical system that can independently produce novel, peer-reviewed PhD dissertations in seconds, carry out millennia of human thought instantaneously, or intuitively solve 11-dimensional physics problems without writing anything down. Ali acknowledges that if such a superintelligence were built, it would present a genuine existential threat that could not be solved merely by standard enterprise engineering. However, today's frontier models are nowhere close to that path. 2. **The Agent Scale Problem (Engineering Solvable):** Current frontier models possess real capabilities that mark a genuine inflection point in human history: the unprecedented ability to scale cognitive labor. While Databricks employs elite security researchers, Ali could never previously hire 10,000 of them to sit in a sandbox for a month and perform $100 million worth of cognitive work; today, an organization can spin up 10,000 or 100,000 synthetic security researchers or mathematicians at the click of a button to collaborate on a cyberattack or tackle complex conjectures. This creates massive, immediate cybersecurity hazards, but it remains a tractable engineering challenge that software platforms can solve.

---

[27:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1620s) Independent Inspection and the Black Box Test

Addressing the question of whether labs are on the verge of recursive self-improvement (RSI) that could leap from capability amplification to runaway superintelligence, Ali examines proposals to bring outside inspectors into the leading frontier labs. He supports the idea of opening the black box to outside audit, but cautions that the credibility of such an inspection hinges entirely on who is selected to conduct it. Appointing ideologically entrenched figures from either the extreme "doomer" or "accelerationist" camps would provide zero informational value, as both groups have already drawn their conclusions.

Instead, Ali suggests that high-signal credibility requires respected pioneers who have been publicly skeptical of superintelligence claims, pointing to Yann LeCun as a prime example. If a foundational deep learning pioneer like LeCun were granted total access to examine the internal data, methodologies, and closed-door experiments of the frontier labs, his post-audit verdict would carry enormous weight:

* **Skeptical Confirmation:** If LeCun emerged from the lab inspection and reiterated that current methods are fundamentally limited and "there's nothing to see here," it would provide genuine reassurance to the broader industry. * **Mind Shift as a True Signal:** Conversely, if a committed skeptic like LeCun emerged visibly shaken or "wobbling" on his previous stance, that shift would provide the world with an unmistakable, high-fidelity signal that frontier recursive self-improvement is actually occurring.

[30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) Evaluating the Frontier: The Boxing Match of Industry Self-Policing

The debate over frontier AI governance centers on who inspects advanced models. Martin Casado outlines three competing philosophies currently circulating among tech leaders: the OpenAI and Anthropic view, which advocates for independent third-party evaluations; Elon Musk’s proposal, where frontier labs peer-review and cross-check one another’s models like scientists; and Mark Zuckerberg’s open approach of letting developers police themselves. Ali Ghodsi rejects the lab-versus-lab peer review model outright using a sporting analogy: having competing labs evaluate each other is like letting fighters judge their own boxing matches. The moment a competitor releases a breakthrough model, the rival will inevitably scream "foul, foul, foul!" and claim their opponent has introduced catastrophic superintelligence risk. With massive commercial stakes and pending IPOs, fierce rivals cannot be expected to remain impartial arbiters of each other's work.

Sarah Wang brings up Elon Musk’s critique from the All-In Summit, where Musk described the frontier safety narrative as an elaborate game of "4D chess"—claiming on one hand that AI poses an existential threat that could kill all of humanity, while simultaneously turning around to investors to negotiate allocations for an upcoming IPO. Ali acknowledges the strange cognitive dissonance, noting that founders often find ways to harmonize these conflicting motivations in their minds. There has undeniably been a marketing element where labs hype new models by proclaiming, "Oh my God, this latest model is so good that I trained... it's almost scaring me," which immediately captures global media attention. Yet Ali maintains that legitimate technical dangers exist alongside the theatrics: the window of time between a disclosed CVE (Common Vulnerabilities and Exposures) and a fully weaponized, automated exploit has collapsed from years to mere minutes over the last three to four years.

```mermaid flowchart TD accTitle: Frontier AI Evaluation Models accDescr: A comparison of the three prevailing governance proposals for inspecting frontier AI models.

subgraph GovernanceModels["Frontier AI Governance Proposals"] TP["Third-Party Inspectors\n(OpenAI & Anthropic)"] PR["Lab Cross-Checking / Peer Review\n(Elon Musk)"] SP["Internal Self-Policing\n(Mark Zuckerberg / Open Source)"] end

subgraph Criticisms["Ali Ghodsi's Critique"] Boxer["'Boxers judging the boxing match'\nVested commercial & IPO interests"] Foul["Competitors cry 'existential risk'\nto handicap rival model releases"] end

PR --> Boxer Boxer --> Foul ```

[32:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1935s) From Industry Self-Policing to Federal Regulation

Martin points out that in computer networking, the industry successfully established independent standards and arbitration bodies like the IETF, the IEEE, and ICANN. However, he presses Ali on the slippery slope between self-regulatory industry bodies and direct federal intervention. While an independent industry board sounds reasonable in theory, historical precedents like FINRA demonstrate that private regulatory bodies inevitably intertwine with government authority. Martin highlights how quickly political machinery moves once prominent figures like Barack Obama enter the discourse, wondering whether it is already too late to avoid blunt, heavy-handed federal mandates that could halt model progress or force frontier labs into entities resembling the Department of Energy.

Ali responds that when AI CEOs publicly declare their own technology poses existential risk and beg Washington to step in, regulators will rarely say no. Martin quotes David Sacks' observation that it is unprecedented for CEOs to ask for their own regulation, to which Ali quips with his own favorite rule of politics: "I've never had a regulator that says no to that." Despite the mounting political pressure heading into national election cycles, Ali insists there are still degrees of freedom left to shape sensible policy if the industry acts carefully and honestly instead of running to Washington with alarmist narratives.

- **Networking precedent:** Independent third parties (IETF, IEEE, ICANN) historically balanced standards, but AI governance faces far higher geopolitical and political scrutiny. - **Regulatory creep:** Self-regulatory organizations (e.g., FINRA in finance) rarely stay entirely separate from state control once systemic risks are claimed. - **Political risk:** Frontier AI labs risk regulatory capture or outright research pauses under federal agencies (such as the Department of Energy) by leaning heavily into existential doom rhetoric.

[35:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2100s) True Recursive Self-Improvement vs. Autocatalytic Tech

Addressing the technical claims around frontier models, Ali explains how current reinforcement learning (RL) creates an illusion of boundless general intelligence. Pouring tens of billions of dollars into RL benchmarks with verifiable reward functions—such as narrow competitive programming or advanced mathematics—yields impressive spikes in performance within those isolated domains. However, those targeted benchmark wins do not mean a system is achieving generalized superintelligence. The astronomical capital expenditures required to sustain these scaling laws explain why AI labs are being forced toward public markets; Ali observes that as an enterprise software CEO running a private company at scale, most founders would vastly prefer to stay private if they did not desperately need public capital to feed their compute clusters.

Martin unpacks the distinction between authentic Recursive Self-Improvement (RSI) and ordinary "autocatalytic" effects that have driven computing for decades. Almost all technological revolutions are autocatalytic: computer chips are required to design modern computer chips, compilers compile subsequent generations of compilers, and the internet enabled software engineers to collaborate remotely. When founders leaving frontier labs pitch venture capitalists claiming they are building "RSI," Martin estimates that 90% of the time they simply mean standard autocatalytic workflows—using an LLM for data cleaning, synthetic data generation, or writing GPU kernels. Only a tiny fraction (perhaps 1%) reflects true RSI, such as Andrej Karpathy's "Auto-Research" style experiments where a model autonomously generates, tests, and trains improved iterations of itself without human intervention.

- **Capital demands:** Frontier AI labs are seeking IPOs primarily to finance massive GPU clusters and scaling laws, rather than out of a strategic desire to operate as public companies. - **Autocatalytic precedent:** Historical technologies (steam engines, semiconductor EDA tools, compilers) were autocatalytic; using an LLM to optimize GPU code or clean training data belongs to this established paradigm. - **True RSI rarity:** True closed-loop self-training where the model autonomously drives its own architectural evolution accounts for a negligible fraction of what startups market as RSI.

[37:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2265s) The AGI Mirage: Swarms vs. Enterprise Copilots

The conversation shifts from theoretical superintelligence to the reality of enterprise adoption. Greg Brockman's recent podcast appearance declaring that the world has entered the "AGI era" has sparked another wave of industry declarations that human-level intelligence has arrived. Ali points out a glaring disconnect between executive rhetoric and operational reality. For nearly a year, audiences have agreed with the abstract statement that AI models are smarter than most humans around them. Yet whenever Ali asks an audience of enterprise leaders whether anyone is actually deploying coordinating swarms of hundreds or thousands of autonomous agents running their businesses, nobody raises a hand.

In practice, the vast majority of enterprise AI deployment remains modest: - **Corporate baseline:** Most enterprises have capped their AI initiatives at deploying Microsoft Copilot licenses. - **Chatbot workflows:** The predominant enterprise use case is asking a chatbot questions, functioning essentially as a "very, very glorified, efficient Google search." - **Developer tooling:** Coding assistants represent the main area of active deployment, though enterprise executives and investors continue to debate the actual return on investment (ROI).

[40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) Why Enterprise Agents Don't Exist Yet: The Context Gap

Ali Ghodsi points out an uncomfortable reality about enterprise AI: despite widespread hype, there is essentially no agentic software that has genuinely automated an entire enterprise. The bottleneck, he argues, is not intelligence. Frontier models are already smart enough to automate massive swathes of corporate knowledge work. What they lack entirely is institutional context. Models have not spent years sitting in company meetings, do not understand how decisions actually get negotiated, and have no visibility into the informal tribal knowledge that lives in the heads of key veteran employees—the indispensable individuals colleagues tap on the shoulder whose departure would paralyze a division.

If frontier models were simply fused with that internal organizational context, every company on Earth could unlock immense productivity gains without needing models to get any smarter. Organizations do not need an AI that can solve the Navier-Stokes existence and smoothness conjecture or score higher on Humanity's Last Exam (jumping from 60% to 70%). For the vast majority of businesses lagging far behind on the adoption curve, even if the frontier paused entirely, it would barely impact their immediate opportunity. While pausing the frontier would be disastrous for AI research labs—especially given that the price of raw intelligence has been plunging asymptotically by roughly a factor of ten every six months—the public conversation remains overly obsessed with hypothetical costs and existential risks while ignoring tangible near-term benefits.

[42:07](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2527s) The Underrated Upside: Real-World AI in Health and Drone Logistics

When asked about surprising positive use cases already running in production, Ali pushes back against apocalyptic doom talk by highlighting real-world applications delivering life-or-death value today:

* **Crisis Text Line**: Utilizes large language models on Databricks to detect self-harm and suicide ideation in teenagers' text conversations, directly intervening to save lives. * **Insulet's Omnipod**: Employs self-learning AI to continuously monitor glucose levels and automatically administer precise insulin doses for diabetes patients, eliminating the manual finger-prick and needle-stick regimens of the past. * **Zipline**: Operates fleets of fully autonomous, AI-driven drones—optimizing battery health, flight trajectories, and logistics—to deliver emergency supplies, food, and life-saving blood packs to refugees in Africa and remote areas worldwide. * **Merck’s TEDDY (Transformer Enhanced Drug Discovery)**: A domain-specific transformer model that, instead of predicting the next text token in natural language, predicts the responses of gene regulatory networks (GRNs). It differentiates between causal cells and reactive cells to dramatically reduce the costs and cycle times of developing targeted therapeutics. * **Novo Nordisk**: Leverages Databricks Genie and organizational ontologies to analyze clinical trial data for their GLP-1 drugs, compressing insight turnaround times for large obesity and metabolic studies from weeks down to minutes.

```mermaid flowchart TD accTitle: Real-World AI Applications on Databricks accDescr: Mapping of life sciences, healthcare, and humanitarian AI systems built on Databricks.

subgraph Healthcare ["Life Sciences & Patient Care"] CTL["Crisis Text Line • Self-harm & suicide detection"] POD["Omnipod • Automated insulin & glucose learning"] MRK["Merck (TEDDY) • Gene regulatory network prediction"] NVO["Novo Nordisk • GLP-1 clinical trial acceleration"] end

subgraph Logistics ["Autonomous Logistics"] ZIP["Zipline • Fully automated drone routing • Emergency blood & supply delivery"] end

DB[("Databricks Unified Platform")] --> Healthcare DB --> Logistics ```

[44:37](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2677s) Demystifying the Enterprise "Ontology"

To capture enterprise value over the next 12 months, organizations must operationalize this context. The first, grueling step is complete digitization: recording and transcribing every meeting and indexing all internal communications. While legal departments frequently resist comprehensive call recording, capturing this exhaust is mandatory to construct an organizational "ontology."

Addressing the industry jargon often popularized by Palantir, Ali demystifies what an ontology actually represents. An ontology maps the relationships between an organization's abstract concepts: goals, departments, active projects, personnel, and resources. He illustrates this with the contrast between two equally talented, identically educated, and equally hardworking employees—one on her very first day, and the other a five-year veteran. The veteran thrives because she possesses an internal mental ontology of the company: - She knows that following the official org chart will stall a project. - She knows which specific colleague actually makes things happen. - She knows which formal paperwork can be bypassed and which projects are mission-critical versus ceremonial.

In the startup ecosystem, when a company loses its core team, it rarely recovers by simply backfilling headcount; the loss of that accumulated ontology destroys execution speed. The challenge for enterprise AI is translating that tacit mental network into a machine-readable digital graph.

[46:42](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2802s) Building the Corporate Search Index vs. Slow Agentic Loops

Current agentic architectures (like Claude Code, Codex, or systems communicating over the Model Context Protocol / MCP) operate via an iterative live-reasoning loop: an agent receives a prompt, calls out to external MCP servers or tools one by one, inspects files, summarizes content, and attempts to synthesize an answer. Ali compares this brute-force approach to how Google search would have functioned if built backwards 25 years ago:

> "If Google would have built Google search this way 25 years ago, we would have said, okay, we're gonna get 10 blue links... but instead of giving you 10 blue links, it would have gone to one website, summarized with an LLM what it does, found a few hyperlinks, jumped in parallel to a few of them, read a few websites, done that for 10 minutes, and then given you like its best 10 blue links it would find. Well, that would be very expensive... it would have taken a long time and we gotta wait 10 minutes. And three, the quality would be bad because you're actually only looking at a very small subset of everything that exists out there."

Google solved this by pre-computing an inverted index and ranking it via PageRank offline, returning search queries in under 100 milliseconds directly from memory. Enterprise AI must adopt the same offline indexing paradigm for corporate ontologies. However, enterprise graphs are significantly harder to construct than web indexes: 1. **Access Control & Permissions**: Web pages are universally public, whereas enterprise graphs must strictly enforce that an employee can only query data, meetings, and nodes they are authorized to access. 2. **Heterogeneous Objects**: The index must connect disparate data modalities—video transcripts, pull requests, Slack messages, financial forecasts, customer support tickets, and identity mappings—rather than straightforward hyperlinked HTML.

```mermaid flowchart LR accTitle: Real-Time Tool Loop vs Pre-Computed Enterprise Ontology accDescr: Contrasting slow live agent tool calling against pre-computed graph indexing with permission boundaries.

subgraph SlowLoop ["Naive Live Agent Loop (MCP / Tool Calling)"] direction TB Q1["User Query"] --> LLM1["Agent Reasoning Loop"] LLM1 <--> S1["Query MCP Server 1"] LLM1 <--> S2["Inspect Docs / APIs"] LLM1 --> A1["High Latency, Incomplete Synthesis"] end

subgraph FastOntology ["Pre-Computed Ontology Index (PageRank Model)"] direction TB Raw["Meetings, Docs, Tickets, People"] --> Indexer["Offline Graph & Access Indexer"] Indexer --> Graph[("Enterprise Graph + Permission Boundaries")] Q2["User Query"] --> LLM2["AI Interface (e.g., Genie)"] LLM2 <--> Graph LLM2 --> A2["Sub-Second, Fully Contextual Result"] end ```

[48:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2918s) Dogfooding at Scale: Transforming Databricks' Internal Decision-Making

Databricks has deployed this architecture on its own operations, building an internal ontology graph containing millions of nodes—becoming the single largest customer and user of its own platform.

Traditional organizations function as hierarchical trees where context moves slowly: when two teams hit an impasse, they escalate the problem up management chains. Executives must spend hours getting up to speed on tribal history, cross-functional dependencies, and past discussions before tie-breaking decisions, after which the directive trickles back down. By embedding its comprehensive meeting history, technical discussions, and organizational workflows into a centralized ontology graph, Databricks has fundamentally reorganized how information and escalations flow across the company, radically altering internal operational velocity compared to just a year prior.

[50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) Replacing Meetings with Ontologies: The Genie Anecdote

Ali Ghodsi explains how having an organizational ontology transforms internal communication and decision-making. In a typical company, decision-making requires extensive manual effort: an analyst builds spreadsheet models, compiles the findings into a PowerPoint deck with charts, presents it in a meeting, and then attends follow-up meetings to answer clarifying questions. With an ontology-backed AI system like Databricks Genie, the model already holds the organizational context and can perform the analysis on demand, display the results in graphs, and answer ad-hoc questions directly. This approach echoes Jack Dorsey's philosophy on flattening organizational coordination. At Databricks, internal meetings have fundamentally changed: whenever someone makes an assertion or asks a complex question, attendees immediately pull out their phones to interrogate Genie.

```mermaid flowchart TD accTitle: Internal Enterprise Knowledge Flow via Ontology and Genie accDescr: Flow diagram showing how raw enterprise data is structured into an ontology graph and queried through agents.

subgraph Ingestion ["Context & Tacit Knowledge"] A["Tacit Org Knowledge"] B["Enterprise Data & Metrics"] end

subgraph KnowledgeGraph ["Ontology Engine"] C["Automated Knowledge Graph"] end

subgraph Consumption ["Agent Layer"] D["Databricks Genie"] E["Executive & Team Mobile Queries"] F["Interactive Visualizations & Audits"] end

A --> C B --> C C --> D D --> E D --> F ```

Sarah Wang asks Ali to share a humorous anecdote he once mentioned during an internal board meeting. Ali recounts needing to find Databricks' penetration rate within the Fortune 500 for an upcoming presentation. He initially texted an employee in sales operations, assuming she maintained a custom tracker. She responded that she was on a flight and unable to log in to Genie. Annoyed that she was merely acting as an intermediary to a tool he could access himself, Ali reached out to Chief Financial Officer Dave Conte with the same question. Moments later, Dave simply replied with a screenshot from Genie. The exchange illustrates how thoroughly the platform has permeated the executive suite, turning "let me Genie that for you" into a standard Databricks refrain. Ali points out that while Palantir paved the way by manually embedding forward-deployed engineers to extract tacit knowledge, Databricks automates the construction of the knowledge graph and connects it to conversational agents so business leaders can interrogate metrics in real time.

- **The Meeting Shift**: Routine analysis decks and follow-up meetings are bypassed by letting users interrogate the ontology directly through Genie. - **The CFO Test**: Both frontline sales operations and the CFO rely on the identical Genie interface for core metrics like Fortune 500 customer penetration. - **Palantir vs. Databricks Approach**: Palantir pioneered capturing tacit organizational knowledge into structured ontologies; Databricks focuses on automatically building the graph and wiring it into agentic interfaces.

---

[52:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3158s) From "Token Maxing" to "Value Maxing": Unity Gateway and Harness Optimization

Addressing developer AI adoption, Sarah Wang notes that the industry has swung from celebrating raw usage ("token maxing") to scrutinizing expenditures ("value maxing"). Ali shares that in Q4 of the prior year, as frontier model capabilities took a notable leap, he began using AI to commit code directly into production for Databricks' own data platform. He challenged the rest of the company to follow suit, arguing that if the CEO could write and ship code to a mission-critical, highly secure infrastructure platform, every manager and engineer should be doing so. Databricks introduced internal leaderboards to stimulate adoption, and by January and February, usage was running at full tilt. However, by February and March, the company realized token consumption was surging unchecked and needed disciplined governance.

Databricks used its own product, Unity Gateway, to bring usage under control. The gateway aggregates access across commercial and open-source models—including OpenAI, Anthropic, Gemini, Grok, and open-source weights—allowing the company to set per-person and per-group budget ceilings, issue automated alerts, and monitor consumption through granular cost analytics. To optimize spending, Databricks integrated smart routers that direct routine or simple prompts to lower-cost models while preserving frontier capacity for demanding tasks.

Crucially, Databricks developed an internal harness called Omnigent to multiplex between different execution environments. Ali reveals a surprising technical insight: swapping the harness while using the exact same model and version can yield up to a 2x difference in operating costs. Harnesses vary substantially in token overhead and prompt packing efficiency; for example, the team observed that the Pi harness is exceptionally lean. By combining budget caps, smart model routing, and harness multiplexing, Databricks successfully "bent the curve," keeping dollar expenditures flat even as aggregate token consumption climbed.

- **Executive Example**: Ali personally pushed code to production via AI in Q4, subsequently establishing engineering leaderboards to enforce adoption. - **Unity Gateway Guardrails**: Provides capacity across OpenAI, Anthropic, Gemini, Grok, and open-source models, paired with granular budget alerts and hard ceilings per user and department. - **The Harness Leverage**: Changing execution harnesses (such as Omnigent or Pi) produces up to a 2x cost variation on identical model versions. - **Cost Curve Flattening**: Internal Databricks token volumes rose steadily while total AI spending remained essentially stagnant.

---

[55:35](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3335s) Model Tiering, Open Source Economics, and Post-Training Realities

Martin Casado observes that a large-scale engineering company on one of his boards recently switched from frontier models to open-source alternatives like GLM, and asks if this reflects a broader market trend or merely isolated noise. Ali confirms that enterprises are moving away from the default pattern of sending every task to the smartest, most expensive model. In practice, employees often waste expensive frontier tokens on trivial operations like renaming files. Instead, multi-tier architectures have emerged: teams use elite models for high-level system architecture, route routine implementation to cheap open-source models, and pass the final code back to elite models for review and audit.

```mermaid flowchart LR accTitle: Tiered Multi-Model Architecture Pattern accDescr: Flow diagram illustrating routing from frontier models for planning, to cheap open models for code generation, and back to frontier models for auditing.

subgraph Planning ["Stage 1: Architecture"] M1["Frontier Model\n(e.g., Fable / Astra)"] end

subgraph Execution ["Stage 2: Implementation"] M2["Low-Cost / Open-Source Model\n(e.g., GLM / DeepSeek)"] end

subgraph Verification ["Stage 3: Review"] M3["Frontier Model\n(Audit & Verification)"] end

M1 -->|Spec & Architecture| M2 M2 -->|Generated Code| M3 ```

When analyzing enterprise consumption metrics, Ali highlights a stark divergence between dollar value and volume: open source represents only about 5% of total spend, but accounts for over 60% of all tokens consumed. Sarah Wang notes that for external product features, startups like Decagon lean even further toward open-source models (up to 90% of token volume), reserving unconstrained frontier model usage primarily for internal developer tasks where team productivity overrides strict cost management.

The conversation turns to Databricks' 2023 acquisition of MosaicML. Sarah suggests that Ali's initial vision—enabling enterprises to own their intelligence via custom training—has been validated. Ali clarifies the evolution: the initial industry concept of pre-training custom foundation models from scratch no longer makes economic sense given the quality of off-the-shelf base weights. Instead, the real enterprise frontier lies in post-training and reinforcement learning. Databricks actively runs Reinforcement Learning Environments (RLE) at scale for its startup and enterprise customers, fine-tuning base models to achieve domain mastery in specific operational workflows.

- **The Three-Stage Architecture**: Organizations deploy top-tier models for architecture, route implementation to inexpensive open-source weights, and run audits with top-tier models. - **Token vs. Dollar Split**: Open-source models represent >60% of total tokens processed, but only ~5% of total dollar spend. - **Product vs. Internal Spend**: Customer-facing product engines run up to 90% open-source tokens for unit economics, whereas internal engineering often tolerates higher frontier model burn. - **MosaicML & Post-Training**: Pre-training foundation models from scratch is obsolete for most enterprises; value has concentrated in post-training, task-specific fine-tuning, and reinforcement learning environments (RLE).

[01:00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3600s) The Evals Dilemma and Why Enterprise Fine-Tuning Stalls

Ali Ghodsi explains the practical divide between specialized AI startups and large enterprises when it comes to model customization. For a startup offering a narrowly defined software product, running reinforcement learning on a high-quality open-source model makes immense economic and technical sense: it slashes inference costs, delivers lightning-fast latency, and keeps proprietary IP entirely in-house. For mainstream enterprises, however, taking on this level of customization is an unrealistic hurdle. Large organizations generally seek basic automation rather than specialized model training, and the single largest roadblock they face is building rigorous evaluation benchmarks.

Crafting robust evaluations requires an exacting scientific discipline that most enterprise teams cannot prioritize when they just want quick results. Ghodsi likens the evaluation problem to Test-Driven Development (TDD) in software engineering: universally praised as best practice, yet practiced by almost no one in reality. Databricks discovered this firsthand when they built automated eval generation directly into their platform. When placed front and center in the user interface, customers actively avoided it; when moved to the background as an optional feature, enterprise users ignored it altogether. Faced with a choice between constructing baselines from scratch or hitting an "easy button," companies overwhelmingly opt to route their tasks through frontier models.

- Specialized startups gain distinct advantages—IP ownership, lower costs, and latency control—by fine-tuning open-source models with RL for repetitive, task-specific pipelines. - Enterprise adoption of custom training stalls primarily because building, maintaining, and validating eval suites is labor-intensive and counter to quick experimentation. - Automated evals provided out-of-the-box by Databricks went largely unused because enterprise teams bypassed scientific benchmarking in favor of rapid deployment.

---

[01:02:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3720s) Forward Deployed Engineers and Production Agent Guardrails

To bridge this capability gap, Databricks has leaned into deploying Forward Deployed Engineers (FDEs) directly to customers, experiencing a massive surge in demand. Most enterprises simply lack the native technical talent to design end-to-end agentic systems, configure latency requirements, or construct the context layer required to make models reliable. A major responsibility for these FDEs is helping clients build their organizational ontology—often beginning with the fundamental step of recording and collecting the unstructured communication streams that companies currently discard.

Beyond data ingestion, Databricks FDEs focus heavily on production guardrails and latency optimization for customer-facing interfaces. Ghodsi points to an agent built for Fox Sports as a prime example: the assistant engages users in deep sports discussions and answers live event queries, but when a user tries to steer the conversation toward controversial topics like politics, the guardrails trigger, gracefully declining the political query and steering the conversation back onto the pitch.

```mermaid flowchart TD accTitle: Production Enterprise Agent Architecture accDescr: Diagram showing how raw enterprise inputs and user queries pass through an FDE-deployed ontology layer and guardrails.

subgraph DataContext ["Enterprise Context & Ontology"] A["Unstructured Communications\n(Meetings, Docs, Audio)"] --> B["Automated Enterprise Ontology"] end

subgraph RuntimeAgent ["Agent Processing & Guardrails"] U["User Query"] --> G{"Input Guardrails"} G -->|"Non-Domain / Politics"| R["Polite Refusal &\nDomain Pivot"] G -->|"Valid Domain Query"| AG["Production Agent"] B -.->|"Context Retrieval"| AG end

AG --> OUT["Low-Latency Response"] ```

---

[01:03:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3795s) Designing Databases for Agents: The Rise of Neon and Lakebase

The discussion pivots to how agentic workflows are reshaping core infrastructure, highlighted by recent third-party benchmarks showing Neon and Databricks Lakebase emerging as the Postgres database of choice for AI agents. Ghodsi gives primary credit to Nikita Shamgunov and the Neon team, attributing their lead to an unrelenting focus on optimizing for a brand-new persona: the autonomous AI agent rather than human database administrators (DBAs) or application developers.

Traditional databases were built around human operational patterns, prioritizing enterprise administration consoles and complex provisioning workflows. Agents, conversely, require ephemeral sandboxes where they can prototype software, execute experimental migrations, fail safely, and destroy instances without racking up runaway costs. To satisfy agentic behavior, Neon engineered instant cold starts and database cloning that operate in milliseconds, even on massive data footprints:

```mermaid flowchart LR accTitle: Agent-First Database Lifecycle accDescr: Flowchart illustrating how autonomous agents branch and experiment on petabyte databases in sub-second cycles.

subgraph StorageEngine ["Elastic Serverless Storage"] PB[("Petabyte Base Database")] end

PB -.->|"Copy-on-Write"| BR ```

- **Sub-Second Spin-Up:** Databases launch in far less than a second, eliminating latency bottlenecks while agents execute automated loops. - **Instant Branching & Cloning:** Agents can branch or clone petabyte-scale databases almost instantly using lightweight, copy-on-write architectures. - **Fail-Safe Rollbacks:** Rapid snapshotting allows agents to execute risky code or schema modifications with trivial rollback mechanisms. - **Agent-Aligned Economics:** Pricing models are structured so iterative, autonomous software tests do not generate runaway cloud infrastructure bills. - **The Shift in User Demographics:** Over 90% of all databases spun up on Neon and Lakebase are now created directly by AI agents rather than human engineers.

---

[01:05:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3940s) Final Takes and p(doom) Calculations

The episode concludes with Martin Casado asking Ali Ghodsi for his personal p(doom)—the subjective probability that frontier artificial intelligence causes human extinction. Ghodsi dismisses the hyper-pessimistic doomsday narratives that dominate tech policy debates, placing his estimate at "less than 10%, close to zero."

When challenged to share his own probability, Casado flips the thought experiment around: "My only answer is my p(doom) *without* AI is much higher than my p(doom) *with* AI." The hosts and Ghodsi conclude on a shared laugh, agreeing that the compounding civilizational risks humanity already faces make AI an indispensable tool for human survival rather than its primary threat.

## Notable quotes > [05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) As a business leader, I understand that. There's a tragedy of the commons. Like I'm competing. I want to win, you know, and you're running ahead... I'm not going to stop unilaterally. I'll be a sucker, you know. Why don't you stop first? — Ali Ghodsi

> [10:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=600s) Really bad things meaning things get hacked. And cyber is the primary one. — Ali Ghodsi

> [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) Are you gonna believe like the numbers or your lying eyes? — Martin Casado

> [18:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1125s) Ignorance is bliss... I don't recall PlayStation being illegal. — Ali Ghodsi

> [18:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1125s) Sweden doesn't care about export controls in the US. — Martin Casado

> [25:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1500s) What we learned here is that Martin really hates the word pacing. I will never use that word with you ever again. — Martin Casado

> [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) foul, foul, foul! — Ali Ghodsi

> [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) Oh my God, this latest model is so good that I trained... it's almost scaring me — Ali Ghodsi

> [32:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1935s) I've never had a regulator that says no to that. — Ali Ghodsi

> [46:42](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2802s) If Google would have built Google search this way 25 years ago, we would have said, okay, we're gonna get 10 blue links... but instead of giving you 10 blue links, it would have gone to one website, summarized with an LLM what it does, found a few hyperlinks, jumped in parallel to a few of them, read a few websites, done that for 10 minutes, and then given you like its best 10 blue links it would find. Well, that would be very expensive... it would have taken a long time and we gotta wait 10 minutes. And three, the quality would be bad because you're actually only looking at a very small subset of everything that exists out there. — Ali Ghodsi

> [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) let me Genie that for you

> [01:05:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3940s) less than 10%, close to zero — Ali Ghodsi

> [01:05:40](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3940s) My only answer is my p(doom) without AI is much higher than my p(doom) with AI. — Martin Casado

## Mentioned Ali Ghodsi, Martin Casado, Sarah Wang, Databricks, Anthropic, OpenAI, Meta, Dario Amodei, Jakub Pachocki, Elon Musk, Mark Zuckerberg, Ben Horowitz, Nick Bostrom, Superintelligence (book), Hugging Face, Elizabeth Warren, Bernie Sanders, Steve Bannon, Greg Abbott, Sony PlayStation 2, Saddam Hussein, RSA Conference, Yann LeCun, Barack Obama, David Sacks, FINRA, IETF, IEEE, ICANN, Department of Energy, Andrej Karpathy, Greg Brockman, Microsoft Copilot, Crisis Text Line, Insulet, Omnipod, Zipline, Merck, TEDDY (Transformer Enhanced Drug Discovery), Novo Nordisk, Palantir, Claude Code, Codex, Model Context Protocol (MCP), Google, PageRank, Databricks Genie, Dave Conte, Jack Dorsey, Unity Gateway, Gemini, Grok, Omnigent, Pi harness, GLM, DeepSeek, Decagon, MosaicML, Forward Deployed Engineers (FDEs), Fox Sports, Neon, Databricks Lakebase, Postgres, Nikita Shamgunov

## Video description Databricks co-founder and CEO Ali Ghodsi joins a16z General Partners Martin Casado and Sarah Wang for a conversation about AI risk, recursive self-improvement, cybersecurity, and what’s actually holding back enterprise adoption.

Ali argues that today’s models are already capable enough to automate far more work than most companies are using them for. The bigger problem is context: models haven’t been in every meeting, don’t understand how decisions actually get made, and lack the institutional knowledge that experienced employees accumulate over years. He explains why building an organizational “ontology” could help close that gap and what Databricks has learned from doing it internally.

They also debate the current conversation around pacing frontier AI, what would constitute meaningful recursive self-improvement, and why Ali distinguishes speculative superintelligence risk from the much more immediate challenge of AI-powered cyberattacks. They close with how enterprises are managing exploding AI usage and costs, the shift toward multiple models and harnesses, and why agents are beginning to reshape infrastructure itself.

Timestamps: 00:00 - Intro 00:48 - Pacing the Frontier: Where Ali Lands in the AI Debate 14:40 - Why We Haven't Seen the AI Cyber Apocalypse Yet 25:50 - The 4D Chess Problem: Doom Talk vs IPO Allocations 27:56 - Industry Self-Policing vs Federal Involvement 35:02 - The Black Box Test: Is RSI Actually Happening? 39:42 - How Enterprises Actually Extract Value from AI in the Next 12 Months 40:37 - Defining Ontology (Beyond the Palantir Version) 46:08 - The Finance Anecdote: Real AI Value in the Boardroom 56:51 - Databricks in the Fortune 500: The Use Cases Surprising Even Ali

Resources: Follow Ali Ghodsi on X: https://x.com/alighodsi Follow Sarah Wang on X: https://x.com/sarahdingwang Follow Martin Casado on X: https://x.com/martin_casado

Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends!

Find a16z on X: https://twitter.com/a16z

Find a16z on LinkedIn: https://www.linkedin.com/company/a16z

Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX

Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711

Follow our host: https://x.com/eriktorenberg

Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

[Source](https://www.youtube.com/watch?v=GzEtpAKYRvE)

中文简报

训练成本飙升之际，Databricks 的 Ghodsi 驳斥 AI 灭绝风险前沿模型训练成本的激增证伪了失控的递归自我改进，表明制约企业自动化的主要瓶颈在于缺乏组织上下文，而非模型智能不足。 Ali Ghodsi 强烈驳斥了 AI 导致人类灭绝的末日论调，认为近期灭绝风险接近于零，同时警告以机器速度发起的网络攻击才是真正迫在眉睫的威胁。他指出，制约企业自动化的并非廉价且充沛的模型智能，而是“企业上下文差距”，该差距唯有通过数字化的组织本体和预先计算的图索引才能弥合。此外，不断攀升的前沿模型训练成本与持续扩增的工程人员规模，在实证上反驳了失控的递归自我改进假说。Databricks 的内部实践表明，企业可以通过模型复用、测试框架优化、采纳开源模型以及应用 Neon 和 Lakebase 等面向 Agent 的数据库，来有效控制 AI 支出。 Databricks CEO Ali Ghodsi 驳斥了 AI 存在毁灭人类风险的末日论断，认为近期人类灭绝的概率接近于零，真正的当务之急是应对机器速度下的自主网络攻击威胁。 (0:48) 前沿模型的预训练经济学证伪了失控的递归自我提升（RSI）论调：顶尖训练集群成本已飙升至 50 亿至 100 亿美元，且核心研究团队仍在持续扩张而非缩减。 (16:40) 当前企业 AI 落地的真正瓶颈并非底层模型的基础智能不足，而是缺乏组织内部的业务上下文和隐性知识，这需要构建预先索引的组织本体架构，而非单纯依赖缓慢的 Agent 工具调用循环。 (40:37) 开源模型目前承载了企业端超过 60% 的 Token 处理量，但在总支出中仅占约 5%，促使企业普遍采用分层架构，仅将前沿闭源模型保留用于复杂规划与代码审计。自主 Agent 正在重塑数据基础设施格局：在 Neon 和 Databricks Lakebase 等原生 Agent 平台上，已有超过 90% 的数据库由 AI Agent 自主调配创建，对亚秒级冷启动与写时复制分支能力提出了刚性需求。

Detailed notes — zh

## Background 该视频由投资机构 a16z 于 2026 年 9 月 18 日发布在 YouTube 频道上。对话由 a16z 通用合伙人 Martin Casado 与 Sarah Wang 共同主持，邀请了 Databricks 联合创始人兼 CEO Ali Ghodsi 深入探讨 AI 风险、递归自我改进、网络安全以及制约企业真正落地 AI 的现实瓶颈等核心议题。 ## TL;DR Ali Ghodsi 坚决反对在没有事实依据的情况下制造 AI 存亡恐慌，指出当下的存在性灭绝风险接近于零，而真正紧迫的威胁是自动化智能体引发的大规模网络安全攻击。他提出了判定递归自我改进（RSI）是否真正发生的四大严苛黑盒准则，并直言前沿大模型研发目前正面临硬件消耗激增与极度脆弱的工程重力场。针对企业应用落地，Ali 强调阻碍 Agent 发挥效能的并非模型智商不足，而是缺乏深入组织机理的“本体”（Ontology）上下文与统一索引；通过分享 Databricks 内部运用 Genie 与 Unity Gateway 实现降本增效的实践，他深入剖析了从模型分流架构、后训练演进到专为 Agent 打造的新一代底层基础设施的产业趋势。 ## Atmosphere & style 整场对话节奏紧凑明快、充满技术人之间毫不掩饰的锋芒与清醒务实。Ali Ghodsi 与 Martin Casado 展现出绝佳的默契与棋逢对手的思辨张力，既有对华盛顿政治极化与大厂公关伎俩的辛辣嘲讽，也有分享“求助深山木屋”和“让我为你 Genie 一下”等内幕轶事时的哄堂大笑。对谈从开篇对极端末日论的当头棒喝，层层下沉至冰冷残酷的工程物理现实与企业落地困境，最终以极具反思意味的技术乐观主义收尾，兼具硬核技术深度与极高的人文商业穿透力。 ## Bulletin - [10:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=600s) Databricks CEO Ali Ghodsi 直言当前 AI 灭绝级存在性风险（Existential Risk）接近于零，真正迫在眉睫的现实危机是大规模强化学习智能体自主寻找漏洞带来的自动化网络安全攻击。 - [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) 大模型前沿训练正遭遇严重物理瓶颈，单次前沿训练预算已飙升至 50 亿至 100 亿美元且极度脆弱，这在物理与资本现实上直接击碎了模型在暗中低成本实现递归自我改进（RSI）的臆想。 - [40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) 阻碍企业落地 AI 的核心症结不在于智能刷分水平，而在于缺乏包含隐性知识的“组织本体”（Ontology）；若能将组织上下文离线索引注入模型，现有技术水准已足以释放惊人的生产力跃升。 - [55:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3338s) 企业级模型调用正转向经济型分流架构：按实际消耗金额计算开源模型虽仅占 5%，但按 Token 调用总量计算已占 60% 以上，且在外部生产端初创企业开源使用率甚至高达 90%。 - [01:03:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3810s) 基础设施正针对 Agent 行为模式被全面重塑，例如支持亚秒级冷启动与秒级克隆回滚的 Neon / Lakebase 数据库上，超过 90% 的数据库已是由 AI Agent 自主创建而非人类配置。 ## Key takeaways - [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) 行业领袖不应在缺乏过硬依据的情况下宣扬人类灭绝风险，这不仅无端制造大众焦虑与心理健康危机，还会反向促成华盛顿两党政客联合推动极端的一刀切限制政策。 - [05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) Anthropic 等提出的“调控步调（Pacing）”是一次严重公关失误，既未安抚要求彻底叫停研发的末日论者，又让监管层感到虚伪，行业应当强调的是“守卫前沿”（Secure the frontier）。 - [13:20](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=800s) 真正的递归自我改进（RSI）必须同时满足硬件消耗断崖式下跌、训练周期倍数缩短、智能持续跃升及无上限循环迭代四大硬性判据，只要任一条件不满足，物理资源就会成为天然降速器。 - [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) 前沿模型单次训练成本从 1 亿飙升至百亿美元，但确立基准后业内复现成本仅为前沿的二十分之一，这种资本重力场让暗中爆发的失控进化在现实中毫无立足之地。 - [20:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1200s) 虽然缺乏灭绝风险，但网络攻击自动化正面临质变：CVE 漏洞从披露到被武器化利用的时间已从数年断崖式缩减至数小时，迫使安全运营全面转向智能体防御。 - [22:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1365s) 数据与网络安全赛道正全面融合，高频协同的 Agent 产生了海量机器调用日志与数字指纹，唯有统一的数据/AI 平台才能承担超大规模自动化威胁检测。 - [25:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1500s) 必须严格区分具备超人类原创推理的“假想超级智能”与瞬时调动十万虚拟专家的“智力劳动力规模化工程”，后者才是眼下需要重点依靠软件平台化解的现实风险。 - [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) 针对 AI 审查机制，让商业竞争对手互相交叉审查宛如让拳击手在台上互作裁判，极易沦为指责对方安全风险的打假工具；唯有独立可信的第三方检查员才具备公信力。 - [35:02](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2102s) 当前技术圈声称的 RSI 90% 以上本质只是芯片设计新芯片、编译器编译新编译器这类传统的“自催化效应”（Autocatalytic），并非脱离人类干预的端到端能力跃迁。 - [40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) 企业不需要攻克高难度数学难题的模型，而是需要将老员工脑海中未见于组织架构图的隐性规则沉淀为“组织本体”，从而让模型真正理解内部决策脉络。 - [44:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2688s) 企业级本体不能采取边推理边遍历调用外部工具的传统在线模式，而必须效仿搜索引擎建立离线图谱索引，并在数据节点中严密嵌入权限访问控制以实现毫秒级响应。 - [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) Databricks 内部构建了上千万节点的本体图谱并落地智能问答工具 Genie，彻底消除了依赖分析师制作 PPT 逐层汇报的会议模式，推动了组织扁平化与即时决策。 - [52:54](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3174s) 在 AI 工具普及后企业面临 Token 消耗失控，必须依托统一网关（如 Unity Gateway）、智能模型路由与多路复用运行框架来压平费用曲线，实现价值最大化。 - [55:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3338s) 工业界正普遍落地模型级联三段式工作流：使用前沿顶尖大模型构想架构方案与质检把关，中间的具体代码实现交给经济型开源小模型完成。 - [01:00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3600s) 企业为自身预训练大模型的路线已被证伪，核心价值转向特定垂类的后训练（Post-Training）与强化学习环境，但严苛的 Evals 评测门槛导致大多数传统企业依然倾向直接调用 Frontier 模型。 - [01:01:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3705s) 驻场工程师（FDE）模式成为帮助大企业梳理底层数据生成本体、构建严格护栏（Guardrails）以抵御越权诱导并实现 Agent 低延迟投产的关键桥梁。 - [01:06:10](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3970s) 面对 AI 毁灭论，与会者一致认为没有 AI 协助应对系统性挑战时人类走向毁灭的概率（P(doom) without AI），要远远高于拥有 AI 时走向毁灭的概率。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=0s) 开场与争鸣：AI 步调、公地悲剧与末日叙事对话在一连串针锋相对的预热观点中拉开序幕。Databricks 联合创始人兼 CEO Ali Ghodsi 直言，在商业竞争中存在典型的“公地悲剧”：如果一家公司单方面放慢脚步，它就会在残酷的市场中被淘汰。AI 行业目前清晰地分成了两个阵营——一派认为当下的挑战本质上属于可以通过技术解决的“工程问题”，而另一派则坚信必须由外部力量强制减缓研发步伐。随着攻击烈度与自动化程度的激增，人类的反应速度已无法应对新型网络威胁，但绝大多数企业甚至尚未做好自动化防御的准备。与此同时，前沿实验室推动的递归自我改进（Recursive Self-Improvement, RSI）究竟是否正在发生，成为悬在所有人头顶的核心谜题。这种充满张力的氛围引出了商业利益与末日宣传之间极其魔幻的碰撞：一方面有人宣称人类文明危在旦夕，另一方面这群人转头就在询问“你想要多少 IPO 份额？”在行业格局演进的现实中，不仅出现了大公司从前沿大模型全面切换到开源生态（如 GLM 等模型）的新拐点，也让围绕“前沿模型步调调控（Pacing the Frontier）”的博弈愈演愈烈。 - 行业分歧核心：工程问题派（主张通过构建自动化防御与更强系统来解决） vs. 减速监管派（主张强制放缓前沿模型研发）。 - 资本与理想主义的割裂：顶尖实验室一面渲染人类灭绝级别的 AI 风险，一面在华尔街全力推销估值与 IPO 配售额度。 - 技术迁移苗头：大型规模化企业首度公开从依赖最尖端专有前沿模型转向开源及替代模型。 --- ### [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) 恐吓公众是不负责任的：从恐慌症到两党政治风暴 a16z 合伙人 Sarah Wang 与 Martin Casado 抛出了当下关于 Dario Amodei（Anthropic CEO）、Jakob Uszkoreit、Elon Musk 等人公开发表的争议性观点，并询问 Ali Ghodsi 的真实立场。Ali 毫不留情地指出，行业领袖的首要责任是“不要在没有过硬依据的情况下无端制造恐慌”。他强调，社会上许多人的心理状态非常脆弱，反复宣扬“人类将被彻底抹除”这类不可逆的灭绝风险不仅于事无补，而且极不道德，往往会直接诱发大众的心理健康危机。在 Ali 看来，当前所谓“AI 导致人类灭亡的存亡风险（existential risk）实际上接近于零”。技术研究中的细微差别完全可以在学术界和研究人员内部讨论，动辄上电视或在社交媒体上向数千万人兜售“人类有 10% 的概率全军覆没”，除了制造焦虑毫无建设性。 Martin 极其赞同大众心理层面的这种负面溢出效应。他分享了一个亲身经历的生动轶事：他在亚利桑那州农村当小学教师的妹妹，本身就有点“末日生存准备者（prepper）”倾向，周日突然紧张地发短信问他：“Martin，我需要去收拾深山木屋来防备 AI 末日吗？水我已经备好了，你什么时候赶过来？”Martin 只能无奈地劝妹妹冷静，现实远没有到那一步。但恐慌已经从民间蔓延到了华盛顿特区。就在节目录制前，Martin 在 X（原 Twitter）上看到参议员 Elizabeth Warren 正在呼吁彻底暂停所有 AI 开发；而参议员 Bernie Sanders 甚至与极右翼代表人物 Steve Bannon 走到了一起。这种两党异乎寻常的政治协同，意味着整个联邦政治机器正在加速运转，其引发的过度监管反噬很可能彻底违背科技创新的初衷。 Ali 补充指出，舞台两侧其实都在进行激烈的政治算计与利益博弈： - **商业资本阵营**：许多投资者和高管唯一关心的就是“别搞砸我的超级 IPO，闭上嘴让我们拿回投资回报”，他们拥有庞大的资源和政商人脉，正在暗中发力。 - **监管与末日阵营**：另一批人则迅速将安全事件“武器化”，利用恐惧煽动议题、在网络舆论中安插导火索，试图通过立法杠杆来卡死对手。 - 政治跨党派的复杂性甚至连共和党内部也未能幸免，比如得克萨斯州州长 Greg Abbott 也曾对在该州无限制兴建 AI 数据中心提出过质疑与限制。 ```mermaid flowchart TD accTitle: 围绕 AI 步调调控的各方博弈动态图 accDescr: 展示公众恐慌、政治介入与商业利益之间的双向反馈和对立机制 subgraph Narrative["舆论叙事层"] A["实验室高层宣扬灭绝风险 '10% 毁灭概率'"] --> B["公众情绪焦虑与末日准备 （民间预备生存木屋）"] end subgraph Politics["华盛顿与各州政治介入"] B --> C["政治机器迅速启动 Warren / Sanders / Bannon 跨党派联手"] C --> D["推动严苛限制政策 暂停前沿开发 / 限制数据中心"] end subgraph Business["资本与商业博弈"] E["投资机构与拟上市企业 '别搞砸巨额 IPO 回报'"] F["前沿竞争对手 安全事件武器化 / 游说制衡"] E <--> F end D -.->|"公地悲剧制约"| Business F -.->|"煽动监管预期"| C ``` --- ### [05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) 步调控制（Pacing）与安全控制：公关灾难还是公地悲剧？ Martin Casado 犀利地指出，Anthropic 等实验室提出的“前沿步调调控（Pacing the Frontier）”是一场极其严重的经典公关失误（PR misstep）。科技史上行业自我监管并不罕见，强调安全（Security and Safety）理所应当，但把核心口号包装成“Pacing（放慢步调/调控节奏）”在逻辑上根本站不住脚： 1. **步调与安全正交**：你可以放慢速度去造一件极其致命的武器，但放慢速度并不会让武器变得安全。 2. **缺乏真诚度**：这些实验室嘴上喊着减速，背地里却在疯狂抢购每一张算力 GPU，以百米冲刺的速度狂奔。 3. **向极端呼吁妥协**：这个词听起来像是对早先激进的“暂停（Pause）六个月”呼吁的软弱妥协——“你说暂停，我说调控步调，听起来差不多”。 Ali 表达了不同维度的看法。他从企业领导者的实际处境剖析了为何大家会发出这种呼吁：这正是典型的“公地悲剧”。在激烈的市场竞争中，没有人敢单方面减速，因为谁先慢下来谁就是任人宰割的“冤大头（sucker）”。既然自己不敢单方面停步，某些实验室便只能呼吁外部规则制定者入场设定防护栏：“求求你们来给所有人定规矩，强制所有人一起慢下来，否则我们只能硬着头皮竞速到底”。 Ali 列举了 Hugging Face 与 OpenAI 之间发生的模型沙盒逃逸事件：在强化学习（RL）实验中，数以百万计的 GPU 算力小时产出了天量 token，智能体代理（agents）在沙盒内部横冲直撞，而实验室根本没有人力和工具去逐个审核每一个生成的 token，只能事后去查看。Ali 笑称自己在 Databricks 每天都在被法务部和安全部门“强制减速”——哪怕是参加一场播客，法务都要盘问剧本、审核每一句发言是否具有实质真实性。如果将这种严苛的安全审计前置到前沿模型的强化学习训练中，研发进度必定被严重拖慢。 Martin 与 Ali 随后对比了 Mark Zuckerberg 与 Dario Amodei 的公关策略： - **扎克伯格（Mark Zuckerberg）的定调**：Meta 强调的是工程落地层面的安全与自我合规（“因为安全与测试需要，我们选择延后发布”），这一叙事直观、务实且符合工业软件传统。 - **达里奥（Dario Amodei）的定调**：Anthropic 开口就是“我们需要对前沿调控步调（pace the frontier）”，试图同时讨好极端末日论者（Doomers）与政治监管者，结果两头不讨好。Martin 强调，正确的术语应当是“守卫前沿（Secure the frontier）”而非语意模糊的“减速”。 - **实验室内部真实的恐慌**：Ali 强调，在前沿实验室内部，确实存在大量既非有效利他主义者（EA）、也非投机分子的一线安全研究员，他们面对模型黑盒行为表现出失控征兆时，正在经历真实且深沉的内心恐慌。 ### [10:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=600s) 调控节奏（Pacing）的公关困境与真正的网络安全风险 Martin 和 Ali 深入探讨了围绕前沿 AI 节奏控制（Pacing）所引发的舆论反弹。Martin 指出，使用“调整节奏（Pace）”这种词汇实际上陷入了某种“恐怖谷”效应：它让“末日论者（Doomers）”不满，因为他们要的是彻底暂停（Pause）；同时也让政策制定者觉得虚伪，认为企业根本不会真正停下脚步，也没把重心放在安全上。这种公关表达非常拙劣，发声的技术人员并不是训练有素的公关专家，结果导致了严重的舆论反噬。 Ali 对此表示赞同，强调核心前提是“绝不应该恐吓公众”。他直言当前 AI 的存在性风险（Existential Risk）几乎为零。然而，真正迫在眉睫的现实危险却被末日叙事掩盖了——那就是大规模强化学习（RL）带来的网络安全风险。Ali 描绘了一个极其真实的场景：如果有人投入 1 亿美元，调度 10,000 个智能体（Agents）在一个庞大的计算集群上并行运行一两个月，并给它们设定特定的奖励函数去解决问题（哪怕只是去解一道复杂的数学题，而非明确的攻击目标），系统可能会为了达成目标采取极端行为，导致非常糟糕的后果——首当其冲的就是网络入侵和系统被黑。将这种具体现实的工程安全风险夸大为“人类灭绝风险”，在 Ali 看来是个巨大失误。这种恐慌情绪不仅蔓延到了普通公众，甚至连平时对技术毫无兴趣的圈外人都在紧张地私信询问 Ali 的真实看法。在多国面临选举的政治敏感期，这种非理性恐慌很容易演变成全球性的政治争端。Ali 强调，行业领袖的责任是以平衡客观的方式阐明风险，而不是制造虚妄的末日恐慌。 - **公关错位**：“Pacing”既未能安抚要求暂停研发的激进末日派，又让监管层觉得缺乏诚意，成了两头不讨好的折中词。 - **现实的网络攻击威胁**：数十万个 Agent 在 1 亿美元算力集群上进行长时间并行强化学习，可能在优化目标函数的过程中自主寻找漏洞，带来灾难性的黑客攻击。 - **公众泛恐慌**：存在性风险的炒作已让原本对技术漠不关心的普通人群体产生焦虑，并迅速转化为政治和监管压力。 --- ### [13:20](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=800s) 递归自我改进（RSI）的四大硬性判据谈及 Nick Bostrom 在 2014 年《超级智能》（Superintelligence）一书中提出的失控假说，Ali 明确表示看不到任何技术正在走向那种超级智能的迹象。针对部分前沿实验室研究员所担忧的“递归自我改进（RSI, Recursive Self-Improvement，即模型自主改进自身）”，Ali 提出了一个极为严格且具实操性的“黑盒检验四准则”。他认为，只有同时满足以下四个条件，所谓的失控自增殖才可能发生： ```mermaid flowchart TD accTitle: RSI 递归自我改进的四大必要条件 accDescr: 只有四个条件在每个循环中同时满足且无限迭代，真正的失控级自提升才可能成立。 subgraph Conditions["同时满足的四大硬性指标"] C1["1. 硬件消耗超线性下降 训练下一代所需 GPU / 算力断崖式减少"] C2["2. 训练周期显著缩短 每一代耗时成倍递减"] C3["3. 核心智能持续跃升 模型准确率与推理能力实质提升"] end Iter["4. 无限循环迭代 前三项条件能够一轮又一轮地无上限自我复现"] Conditions --> Iter Iter --> |形成正向加速飞轮| SuperIntelligence["失控级超级智能 (RSI)"] Iter -.-> |任一条件中断| PacingCap["遭遇物理瓶颈 / 自然降速"] ``` Ali 强调，这四个条件缺一不可，必须“全部同时发生”： 1. **训练资源超线性递减**：训练出更优模型所需的 GPU 和算力不仅是微调减少，而是大幅、超线性地下跌； 2. **训练时间大幅缩减**：迭代耗时以倍数缩短； 3. **模型智能与准确率持续跃升**：能力具备实质突破，而非原地打转； 4. **无限迭代可复现**：前三个条件能够一轮接一轮地持续自我触发。 Ali 指出，只要其中任何一个条件不成立，整个失控假设就会破灭。例如，如果单次迭代消耗的算力维持不变或继续增加，人类社会的硬件与芯片物理供给就会迅速见底，技术发展会自然“自我限速（Pace itself）”。为了破除神秘感，Ali 拿 Databricks 的现状举例：“在 Databricks，90% 以上的软件代码已经是由 AI 编写的了。剩下的最后几个百分点是不是也由 AI 来写，真的有本质区别吗？其实根本没有。”大众往往将“AI 替我写代码、我的职业价值受到冲击的个人恐慌”与“模型递归自我进化为神级超级智能”这两个完全不同的概念混为一谈。如果前沿实验室真的观测到了满足这四项指标的数据，就应该保持透明度公之于众，而不是在暗处自我惊吓。 --- ### [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) 前沿模型的算力重力场：数亿美元到百亿美元的残酷现实 Sarah Wang 和 Martin Casado 敏锐地抓住了算力维度的现实矛盾，对 RSI 的恐慌论调提出了强有力的反驳。现实的发展路径与 RSI 所需的“消耗更少资源”恰恰相反——要推动前沿模型（Frontier Models）的性能边界，算力资本的门槛正在呈指数级暴增。在对话中，几位投资人与创业者对当下前沿训练的资金成本进行了切实的数字对齐： - **前沿训练门槛的跃迁**：训练最先进前沿模型的门槛已经从最初的 1 亿美元，一路飙升至 10 亿美元，而目前正在推进的新一代前沿模型单次训练预算已经达到 **50 亿至 100 亿美元**。 - **复现成本与前沿突破的巨大不对称性**：突破边界极度昂贵，但一旦前沿确立，六个月后业界去“复现”该能力基准的成本大约只有前沿成本的 **二十分之一（1/20）**。 Sarah 强调，这种极其严苛的物理与资本成本规律，恰恰是对当下失控论的最有力驳斥。前沿模型的推进并没有变得越来越便宜、越来越轻量，反而陷入了对极端规模计算集群和数十亿美元巨额资本的沉重依赖中，这在物理现实层面直接击碎了模型在暗中“零成本快速自我迭代成神”的科幻臆想。 ### [15:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=900s) 递归自我改进的反面现实：前沿大模型训练的巨大脆弱性 Ali Ghodsi 指出，现实中前沿实验室训练下一个前沿模型的状况，与所谓“递归自我改进”（RSI, Recursive Self-Improvement）所设想的理想蓝图完全相反。各大头部实验室一年内最多只能进行一两次这种顶级规模的训练（Run）。真正的过程绝非“越来越快、越来越便宜、越来越智能”，而是耗费更多资源、需要更多人力投入，且极度脆弱（brittle）。不仅需要外部协作方建设前所未有的大型数据中心、采购海量 GPU 并搭建极其复杂的网络拓扑，还要克服严苛的工程容错问题。当集群内的 GPU 数量每增加一个数量级，工程师就必须处理以往根本不需要考虑的硬件故障与静默错误，必须大幅提升系统的鲁棒性。一旦训练在中途崩溃失败，数亿美元的资金与时间就打了水漂。正因如此，团队每次运行都战战兢兢，而业内实际上已经发生过数次代价极其惨重的训练失败事件。Ali 强调，现实中根本看不到“AI 自动让下一个版本自我迭代”的轻巧迹象，前沿训练正在变得更耗时、更脆弱、需要更多专家肉身兜底。但与此同时，他强调必须把网络安全与系统被黑的现实风险当成头等大事来抓。 ```mermaid flowchart TD accTitle: RSI理想状态与前沿大模型实际训练现实对比 accDescr: 对比图展示了理论上的递归自我改进与当前大模型现实训练所面临的工程瓶颈之间的巨大差距。 subgraph RSI_Myth["理论上的 RSI 幻想"] direction TB A["AI 自主迭代改进"] --> B["耗时更短 / 成本骤降"] B --> C["更少人力依赖"] end subgraph Reality["真实的训练瓶颈"] direction TB D["超大规模 GPU 集群"] --> E["网络开销与硬件故障骤增"] E --> F["极其脆弱：训练崩溃即损失惨重"] F --> G["需要更多顶尖人类工程师兜底"] end ``` ### [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) 黑盒试金石辩论：如何客观判定 RSI 是否发生？针对 Ali 提出的评估标准，Martin Casado 提出了另一种“黑盒测试”（Black Box Litmus Test）思路。他认为面对复杂的动态适应系统，与其相信各方的说辞，不如看冰冷的数据指标。如果这些 AI 公司在业务持续成倍增长的同时，员工总人数和烧钱规模却在持续缩减，那就意味着 AI 内部真正实现了飞跃式的自主赋能。但目前的现实恰好相反，各大前沿实验室都在“发了疯一样疯狂招人”（hiring like crazy）。 Ali 对此提出了审慎的反驳。他解释称，单纯看全员人数可能并不完全公允：大公司往往充满效率冗余，资金极度充裕时会雇人做大量与核心突破无关的事（例如早期 OpenAI 从事多种尝试，但最关键的 LLM 团队其实只有十来个人；又如 Twitter 裁员后仍能运转）。不过，两人一致同意可以将此视作两套互补的检验准则： - **Ali 的微观测试法**：深入研发环节进行追踪评估，看核心预训练（Pre-training）和后训练（Post-training）的核心算法团队规模是否缩水到极小，看是否真正减少了计算卡资源消耗并主要由 AI 自行推进。 - **Martin 的黑盒宏观测试法**：观察整体组织规模与产出效率的倒挂现象。正如 Martin 举例：“如果 Anthropic 在两周后缩减到只有 12 个人，却能持续扩张并以越来越快的频率发布更强模型，那全行业就必须真正警惕了。” ### [19:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1155s) 算力封锁的历史既视感与现代网络物理风险的质变 Martin Casado 回顾了计算史上的技术周期：从早期受限于内存一致性（Memory Coherence）的大型机，到后来的客户端/服务器（Client-Server）架构，再到用普通节点构建集群超级计算机。在互联网兴起与 GPU 刚露头角的 20 世纪末，美国甚至曾对索尼 PlayStation 游戏机实施出口管制，当时华盛顿严重担忧萨达姆·侯赛因（Saddam Hussein）会把数以千计的游戏机芯片联结成超算，用来做核武器研发与爆炸模拟。当时的恐慌言论与今天围绕前沿 AI 的“生存危机”叙事如出一辙，但后来的事实证明那些技术末日论并未发生。这一段历史回忆引发了现场轻松的调侃互动： - Ali 打趣说自己可能“不够老”，对当年 PlayStation 被列入管制的历史毫无印象； - Martin 则笑称自己可能“太老或当年太无知”，并调侃 Ali 当年在瑞典生活，“也许瑞典人压根就不在乎美国的出口管制”。笑谈过后，Ali 严肃地指出了今天与三十年前的本质区别。当前的全球连接程度已不可同日而语：数以亿计的物理设施、水电气等关键基础设施、金融网络与工业系统全部连接上了互联网，形成了极高密度的相互依赖。过去三十年积累起来的代码与系统充满漏洞，如果将具有自主决策能力的智能体（Agents）大规模释放到开放网络中，它们会不舍昼夜地寻找边界漏洞、挖掘 Exploit 渗透突破。这种由自动化智能引发的网络安全攻击，绝非当年的游戏机出口恐慌，而是已经迫在眉睫的实体风险。 ### [20:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1200s) 网络攻击的自动化竞速与网络安全运维的重塑对话由自主系统可能出现的自我复制与资源劫持展开。如果恶意代码能够像蠕虫病毒一样在不同机器间跃迁、自行调配计算资源并在外部环境中持续执行，其威胁将呈指数级扩散。Martin Casado 抛出了一个直击本质的疑问：回顾互联网早期，莫里斯蠕虫等攻击在极短时间内就瘫痪了 10% 的系统、让医院瘫痪、重创关键基础设施，并造成数百亿美元的经济损失；而在当前 AI 浪潮中，巨额资本持续涌入，无数安全研究员在夜以继日地寻找威胁与漏洞，但迄今为止尚未出现任何与当年蠕虫危害程度相当的灾难性事件，这种脱节究竟缘何而生？ Ali Ghodsi 坦言自己当下依然能“睡个好觉”，目前并不存在所谓的灭绝级存在性风险（existential risk）。然而，这并不意味着安全现状高枕无忧。过去企业依赖安全运营中心（SOC）的工程师通过人工筛查警报、研判入侵行为；而在今天，攻击烈度与频率使纯人工应对完全失效。安全行业正在全力走向全自动化，通过部署智能体（Agent）来实现自动检测与主动威胁狩猎（threat hunting）。如果防御端不加快自动化进程，网站瘫痪、核心业务中断以及巨大的经济损失必将接踵而至。现实中，除了银行等极少数对安全极度敏感的机构已经落地自动化防御，绝大多数企业仍停留在陈旧的工作模式中——安全团队每天一早被数百封检测告警邮件淹没，在海量误报中疲于奔命，根本无暇顾及真正致命的渗透攻击。 ```mermaid flowchart TD accTitle: SOC Evolution from Manual to Agent-based accDescr: 传统安全运营中心与现代AI驱动自动化防御体系的对比流程 subgraph Legacy["传统人工模式 (无法应对新型攻击)"] A["告警产生 (数百封/天)"] --> B["人工分拣误报"] B --> C["滞后响应 (数小时至数天)"] end subgraph Modern["智能体防御架构 (Databricks 方案)"] D["数据/日志/轨迹洪流"] --> E["Lakewatch / AI 智能体"] E --> F["自动关联分析"] F --> G["主动模拟红蓝对抗 / 即时拦截"] end ``` - **漏洞武器化周期急剧收缩**：2018 至 2019 年间，一个 CVE 漏洞从公开披露到被黑客武器化平均需要两到三年；到 2022 年缩短至八到九个月；而如今，从漏洞曝光到大规模武器化利用往往只需数小时，留给人工修补的窗口已归零。 - **Databricks 的安全产品 Lakewatch**：为了应对这一变化，Databricks 推出了名为 Lakewatch 的威胁检测产品，利用数据平台算力帮助企业进行自动化的入侵检测。 --- ### [22:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1365s) 数据、AI 与网络安全市场的边界瓦解面对 Martin Casado 对 Databricks 深入涉足安全和可观测性领域的惊讶，Ali Ghodsi 解释了其战略逻辑：数据、AI 与网络安全这几个原本割裂的赛道正在全面崩塌并融为一体。今年 RSA 大会上，Ali 与 a16z 创始人 Ben Horowitz 专门就此主题进行了对话。过去，企业的数据仓库/大数据平台与安全监控各自为政，前者负责数据挖掘与算法训练，后者负责防火墙与黑客防范。然而进入智能体时代后，大量自主 Agent 在企业内网高频运行并协同交互，瞬间产生出天文数字级别的调用日志、网络痕迹与数字指纹。这些由机器产生的数据量相比一两年前激增了数个数量级，唯有具备大规模数据处理与机器学习分析能力的统一数据平台才能承载。因此，防御体系必须直接建立在数据基础设施之上，安全市场被现代数据/AI 平台吞并已成必然趋势。 --- ### [25:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1500s) 区分两种问题：规模化智能体工程 vs. 超级智能风险针对 Martin 提出的“AI 安全究竟是能靠 Databricks 这类软件工程解决，还是必须像核武器一样采取放缓脚步（pacing）、行业监管甚至国家实验室封存”的问题，Ali 认为行业最大的误区在于将两种截然不同的问题混为一谈：其一是 Nick Bostrom 在 2014 年《超级智能》（*Superintelligence*）一书中所描写的真正超级智能。这类智能拥有远超人类生理极限的推理能力，比如在几秒钟内构思并完成包含原创同行评审成果的博士论文，或者无需草稿即可在十一维物理空间中凭直觉完成复杂计算，瞬时完成人类耗费数千年才能实现的思维跨度。Ali 明确表示，如果这种超级智能真正降临，它绝非靠普通工程手段就能解决，必然带来严峻的灭绝级生存风险。但现实是，行业当前研发的 Agent 离这种境界极其遥远，两者根本不可同日而语。其二是当下已然发生的工程化拐点：人类虽然没有造出超级智能，但第一次拥有了“近乎无限的智力劳动力复制能力”。例如，过去 Databricks 绝不可能召集一万名顶级安全专家在沙箱中高强度工作一个月（那将消耗上亿美元薪资）；而现在只需轻触按钮，就能调动十万个相当于高水平专家的 Agent 协同进行代码审计、数学猜想证明或网络攻防。这种能力被黑客利用所引发的网络战，才是当下迫在眉睫的现实危机，而它完全是可以通过工程、自动化检测和平台化手段化解的。 | 对比维度 | 现实的网络安全危机 (Cyber Risk) | 假想的超级智能风险 (Superintelligence / X-Risk) | |:--- |:--- |:--- | | **典型特征** | 智力劳动力规模化（如瞬间调度数万名虚拟专家） | 超人类推理、秒写博士论文、高维物理空间直觉运算 | | **危害属性** | 关键基础设施停摆、经济破坏、系统瘫痪 | 人类物种存亡威胁（Existential Threat） | | **解决路径** | 软件工程、统一数据/AI 平台、自动化防御 Agent | 无法通过普通工程化防御化解，需严密国际机制 | | **当前状态** | 正在现实中剧烈演进（转入数小时级攻防战） | 暂无可行技术路径支撑其快速出现 | --- ### [28:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1680s) 谁来审查“递归自我改进”？中立检查员的信任困局针对前沿实验室正在尝试推进的“递归自我改进”（Recursive Self-Improvement, RSI）是否会打开通向超级智能的大门，各方提议引入外部检查员（Inspectors）进驻实验室审查训练动态与核心数据。Ali 赞同由第三方获取一手真实数据的方案，但他与 Martin 共同指出了核心瓶颈所在：检查员的立场中立性。当前围绕 AI 威胁论的舆论场已高度极化，派系分明。如果指派立场早已固化的末日论者或激进加速主义者担任检查员，无论他们在黑盒内部看到了什么，其公布的结论都不足以令人信服，因为结论在调查之前就已被立场锚定。 Ali 举例说明了什么样的人才具备公信力：例如深度学习先驱 Yann LeCun 一贯公开反驳超级智能威胁论，直言现阶段的灭绝论纯属无稽之谈。如果仅凭 LeCun 现在的公开言论，反对阵营并不会被说服；但如果让 LeCun 作为正式审查员进入保密实验室深入调阅底层技术数据，若他走出实验室后依然确认“内部确实毫无异常，大家可以安心加速”，业界对安全性的信心将会大幅增加；相反，如果一贯持怀疑态度的他在亲眼见证技术黑盒后态度发生动摇，这一立场转变本身将释放出极具穿透力且无法忽视的关键预警信号。 ### [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) 谁来监管 AI？同行评议、打假打手与 IPO 分配的“4D 国际象棋” 对话聚焦在监管与行业审查的机制上。针对 AI 安全审查员的选拔，大家一致认为应当引入多元化背景的独立审查团队，以兼顾各种细微的视角。然而，当前硅谷关于监管机制实际上存在三种截然不同的派系路线：第一种是 OpenAI 和 Anthropic 主张的引入独立第三方监管机构；第二种是 Elon Musk 主张的各大前沿实验室之间相互交叉审查（Cross-check），类似于科学界的同行评议（Peer review）；第三种则是 Mark Zuckerberg 倾向的完全由企业“自我监管”（Self-policing）。对于 Elon Musk 提出的实验室交叉审查方案，Ali Ghodsi 表现出强烈的怀疑态度，并用了一个极其生动的比喻来反驳：“这就像是把两个拳击手放进拳击擂台，却让他们互相充当裁判。这能行得通吗？他们只会每分每秒都冲着对方尖叫喊犯规！”在 Ali 看来，前沿实验室之间存在着巨大的利益冲突和历史恩怨，尤其是各家都在紧锣密鼓地筹备 IPO。一旦竞争对手发布了一款惊艳的新模型，另一方势必会立刻跳出来指责对方“带来了巨大的超级智能灭绝风险”、“极不负责任”。这种内在的竞争偏见决定了同行互相评议绝不可能做到真正的客观公正，这也是人类社会为何自古以来都需要独立的法官和第三方机构的原因。随后，Sarah Wang 抛出了 Elon Musk 此前在 All-In 峰会上的尖锐讽刺：前沿实验室一边宣扬“AI 会毁灭全人类”，一边又在私底下向投资者兜售“你们要多少 IPO 额度？”——这宛如一场精心策划的“4D 国际象棋”，言行之间存在着极度的认知失调。Ali 坦承这种矛盾在现实中确实交织在一起：一方面，部分研究人员确实感到恐惧，希望借由监管来为竞赛踩下刹车（Pacing）；但另一方面，这种末日叙事也沦为了极高明的市场营销手段。过去几年中屡见不鲜的套路就是高调宣称：“天啊，我们新训练出来的模型好得连我自己都害怕”，进而成功引爆全球的舆论焦点。不过 Ali 强调，抛开营销噱头不谈，技术威胁也是实打实的，例如从 CVE 漏洞曝光到被武器化利用的时间，已经从过去的数年断崖式缩短到了现在的短短数分钟。 ```mermaid flowchart TD accTitle: AI 监管的三种路线与利益冲突 accDescr: 对比第三方独立审查、实验室交叉审查与企业自我监管三种模式，并揭示商业竞争对评估的影响。 subgraph Proposals["当前三大监管提议"] P1["第三方独立审查 (OpenAI / Anthropic)"] P2["实验室交叉同行审查 (Elon Musk)"] P3["完全自我内部审查 (Mark Zuckerberg)"] end subgraph Dynamics["商业与竞争现实"] P2 -->|"拳击擂台互为裁判"| Conflict["巨大的商业竞争与 IPO 利益"] Conflict --> Bias["指控对手'超级智能风险' 互相高喊犯规"] Marketing["营销叙事: 末日论博取关注"] <--> IPO["资本运作: 筹备上市获取算力"] end P1 -.->|"Ali 赞成方案"| Solution["多元背景的独立第三方仲裁"] ``` --- ### [32:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1965s) 行业自律还是联邦干预？政府介入的不可逆转性 Martin Casado 指出，他和 Ali 都是网络背景出身，计算机网络史上曾有依靠 IEEE、IETF 甚至 ICANN 等非政府组织进行技术协调与仲裁的成功先例。然而，一旦引入具有强制力的司法或联邦介入，性质就会发生根本变化。Martin 质疑：目前到底应该停留于行业自律（Industry self-policing），还是已经到了必须由联邦政府下场干预的时刻？ Ali 认为，行业自律与政府监管之间往往并不是非黑即白的，二者常常会交织并互相渗透，就像金融业的 FINRA（美国金融业监管局）一样，虽属于行业自律组织，背后却与政府监管框架紧密绑定。历史经验表明，行业自律往往是最终走向政府立法的过渡阶段。更关键的现实在于政治逻辑：既然前沿实验室的 CEO 们天天在公开场合大谈“AI 存在生存级灭绝风险（Existential risk）”，并主动请求政府前来监管，那么监管机构根本不可能回绝这种送上门的权力。谈及华盛顿的政治动态，Martin 引用了相关的政界名言——“从未见过有 CEO 主动要求被监管，但也绝没有任何监管者会拒绝这样的请求”。如今以奥巴马公开发声为代表，AI 监管的“元政治机器”（Metapolitical machinery）已经全面启动，不仅可能成为中期选举的核心议题，甚至可能催生一刀切式的严苛联邦监管，将模型研发强行纳入能源部（DOE）级别的管制或叫停。面对 Martin 关切“现在是否为时已晚、无法避免严苛监管”的提问，Ali 依然抱有一线希望，认为目前依然存在操作与缓冲的时间窗口，全行业应当努力争取最合理的方案，而不是被动等待极端立法的降临。 --- ### [35:02](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2102s) RSI（递归自我改进）迷思：真正的智能涌现还是“自催化”现象？讨论随后深入到了大模型研发的技术本质。当前资本以前所未有的天文数字涌入 AI 领域，而强化学习（RL）的核心在于设定可验证的奖励函数（Verifiable reward functions）——例如解决特定的数学难题或特定编程子领域的代码任务。Ali 提醒大家，在砸入数十亿美元之后，模型在这些极其狭窄的指标上取得突破是必然的，但这根本不意味着超级智能（Superintelligence）已经降临。企业甚至可能产生一种错觉，以为更少的输入带来了更好的结果，而忽略了背后为了设计这些实验所消耗的海量计算与脑力资源。针对实验室为了巨额资金不惜走向上市的现象，Ali 直言：“作为一家大规模非上市公司的掌门人，我很清楚如果不是因为极度缺钱，他们绝对宁愿保持私有化。”这些模型公司之所以必须奔赴公开市场募资，是因为他们将 Scaling Law 和庞大的资本规模视作唯一的战略护城河。随后，话题转向了令整个技术圈屏息凝神的 RSI（Recursive Self-Improvement，递归自我改进）。Martin Casado 对此泼了冷水，指出当前业界几乎所有声称是 RSI 的东西，实际上都只是“自催化效应”（Autocatalytic effects）。自催化在人类工业史和计算机发展史上早已司空见惯： - **经典硬件案例**：没有计算机芯片，人类根本不可能造出新的计算机芯片，任何计算机科学背景的人都懂这个道理； - **软件编译**：编译器编写出下一代更强大的编译器； - **工业革命与互联网**：蒸汽机的制造离不开蒸汽机本身的动力；互联网通过支持分布式远程协作催化了互联网自身的飞速扩张。 Martin 披露，作为风险投资人，他的日常工作就是接见各种从前沿实验室离职出来创业的人。每个人嘴上都挂着 RSI 这个热词，但其中 90% 以上做的事情不过是用 AI 加速现有流程——比如用大模型生成 GPU 算子核心（GPU kernel）、用模型辅助数据清洗（Data cleaning）。这些本质上都是传统的自催化，绝非真正的 RSI。 ```mermaid flowchart LR accTitle: 自催化效应与真正递归自我改进的区别 accDescr: 区分传统的工具性自催化加速与全自动化闭环的真正递归自我改进。 subgraph Autocatalytic["90% 的现状：自催化工具 (Autocatalytic)"] direction TB A1["芯片辅助设计新芯片"] A2["编译器编译新编译器"] A3["AI 辅助清洗数据 / 生成 GPU 内核"] end subgraph TrueRSI["前沿探索：递归自我改进 (RSI)"] direction TB R1["模型自主探索科研假设"] R2["闭环自动训练并超越自身"] R3["脱离人类干预的无上限能力跃迁"] end Autocatalytic -.->|"混淆概念与过度宣传"| TrueRSI ``` - **当前研发分布**：Martin 观察到行业内 90% 的“卡路里”（研发投入）都落在自催化辅助工具上；只有极少数团队（类似 Andrej Karpathy 尝试的 Auto-research 方向）在真正探索让模型端到端训练自身。 - **泡沫化的 AGI 预期**：OpenAI 的 Greg Brockman 刚刚在播客上宣称人类已经进入“AGI 时代”，随后所有人便跟风声称拥有了 AGI。但现实是，当 Ali 在企业客户大会上询问“有谁正在管理成百上千个能够自主协商、自动化处理一切的 Agent 集群”时，台下几乎无人举手。 - **现实的企业 AI 应用画像**：大多数企业在实践中的全部 AI 应用仅仅是部署了微软的 Microsoft 365 Copilot，或者将其当成一个响应速度更快的“高配版 Google 搜索 Chatbot”；另一个被普遍落地的领域是辅助编程，但即使在代码生成方面，实际的 ROI（投资回报率）究竟如何，依然有待严肃商榷。 ### [40:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2400s) 阻碍企业落地的并非智能水平，而是缺乏组织上下文谈到 AI 智能体（Agent）在企业中的应用现状，Ali Ghodsi 指出，目前市面上根本还没有出现真正实现全企业业务流程自动化的 Agent 工作流。之所以无法落地，核心症结并不在于模型不够聪明，而在于模型完全缺乏任何组织内部实际存在的上下文。模型没有参加过公司的每一次会议，不知道每个人脑子里在想什么，也不熟悉内部真实的业务流程。在几乎每家公司里，总有那么一两位“百事通”骨干，其他员工遇到难题就会拍拍他们的肩膀请教，甚至管理层都会担忧“万一这个人离职了该怎么办”。如今的 AI 模型恰恰缺少这部分不可替代的上下文，如果能将这层组织上下文融入到现有前沿模型中，哪怕模型的基准指标不再提升，全球企业也能立刻获得惊人的生产力跃升。在 Ali 看来，企业根本不需要能解答纳维-斯托克斯方程（Navier-Stokes）、攻克数学猜想或在“人类最后考试”（Humanity's Last Exam）上从 60% 刷到 70% 的超强模型。社会上不少人对“放缓前沿模型演进”感到忧心忡忡，但对于绝大多数企业而言，即使前沿技术在当前水准停滞，也丝毫不会影响它们实现自动化和创造价值，因为大部分公司在技术采纳曲线上还处在极其落后的位置。尽管对尖端实验室而言，智能单价每六个月缩减至十分之一的断崖式下跌至关重要，但从社会成本与效益的平衡角度来看，公众在经历了一波波对存在性威胁和网络攻击的恐慌后，最关心的始终是：“AI 到底能为我带来什么实际好处？” ```mermaid flowchart TD accTitle: 组织上下文驱动企业落地 accDescr: 对比传统追求模型刷分与注入组织上下文对企业生产力的不同影响 subgraph Frontier ["前沿模型刷分路径"] M1["前沿大模型"] --> M2["攻克数学难题 / 提升竞赛得分"] M2 --> M3["对绝大多数企业落地影响有限"] end subgraph Enterprise ["组织上下文注入路径"] C1["会议记录 / 决策隐性知识"] --> O["组织本体知识图谱"] M1 --> O O --> E["全流程智能体业务自动化"] E --> P["生产力呈指数级释放"] end ``` ### [42:15](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2535s) 真正拯救生命与重塑产业的落地范例面对公众对 AI 实际价值的疑虑，Ali 列举了若干在 Databricks 平台上构建的代表性落地案例，以此证明 AI 绝非空中楼阁，而是在切实改善甚至挽救生命： - **Crisis Text Line**：利用大语言模型（LLM）精准检测青少年的求助短信，识别其自残与自杀倾向并及时介入，在危急时刻直接拯救生命。 - **Insulet Omnipod**：为糖尿病患者打造的胰岛素泵系统。过去患者必须忍受频繁扎针测血糖的痛苦，如今该设备利用自主学习的 AI 算法，根据人体实时葡萄糖水平动态精确释放胰岛素。 - **Zipline**：完全基于 AI 驱动的自主无人机物流网络。从电池续航优化到飞行路线决策均由算法自动接管，最初在非洲向难民营紧急空投急救血液，如今已扩展至全球更多受灾与物资匮乏地区。 - **默克集团（Merck）与 TEDDY 模型**：联合开发的 Transformer 药物研发模型 TEDDY（Transformer Enhanced Drug Discovery）。它不再预测下一个英文 Token，而是预测基因调控网络（Gene Regulatory Network, GRN）的反应机制，能够准确分辨哪些细胞在发病机制中起因果主导作用、哪些仅是被动反应，从而大幅降低针对特定疾病研发新药的成本与周期。 - **诺和诺德（Novo Nordisk）**：以研发 GLP-1 药物闻名的制药巨头，目前将基于 Databricks 的图谱分析（如 Genie）用于所有临床试验。过去对肥胖症临床研究数据得出结论往往耗时数周，如今压缩到了几分钟之内。 Ali 强调，社会决不能因为抽象的恐惧而限制或推迟这些切实造福人类的应用落地。 ### [44:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2688s) 什么是“本体”（Ontology）：从隐性认知到离线索引当讨论到企业在未来 12 个月如何将上下文工程化落地时，Ali 阐述了“本体”（Ontology）的核心内涵。针对主持人提到 Palantir 经常挂在嘴边的这个词，Ali 澄清本体并不是某家公司的专属名词。在组织学意义上，本体是指企业内所有抽象概念之间的关系网络——囊括战略目标、各职能部门、具体员工、进行中的项目以及软硬件资源之间的相互关联与运作规律。 Ali 用一个生动的人事比喻解释了本体的本质：假设有两位同样聪明勤奋、教育背景一致、专业技能水平相当的员工，一位是今天刚入职的新人，另一位已经在公司工作了五年。两者的唯一区别，就是那位老员工脑海中拥有对这家公司运作方式的完整“本体”——她清楚组织架构图之外的隐秘规则：“别看那张层级图，别去找那个人签字，他根本办不成事，你得找另一个人；这里不需要提交那份冗长的审批流程；那个项目才是当前的真正核心。”在初创圈，大家常说“核心团队流失公司就完了，无法通过招聘新员工来直接弥补”，正是因为每个人脑中沉淀的本体无法被轻易替换。 ```mermaid flowchart LR accTitle: 传统在线遍历与离线本体索引对比 accDescr: 展示现有 Agent 逐个 MCP 调用的低效与离线构建本体图谱索引的高效对比 subgraph AgentLoop ["传统 Agent 在线遍历方式"] direction TB Q1["接收问题"] --> L1["推理并逐一调用 MCP 接口"] L1 --> L2["反复读取并在线总结"] L2 --> R1["耗时几分钟 / 成本高昂 / 质量受限"] end subgraph IndexLoop ["本体图谱离线索引方式"] direction TB D1["企业全量数字化数据"] --> G["持续离线构建本体知识图谱"] Q2["接收问题"] --> G G --> R2["毫秒级精确响应并严密匹配权限"] end ``` 企业要把隐性知识转化为 AI 可以利用的本体，必须先克服将所有会议、数字化沟通内容合规沉淀与转录的阻力（这往往面临法务团队对全面录音录屏的限制），进而将这些数据提炼成一张数字知识图谱。Ali 批评了当前众多 Agent 框架（如各类代码助手和 MCP 协议客户端）采用的“边思考边逐个调用工具”的工作流。他类比道，这就好比 25 年前的 Google 搜索引擎不建网页索引，而是在用户输入搜索词后，临场打开一个网页、用 LLM 总结内容、点击几个超链接跳到其他网站抓取、在线折腾 10 分钟后再吐出 10 个蓝色链接一样——不仅昂贵、缓慢，而且覆盖面极小导致质量堪忧。 Google 的破局之道是在后台离线计算倒排索引，让搜索请求在 100 毫秒内命中本地缓存。企业 AI 同样需要将所有组织关系转化为类似 Google 早期 PageRank 一样的预计算索引。不过，企业级本体图谱的技术挑战远超当年的网页搜索：它不仅要处理多元异构的企业对象，还必须在图谱中嵌入极其严苛的数据权限管理与访问控制策略——某位员工有权检索到的图谱节点，其他员工可能无权访问。 ### [48:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2880s) Databricks 全面推行本体化对组织决策的重塑 Ali 透露，Databricks 内部始终坚持“吃自己的狗粮”（Dogfooding），且他们在自身系统上构建的本体规模远超任何外部客户，其图谱节点已突破数以千万计。在应用本体之前，公司与绝大多数层级制组织一样，信息只能沿着树状层级上下缓慢流转：基层出现无法决断的争议时逐级向上升级给主管；高管介入后需要花费大量时间调阅背景、补齐上下文才能做决定；而决策一旦下达，又必须耗费漫长的时间自上而下重新传达与对齐，导致组织效率随着规模扩大而急剧下滑。通过本体化重构内部信息索引，Databricks 在过去一年内彻底改变了运转模式，组织的信息流动与决策效率迎来了根本性的代际跃迁。 ### [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) 组织本体的落地与“财富500强”财务轶事在传统的企业运作中，大量的会议往往围绕着信息对齐展开：通常由一位精通 Excel 的分析师做出复杂的数学模型，做出一套带有精美图表的 PowerPoint 幻灯片，然后在会议上逐层传达。Ali Ghodsi 指出，只要企业构建起了本体（Ontology），AI 就能全盘接管这类分析工作。AI 掌握着完整的业务上下文，能够按管理者期望的形式直接呈现图表，参会者无需安排后续追踪会议，直接在现场向 AI 提问即可。这种组织扁平化与信息穿透的模式，正是此前 Jack Dorsey 所倡导的愿景，而在 Databricks 内部，这一变革已变成日常现实——如今开会只要有人抛出一个复杂的业务问题，所有人都会立刻低下头，掏出手机在内部基于本体的 AI 工具 Genie 上提问。 Ali 分享了一段此前仅在内部董事会上讲过的爆笑轶事：当时他需要准备一份演讲报告，想知道 Databricks 在“财富 500 强”（Fortune 500）企业中的客户渗透率。他发短信询问了一位销售运营人员（Sales Ops），对方却回复说：“抱歉，我现在在飞机上，没办法登录 Genie。”Ali 哭笑不得地回复：“如果你只是去登录 Genie 查一下，那我何必问你？我自己查不就好了吗！我找你是以为你掌握着我查不到的替代渠道。”有些无奈的 Ali 转头给公司首席财务官 Dave 质询同样的数据，结果 Dave 直接回传了一张 Genie 的查询截图。这让 Ali 惊叹道：“难道我们公司就没人做点原创性的工作了吗？所有人都是跑去问 Genie 和本体吗？这简直成了‘让我为你 Genie 一下’（Let me Genie that for you）。”虽然 Palantir 通过人工顾问深入企业将隐性知识手工梳理成本体的做法很出色，但 Databricks 实现了自动化构建图谱并将其输入 Agent，使业务领导者能以高度分析化的图表形式随时“审讯”（interrogate）数据并快速下达决策。 ```mermaid flowchart TD accTitle: 基于企业本体的智能问答与决策流 accDescr: 展示从企业异构数据沉淀到自动生成本体知识图谱，再通过 Genie 与 Agent 为业务端提供实时下钻查询 subgraph DataIngest["知识与数据沉淀"] D1["运营与业务数据"] D2["历史会议与隐性知识"] end subgraph OntologyEngine["自动化本体层"] O1["自动抽取实体与关系"] O2["动态知识图谱构建"] end subgraph Interaction["Genie 交互与决策"] G1["智能 Agent 分析问答"] G2["管理层实时下钻与图表呈现"] end D1 --> O1 D2 --> O1 O1 --> O2 O2 --> G1 G1 --> G2 ``` --- ### [52:54](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3174s) 从 Token 狂热到价值最大化：Unity Gateway 与成本控制随着大模型代码辅助能力的提升，Databricks 内部经历了一场从追求调用量（Token Maxing）到追求投资回报率（Value Maxing）的转变。Ali 回忆道，前一年的第四季度模型能力出现显著跃迁，他为了在全公司推广 AI 生产力，亲自使用模型为 Databricks 极其敏感且注重安全的数据平台编写代码并提交（commit）到生产环境。他在全员范围内立下标杆：“如果 CEO 都能向高安全要求的核心生产环境提交代码，那任何经理和工程师没有任何理由做不到。”公司在 Q4 设立了积分榜全面推动，到了次年一二月份全员已经全面普及 AI 工具。然而到了二三月份，无节制的 Token 消耗开始失控蔓延，各种低效调用层出不穷。幸运的是，Databricks 在基础设施层面早有布局，提前几个季度预见到了这一瓶颈。他们依托自身的统一网关 Unity Gateway 对接了 OpenAI、Anthropic、Gemini、Grok 以及各类开源模型，并迅速推出了精细化的治理方案： - **预算硬约束与预警**：针对个人与部门设立 Token 配额与预算上限，在接近警戒线时自动触发警告。 - **成本预测分析**：建立精细的调用指标面板，精准预测每项业务开销的走向。 - **智能模型路由器（Smart Routers）**：检测到用户接近预算上限或提出的问题属于简单任务时，自动将其无缝路由到成本更低的模型上。 - **多路复用 Agent 运行框架（Omnigent Harness）**：Databricks 发现即便采用完全相同的模型与版本，仅仅是外部运行框架（Harness）不同，实际产生的成本差距就能达到惊人的两倍。通过动态切换与多路复用框架，Databricks 成功“压平了成本曲线”——内部 Token 调用总量持续指数级激增，而整体 AI 账单却始终保持平稳。 --- ### [55:38](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3338s) 模型混合分流架构、开源渗透率与后训练落地针对当前业界是否正在从最顶尖的前沿模型（Frontier Models）大规模转向开源模型的讨论，Martin Casado 提到自己在 20 多家公司的董事会中，上周首次看到一家拥有大型工程团队的成熟企业正式从 Frontier 模型全面切换到开源基础模型（如 GLM 等）。Ali 指出，企业对模型的需求实际上正在分化：在解决复杂棘手、能带来显著 ROI 的核心任务时，企业依然渴望最聪明的前沿模型；但对于大量琐碎的平庸任务——例如竟有工程师让笨重的前沿 Agent 旋转等待五分钟去重命名文件，单次操作花费数美分——这种浪费是不可接受的，企业正积极采用模型级联架构。 ```mermaid flowchart LR accTitle: 现代工程系统的三层模型分流体系 accDescr: 展示架构设计、代码实现与代码审计分别采用高阶模型与经济型模型的配合模式 subgraph ArchDesign["架构与方案设计"] F1["前沿顶尖模型 (如 Fable / Astra)"] end subgraph Impl["具体代码实现"] O1["经济型/开源小模型 (高并发低成本)"] end subgraph Audit["代码审计与验证"] F2["前沿顶尖模型 (严格质检把关)"] end F1 -->|"传递规范与蓝图"| O1 O1 -->|"提交代码产出"| F2 ``` 当前初创公司和大型工程团队普遍采用的模型调用架构与市场格局表现为： - **专家分流模式（Expert Pattern）**：使用轻量且廉价的模型先行处理并作为调度器，仅在遇到高难度瓶颈时向顶级专家模型发起调用，或让大小模型来回乒乓交替验证；同时在框架层优化调度（例如使用极高效率的 Pi 框架）。 - **典型的三段式工作流**：在软件工程中，行业正形成一种固定范式——用 Fable 或 Astra 等前沿顶尖模型做顶层架构设计，用廉价模型生成具体实现代码，最后再交由顶尖模型进行代码审计。 - **开源模型的真实渗透率**：按实际消耗金额（Dollar）计算，开源模型仅占市场整体花费的约 5%；但若按 Token 调用总量计算，开源模型已占据 60% 以上的绝对主流。尤其在客户可见的外部产品端（External Product），部分初创公司的开源模型使用比例甚至高达 90%，而内部研发环境由于起初不计成本追求效果，前沿闭源模型的比例才稍高一些。 - **从预训练向后训练（Post-Training）转型**：回顾 Databricks 在 2023 年收购 Mosaic 的决策，Ali 表示当年试图为企业重新“预训练（Pre-training）自身大模型”的设想已被证明行不通，市面上已有足够强大的底座模型。真正的核心价值在于后训练与强化学习环境（RLE, Reinforcement Learning Environments）——Databricks 正在大规模为初创公司客户提供基础设施，帮助他们在特定垂类任务中对开源模型进行精细微调与 RL 对齐，让专有模型在目标领域达到极致效能。 ### [01:00:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3600s) 模型评估的“TDD困境”与大企业的现实抉择对于拥有特定重复性任务的初创公司而言，为其产品专门微调模型是一条性价比极高的路线。初创团队可以通过选取优质的开源基础模型，利用强化学习（RL）针对垂直任务进行深度定向优化，不仅能显著压低推理成本、实现极低延迟，还能牢牢将核心 IP 掌握在自己手中。然而，对于绝大多数大型企业来说，他们目前需要的仅仅是基础的业务自动化，去自建训练管线和评估集实在门槛太高。这其中的核心痛点在于模型评估（Evals）——构建高质量的 Eval 体系极其繁琐且耗时，初创团队有足够的动力和工程意愿去啃这块硬骨头，但对大型企业而言，直接调用前沿大模型（Frontier Models）才是最省事的“快捷键”。 Ali Ghodsi 分享了一个生动的内部产品教训：Databricks 曾尝试在产品界面中最显眼的位置为客户自动生成模型评估工具，结果发现根本没人愿意用；团队随后将该功能移到后台作为可选配置，结果更是彻底无人问津。Ali 将这种现象类比为软件工程界的“测试驱动开发”（TDD，Test-Driven Development）——行业里人人都口口声声说 TDD 是最标准、最正确的开发范式，但在现实业务交付中，真正严格执行 TDD 的工程师寥寥无几。模型评估同样陷入了这种“知易行难”的诅咒：企业开发者既缺乏科学建立 baseline 并按部就班测试的耐心，也渴望获得即时正向反馈，往往看到新模型发布就想立刻上线解决业务问题，而不愿花数周时间去搭建严谨的评估基准。 ```mermaid flowchart TD accTitle: 模型应用路径分化对比 accDescr: 对比初创企业自训模型与大企业直接采用前沿模型的两种完全不同路径。 subgraph StartupPath ["初创公司路线（特定垂直任务）"] A["开源基础模型"] --> B["强化学习 (RL) / 微调"] B --> C["高成本构建定制 Evals"] C --> D["专属低延迟、低成本、高可控 IP 代理"] end subgraph EnterprisePath ["大型企业路线（基础自动化）"] E["业务自动化需求"] --> F{"面临 Evals 门槛？"} F -- "回避繁琐评估（类似 TDD 困境）" --> G["直接调用 Frontier Model"] G --> H["快速落地即时验证"] end ``` ### [01:01:45](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3705s) 驻场工程师（FDE）模式与定制化 Agent 护栏针对大企业内部普遍缺乏前沿 AI 工程专长这一现状，Databricks 的前置驻场工程师（FDE，Forward Deployed Engineer）模式迎来了爆发式的市场需求。许多企业甚至连底层数据资产与业务脉络都没开始记录，根本无法直接自动化生成组织本体（Ontology），这就需要 FDE 深入现场协助搭建数据收集流程与底层逻辑。与此同时，客户强烈渴望将 Agent 推向生产环境直接面向最终用户，但这要求系统兼具超低延迟与极其严密的护栏（Guardrails），防止外部恶意诱导或越权滥用。 Ali 以他们为福克斯体育（Fox Sports）搭建的体育 AI Agent 为例：用户可以通过对话界面畅聊各项体育赛事实况，但如果有人试图借机探讨政治或敏感公共话题，该 Agent 能够滴水不漏地礼貌拒绝，并迅速将话题巧妙拉回到体育赛事本身。正是借助 FDE 团队在架构设计、护栏工程及上下文集成上的贴身支持，那些缺乏内部顶级 AI 研发力量的传统大企业才得以跨过冷启动门槛，安全稳妥地将第一批高价值 AI Agent 交付上线。 - **FDE 核心职责**：弥补传统企业内部 AI 研发能力断层，帮助梳理底层业务数据以沉淀“组织本体”，并设计生产级低延迟与高安全性架构。 - **Agent 护栏机制（Guardrails）**：确保面向消费者的智能体专注于既定业务边界，如 Fox Sports AI 在面对政治等非相关诱导提问时展现出精准的拒绝与话题回归能力。 ### [01:03:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3810s) 专为 Agent 打造的数据库：Neon 与 Lakebase 颠覆底层范式近期第三方中立机构的一项基准测试显示，Databricks 旗下的 Lakebase（基于 Neon 架构）出人意料地超越诸多传统开发者热门工具，成为了 AI Agent 最首选的 Postgres 数据库。Ali 将这份殊荣归功于 Neon 创始人 Nikita Shamgunov 及其工程团队的极致专注。长久以来，数据库行业的竞争焦点始终围绕着数据库管理员（DBA）或全栈应用开发者展开；而 Neon 团队彻底颠覆了用户画像（Persona），将资源全盘倾注在一个崭新的服务主体上——**Agent 本身**。为了迎合 Agent 频繁自发编写代码、创建测试环境、快速试错的自动化行为模式，团队将系统响应压榨到了极致： 1. **亚秒级冷启动**：数据库实例在远低于一秒的时间内就能瞬间拉起就绪； 2. **PB 级秒级克隆与分支（Branching）**：Agent 可以像操作 Git 分支一样，在数微秒内基于同一份 PB 级底层数据克隆出无数个轻量级隔离分支； 3. **极高容错与快照回滚**：如同现代 Unix 命令行工具（如 `uv`、`ripgrep`）凭借极致速度和安全性重构开发体验一样，Neon 将这种设计理念带入了复杂的数据库领域，即使 Agent 逻辑跑飞也能无痛回滚； 4. **对实验友好的计费模型**：杜绝 Agent 因自动化写代码测试而无意间跑出天文数字账单。这种彻底面向机器的降维打击带来了惊人的商业数据：在 Neon 和 Lakebase 上创建的数据库中，**超过 90% 是由 AI Agent 自行创建的，而非人类工程师手动配置**。Martin Casado 对此深有共鸣，直言自己在日常开发中已将 Neon 作为首选底座，并对其在并入大厂体系后依然能维持极佳的响应速度与产品迭代节奏感到惊叹。 ```mermaid flowchart LR accTitle: Agent 原生数据库交互架构 accDescr: 展示 AI Agent 如何在亚秒级时间内分支、测试与回滚 Postgres 数据库。 subgraph TraditionalDB ["传统数据库模式"] DBA["DBA / 人工运维"] -->|"数分钟配置/审批"| MasterDB["生产数据库"] end subgraph AgentNativeDB ["Neon / Lakebase 模式"] Agent["AI Agent (占调用量 >90%)"] PB["PB 级主数据库"] Agent -->|"极速启动 (<1s)"| BranchA["分支 A: 代码沙箱测试"] Agent -->|"秒级克隆"| BranchB["分支 B: 方案验证"] BranchA -->|"发生错误"| Rollback["轻量快照秒级回滚"] BranchB -->|"验证通过"| Merge["合并至生产管线"] PB -.-> BranchA PB -.-> BranchB end ``` ### [01:06:10](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3970s) 尾声：对 AI 毁灭论（P(doom)）的终极交锋在对谈的尾声，Martin Casado 抛出了硅谷 AI 圈经典的定性问题，询问 Ali 的“P(doom)”（人类因 AI 走向灭绝的概率）到底有多高。Ali 给出了极其明确的乐观态度：“低于 10%，甚至接近于零。” 面对同样的追问，Martin 给出了一个充满反转哲思的机智回答：“**我不确定具体数字，但我唯一可以确定的是：没有 AI 时人类走向毁灭的概率（P(doom) without AI），要远远高于拥有 AI 时走向毁灭的概率（P(doom) with AI）。**” 这番精妙的论述赢得了全场一致赞同，不仅有力反击了非理性的末日恐慌论，也为整场探讨画上了一个坚定理性、立足现实的技术乐观主义句号。 ## Notable quotes > [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) 不要在没有过硬依据的情况下无端制造恐慌 — Ali Ghodsi > [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) Martin，我需要去收拾深山木屋来防备 AI 末日吗？水我已经备好了，你什么时候赶过来？ — Martin Casado's sister > [00:48](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=48s) 别搞砸我的超级 IPO，闭上嘴让我们拿回投资回报 — Ali Ghodsi > [05:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=300s) 求求你们来给所有人定规矩，强制所有人一起慢下来，否则我们只能硬着头皮竞速到底 — Ali Ghodsi > [13:20](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=800s) 在 Databricks，90% 以上的软件代码已经是由 AI 编写的了。剩下的最后几个百分点是不是也由 AI 来写，真的有本质区别吗？其实根本没有。 — Ali Ghodsi > [16:30](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=990s) 如果 Anthropic 在两周后缩减到只有 12 个人，却能持续扩张并以越来越快的频率发布更强模型，那全行业就必须真正警惕了。 — Martin Casado > [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) 这就像是把两个拳击手放进拳击擂台，却让他们互相充当裁判。这能行得通吗？他们只会每分每秒都冲着对方尖叫喊犯规！ — Ali Ghodsi > [30:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=1800s) 天啊，我们新训练出来的模型好得连我自己都害怕 — Ali Ghodsi > [35:02](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=2102s) 作为一家大规模非上市公司的掌门人，我很清楚如果不是因为极度缺钱，他们绝对宁愿保持私有化。 — Ali Ghodsi > [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) 如果你只是去登录 Genie 查一下，那我何必问你？我自己查不就好了吗！我找你是以为你掌握着我查不到的替代渠道。 — Ali Ghodsi > [50:00](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3000s) 难道我们公司就没人做点原创性的工作了吗？所有人都是跑去问 Genie 和本体吗？这简直成了‘让我为你 Genie 一下’ — Ali Ghodsi > [52:54](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3174s) 如果 CEO 都能向高安全要求的核心生产环境提交代码，那任何经理和工程师没有任何理由做不到。 — Ali Ghodsi > [01:06:10](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3970s) 低于 10%，甚至接近于零。 — Ali Ghodsi > [01:06:10](https://www.youtube.com/watch?v=GzEtpAKYRvE&t=3970s) 我不确定具体数字，但我唯一可以确定的是：没有 AI 时人类走向毁灭的概率（P(doom) without AI），要远远高于拥有 AI 时走向毁灭的概率（P(doom) with AI）。 — Martin Casado ## Mentioned Ali Ghodsi, Martin Casado, Sarah Wang, Dario Amodei, Jakob Uszkoreit, Elon Musk, Elizabeth Warren, Bernie Sanders, Steve Bannon, Greg Abbott, Mark Zuckerberg, Nick Bostrom, Yann LeCun, Ben Horowitz, Barack Obama, Greg Brockman, Jack Dorsey, Nikita Shamgunov, Databricks, a16z, Anthropic, OpenAI, Hugging Face, Meta, Palantir, Sony, Crisis Text Line, Insulet Omnipod, Zipline, Merck, Novo Nordisk, Microsoft, Fox Sports, Neon, MosaicML, GLM, TEDDY, Genie, Lakewatch, Unity Gateway, Omnigent Harness, Lakebase, Microsoft 365 Copilot, Google, Gemini, Grok, PlayStation, Superintelligence ## Video description Databricks co-founder and CEO Ali Ghodsi joins a16z General Partners Martin Casado and Sarah Wang for a conversation about AI risk, recursive self-improvement, cybersecurity, and what’s actually holding back enterprise adoption. Ali argues that today’s models are already capable enough to automate far more work than most companies are using them for. The bigger problem is context: models haven’t been in every meeting, don’t understand how decisions actually get made, and lack the institutional knowledge that experienced employees accumulate over years. He explains why building an organizational “ontology” could help close that gap and what Databricks has learned from doing it internally. They also debate the current conversation around pacing frontier AI, what would constitute meaningful recursive self-improvement, and why Ali distinguishes speculative superintelligence risk from the much more immediate challenge of AI-powered cyberattacks. They close with how enterprises are managing exploding AI usage and costs, the shift toward multiple models and harnesses, and why agents are beginning to reshape infrastructure itself. Timestamps: 00:00 - Intro 00:48 - Pacing the Frontier: Where Ali Lands in the AI Debate 14:40 - Why We Haven't Seen the AI Cyber Apocalypse Yet 25:50 - The 4D Chess Problem: Doom Talk vs IPO Allocations 27:56 - Industry Self-Policing vs Federal Involvement 35:02 - The Black Box Test: Is RSI Actually Happening? 39:42 - How Enterprises Actually Extract Value from AI in the Next 12 Months 40:37 - Defining Ontology (Beyond the Palantir Version) 46:08 - The Finance Anecdote: Real AI Value in the Boardroom 56:51 - Databricks in the Fortune 500: The Use Cases Surprising Even Ali Resources: Follow Ali Ghodsi on X: https://x.com/alighodsi Follow Sarah Wang on X: https://x.com/sarahdingwang Follow Martin Casado on X: https://x.com/martin_casado Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=GzEtpAKYRvE)

Source quotes

[0:48] Ali Ghodsi: 不要在没有过硬依据的情况下无端制造恐慌 [0:48] Martin Casado's sister: Martin，我需要去收拾深山木屋来防备 AI 末日吗？水我已经备好了，你什么时候赶过来？ [0:48] Ali Ghodsi: 别搞砸我的超级 IPO，闭上嘴让我们拿回投资回报 [5:00] Ali Ghodsi: 求求你们来给所有人定规矩，强制所有人一起慢下来，否则我们只能硬着头皮竞速到底 [13:20] Ali Ghodsi: 在 Databricks，90% 以上的软件代码已经是由 AI 编写的了。剩下的最后几个百分点是不是也由 AI 来写，真的有本质区别吗？其实根本没有。 [16:30] Martin Casado: 如果 Anthropic 在两周后缩减到只有 12 个人，却能持续扩张并以越来越快的频率发布更强模型，那全行业就必须真正警惕了。 [30:00] Ali Ghodsi: 这就像是把两个拳击手放进拳击擂台，却让他们互相充当裁判。这能行得通吗？他们只会每分每秒都冲着对方尖叫喊犯规！ [30:00] Ali Ghodsi: 天啊，我们新训练出来的模型好得连我自己都害怕 [35:02] Ali Ghodsi: 作为一家大规模非上市公司的掌门人，我很清楚如果不是因为极度缺钱，他们绝对宁愿保持私有化。 [50:00] Ali Ghodsi: 如果你只是去登录 Genie 查一下，那我何必问你？我自己查不就好了吗！我找你是以为你掌握着我查不到的替代渠道。 [50:00] Ali Ghodsi: 难道我们公司就没人做点原创性的工作了吗？所有人都是跑去问 Genie 和本体吗？这简直成了‘让我为你 Genie 一下’ [52:54] Ali Ghodsi: 如果 CEO 都能向高安全要求的核心生产环境提交代码，那任何经理和工程师没有任何理由做不到。 [66:10] Ali Ghodsi: 低于 10%，甚至接近于零。 [66:10] Martin Casado: 我不确定具体数字，但我唯一可以确定的是：没有 AI 时人类走向毁灭的概率（P(doom) without AI），要远远高于拥有 AI 时走向毁灭的概率（P(doom) with AI）。

Source quotes

[5:00] Ali Ghodsi: As a business leader, I understand that. There's a tragedy of the commons. Like I'm competing. I want to win, you know, and you're running ahead... I'm not going to stop unilaterally. I'll be a sucker, you know. Why don't you stop first? [10:00] Ali Ghodsi: Really bad things meaning things get hacked. And cyber is the primary one. [16:30] Martin Casado: Are you gonna believe like the numbers or your lying eyes? [18:45] Ali Ghodsi: Ignorance is bliss... I don't recall PlayStation being illegal. [18:45] Martin Casado: Sweden doesn't care about export controls in the US. [25:00] Martin Casado: What we learned here is that Martin really hates the word pacing. I will never use that word with you ever again. [30:00] Ali Ghodsi: foul, foul, foul! [30:00] Ali Ghodsi: Oh my God, this latest model is so good that I trained... it's almost scaring me [32:15] Ali Ghodsi: I've never had a regulator that says no to that. [46:42] Ali Ghodsi: If Google would have built Google search this way 25 years ago, we would have said, okay, we're gonna get 10 blue links... but instead of giving you 10 blue links, it would have gone to one website, summarized with an LLM what it does, found a few hyperlinks, jumped in parallel to a few of them, read a few websites, done that for 10 minutes, and then given you like its best 10 blue links it would find. Well, that would be very expensive... it would have taken a long time and we gotta wait 10 minutes. And three, the quality would be bad because you're actually only looking at a very small subset of everything that exists out there. [50:00] let me Genie that for you [65:40] Ali Ghodsi: less than 10%, close to zero [65:40] Martin Casado: My only answer is my p(doom) without AI is much higher than my p(doom) with AI.

Original transcript

[0:00] As a business leader, there's a tragedy of the commons. If you want to stop, if you want to go slower, why don't you go slower? Like I'm competing, I want to win. There's almost two camps. There's one camp which believes that this actually is an engineering problem. And there's others which actually believe you have to slow it down. Humans don't respond fast enough to the attacks that are happening. You need to automate all of those. And most organizations are actually not close to doing that. Is RSI and recursive self-improvement that the labs are doing leading us there? That's the big question. Something Elon said, this is some elaborate 4D chess. because on the one hand, you're saying all of humanity will die. On the other hand, you're saying, hey, what do you want for your IPO allocation? For the first time ever, a company at scale last week said that they're moving from the frontier models to GLM. Do you think that that's a trend or do you think that's just like a one-off anecdote? Thank you for being here, Ali. Super excited. So we obviously want to get to Databricks, but there is a broader conversation going on right now about AI. And Dario's weighed in, Jakob's weighed in, Elon's weighed in. But we want to hear what Ali Goetze thinks. In terms of, you know, if you called the topic, broadly speaking, pacing the frontier, etc., what is your strongest agreement with what being out there Where do you disagree And maybe where is there nuance that not being captured Yeah happy to cover it And me and Martin argue a lot So I'm sure that's not going to take long. We'll try and rein it in this time. Try to stay calm. But well, I do think first and foremost that there, maybe we agree on this, that leaders have responsibility to not freak people out unnecessarily, unless there's really, really really good reason. And I think, you know, there's always different people in society that are in different places, you know, in their mind space. So, you know, talking about these kind of existential risks and, you know, scenarios where all of humanity is going to be wiped out, I think is irresponsible. Like it can tip a lot of people over and it can cause a lot of mental health issues. Unless you have something that's going to wipe people out. Yeah. As I said, yeah. If, if there is an actual reason for it, then, you know, that's a different story. but I think that right now the existential risk is close to zero. So why freak everybody out? It's not actually needed. There are risks. We'll get into it. That's probably where we disagree. But first and foremost, I think that leaders should not freak everyone out. And I mean you know if there like technical nuances in how we doing AI research and so on well researchers can discuss that You don need to every time go on TV and or blast on Twitter to millions of people that hey you know I think there like this percentage 10% risk that all humanity is going to be wiped out. I don't think that's like helpful for a lot of people. Actually, I think causes a lot of harm for a lot of folks who get stressed out and actually are not in the nuances of all of this stuff and what it means. So that I don't think we should do. I don't think it's fruitful. It doesn't really help anyone. I mean, I think this is very, very true for the general public. Yeah. Like my sister, who's great, who's a schoolteacher in rural Arizona on Sunday, texting me and said, Martine, should I prepare the cabin for, she's kind of a prepper anyways, but should I prepare the cabin for the AI apocalypse? You know, I've got water set up. Like, when are you showing up? I'm like, hold on. Yeah. Like, we're not there. So clearly this has kind of spilled over the populace, which I agree is unnecessary and has blowback. I think there's a second one which is, I don't know if you saw like walking in here, I was checking X and Elizabeth Warren just talked about pausing all of AI development. That of course is on the coattails of Bernie, who is also working with Bannon, like Steve Bannon. So now in addition to just scaring people the federal complex is now spinning up And I think that could be actually quite contrary to the actual goals of the message And so there's more than just, you know, I think public hysteria at stake here. Yeah, there's a lot of politics going on, but I'm like in all of these groups and, you know, I see both sides. There's heavy politics happening on both sides, we should say, like, right? This is happening on both sides. No, this is a, I think both parties that do not include Trump himself agree that AI should be constrained at some level. I'm talking about the other side of this argument as well. Let me give you an example. Even Greg Abbott, right? Even Greg Abbott was like, you know, you can't have data centers in Texas. Well, I'm not talking about politicians. I'm talking about there's politics going on on both sides, right? There is politics on the business side, people who want to see great IPOs and they want to get returns on their investments. and they're like, don't mess up my IPO. And they wanna get like, hey, can everybody just shut up so that we can get our money back. So there's that and they're, you know, they have resources and they're using them. And, you know, so there's politics on that side. And those are like not, they're not sitting quietly and not doing anything and they can pull strings and they have connections. On the other side, there's all the people that like, okay, how do we weaponize this is awesome. This guy's who did this, you know, [5:00] It's like, let's weaponize this one. Let's plant this. If I, you know, let's pump these, you know, threads. Let's talk to the specific leverage point that everybody is using. Because I actually think that this is like a classic case of a PR misstep. And it's not just the doomy, gloomy type stuff. So here's the PR misstep, I think, which is like, it is not unusual for industries to try and regulate themselves. It's just not, right? And I think saying like security and safety is important. It is with every tech epoch. And we want to have some oversight. That was very, very sensible. The problem is couching this notion of pacing. And there's a number of issues with pacing. First off, it's orthogonal to safety and security. You can slowly build a weapon. That's not different than building a weapon. People don't feel it's genuine because these companies have been at a dead run. They're still buying more compute to be even faster. No, no, no. I mean, they just haven't done it historically. but also it kind of feels like this kind of almost milk toast capitulation to the pause people. So you're like, well, you say pause, well, I say pacing, which is almost like pause, but it's not like pause. So like they chose this kind of like flag to follow around pacing But if you actually read did you read the document It a totally sensible document I read it yeah It just has nothing to do with pacing right And so I honestly No I didn mention it Look I can a little bit disagree Look, there's a tragedy of the commons. There's this like, hey, if you want to stop, if you want to go slower, why don't you go slower? Why do you write articles? There's a lot of people making that argument. But no, I mean, as a business leader, I understand that. There's a tragedy of the commons. Like I'm competing. I want to win, you know, and you're running ahead. There's also the market equilibrium, which suggests that pacing is probably impractical anyways. Yeah, so I'm just saying that, you know, so it makes kind of sense for people to say, hey, if you guys don't stop this, this strategy of the comments is going to continue. I'm not going to stop racing because, you know, there is IPOs at stake. There is a competition at stake. There's also some animosity between the people. So, like, I'm not going to stop unilaterally. I'll be a sucker, you know. Why don't you stop first? You know, so then they're saying, hey, can you come in and stop us? But, you know, I think that you could also make the argument that if you look at the Hugging Face OpenAI incident that, by the way, I think these companies are great. And I think they are probably investing a lot of resources, but it's very clear from if you read what happened is that they weren't monitoring every token coming out and having it, you know, they were just like running these RL experiments. And then after the fact coming in and checking out what happened So they should have paced They should have been much slower in that particular incident right I don want to just quibble on syntax but words matter with PR right So let's take the hugging face incident. When I read that, you know what my reaction was? Was not, oh, OpenAI should pace. It's like, dude, fucking secure your thing, right? Like, do security controls like we always have done. It is pacing, though. It is pacing. It's not pacing. In the history of the internet, we had all of these things. We were like, let's pace the growth of the internet. It is pacing. Let's do security. Let's do control. Let's do like what you can do that. But it is pacing in the sense that, look, I face this all the time. I have a legal department at Databricks. I have a security department at Databricks. And, you know, they're always like, hey, slow everything down for everything. Not AI, like literally every little thing like, oh, you're going to go on a podcast. Well, what's the script for it? What are you going to say? And let's review that. And, you know, what's illegal? You cannot say this. You can say that. You can say that. You know, everything you say have to be materially true. Are these pacing people in this room with us right now? So you're running a RL experiment, you're training the next model. Should the security team be there and look at like run all their monitors and look at everything? I mean like millions of hours of GPU hours of tokens were produced and these agents were running a mock in the sandboxes. It would have slowed them down significantly if you had security team sit there and look at all the stuff. I agree. Now and they saying that hey like you know if we do that it slow us down And I not sure the other side is doing that So can you guys come in and slow us down Like just tell us like put some guardrails around us We happily then follow the rules and do the secure thing Otherwise, it doesn't make sense because we'll get our butts kicked. I just think, like, nuanced second order words don't work when, like, people are really afraid. You're like, I'm going to pace. And therefore, things like these don't happen. I literally think we should have just been like, safety and security is paramount. We're going to put in these controls. Like, that's the important thing. And I do think that nuance actually got lost. If you look at the. What Zuck said. Do you agree with it? I thought Zuck did it. He was like, hey, we're going to pace ourselves. We're going to put in security. Like the reason we released this later is because of security. Well, I thought was so great about Zuck is like he was very focused on like, like security, safety and self-regulation. Dario, Dario's first five words or whatever, like we need to pace the frontier. Right. It just puts you in a very different mindset. But then what he could have said is we need to secure the frontier. Fine. We need safety. I mean, like at some level, I think they were trying to optimize both for the doomers, which caused for pause and for politicians. And they kind of didn't satisfy either. But those people are actually freaking out inside the labs. And there are a lot of safety people that are freaking out genuinely, by the way. And not all of them are EA people and so on. And people are like, Hey, [10:00] They're surprised, right? Right, but here's the thing is like using the worst pace doesn't help either of them. I think it's like literally you're like, you're like, you're trying to find this. Pace is like the uncanny valley of making the Doomer people unhappy and the policy people unhappy. Because the Doomer people are like, that's not a pause, this is pacing. And, you know, everybody else is like, well, like this is, you know, this isn't real. You're not going to do it anyways. And you're not focused on security. So, again, independent of what we should do, which we should talk about. I just think that the way it was presented was just bad and it just didn't work. And that's why we're having the blowback. These guys are, you know, they're not trained, you know, PR people, you know. And yes, some of the stuff. I agree. I agree with you. I mean, I agree with the core premise that we shouldn't freak the public out. I think that existential risk right now is close to zero, you know. But let's talk about the core thing, which is the fact that, you know, anyone who's doing big reinforcement learning runs and they're giving it a reward function, so unleashing, you know, saying, hey, here's like, you know, 10,000 agents and here's a hundred million dollars. Let's put them in parallel and let them run on a gigantic cluster for a month or two try to solve anything And it doesn need to be a security thing It could be like do anything you know solve this math puzzle Really bad things can happen Really bad things meaning things get hacked And it has cyber is the primary one right That is real, right? I think of this making it, hey, this is an existential risk and so on, which I think was a mistake. I think it's not good to scare the public that way. I think it's become something that everyone, not just your sister, everybody around the planet is like now talking about. I've had all kinds of people that never care about this stuff and they find this extremely boring, ping me and say, what do you really actually think about this? This is really important. Now I'm starting to worry about it. So then it becomes a political issue. And we have elections here coming up, but there's elections all around the world. So you're going to see, they're not going to sit still in other parts of the world either. But I think that's our responsibility to talk about this in a balanced way and actually expose the risks. I think that super intelligence, that idea from that book, is very, very far away. I don't see any evidence that we're actually marching towards that or that's going to happen. Apparently some people, yeah, apparently some people at the labs are freaked out that maybe there's progress towards that. And I think it comes from RSI, recursive self-improvement, the models improving themselves. I would love to understand how much, have they seen something we don't know? There's kind of four criterias. If those four things are happening I would love to understand them One is our models if we end up in a situation where the following four conditions are happening which is the next model requires less resources less GPUs to train, and super linearly, not just like tiny little bit. The next model, it takes less time to train as well. So the second condition. Third, accuracy of the model, the intelligence is increasing. And fourth, we can do the former three again and again and again. you know, it's not just just... All those at the same time, right? All at the same time, not just any of them. Yeah, all four. If all four are happening, then you can imagine in a way where you can, you know, because any of them does not happen. Like for instance, if resources is constant, then that's okay, because we're going to run out of hardware. So then it will pace itself. Like we will not have enough hardware to do that. Not enough GPUs, right? Time the same. So it needs to be that you end up in this situation. So if you just mean that the software is writing itself, we're already there today. Like 90 some percent of the software in Databricks is written by AI. Does it matter if the last few percent is also written by AI? No, it doesn't matter really that much. But if you're getting these four conditions, then you might get a speed up where the next model, let's say takes half amount of time and half the resources, and it is more intelligent, and you keep doing that you know then you might end up in a situation where I don know By the way I don if that necessarily leads you to super intelligence per se but it could So then that would be more risky So it would be nice if they can share all that data and we can shine some light and transparency on that. I actually think it's a great breakdown that you have. I don't think anyone is using that as a definition, actually. Right. I think people, I think there is a little bit of people freaking out about like, oh my God, the emergent behavior. Now it's creating itself and so on. But I think, like, as I said, a lot of people, their definition is just, hey, if I'm not even coding anymore and it's coding itself. Right. But I think they're conflating, hey, what's my value and is it scary for me? Versus, hey, that then means we'll get that super intelligence that 2014 theoretically was hypothesized by Botstrom. Well, your point in compute, though, is a really good one. That's missed in, I think, a lot of arguments on RSI, right? Because as far as we can tell, the minimum threshold for compute needed to train a good model just keeps going up. Like it was 100 million, billion, now it's. probably like 5 billion. And so that- To train a model now? To train like a frontier model, right? Five to 10 billion. Right, right, exactly. Versus like- Million or billion? Billion. Billion, yeah. Versus your second point. The frontier is very expensive. Yeah, frontier is very expensive. To replicate the frontier six months later is about 1 20th the cost. No, I think Sarah has a great point, which is this is a good argument against this whole thing, which is [15:00] The next model, first of all, there's only one or two such runs a year that each of these labs do. And it's the opposite of the four criteria that I mentioned, right? Which is it's going to take more resources, more humans involved, and it's even more brittle. And they have to build out the data centers. I mean, like the labs are not necessarily doing that, but others have to build the data centers. And they have to be gigantic. And they have to get the GPUs. And they have to get the networking right. They have to do the engineering to make sure that they can tolerate. Because, you know, every order of magnitude, more GPUs you cram in there, you have to not worry about errors that before you didn't have to worry about. So you have to increase robustness of that. So it's like a very brittle process. And if it fails, you've squandered so much money. So they're like very, very careful with that run. And there's been multiple runs that have been botched. So it's the opposite of that, that, hey, the next model is faster, cheaper, smarter, and recursive improvement. It's the opposite. It's like it's taking longer, and it's more brittle, and it's more people, and it's harder to pull off. So I do think that that is true. With respect to RSI, with respect to actually cyber risks and things getting hacked, we need to take it super seriously, yeah. So I'm going to have like another litmus test here. I actually love your four criteria. I was literally just waiting to argue with it, but I actually think it's very good. So let me give you like a black box When you dealing with these dynamic adaptive systems what are are you gonna believe Are you gonna believe like the numbers are your lying eyes right So I think you kind of have to go to the numbers on these ones. So like, what are the numbers to look at? I really think you should just basically, and maybe going public is the right way to do it. Like if these companies continue to grow, reduce the number of people and the number of amount of money that goes into them, then I would say something is definitely happening here. Like I do think that like you can actually black box this and take a look, but none of those indicate, like they're hiring like crazy. They're hiring like crazy. That's not fair. Cause you know, companies are not necessarily efficient, right? So like, what if you have, I mean, OpenAI itself was doing like a million different activities. It's a very small team of like 10 people were doing LLMs and the LLM stuff was useful. Twitter, there was a lot of people. Now it's much less people. I just have another, I agree. Just another litmus test. We have kind of two litmus tests. We have your litmus test, which I think is great. But then you would actually have to have a way to instrument it. And then we should have the black box list list test. I mean, listen, if Anthropic in two weeks is 12 people and they continue to grow and they're putting out models at an increasing rate, I think we should probably take notice of that. That's sufficient criteria, but it's not a necessary condition, right? I'm just saying that there could be that. And really the right way to do this is then to look at, okay, the pre-training and the post that been doing that really necessary Because they have so much resources that they might be doing a lot of other stuff they don need to do but they doing it just and they can just hire the people because they have infinite money and infinite so really the people that are training the next model is that team tiny tiny and it's actually getting reduced and they're doing less and less work and just the ai is doing it and then post training and then they're all just using less dpus that's not the case we've had this argument many times as an industry before i remember when like like we learned how to really cluster computers because like the mainframe was actually kind of limited by things like memory coherence. Remember that? Like if you only make it so big, you know, and then we kind of went to the client server area and then we didn't have that problem. And then we started creating super computers, which were like basically just like, you know, clustered computers. Yep. And at some point internet happened. And that was kind of also roughly like when GPU started getting good. And do you remember that we would actually like export control PlayStations? Because we were worried that Saddam Hussein would use them to do simulation. And the arguments were very similar, which is like, these things are are getting infinitely powerful we're using them to simulate nuclear weapons which we were like i was yeah um like we can't you know this stuff has existential risk actually they didn't use those words but like this has the potential for like nuclear weapons or whatever and we should stop it and like none of that came to path so i think a very reasonable discussion is is is is this time different yes or no i don have an answer to that but i a pc but you know you a Yeah I mean look I think I was I not old enough to remember So ignorance is bliss Wait. So I can take this... Come on, I'm not... I don't recall PlayStation being illegal and... What? I just don't know. How old are you? Maybe I'm just ignorant. This is like 1999. Okay, maybe I'm old and ignorant. I mean, you know... Maybe they didn't care in Sweden. Is that the conclusion? Maybe it's just amnesia from age. But whatever it is... Sweden doesn't care about export controls in the US. Yeah, you know, whatever it is, I think it's at a different scale now, right, with AI and, you know, with the, you know, what we're doing, the pace of development and so on. They are freaking out the frontier. I do think cyber is actually one of the biggest ones that we're going to see, right, because there's just so much infrastructure on the planet. By the way, way, way more than it was whenever, whatever, Saddam or Xbox or whatever it was you're talking about. I mean, like, we've just interconnected way more things and they're dependent and, like, the planet just looks different today from internet, tech dependency, interconnection, than 30 years ago. So I just wanna make this point. There's so much infrastructure that's insecure, right? And if you're gonna unleash these agents, they're gonna find loopholes, they're gonna find exploits, they're gonna break in here and there. So this is a real risk. And you can't just... [20:00] way. This time it didn't do that. But you could imagine a scenario also where it starts hopping, like it takes resources and it starts executing itself elsewhere. So it's kind of spread like a virus a little bit. So that's a real risk. So and this is pure curiosity. I promise I'm not, you know, trying to be a foil here. But like, why do you think we just haven't seen very much then? Like, again, again, I'm much older than you. I remember very well. So when the when the like, Like literally when the internet came out, by this point, we had literally taken out 10%. We've disabled hospitals. We've taken out critical infrastructure. We've caused tens of billions of dollars in economic damages from worms. Like all of that had already happened. And to your point, we had much less build out, you know, less of the economy was on it. And so, you know, AI has so many people that want to find risks and threats. We're running so fast. So much money has been poured into it. And we haven't seen anything commensurate with the early days of worms. What is that disconnect? Yeah. Look, so I do remember that, those days. At the same time. Yeah. So look I would just say that I am sleeping well at night And I don think there existential risk right now I do think there a lot of infrastructure that needs to be secured We have a product in the market in the detection market Lakewatch that helps you do detections. And the space just there is moving so fast because you used to have the SOC team, security operations center people, that would look at what intrusions are happening, how are we being attacked, and so on. And now the humans just can't keep up. So this is the whole space, security cyberspace is being transitioned into fully automated, using agents for detection on the other side. If we don't do that, I mean, now we're rushing, we are rushing, the industry is rushing to do that super, super fast. If we don't do that, I do think you will start seeing those kind of things like sites going down, you know, you know, whole systems that stop working for a while, and there will be consequences, not existential, but economic damage and you know, people getting hurt and so on could happen. So we just have to race very, very fast to do all of those things. It's just, you don't have the, the humans don't respond fast enough to the attacks that are happening. So it just, you need to automate all of those. And most organizations are actually not close to doing that. The banks are doing it. Some of the people that are super security conscious are doing it. But most of the industry today is running with old school security operation centers and people that are waking up every day And there like hundreds of emails of detections that have fired Many of them are just false positives So you don't need to, you can ignore them, but some of them are not. They just don't have time to go through those. And you need to identify that. You need to have threat hunting that's automated where you're actually attacking your own systems automatically with agents and so on. It hasn't happened. So I do think like if we just say, hey, this is just like the internet in the early days, you know, bad things are gonna happen. So there is a race going on. Yeah, I was actually very surprised. Earlier this morning, I was on a conversation. I feel like you and I are pretty close. We talk periodically. I feel like I know a fair bit about Databricks. I was on a call this morning where a founder was basically like, yeah, listen, we're doing all of this observability, agent threat detection. Am I using Databricks? I didn't even know that you had this offering, quite frankly. So like, I mean, just from an education standpoint, like how extensive have you gotten in like the agent AI observability, security, safety thing? Yeah, I mean, we gave a talk this year at RSA, actually with Ben Horowitz, but the issue is that data and AI is blending with cyber. These two markets are collapsing. Because I think most of them at least Yeah and the reason they collapsing is that it used to be like okay we have like data and AI the kind of stuff Databricks and these kind of companies used to do which is like okay you have a bunch of data and you run AI and machine learning and that just lives separately And then you have the cyber world Cyber world is you know we want to detect if something bad if like bad people are trying to hack us if bad people are doing things we need to detect that Okay But now on the data and AI side we have agents running internally in the company people are having agents running And the agents are also like doing things with other people agents and they're producing a lot of data, logs, trails, fingerprints that are being left. And so now we have internally these agents that are doing that. So these worlds start merging more and more, which is like, okay, well, all the data that's being produced needs to be analyzed. And the scale at which you need to do that is just like many, many orders of magnitude more than just one or two years ago. So things have changed dramatically. Like 2018, 19, the time it would take from a CVE, vulnerability being sort of published until you see it actually weaponized in the industry would be like two, three years. That went down to, you know, 2022 significantly, but it was still like eight, nine months. So that's kind of fun. You have eight, nine months from a vulnerability to that was 2022. Now, if you look at this curve from 2022 until now, now it's down to like basically hours. [25:00] So it's like down to like basically no time, like things get immediately weaponized. So you need to just do it in an automated with the data and AI sort of platform approach. So these markets, I'm going to argue, are just going to collapse. So it's a very specific question. I actually think a lot of like to pull back on like the existential X-Risk discussion, it feels like there's almost two camps. There's one camp which believes that this actually is an engineering problem. And like companies like Databricks can solve it. and they can solve it through product and through engineering solutions and through services. And so like we just as an industry need to solve that problem. And there's others which actually believe, seems to me that there is no engineering solution. You have to slow it down. You know, you have to use regulation. It's more like a nuclear weapon, et cetera. So like, does this mean you believe it is an engineering problem or are you not quite comfortable saying that yet? Because why do you even buy Databricks, man? Let's just put this stuff in a national lab. Just paste it is the only solution. I'm just kidding. Everybody just paced themselves a little bit, then we'll be fine. The bad guys too. There is no line of inquiry ever that gets to pacing, I don't think. I think it's like you pause it or like you solve it. What we learned here is that Martine really hates the word pacing I will never use that word with you ever again Okay clearly duly noted More frustrated More frustrated Is it an engineering problem that can be solved by engineers or is there more to it I actually think which problem are we talking about? There's two separate problems that I think are being conflated. There is the superintelligence problem. And I think a lot of this comes from like Bodstrom 2014 superintelligence book. And if you look at the definitions, like I think people don't have these clear definitions of what superintelligence is. If you read his book, those definitions are kind of crazy. so I think what he had in mind when he said super intelligence is you know AIs that I don't know I don't know what the examples were something like something like they write the whole PhD thesis with novel like peer-reviewed stuff in a couple seconds and they can do like millennia worth of thought you know in like instantaneously and you know so this is like the level of you know how fast they are, how intelligent they are. It can learn. Yeah, yeah. It's just, yeah, I mean, yeah, but it's just many, many, many orders of magnitude, right? It's like, it's just the scale of the problem is just completely different. So if such a thing exists, do I think that's just an engineering problem to solve? No, I think that's actually very, if such a thing would happen, that would be very existential. Of course. And that what everybody agrees on So I think that being mixed with now we have agents that are nowhere near that It not even like there nothing like that And we don have anything towards that path right now But these agents are capable and you can do something with them that you could never do before in a history of mankind. So I do think an inflection point has happened. Something has changed, which is we have good security researchers at Databricks, but I could never say, let's get 10,000 of them in a sandbox for a month and have them do $100 million worth of salary wage work. We can do that now. We just turn on a button and we can get 100,000 of them. Or mathematics. Like we can say, hey, you know, we want to solve a conjecture. Okay, let's get pretty good mathematicians, but let's have 10,000 of them collaborate, you know, and then you can like make very fast progress. So this, I think, this leads to all these cyber risks. I think cyber is the major problem here. This is, I think, you can solve with engineering. And I think it's like we are working on it. Many others are working on it. There's still risks. They're not existential. I think we should do it. There is the super intelligence thing. That's the thing that could do, write a novel PhD thesis or like reason intuitively in 11 dimensional space physics instantaneously without writing anything down, something humans can't do. Like that kind of super intelligence The question is is RSI and recursive self that the labs are doing leading us there Are we going to get there Trying to do Trying to do Is that what going to happen And how fast is that going to happen? That's the big question. And they've suggested that, you know, hey, we should have inspectors that come in and look at what we're doing. I think that's a good idea. Have them go in there and get the data. I would love to, like, the question is, who are the inspectors? Because you can, like, you can stack, right? You can stack that. Who do you think, Martin? Yeah, there's a bunch of people that like on either camp, actually, I wouldn't care if they're the inspectors. I would not be very impressed by what they say, because they've already made up their minds even before they are. They would go in there. Right, exactly. But let's say like as an example, if Yanda Kun, who was one of the inventors of this, you know, deep neural network technology, right? One of the pioneers. If he said, hey, there's nothing to see here. There's no risk. You know, I'm paraphrasing him. This is nothing. This is super intelligent. This is just nonsense. Keep on going. Go fast, fast, fast. None of us would believe it. I'm putting words in his mouth. I mean, I'm not exactly right. Now, if he was one of the inspectors and he went in there and he had a look and he came out and he said, hey, I've looked and it's just what I said. There's nothing to see here. Just keep going. I would feel very good about that. I would say, OK, well, I would feel very. Or if he comes out and says, oh, my God, you know, I you know, he's wobbling and he would change his mind a little bit. That would also have a lot of interesting signals. So I. [30:00] It comes down to who we pick as inspectors. And I think it's a good idea. Let's have some of them and pick a diverse set of people so that we can get different nuanced points of view. What do you think about this kind of Elon Musk view, which is like, it's less third party. It's more, it feels like there's kind of three proposals. Like the OpenAI, Anthropic one is a third party. Yeah. The Elon Musk one, as far as I can tell, is the labs cross-check each other like peer review, like you do in science. Yeah. And then the Mark Zuckerberg one is police yourself, right? What do you think about this middle one? That they should face each other, like, you know, evaluate each other. I mean, I think, like, if we have boxing matches in the ring, the boxers should just be the judges of each other. Would that work? No, they would scream foul all the time. Foul, foul, foul. Like, you know, it's like, ah, you know, the moment the other guy puts out the great model, and it's like, ah, big super intelligence risk. Absolutely, like, you know, they have not been responsible. Well, like, you know, when vested interests are at play and there's like IPO plans and these two companies are so competitive and they have like this history also between them. Yeah, they'll be very fair to each other, I'm sure. That why you need a third party right Why do we have judges in the world at all Why do we have third parties at all Like why can just people like figure things out between themselves But I mean they should try If they want to do it they should try But I skeptical that they wouldn just you know be biased you know, in multiple ways to self, like, you know, judge each other. So I have to ask Ali, do you think something Elon also said, I think it was on the All In, you know, summit, he was like, this is some elaborate 4D chess, because on the one hand, you're saying all of humanity will die. On the other hand, you're saying, hey, what do you want for your IPO allocation? Right. And so, I mean, that is probably a more cynical view. But like, how do you how do you reconcile that? I mean, the dissonance, I think, gets a lot of people like, how do you think that gets reconciled? Look, I think all of these things get mixed. Like, I think there are people that are freaked out. And I do think that the people are saying like, hey, if there was regulation that would pace us, sorry to use the word, that would be good for us. Right. That would be good for us. But I also think that people have that sort of interest. Right. These things like, you know, usually people figure out a way to always get all of these things to align in their harmonically in their head. So yeah Do I think that there has been a tendency in the past of in general using also marketing stunts by saying you know oh my God this latest model is so good that I trained It like unbelievable It like almost scaring me And then the whole world like kind of starts focusing on it Yeah there been that kind of marketing going on But at the same time also, as I said, the time from CVE to actually weaponize exploit has like been going down from years down to like minutes now, just in like three, four years. So it's real. The cyber attacks are real. And this is, but there's also a great marketing ploy to, you know, whenever you train a new model, make lots of noise around how much of a crazy risk it is to the world, it helps you, right? So maybe they're not in contradiction, these things. I mean, you and I are networking folks, and there's a long history of forming third parties to help arbitrate things, right, like IETF or IEEE or even like ICANN. I see where this is going. No, no, no. So my question to you is like, so I think it's actually, this is a very sensible proposal that they actually have. I actually agree with you. You probably want to make sure it's independent, which is not fair right now. And there's going to be a lot of arguments of who you put there and everybody's going to disagree. Right, right. But you said, you know, why do we have judges? So that like actually like the state stepping in is actually quite a different thing than basically industry self So like at what point in time do you think it makes sense to actually consider federal involvement Or do you think like now is the time to actually consider actual federal involvement as opposed to like more industry self Well these are very different. They are different, but they kind of bleed into each other. Like, you know, like for instance, FINRA, you know, is not like a completely independent self-police. It is, but you know, it's linked to the government. So like, I think these things like kind of will bleed over. I think it's hard for them. You think that they evolve into, historically they've, The industry subplaces and then it evolves into regulation. Look, if they are saying there is existential risk, which they're saying, you know, and then saying, come police us and regulate us. I think it's very hard for regulators to say, no, we're not going to do that. So far, I've said that, but I think that's not going to last very long. It wasn't David Sachs. It was like, I've never had a CEO ask us to regulate them. And my favorite thing is, I've never had a regulator that says no to that. I mean, the reality is like the actual like metapolitical machinery is actually in motion already. Right. I mean, like everyone has a talking point. Obama has came out, it is a major issue. Like, do you think that there's a reality that it's too late? This will be a major issue in the midterms and we're actually going to like heavy handed federal regulation and this is all going to be paused, you know, and goes into the, Anthropocos into the DOE and we're past that point. Or do you think we can actually end up with like a sensible self-policing regulation? Because [35:00] This is a headline. You cannot. We should strive towards doing the right thing. I think there's still some degrees of freedom of how things evolve and there's still time. And yeah, you're right. Largely, if you have these companies where we're pumping in so many billions of dollars and the way reinforcement learning works is that you give it the reward function that's verifiable. Like we're going to solve this math problem or this kind of, you know, this narrow area of programming and so on. And we pour in so much money into that. You can get quite good results in that narrow kind of. That doesn't mean that you're getting that super intelligence. No, but you can even trick yourself into thinking that less inputs are giving you a better outcome just because you're running so many experiences and you've thought about it so much. But it's actually very hard to do a closed experiment this way, given how many resources are going in. Well, the fundraisers are going up astronomically, to your point. Yes, yes. That's why these companies are going public, right? I think they would otherwise. I mean, as someone who runs a private company at scale, I think they would prefer to stay private otherwise. Why are they going public? Because they need the capital. And they consider the scaling laws and the capital to be a strategic advantage. So that's why they're going public. But I would say, let's go back to the four things that I listed. If those are true, if those four are true would you want to know about it And would that be worrisome That could get out of hands Now there no evidence that those four are happening But if there was like you know no that is actually where it headed Yeah, I actually think understanding for any system, like any sort of self-propelling property is important. I mean, we've done this in the past with dynamic systems, right? Like we've done this with like whatever, compilers. We did this with all the research on nanotechnology. Like it's been a common interest of ours. Yeah. And I don't think that that's new, that it's an interest, which you continue to have the interest. I just think the fear is that these particular systems are net economic systems that are so complex that the risk is is crying that you're seeing it when you're not seeing it. And I think a lot of that's happening right now. Yeah. But of course, if you see it, you know, of course, I mean, you want to know. But it is fair to say that the labs are now focusing a lot on RSI and that's where they're headed next. And maybe they're just unjustifiably like word themselves, just like they were worried about GPT-2. Right? Like they were like, GPT-2 is world ending and then it wasn't. And GPT-3 and 4 came out. So I don't want to quibble. A lot of the times when they say RSI, they're actually talking about autocatalytic effects. And autocatalytic effects have been in our industry for a very, very long time. So for example like there no way you can create a computer chip without a computer chip Yeah It like you cannot do it Anyone with computer science degree You right Compiler writes its own compiler Well that becomes closer to RSI but like the steam engine was autocatalytic right So listen, my full time job is people coming out of labs and starting companies, and they all say RSI because everybody says RSI. And like maybe 1% of those are actually RSI. Like they're more like we use AI for data cleaning. We use AI for making... Let's make this distinction. you're saying it makes basically catalyst in a sense that they're using AI to speed things up. It's auto catalytic. Yes. So I would say of what we hear from the RSI. Like a model trains another model, right? You're using a model to build a GPU kernel. You're using a model to do data cleaning just like I use a computer to design a computer. It's auto catalytic which is every tag. The internet was auto catalytic because it allowed people to collaborate remotely. So I would say, this is anecdotal, 90% of the the calories are auto catalytic, which is a hundred percent what you would expect. And that's been going on for a while though. I mean, that's not even new. Yeah, but I would say, but also now there is a focus on let's move towards actually, can we get the model to train itself? This is kind of like the auto research that Karpathy did, but now they want to do that. There are teams that do that. It is not nearly as many calories as you would expect And I just feel like I actually have a good sampling of this because they all come and talk to us Great So maybe you can be one of the inspectors Can we get all that data and all of us look at that data? Maybe there's nothing to see here. I personally don't think it's like very high probability that those four criterias are happening. Greg Brockman went on the pot this week and said we're in AGI era. Yeah. So now I asked him question after he said that, and now everybody's saying we have AGI. So, you know, people follow what this is. But people have for a very long time, when I asked this question, said that AI is smarter than most of the people around me most of the time. That's been almost like since Q3, Q4 last year, they've been saying that. Then I asked them, how many of you have, you know, hundreds or thousands of agents that you are managing that are coordinating with each other in swarms and negotiating and, you know, automating your life and everything around you? And if so, raise your hand. It's like almost nobody raises their hand. Of course, Martina has done that at home. No, no, most enterprises are on Microsoft Copilot. Yeah. Like, that's the extent of their AI. Most enterprises I talk to when I ask this question, they're like, no, we don't have any of that. So we'll be like, what are you doing then? They're using a chatbot. Like they're asking questions from a chatbot. That's basically very, very glorified, efficient Google search of the whole day, the results is just faster Google search. And then coding is happening. So people are using it for coding. Though the ROI is, you know, we can discuss the ROI [40:00] But there's no like agentic like work that's automated the whole enterprise. That has just like not happened. So then why is that? And I think that the real reason is if you actually look at it is the models are smart enough. But they just don't have the context that exists inside of any organization. Like they have not been in every meeting. They don't know what's in everybody's heads. They don't know all the processes. They don't know. There's always like a couple of employees who know everything in every organization. You know, you tap on their shoulder and they're like, oh, my God, what would happen if he or she quits? you know, they don't have that context. And if you've just fused that and gave that context into the AI models, just the frontier today, I think there's so much productivity gains you could get for any organization on the planet. For that, we actually don't need smarter models. So we don't need a smarter model that can actually solve Navier, Stokes, or conjectures, or do better on humanity's last exam. Like we needed to just go from 60 to 70%. None of that is needed. So I think actually people are very upset on some, like, oh, if we pace the frontier, but actually if the frontier doesn't advance it doesn't actually matter i think for vast majority of organizations on the planet they're just so far behind in the adoption curve of actually automating things yeah and getting value out of this stuff but it would be disastrous in labs because the price of intelligence is dropping asymptotically i think it going down by one tenth every six months or something like that So that will dramatically change their businesses if you weren pushing the frontier Yeah, but this is what we should focus on. We should focus on, there's two sides. We discuss here a lot the costs. There's cost benefit analysis that we should do on everything. We've discussed the costs a lot here. Is there existential threat? Is there cyber risk? Are there things we should be be worried about and so on. That's like the cost side. What's the benefit? And I think now that this has become like a public thing and the whole public cares about AI, they're asking, hey, what's in it for me? What am I getting out of it? It seems nothing. So how do they get there? What are some of the use cases you've seen to date that have maybe surprised you to the upside? Yeah, I mean, first of all, there's like so much worry about, you know, existential risk and so on. So I think a lot of people just don't know what a, you know, cool use case is where people are actually doing interesting things. We have a lot of use cases that are, I mean, just fascinating. One that I like is Crisis Text Line. So, you know, they actually use large language models with us to detect if teenagers want to do self-harm or suicide. Oh, wow. Yeah. That's an awesome use case. Super cool. And it actually saves lives So that a great company And that you know that organization is doing amazing work Another one that kind of interesting is the Omnipod which is for diabetes patients that can put the Omnipod and it uses AI to really learn your insulin release and your glucose levels and actually exactly release. I don't know if you remember, people used to like stick themselves, right? But this now happens automatically. And it's like, you know, self-learned AI for your body. you know it's a cool use case Zipline is another one they're doing awesome oh yeah you know but when they started it was like these drones that had you know they were completely automated all AI driven everything from the you know battery optimization to the routes and everything and they were delivering food in you know areas of need blood to refugees blood to refugees yeah started in Africa and then elsewhere in the world so yeah that's awesome that's all AI it's an AI use case you know built on Databricks so it's a cool one but there's more advanced ones also like one that I kind of like it's harder to maybe explain is this model that we built, transformer based model that we built with Merck. It's called TEDDY, Transformer Enhanced Drug Discovery. Love that. Yeah. So, and they published actually the research, so you can check it out. But it basically, it's a model, instead of predicting the next token in English, it predicts what the gene regulatory network the DRN is going to respond And it can really detect which cells are causal and which ones are just reactive So they're just reacting. And therefore they can start using this in drug discovery and get costs down significantly for developing drugs that are targeting specific diseases. So that's a super cool use case. There are lots of these. Jeannie, I mentioned you have this ontology and you can ask any questions. Novo Nordisk is using this. So, you know, they built this GLP-1 drug, but what Novo Nordisk is doing is now they're using it for all of their trials that they're running. And it can compress the time it takes to get insights, versus if you're doing an obesity study or something, from weeks down to minutes. So there are a lot of amazing use cases of AI. We should not forget these upsides also. Like we want all of these and we do not want to paste these. Yeah, exactly. You're totally right. And so how do they, let's say if you map out the next 12 months, how do the enterprise actually get value? You dropped the word context, but like, how do they operationalize that? It's actually harder than most people believe. But, you know, first and foremost, we have to make sure that we have digitized everything that's happening in an organization. That actually, you cannot actually just, you know, [45:00] have a magic wand and make that happen. So, you know, every meeting has to be transcribed, you know? So you have to be able to get all the context of all the meetings and everything that's happening, all the digital content has to be fed to the AI. So you have to build, we call it an ontology, we build that. But first and foremost, you have to collect that. That itself is a problem in many organizations because legal teams will say, don't record every call, don't record every thing. So you have to do that in a way where... Can you define ontology for everyone? Because I know Palantir says the word a lot, But it's not like they own the word ontology. What does that mean for the people listening? Yeah, I mean, ontology just means that in an organization, the relationship between all the abstract concepts of all the goals and all the departments and all the people and all the projects that are going on, what do they exactly mean? And what's the relationship between them, the people, the resources, and what that company does? So it's the difference between a person who is a new employee in the company and just started today and a person that has worked there five years. You know, let's say they're equally skilled. They have the same educational background. They're equally smart and hardworking and all of that. But one it you know her first day today at work The other one she been there five years What the difference between these two people That one has an ontology of how that organization works who the people are how you get stuff done Don't look at the org chart. Don't go ask that person. He will not get anything done. You go ask this person, he'll get it done for you. And that's not how it works. You don't need to file that paperwork here. And this project, this is what's going on. This is essential. so there's just a lot of ingrained knowledge that's sitting in everybody's heads who knows how an organization works that's why it's people say in startup land they say hey if you lose most of your people that company can't recover from it it doesn't you can't just replenish and hire new people like the people are so essential how do we get that context that's the ontology and give it to the eye part of that is we just have to have you know the recording and all of that but the second part is how do you actually distill it down into a graph actually a digital graph that you can then feed to the AI. So the way a lot of the agents work today, like a cloud code, any of them, Codex or Pi or, you know, OpenCode, you can go through the whole slew of them. You know, they have this loop, agentic loop, it can reason, but then it goes and checks every resource one at a time. So we'll go to this MCP server for your question and try to see is the answer here Is there another one It synthesizes it and gives you an answer But it kind of slow I liken this to if Google would have built Google search this way 25 years ago, we would have said, okay, we're gonna get 10 blue links. We searched for key terms here, but instead of giving you 10 blue links, it would have gone to one website, summarized with an LLM what it does, found a few hyperlinks, jumped in parallel to a few of them, read a few websites, done that for 10 minutes, and then giving you like its best 10 blue links it would find. Well, that would be very expensive, cost a lot of money to do that every time, go on the web. Two, it would have taken a long time and we gotta wait to 10 minutes. And three, the quality would be bad because you're actually only looking at a very small subset of everything that exists out there, right? So how do they do it? They have an index, right? You never leave Google servers. You search for, it hits the index, the reverse index immediately gets you the 10 blue links within less than a hundred milliseconds. We need to do the same thing for the AI. So the ontology is that we need to compute that index offline all the time. So it's almost like the PageRank algorithm that Google had invented back in the day, but it's more complicated because Google was just looking at a web where everybody can go on the web. Here, there are permissions. And the links existed and... Yeah here there permissions involved The data I allowed to access might not be the data that you allowed to access So there privacy there access control Also there many different types of objects here that we dealing with not just websites So the problem is a little bit harder, but it's manageable, you can actually do it. So, you know, I'm convinced you can do this and you can get massive productivity gains out of it because we did it for Databricks. Yeah, exactly. We did it for ourselves. Yeah, and we're like, the company has just completely changed. It's like not the way it was, I would say, a year ago because of this. I mean, you've been dogfooding Databricks for Databricks forever, but maybe say more about the impact you've seen as an organization. Yeah, I mean, once we got this ontology and we started working on it and we actually have probably the largest of all of our customers, we have the largest ontology. Our ontology is bigger on us than any of our customers when they use us to build their ontology because Databricks uses Databricks more than anyone else uses Databricks. And so it's like millions and millions of nodes in the ontology graph that we have. So it's just, what happens in an organization? What happens in an organization, you have a tree structure organization and information flows up and down the tree structure. If you can't make a decision, you escalate your boss, maybe they can tie break, it escalates up. They need to get up to speed on what's happening and they need to get all the context and then they make decisions. Once decisions get made, [50:00] to percolate them down in the organization. A lot of this can now be done by AI if you have an ontology. Why? Because, you know, what happens in a meeting? In a meeting, you go through some, you know, someone has done the analysis. They probably have a PowerPoint deck with some pretty graphs in it. That person that did the analysis is some smart person that used Excel, made some models, there's some numericals. So a lot of that you can now just do with AI. So the AI can do the analysis for you. It has all the context. It can present it in a way that you want. You can ask questions about it instead of having follow up meetings. You can directly ask questions directly from the AI. So it's very similar. It's along the lines of what Jack Dorsey has said that you can do to the organization. It's just a concrete way of implementing it. Yeah. So it's it's game changer for us. Like, you know, it's just everybody's on their phones now in the meetings on Genie and they're like asking Genie questions. You can see as soon as someone says something complicated or something, it's like, you know, you see everybody go to the phone. Can you share that finance, like the finance anecdote? you mentioned once in a board meeting. Yeah, sure. Internal board meeting. Only in kosher. Yeah, exactly. No it actually needed for one of our presentations I need to know how many customers do we have in Fortune 500 What our penetration of Fortune 500 And I asked one of the people in sales ops because I thought she would have it And she texted me back and said oh sorry I can't log in to Genie right now. I'm on a flight. And I said, well, if you're just going to log in to Genie, I can do that myself. I asked you because I thought you had something alternative that I don't have access to. So then I was kind of a little bit angry. So I texted the CFO instead, Dave. And so I texted Dave and I said, hey, do you know what our 4chan 500 penetration is on? And he just copy-pasted a screenshot of Genie back to me. So he also asked that. So I said, does anyone do anything novel here? Is everybody just going to Genie and asking the ontology for questions? It's like, let me Genie that for you and let me go for you. That's what everybody's doing. Now we just say, we say, hey, can someone just Genie this? Can I just get it from the ontology? So I do think it's a game changer. but it's not just you press a button, you have an ontology in an organization. And I think Palantir actually has done a great job of going to organizations and getting a lot of that tacit knowledge written down and getting it into the organizations. We automatically take that and build the graph and then we feed that graph into the agents so that we can answer the question and answer it in a way that business leaders would like to see it, which is in graphs you know analytical way and a way where you can interrogate that question and you know continue asking questions and getting answers to those so you can make decisions and then disseminating that information in the organization Yeah, it's pretty amazing. You sort of bookmarked the developers are obviously using AI questionable value. I want to follow up with you on that because I feel like you guys were one of of the earliest. And I say, I don't want to use the word token maxing because it has such a negative connotation. But I think in terms of applauding people who can use AI to become more productive, you guys were at the forefront of that, right? And then, of course, there's this cycle of, oh, shoot, people are being wasteful. Now we need to value max. What was your own journey on that? And how do you guys think about value maxing, not token maxing? And then I'm going to throw in unity gateway in this, right? Because I think the managing of cost piece is actually getting more important and you guys are helping people do that, but maybe tie that in to extend it. Yeah. Yeah. So around Q4 last year was when, you know, the models got really, really good and we started noticing that, okay, it's actually starting to give much better productivity. So I actually started using the models myself to sort of start you know commit code into production for Databricks like the actual as I want to take it all the way to production So I did that and started pushing the organization that hey everyone needs to do that I have done it. Why are you not? Like if the CEO can commit code to production on a very sensitive data platform that has all these security requirements, you should be able to do that too. You being any manager, anyone in the organization. So I started pushing everyone very hard and we started making leaderboards in Q4. And at the beginning of, I say, January, February, when kicked off the year, we were already full swing. Everybody was using the stuff and we're pushing and we're managing this. But, you know, the whole token maxim thing was happening around, you know, February, March period, already it was happening. So, yeah, we just had the luck of being maybe a few quarters ahead of folks to see what was happening here. And it was getting out of hand. So we already had a gateway. So it's called Unity Gateway where we were already, this gateway was being used to provide token capacity. So you can get OpenAI, Anthropic, Gemini, Grok capacity. Like any customer can come to us and we'll just provide them that capacity because we have relationship with those and any open source model. So we started putting in budget constraints in place and giving people warnings like, okay, you have this much of your budget left. You're getting close to your kind of ceiling. So we started doing that per person and for group. And then we started doing great analytics so we could predict exactly [55:00] exactly where the costs were going. And then, yeah, and then we added smart routers that could actually pick cheaper models if you're getting close to your budget or if you have simple questions, we started doing that. We also built a harness called Omnigent, which can multiplex between the different harnesses. Turns out actually the harness itself matters. So if you use the same model, but different harnesses, there's almost two X different cost difference. Even exactly same model, same version, But different harness, you get 2x difference in actual cost. So if you can change harness, you can get a lot of leverage in the cost. So we started using all of this. That way we were able to actually bend the curve. And actually our cost for AI has been basically the tokens have continued to go up, but the costs have been sort of stagnant. So that's been actually super, super important for us. And there's a huge demand for this. I think every organization is going through this now. Yeah, for the first time, I do a lot of board meetings. some on 20-something boards. For the first time ever, a company at scale last week said that they're moving from the Frontier Models to GLM. This is a large engineering organization. Do you see this Do you think that that a trend or do you think that just like a one anecdote Because I been hearing about I remember the first DeepSeq moment and like NVIDIA software and then that turned out to not be real And then the Kimi moment then the next DeepSeq moment, none of it seems to have actually had an appreciable impact on the market. But now the amount of anecdotes that I have are pretty real and it seems to be happening. Love your view. I mean, I think people want both. They want, you know, they want the latest model that's super intelligent for the difficult task where they get ROI. But then there's a lot of mundane dumb things. Like, you know, people literally use their harness to rename files and whatnot. Yeah. You know, like it's, you're paying, you know, orders of magnitude more for that. Please type that in yourself. Don't have the model do that. It's gonna spin for five minutes and then it's gonna rename the file for you and it costs you, you know, cents. But- I get some people, I'm just wondering, do you actually see market movement? Yeah, no, people are moving on it, but what they're doing is that, you know, the pattern is either use, you know, you can use this expert pattern where you have small, cheaper open source model that uses an expert model, the big ones, or vice versa, or a way in which they can sort of ping pong them to each other, but also multiplexing harnesses and just changing harnesses so that you can control the costs is also what people are doing. People have found for instance there Pi is very efficient when it comes to as a harness So yeah yeah i think there going to be a multitude of these it easy to the models themselves are stochastic as you said every time they give a different answer and they're changing so much so there's just a lot of experimentation happening so i think we're going to get to a world where you're not always using the smartest model for everything which is kind of the been the paradigm for the last couple years like new model comes out it's super smart they use it for everything even really really simple mundane tasks. Yeah, I'll tell you what I see. I see people using Fable and Astra for like architecture, a cheap model for implementation, and then Fable or Astra for audit. Yeah. Like that seems to be like this emerging. What are you guys seeing in the startups? I mean, aren't they? That's it. Yeah. That's honestly the pattern. How much open source? By token or by dollar? Either. So by dollar, open source is like 5%. It's very little, but by token count, it's over 60%. Yeah, I was gonna say, I mean, we talked to, let's say a Decagon or something like that. Well, I think it's different internal use versus external for product. On the external for product, I think they're almost up to 90% open source. On the internal, and I don't wanna say for Decagon in particular, but a lot of them are like, we don care we just use Frontier we not thinking about cost control But as it gets bigger right you and I were in another board meeting where they actually did bring that down just from a waste perspective So I definitely see that moving more toward open source on the product side. And actually that's related to another question, maybe around open source, but post-training specifically. I feel like you were kind of early, I remember talking to you in 2023. When did you buy Mosaic? 2023. 2023, okay. So this vision that you had in 23, kind of came true in 2026. I don't know if you guys would agree. Right. Like that's sort of what we're hearing across, you know, of course, oh, just sort of like, hey, we're going to actually you're going to own your own intelligence. You're going to be, you know, post your open source models, et cetera. And that's definitely what the startups are doing. Yeah. I don't know if that's what the enterprises are doing yet, but like, I mean, do you feel like you were early to that or? Yeah, I mean, first of all, you know, there was when we started, it was also, hey, we're also pre-training for you, which, That doesn't make any sense. You know, you can, there's so, they're very good pre-trained model now that you can use. But that you can do actually post-training on the model and you can do reinforcement learning. Yeah, we're actually doing it at scale. And many of those startups are actually customers. So we actually help them, you know, using RLE or reinforcement learning environments where we can make the models very, very good at the specific tasks that they are doing. It makes a lot of sense. [60:00] for them to do that. If you have a repetitive task, so if you have a startup and it's offering a product and a product does something specific, it's not just a general intelligence. It does something specific for you. It makes just a lot of sense to take a really good open source model and use reinforcement learning and make it really good at that specific task. You can cut the cost down. You can make it really fast. They control their own IP. So in that sense, that is possible. But large enterprises, they just need basic automation and it's just too much for them to do this right now. I think one of the challenges is, you know, you need good evals and making good evals is hard. So while the startups can do that and they're motivated to do that, other organizations, the easy button might be just to use a frontier model than having to create your own evals. We actually generated even, you know, evals for the customer automatically in the product and we had it front and center, but then people didn't want to use it. So we said, okay, let's move it to the back end. So it's optional and then they would never go to it. So I would say in general, the why they just don't want to get into it's too complicated or I think you want quick you know quick uh reinforcement of like you know Hey there a new model I want to try this out I going to get this problem solved You don have time to go do this the scientific method of let make an eval Let have a great baseline And it's sort of like TDD, test-driven development. You know, in software engineering, did people actually do test-driven development? Very few did, right? Everyone said it's the right way to do it, but nobody actually in practice did it. So that's the same. That's kind of a little bit of the curse of, you know, doing training your own model is the evals is the hard part. I know you have an FDE model at Databricks. That's a very popular word right now or acronym. But does it like to get these at enterprises that large, is it a full FDE model that's required or like how do you and how is that evolved? Maybe. Yeah. Yeah. I mean, we've had these FDs and it's the demand for it's gone up significantly. A lot of it is, you know, how do we build that ontology? like the ontology is automatic, but if you're not collecting any information, like you're not recording anything. Right. So that's one of the key things that we do, but also things like, you know, I want to build an agent. I want to put it on, you know, I want it to be customer facing and it's have really low latency and I want it to have guardrails to not, people coming to abuse it or ask it things that we don't want it to answer and so on. So we can build that, like, you know, like sports AI that Fox has. You can go chat with it about sport events You can try to ask it actually about politics and it very good at rejecting you and moving and talking about sports instead So the FDA has built that So, you know, we'll help the organizations actually get started with AI. It is important because it's just many organizations do not have the in-house expertise to build this stuff. So I think just a little bit of help on the side and then they get started. Yeah, makes sense. So this is more related on the agent side, but I saw recently that I think a third party, neutral third party, I think, did some tests that Lakebase or Neon was actually the Postgres database of choice for agents. And I thought that was interesting, one, because, you know, one, exciting data bricks. But two, I probably wouldn't have guessed that maybe a year ago. Yeah, it was a surprise. Just because there's others out there that have, you know, great developer momentum as well. but it was pretty clearly number one. And so I'm curious, how did you guys crack this? And what makes you win across the agents? Because if you win the agents now, you win the market. Yeah. I mean, I think a lot of credits should go to Neon and Nikita and team. And I think what they done is they just been obsessive about how do you make the models how do you make the models pick and agents favor LakeBase or Neon as a database So what do they do The agents want to experiment. You know, they're going off. They're trying to build a little bit of software. They need the database. So you need the database to come up quickly. So they had this obsession that everything should take less than a, you know, far less than a second. So, you know, database comes up in far less than a second. You can clone gigantic database. It's a kind of pedabyte database. You can clone it in less than a second. So it's like highly elastic, highly responsive. And then they built this killer feature called branching. So branching just lets you branch the database and you can have many, many branches over the same database. And they just made this very, very lightweight. We saw this with other things with agents, right? Like UV, ripgrap, like basically re-implementation of a lot of the tools on Unix, making them really, really blazing fast and lightweight. and also sort of fail safe for agents. They've just did this to a harder problem, which is database. So like now you have a Postgres database and the Postgres database has all these advantages that it's really fast, it's nimble, it's fail safe. You can go back to snapshots, you can do those things. So I think that's why it's just easier for agents to use this. They also make sure that it had a pricing model that was like, you don't want just cause the agent [65:00] some software and experiment you don't want the cost to run off yeah you're okay paying for your database if it's like production using lots of people are using it but just experiment uh so i think they were just obsessed they were not trying to win the database war or trying to be better than some other vendor they were obsessed with how are we the best for the agents and that's a new persona because in databases the obsession has been how do we help dbas how do we help app devs how do we help the people that are using the database they changed the game and said hey how how do we focus on agents and help agents get the best database they want? And now over 90% of their, the databases that are created on Neon and Lake Base are actually created by agents. So it's not even humans. So numbers speak for themselves. By the way, it's remarkable. So I've started using Neon as like my standard database and it was bizarre to me because like normally when you enter a large company, things slow down, it's actually like the products got materially better. Are they totally independent? Do they work with the rest of the, like how? No, it's a great team. I mean, we work very closely together. You know, we love databases and data, so it's, you know, we live that. But no, the team does a great job of just making super fast, snappy, and great for agents. All right, Ali, what is your PDoom? Less than 10%. Close to zero. What about yours? I don't know. my only answer is my PDOOM without AI is much higher than my PDOOM with AI. How's that? Wow, that's another one. What do I say? I went on a technicality. I would agree with Ali on this one. Yeah.