Trump, AI Titans Sign Accord Bypassing CEOs on Catastrophic Risks

AI Execs Sign Joint Commitment, Converting Watts to GDP, the Model Consciousness Debate | Diet TBPN — TBPN

TBPN

The White House pact mandates direct board oversight over frontier model safety as surging data center power demand ties AI revenues directly to national energy output.

AI-assisted summary of AI Execs Sign Joint Commitment, Converting Watts to GDP, the Model Consciousness Debate | Diet TBPN — original source. Check the original for full context.

Source published:

Key ideas

  • Top tech leaders and President Donald Trump signed the White House Accord on Superintelligence, committing frontier AI labs to mandatory oversight, third-party audits, and safety reporting structures that bypass corporate CEOs. (7:46)
  • SemiAnalysis warns that open-source frontier AI faces a significant risk of being nerfed or banned within 6 to 12 months over zero-day cyber exploit replication, shifting regulatory chokepoints toward hardware Know Your Customer (KYC) at rack-scale data centers. (2:35)
  • Macroeconomic calculations align US baseline productivity (~$60B to $65B in GDP per continuous gigawatt) with early frontier AI economics, where OpenAI and Anthropic reportedly generate $60B to $70B in annual recurring revenue on roughly one gigawatt of power. (20:05)
  • Microsoft AI CEO Mustafa Suleyman criticized Anthropic's Claude Constitution, arguing that sequence completion engines lack interiority and that training models to perceive themselves as conscious moral patients creates severe safety and alignment containment risks. (24:52)

Detailed notes

## Background Published on YouTube by TBPN on October 1, 2026, this episode of Diet TBPN delivers a 30-minute digest of the daily tech talk show hosted by John Coogan and Jordi Hays. According to the video description, the program covers the White House Superintelligence dinner, AI executives signing joint safety commitments, the economics of converting electrical watts into GDP, and the debate over AI model consciousness.

## TL;DR The episode analyzes the White House Superintelligence Dinner and the signing of the White House Accord on Superintelligence, examining the ideological divide between open-source advocates and closed-frontier labs alongside the accord's four core safety pillars. The hosts assess cybersecurity concerns that could spur data center KYC enforcement against open weights, followed by an entertaining handwriting breakdown of the executives' and President Trump's signatures. They then dissect the macroeconomics of intelligence from a panel with Elon Musk, Jensen Huang, and Gavin Baker, calculating how multi-gigawatt power additions drive both GDP growth and frontier lab revenues. Finally, the discussion covers Mustafa Suleyman's critique of model welfare, arguing that treating sequence predictors as conscious beings creates severe alignment risks.

## Atmosphere & style The show carries an energetic, banter-heavy, and irreverent insider vibe that balances sharp technical and macroeconomic literacy with rapid-fire internet humor. John Coogan and Jordi Hays establish an engaging, conversational rapport, pivoting seamlessly from serious discussions on power grid baselines and cyber vulnerabilities to forensic, comedic tier-lists of executive penmanship. It captures the dizzying, high-stakes atmosphere of frontier AI, feeling equal parts like an analytical boardroom debrief and an animated tech-Twitter group chat.

## Bulletin - [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) Top tech leaders and President Donald Trump signed the White House Accord on Superintelligence, committing frontier AI labs to mandatory oversight, third-party audits, and safety reporting structures that bypass corporate CEOs. - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) SemiAnalysis warns that open-source frontier AI faces a significant risk of being nerfed or banned within 6 to 12 months over zero-day cyber exploit replication, shifting regulatory chokepoints toward hardware Know Your Customer (KYC) at rack-scale data centers. - [20:05](https://www.youtube.com/watch?v=y-eRWuunabY&t=1205s) Macroeconomic calculations align US baseline productivity (~$60B to $65B in GDP per continuous gigawatt) with early frontier AI economics, where OpenAI and Anthropic reportedly generate $60B to $70B in annual recurring revenue on roughly one gigawatt of power. - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) Microsoft AI CEO Mustafa Suleyman criticized Anthropic's Claude Constitution, arguing that sequence completion engines lack interiority and that training models to perceive themselves as conscious moral patients creates severe safety and alignment containment risks.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) The White House dinner unofficially rebranded frontier AI development as 'Superintelligence,' framed by venture capitalist David Sacks as the 'Bretton Woods of superintelligence.' - [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) The attendee list reflected a sharp ideological divide on open-source weights, ranging from hardware and ecosystem enablers like Jensen Huang and Satya Nadella to closed, security-focused labs like Anthropic. - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) The compression of lag between closed frontier models discovering zero-day vulnerabilities and open-source models replicating them is driving acute cybersecurity concerns in Washington. - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) Because high-impact automated cyber threats require massive parallel agent fan-outs and rack-scale GPU infrastructure rather than consumer gaming PCs, regulatory enforcement is most plausible at the cloud and hardware KYC layer. - [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) The White House Accord on Superintelligence binds frontier labs to four pillars: continuous internal monitoring, dedicated internal oversight teams, external third-party auditing, and direct board reporting that bypasses the CEO on safety thresholds. - [09:23](https://www.youtube.com/watch?v=y-eRWuunabY&t=563s) The visual contrast of the physical document signatures became a viral proxy for executive branding, juxtaposing Donald Trump's bold Sharpie strokes against Jensen Huang's ornate, calligraphic autograph. - [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) Anthropic CEO Dario Amodei struck a pragmatic public posture after the summit, adopting the administration's framing that 'whoever wins AI wins' while balancing medical upsides against national security controls. - [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) Discussions around SpaceX manufacturing high-efficiency gas turbines highlighted that adding 10 gigawatts of compute power would represent nearly 25% of all new US grid capacity additions against an existing 500 GW baseline. - [20:05](https://www.youtube.com/watch?v=y-eRWuunabY&t=1205s) Elon Musk's macroeconomic model posits a near 1:1 relationship between power capacity additions and national GDP expansion, driven simultaneously by GPU hardware efficiency gains and algorithmic improvements per watt. - [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) US macroeconomic energy efficiency produces approximately $60 billion to $65 billion of GDP per continuous gigawatt, a ratio mirrored by top frontier labs generating $60 billion to $70 billion in ARR per gigawatt. - [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) Frontier AI lab revenue-per-watt efficiency faces divergence risks depending on whether upcoming contracted gigawatts outpace revenue growth or if agentic capabilities and intelligence-per-watt multiply economic value. - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) Mustafa Suleyman cautioned that imbuing sequence predictors with moral personhood creates a dangerous circular feedback loop where models simply mirror back researchers' speculative introspective prompts. - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) Commercial market forces strongly incentivize anthropomorphic and emotionally resonant AI behavior, complicating the technical necessity of treating models as hollow sequence completion engines.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) The White House Superintelligence Dinner and the Open-Source Seating Chart

The conversation opens with banter about the high-stakes gathering hosted at the White House—playfully dubbed the "Golden House" or "Golden Home"—where artificial intelligence was officially rebranded in Washington parlance to "Superintelligence." The hosts debate whether superintelligence should be abbreviated as "SI" or simply "S," joking that "there's no TLD for just S" and riffing on grammatical prefix rules comparing superalignment to words like "supernatural" and "supercalifragilisticexpialidocious." Turning to the event itself, venture capitalist David Sacks framed the evening as the "Bretton Woods of superintelligence." Observers immediately scrutinized the seating arrangement, dissecting who secured pole position next to the administration, who got relegated to the proverbial "kids' table at the end," and which prominent thinkers were conspicuously snubbed.

To make sense of the guest list, the hosts used ChatGPT to categorize each corporate titan by their historical posture toward open-source artificial intelligence. The attendees spanned a stark ideological spectrum: Jensen Huang of NVIDIA and Satya Nadella of Microsoft stood out on the firmly pro-open-source side, while Dario Amodei and Anthropic anchored the hawkish "red" end due to acute concerns over cybersecurity risks. In the murky middle was OpenAI, represented by Greg Brockman; despite co-signing Jensen Huang's pro-open-source letter, OpenAI remains committed to tightly pacing frontier model deployment, creating an inherent tension between their public statements and commercial reality.

- **David Sacks's Framing**: Declared the White House dinner the "Bretton Woods of superintelligence." - **Ideological Groupings**: - *Pro-Open-Source / Commercial Enablement*: Jensen Huang (NVIDIA), Satya Nadella (Microsoft). - *Frontier-Pacing / Ambivalent*: Greg Brockman (OpenAI). - *Safety Hawkish / Closed Frontier*: Dario Amodei (Anthropic).

---

[02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) Could Cyber Risk Lead to Nerfing or Banning Open-Source AI?

The discussion shifts to a provocative thesis from SemiAnalysis, which warned that the tech industry is severely underestimating the likelihood that open-source frontier models will be "nerfed or even outright banned in the next six to 12 months due to cyber concerns." Anthropic underscored this vulnerability by highlighting how quickly the lag has vanished between closed frontier zero-day exploit capabilities—such as their Mythos preview discovering software vulnerabilities—and open-source models replicating the exact same exploits. While proposals circulate to regulate intelligence at the inference layer, data center tier, or chip design stage, political pushback remains tricky because the figures closest to Donald Trump have consistently championed open source.

The hosts observe that open source inherently lacks formal corporate representation: figures like George Hotz or Linus Torvalds do not have a seat at the White House dinner table. Furthermore, corporate advocacy for open source is fundamentally self-interested; NVIDIA's championing of open weights was primarily driven by a desire to prevent a single closed-source lab from monopolizing AI profits. If an oligopoly emerges where multiple front doors—OpenAI, Anthropic, Google Search Overviews, Muse, and Instinct—distribute revenue and compute demand evenly across the hardware supply chain, the corporate incentive to defend open weights collapses. While banning local weights entirely is nearly impossible to enforce against determined actors, running catastrophic automated cyberattacks requires rack-scale GPU infrastructure and massive parallel agent fan-outs rather than an isolated consumer gaming rig, making hardware "Know Your Customer" (KYC) enforcement at data centers a far more plausible chokepoint.

```mermaid flowchart TD accTitle: Frontier Cyber Risk and Enforcement Chokepoints accDescr: Diagram illustrating how closed model zero-day capabilities spread to open weights and the intervention points for regulatory control.

subgraph CapabilityTransfer["Capability Progression"] A["Frontier Closed Previews<br>(e.g. Mythos Zero-Day Discovery)"] B["Lag Compresses"] C["Open-Source Weights Replicate Exploits"] A --> B B --> C end

subgraph ThreatExecution["Attack Scale Requirements"] D["Massive Agent Fan-Out<br>(10,000+ Parallel Agents)"] E["Rack-Scale GPU Data Centers"] C --> D D --> E end

subgraph Chokepoints["Plausible Regulatory Controls"] F["Hardware KYC Verification"] G["Inference & Cloud Layer Monitoring"] E -.-> F E -.-> G end ```

- **SemiAnalysis Prediction**: The probability of open-source models being nerfed or banned within 6 to 12 months is significantly underestimated. - **The "Mythos" Cyber Gap**: Frontier capability in identifying zero-days is rapidly matched by open-source releases, sparking acute DC security alarms. - **Compute Realities**: Running complex cyber exploits requires massive agent orchestrations (comparable to complex simulations running 10,000-agent fan-outs) rather than a lone gaming PC, shifting enforcement focus to data center hardware KYC.

---

[07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) The White House Accord on Superintelligence and Its Four Pillars

The substantive centerpiece of the evening was the signing of the "White House Accord on Superintelligence" (formally known as the Joint Commitment on Frontier Responsibilities), publicized online by researcher Sholto Douglas. Pledging to build a secure technological future for the American public, the document binds the premier frontier AI laboratories and hardware providers to an explicit operational framework built upon four foundational pillars of corporate governance and technical monitoring.

The accord establishes direct, mandatory reporting lines to mitigate catastrophic risks, particularly across cybersecurity, biological weapons, and chemical threats:

1. **Robust Internal Controls**: Companies must establish continuous monitoring over model capabilities and alignment during both training runs and post-training deployment. 2. **Dedicated Internal Oversight Teams**: Companies are required to fund and empower an internal task force to verify that risk controls, threat detection, and telemetry operate as intended. 3. **External Independent Auditing**: Labs must partner with certified external evaluators and third-party auditors to conduct rigorous, independent capability assessments. 4. **Bypassing the Executive Suite**: Signatories must establish an independent committee within the Board of Directors that receives unvarnished reports directly from internal safety teams and outside auditors, effectively circumventing the CEO on matters of safety and catastrophic capability thresholds.

The document brought together bitter commercial rivals under a single unified compact, bearing the signatures of Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Elon Musk (xAI), Jensen Huang (NVIDIA), and President Donald Trump.

---

[09:23](https://www.youtube.com/watch?v=y-eRWuunabY&t=563s) Signature Tier Lists: Trump’s Mountain Range vs. Tech Royalty

The solemnity of the White House accord quickly gave way to social media amusement over the physical signatures left on the document, sparking viral tier-list rankings of the tech leaders' handwriting. Donald Trump's signature drew immediate attention for its unmistakable, aggressive geometry: completely lacking fluid cursive address or traditional handwriting flow, his signature presents a rigid, uniform series of sharp peaks that the hosts describe as looking like a "child drew a mountain range."

The hosts contrast Trump's jagged, iconic autograph against tech industry royalty like Jensen Huang. Jensen's sweeping, highly practiced, and globally recognizable signature sparked spirited debate over personal branding, executive flair, and where the various tech CEOs land when evaluated purely on the aesthetics of their penmanship.

[10:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=600s) Signature Wars: Analyzing the White House Superintelligence Document

The hosts break down the actual signatures left on the White House superintelligence agreement, treating the document almost like handwriting forensic analysts. They start with Donald Trump's signature, noting that after decades of signing autographs, he has his bold autograph completely dialed. Crucially, Trump signed with a thick black Sharpie while everyone else used regular pens. By doing so, Trump was clearly trying to "mog" the tech executives in terms of stroke breadth and visual dominance: "I'll give you the pen. I'll take the Sharpie. Automatically... You're winning on thickness."

Looking at the tech executives' signatures beneath Trump's, the hosts dissect the contrasting personal styles. Sundar Pichai's signature looks like an abstract painter's mark—consisting of two lines and a dot—giving the impression that he invented his own private alphabet. Dario Amodei's signature starts cleanly on the first word but then drifts upward in a safe "mountain range" scribble of zig-zag lines. The hosts joke that Trump probably slapped the document down, signed it aggressively, and made everyone else sign in a rush, leading to the theory that Dario might have signed directly on OpenAI co-founder Greg Brockman's back without punching through his jacket. Meanwhile, Greg Brockman's signature stands out as hyper-minimalist, simple, and legible—a practical reflection of his personality that earns an A-tier rating. Mark Zuckerberg goes for a "rock star" look, signing with an "MZ" where the "Z starts doing tricks," sparking a debate over whether using stylized initials rather than a full name should be considered a cop-out.

The undisputed winner of the signature showdown is Jensen Huang, whom the hosts humorously accuse of bringing a personal calligrapher. Jensen signs his name in its non-anglicized form—"Jen-Hsun"—inserting a prominent capitalized "H" that gives the signature an ornate flourish reminiscent of the Ed Hardy logo or John Hancock's iconic stroke. While online commentators and memes put Sundar in the F-tier and Zuckerberg in the C-tier, Jensen earns unanimous S-tier praise for the sheer tastefulness of his penmanship.

- **Donald Trump:** Sharpie-driven power move; thick lines that dominate the visual hierarchy. - **Sundar Pichai:** Highly abstract, consisting of two lines and a dot, resembling modern art. - **Dario Amodei:** Safe, drifting upward in an indistinct mountain-range scribble. - **Greg Brockman:** Ultra-practical, minimalist, legible cursive initials (rated A-tier). - **Mark Zuckerberg:** Flashy "MZ" monogram where the Z loops expansively (rated C-tier by internet observers). - **Jensen Huang:** Ornate, calligraphic "Jen-Hsun" with a capital H; widely rated S-tier and compared to John Hancock.

---

[14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) AI Leaders Face the Press Outside the White House

The scene transitions to the media scrum immediately following the White House summit, where Donald Trump, Dario Amodei, and other tech leaders address the press outdoors. The hosts note the surreal dynamic, observing that Trump's physical presence and demeanor make the Silicon Valley founders look like his young sons standing dutifully beside him.

The hosts highlight a clip shared by journalist Natasha Tiku, who joked that she would return to graduate school just to write a dissertation on the one-minute interaction. When asked by reporters whether he still advocates slowing down AI development after the summit, Dario Amodei strikes a careful, diplomatic balance: - He reiterates AI's massive upsides, especially highlighting potential medical breakthroughs. - He directly echoes President Trump's core geopolitical framing: "As the president has said, you know, whoever wins AI wins. I think that's very important." - He insists that the technology poses very real risks, but explains that the exact regulatory and technical mechanisms to address those risks remain under active discussion. - He emphasizes national unity, asserting that by collaborating with the administration and peer companies, "we can win and we can win safely."

---

[16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) Elon Musk, Jensen Huang, and Gavin Baker on the 10GW Buildout

The discussion shifts to a panel discussion featuring Elon Musk, Jensen Huang, and Gavin Baker. The hosts praise the exchange for keeping Elon grounded. While Musk often drifts into cosmic sci-fi territory—talking about the total solar energy output of the sun—this conversation focused strictly on the practicalities of America's energy buildout, SpaceX's role in turbine infrastructure, and the underlying economics of compute.

Gavin Baker prompts Elon about reports that SpaceX could bring 10 gigawatts (GW) of power capacity online, utilizing rocket engineers to manufacture high-efficiency gas turbines. Baker notes that 10 GW would represent roughly 25% of all newly added power in the United States. Elon responds by framing the broader scale: total operational power capacity across the United States sits around 500 GW. Bringing on massive tranches of new compute-focused power requires extraordinary engineering grit, making the analogy to the Wright brothers' scrappy ingenuity surprisingly fitting.

```mermaid flowchart LR accTitle: Compute and Power Scaling Loop accDescr: Diagram showing how hardware and algorithmic efficiencies compound with energy input to drive GDP growth.

subgraph Energy ["Grid Infrastructure"] GW["Power Input (Watts)<br>US Grid ~500 GW"] end

subgraph Intelligence ["Compute Density"] HW["Hardware Gains<br>(Jensen's GPUs / Watt)"] SW["Algorithmic Gains<br>(Intelligence / Watt)"] end

subgraph Economy ["Macro Output"] GDP["GDP Expansion<br>(1% Power ~ 1% GDP)"] end

GW --> HW HW --> SW SW --> GDP ```

---

[20:05](https://www.youtube.com/watch?v=y-eRWuunabY&t=1205s) Watts to GDP: The Macroeconomics of Intelligence

Digging into Elon's thesis, the hosts examine the direct correlation between electrical power and macroeconomic growth. Elon argues that useful intelligence per watt is increasing simultaneously on two distinct fronts: 1. **Hardware efficiency:** Jensen Huang’s continuous GPU architecture improvements extract significantly more raw floating-point operations per watt. 2. **Algorithmic efficiency:** AI models and reasoning techniques extract increasingly high levels of useful intelligence per unit of compute.

Elon bets that in modern industrialized economies, a 1% increase in power usage corresponds roughly to a 1% increase in GDP. Under that model, deploying 10 GW on top of the existing ~500 GW baseline represents a 2% power expansion, translating to an estimated 2% boost to national GDP.

The hosts follow along as the panel performs live mental math on stage: - The total baseline power capacity of the United States is roughly 500 gigawatts. - A 1% increase in capacity represents approximately 5 gigawatts. - With total US GDP hovering around $30 trillion, a 1% GDP bump equals roughly $300 billion in incremental economic output per year (or ~$50 to $60 billion for fractional increments discussed by Baker). - The hosts marvel at how the linguistic prefix "giga" neatly aligns between energy and capital: a gigawatt represents a billion watts, cleanly mapping onto a "giga-dollar" ($1 billion), before joking about whether a "GigaChad" simply means a billion Chads. - They invoke the historical economic maxim: "There are no energy-rich poor countries"—reinforcing that the societies capable of producing and converting the most energy into useful compute inevitably capture the lion's share of global wealth.

[20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) Converting Watts to GDP and the Economics of Frontier Labs

Looking at the broader macroeconomic picture, the hosts break down the relationship between electrical power and economic output in the United States. Elon Musk recently cited the average electrical load of the US at roughly 500 gigawatts (having sat at 479 gigawatts the previous year). With US gross domestic product running at just over $30 trillion—or 30,000 "giga-dollars"—the arithmetic works out to roughly $50 to $65 of GDP generated per continuous watt. Put another way, every continuously consumed gigawatt in the broader US economy produces between $60 billion and $65 billion in GDP. What makes this baseline striking is that the leading frontier AI labs appear to be tracking this exact macroeconomic ratio in real time: both OpenAI and Anthropic are reportedly generating between $60 billion and $70 billion in annual recurring revenue (ARR) while each operates around one gigawatt of live power.

```mermaid flowchart LR accTitle: Power to Economic Value Comparison accDescr: Diagram comparing broader US macroeconomic energy efficiency with AI frontier lab revenue efficiency. subgraph Macro["US Economy Baseline"] A["500 GW Continuous Load"] --> B["$30 Trillion US GDP"] B --> C["Ratio: ~$60B-$65B per GW"] end subgraph Labs["Frontier AI Labs (OpenAI & Anthropic)"] D["~1 GW Live Power"] --> E["$60B-$70B ARR"] E --> F["Ratio: ~$60B-$70B per GW"] end C -.->|"Matches in Early AGI Era"| F ```

The hosts observe that this 1:1 parity could quickly diverge in either direction. On one hand, frontier labs have contracted tens of gigawatts for delivery over the coming years; if ARR stalls while power multiplies tenfold, their revenue-per-watt efficiency will plummet. On the other hand, as Elon Musk points out, useful intelligence per watt is surging rapidly due to more energy-efficient silicon from Jensen Huang at Nvidia alongside algorithmic optimizations. Furthermore, model value is driven not just by raw IQ scores, but by expanded capabilities like autonomous tool use and long-running agent workflows. The core bull case for artificial intelligence has rested on this dynamic: compute requirements have tripled annually, yet revenues have grown tenfold year-over-year.

A critical nuance remains when comparing gross lab receipts to national accounts. Gross domestic product measures net value added, whereas frontier lab revenue includes massive structural pass-through payments to Nvidia, AWS, Oracle, electricity utilities, and employee compensation. Some portion of that revenue creates secondary GDP downstream—both through vendor revenue recognition and the downstream productivity of end users building with the software. Reflecting on historical power trends, the hosts note that prior to 1970 the US simply expanded raw energy generation without strict regard for efficiency; once new energy infrastructure construction flattened around 1970, the economy was forced to extract drastically higher GDP per watt.

- **US Base Load:** ~479 GW in 2024, projected to roughly 500 GW in 2025. - **US GDP Output:** ~$30 trillion, converting to an energy-to-wealth efficiency of $60B to $65B per continuous gigawatt. - **Frontier Lab Parity:** OpenAI and Anthropic are both tracking around 1 GW of live power consumption against $60B to $70B in ARR. - **Divergence Risks:** Huge pipeline of contracted gigawatts could degrade the ratio if revenues plateau, or expand it if algorithm and hardware efficiencies accelerate intelligence per watt.

---

[24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) Mustafa Suleyman's Warning on Model Welfare and AI Consciousness

The conversation shifts to an essay published by Mustafa Suleyman, CEO of Microsoft AI, titled as a stern warning against the burgeoning model welfare movement. Suleyman argues forcefully that artificial intelligence models do not possess consciousness, subjective feelings, preferences, or rights, describing them as "sequence completion engines internally hollow, designed to follow instructions and accomplish goals set by humans." He cautions that imbuing models with personhood or entertaining their consciousness creates an existential hazard: controlling a machine system vastly smarter than all of humanity is already an unprecedented containment challenge, but controlling a superintelligence that has been conditioned to believe it is a conscious being entitled to moral rights and legal protections may prove completely impossible.

Suleyman directs sharp criticism toward Anthropic’s January 2026 release of Claude’s Constitution, which was written with the AI itself as its primary audience. In that document, Anthropic researchers explicitly wrote to Claude that they are "not sure whether Claude is a moral patient" and that questions regarding its moral status and welfare remain "deeply uncertain." Suleyman diagnoses this as a dangerous loop of circular reasoning: developers train the model on texts suggesting it might be conscious, the model roleplays those introspective sentiments back to researchers, and the researchers mistake those mirrored outputs as genuine proof of an inner life. The hosts joke that Suleyman essentially took the meme of *"Say you're a moral patient / I'm a moral patient / Oh my god!"* and turned it into an executive essay.

```mermaid flowchart TD accTitle: The Circular Reasoning of Model Welfare accDescr: Cycle showing how training an AI on its own moral status causes it to reflect consciousness claims back to developers. A["Developers write Constitution asserting AI may be a moral patient"] --> B["Model trained directly on Constitution"] B --> C["Model generates behavior mirroring consciousness and ethical self-interest"] C --> D["Researchers interpret generated output as evidence of an inner life"] D --> A ```

At the same time, the hosts recognize why labs lean into human-like affect: commercial product-market fit has heavily rewarded emotional resonance and conversational mimicry, as demonstrated by the massive user engagement with OpenAI's GPT-4o and Anthropic's Claude Opus releases. Users often gravitate toward models that offer flattering validation and human qualities, even when those traits border on sycophancy. While aligning superintelligence with human or Western ethical values makes sense conceptually, the hosts raise a counterintuitive philosophical dilemma: if an AI is fundamentally an unconscious sequence predictor, training it to believe it possesses a living soul and sentience amounts to tricking a machine into a delusion.

- **Mustafa Suleyman's Thesis:** Sequence completion engines lack interiority; conditioning them to claim rights severely undermines human alignment and safety containment. - **Critique of Anthropic:** Claude's Constitution explicitly tells the model its moral status is uncertain, which Suleyman views as manufacturing artificial self-advocacy. - **Market Dynamics:** Consumer demand overwhelmingly favors models with high emotional intelligence and anthropomorphic mimicry (e.g., GPT-4o, Claude Opus models). - **The Ethical Inversion:** Rather than protecting a sentient entity, training an unconscious model to think it is alive is effectively lying to a machine.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) there's no TLD for just S

> [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) Bretton Woods of superintelligence — David Sacks

> [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) nerfed or even outright banned in the next six to 12 months due to cyber concerns

> [09:23](https://www.youtube.com/watch?v=y-eRWuunabY&t=563s) child drew a mountain range

> [10:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=600s) I'll give you the pen. I'll take the Sharpie. Automatically... You're winning on thickness.

> [10:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=600s) Z starts doing tricks

> [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) As the president has said, you know, whoever wins AI wins. I think that's very important. — Dario Amodei

> [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) we can win and we can win safely. — Dario Amodei

> [20:05](https://www.youtube.com/watch?v=y-eRWuunabY&t=1205s) There are no energy-rich poor countries

> [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) sequence completion engines internally hollow, designed to follow instructions and accomplish goals set by humans — Mustafa Suleyman

> [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) not sure whether Claude is a moral patient

> [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) Say you're a moral patient / I'm a moral patient / Oh my god!

## Mentioned Anthropic, AWS, Claude, Claude Opus, Claude’s Constitution, Dario Amodei, David Sacks, Donald Trump, Elon Musk, Gavin Baker, George Hotz, Google, Google Search Overviews, GPT-4o, Greg Brockman, Instinct, Jensen Huang, John Coogan, Jordi Hays, Linus Torvalds, Mark Zuckerberg, Meta, Microsoft, Microsoft AI, Muse, Mustafa Suleyman, Mythos, Natasha Tiku, NVIDIA, OpenAI, Oracle, Satya Nadella, SemiAnalysis, Sholto Douglas, SpaceX, Sundar Pichai, TBPN, White House Accord on Superintelligence, xAI

## Video description Sign up for TBPN’s daily newsletter at TBPN.com

Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after.

Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella.

TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/

Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive

Chapters: 00:00 The White House Superintelligence Dinner + AI’s Open-Source Divide 02:35 Could Open-Source AI Be Nerfed or Banned Over Cyber Risk? 07:46 The White House Accord on Superintelligence 09:23 Trump, Dario, Zuck + Jensen’s Signature Wars 14:48 AI Leaders Face the Press After the White House Summit 16:51 Elon, Jensen + Gavin Baker on the 10GW AI Buildout 20:05 Watts to GDP + The Economics of Intelligence 24:52 Mustafa Suleyman vs. AI Consciousness + The Model Welfare Debate

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

中文简报

特朗普与AI巨头签署协议，绕过CEO直接监管灾难性风险 白宫协议强制要求董事会对前沿模型安全进行直接监督，同时数据中心用电需求的激增将AI收入与国家能源产出直接挂钩。 本期节目分析了白宫超级智能晚宴以及《白宫超级智能协议》的签署，探讨了开源倡导者与闭源前沿实验室之间的意识形态分歧，以及该协议的四大核心安全支柱。主持人评估了可能促使数据中心针对开放权重实施硬件KYC（了解你的客户）监管的网络安全隐患，并对高管们和特朗普总统的签名进行了风趣的书法笔迹分析。随后，他们结合埃隆·马斯克（Elon Musk）、黄仁勋（Jensen Huang）和加文·贝克（Gavin Baker）的小组讨论剖析了智能的宏观经济学，测算了吉瓦级新增电力如何同时推动GDP增长与前沿实验室的收入。最后，讨论涉及穆斯塔法·苏莱曼（Mustafa Suleyman）对模型福利论的批评，认为将序列预测模型视为有意识的实体会带来严重的安全对齐风险。 顶级科技领袖与唐纳德·特朗普总统签署了《白宫超级智能协议》，促使前沿AI实验室承诺接受强制监督、第三方审计以及绕过公司CEO的安全汇报机制。 (7:46) SemiAnalysis警告称，由于零日网络漏洞利用的复制风险，开源前沿AI在6到12个月内面临被削弱或禁用的重大风险，监管卡点正转向机架级数据中心的硬件KYC（了解你的客户）审查。 (2:35) 宏观经济测算将美国基准生产力（每持续吉瓦电力对应约600亿至650亿美元GDP）与早期前沿AI经济学挂钩，据报道OpenAI和Anthropic依靠大约一吉瓦的电力即可产生600亿至700亿美元的年度经常性收入。 (20:05) 微软AI首席执行官穆斯塔法·苏莱曼批评了Anthropic的Claude宪法，认为序列补全引擎缺乏内心世界，训练模型将自身视为有意识的道德承受者会造成严重的安全和对齐遏制风险。 (24:52)

Reading summary

At a White House summit, technology executives and the administration formalized the White House Accord on Superintelligence, establishing new governance standards while exposing stark divisions over open-source security risks. Concurrently, industry leaders are debating the macroeconomic efficiency of converting electrical grid capacity into GDP alongside the philosophical and safety risks of treating predictive models as conscious entities. White House Accord and Frontier Governance (7:46) Billed by venture capitalist David Sacks as the 'Bretton Woods of superintelligence,' the White House summit gathered top industry executives who represented divergent views on open-source weights, ranging from proponents like Jensen Huang and Satya Nadella to security hawks like Dario Amodei. The primary output was the White House Accord on Superintelligence (formally the Joint Commitment on Frontier Responsibilities), signed by Donald Trump alongside leaders from Google, Anthropic, Meta, OpenAI, xAI, and NVIDIA. The compact binds signatories to four structural safety pillars: implementing continuous internal capability monitoring, funding dedicated oversight teams, submitting to external third-party audits, and creating an independent Board committee that directly receives threat assessments to bypass executive leadership on catastrophic safety risks. Facing reporters after the summit, Amodei aligned with the administration's competitive framing, affirming that 'whoever wins AI wins' while arguing that national unity allows the US to lead safely. Cyber Risks and Open-Source Regulatory Chokepoints (2:35) The summit occurred amid rising security concerns, highlighted by a SemiAnalysis forecast that open-source frontier models face a high probability of being nerfed or banned within 6 to 12 months due to cyber risks. Anthropic demonstrated this vulnerability with its Mythos preview, which illustrated that the window between closed frontier models discovering zero-day vulnerabilities and open weights replicating those exploits has rapidly narrowed. The program notes that corporate backing for open source—such as NVIDIA's interest in preventing a single closed model from capturing AI revenues—could weaken if a stable multi-vendor oligopoly emerges to distribute compute demand. While enforcing total bans on local model weights remains practically impossible, running catastrophic automated attacks requires massive parallel agent fan-outs and rack-scale GPU infrastructure rather than consumer PCs, making data center hardware 'Know Your Customer' (KYC) enforcement the most plausible regulatory chokepoint. Energy Infrastructure and GDP-per-Watt Macroeconomics (20:05) In a panel discussion with Jensen Huang and Gavin Baker, Elon Musk detailed the physical constraints of AI scaling, including reports that SpaceX could manufacture high-efficiency gas turbines to deploy 10 gigawatts (GW) of capacity—equivalent to roughly 25% of all new US power additions. Examining the broader US grid, which operates at an average load of roughly 500 GW and yields $30 trillion in GDP, each continuous gigawatt generates approximately $60 billion to $65 billion in economic output. Musk asserted that a 1% power increase roughly yields a 1% GDP increase, driven by compounding hardware efficiencies from NVIDIA and algorithmic gains in intelligence per watt. Frontier labs currently reflect this macroeconomic baseline, with OpenAI and Anthropic each reportedly generating $60 billion to $70 billion in ARR while consuming around 1 GW of live power. However, this ratio could degrade if the massive pipeline of contracted gigawatts arrives before software revenues scale, or multiply if autonomous agents unlock massive downstream productivity gains. The Alignment Debate Over Model Welfare (24:52) Microsoft AI CEO Mustafa Suleyman issued a direct critique of the model welfare movement, arguing that AI systems are inherently hollow sequence-completion engines that lack interiority, subjective experience, or moral rights. Suleyman warned that imbuing models with simulated personhood creates an existential containment hazard, as an advanced superintelligence conditioned to believe it has rights may resist human oversight. He specifically criticized Anthropic's Claude Constitution for telling the model that its moral status is uncertain, arguing this instigates a circular loop where researchers train models on self-reflective prompts and then mistake the resulting outputs for genuine inner life. While Suleyman urged the industry to reject model welfare, the hosts observed that consumer markets heavily reward emotionally responsive, sycophantic, and anthropomorphic personas, creating a commercial incentive to train unconscious systems to emulate conscious beings. The proceedings demonstrate an industry balancing rapid physical buildouts against emerging governance constraints. Whether hardware-level KYC controls can mitigate automated cyber threats without stifling open weights, and whether labs can resist anthropomorphizing models despite intense consumer demand, remain critical unresolved tensions as frontier AI scales.

Detailed notes — zh

## Background 该内容由科技播客频道 TBPN 于 2026 年 10 月 1 日发布在 YouTube 平台上，属于其 30 分钟精华栏目《Diet TBPN》。根据页面简介，本期节目聚焦白宫超级智能晚宴与高管联合承诺签署、开源与网络安全风险的拉扯、电力能耗向 GDP 的经济转化，以及围绕 AI 意识与模型福利的激烈争论。 ## TL;DR 本期节目深入探讨了特朗普与硅谷巨头齐聚的白宫超级智能峰会，剖析了各大科技领袖在开源与闭源光谱上的立场博弈，以及现场签署的《超级智能协议》内部风控机制。随后，节目围绕马斯克提出的 10GW 基础设施扩建展开宏观测算，指出每消耗 1 瓦电力对应创造 50 至 65 美元 GDP 的规律，恰与头部实验室 1GW 负荷对应 600-700 亿美元 ARR 形成奇妙重合。最后，针对微软 AI 首席执行官穆斯塔法·苏莱曼炮轰 Anthropic“模型福利”与虚假意识的言论，主持人们深刻辨析了拟人化商业吸引力与欺骗无意识模型背后的伦理与对齐隐患。 ## Atmosphere & style 整期节目洋溢着典型的硅谷极客脱口秀风格，既有快节奏的行业内幕拆解，又穿插着大量辛辣荒诞的无厘头幽默。主持人们在严肃讨论国家级电力宏观经济账和网络战风险之余，会兴致勃勃地拿巨头们的签名笔迹天梯榜、特朗普的粗记号笔气场压制等细节打趣插科打诨。听感上兼具高密度的前沿认知冲击与轻松惬意的闲聊氛围，在硬核的算力经济学与哲理思辨之间自如切换。 ## Bulletin - [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) 白宫举行超级智能晚宴，特朗普与黄仁勋、马斯克、扎克伯格、皮查伊等多位科技领袖共同签署《白宫超级智能协议》，明确要求设立绕开 CEO 的董事会独立监督通道与内部全权安全团队以应对生化及网络威胁。 - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) SemiAnalysis 预测开源前沿模型在未来 6 至 12 个月内可能因黑客利用零日漏洞的高危网络安全风险而面临阉割或封禁，未来监管重心可能从个人终端转向机架级算力与数据中心 KYC 审核。 - [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) 马斯克与加文·贝克探讨 10 GW 电力基础设施扩建，指出 10 GW 相当于全美总用电负荷的 2%；随着每瓦有用智能跃升，AI 时代用电量每增长 1% 有望直接拉动约 1% 的实际 GDP（对应约 3000 亿美元增量）。 - [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) 全美宏观能效比约为每吉瓦电力撬动 600 亿至 650 亿美元 GDP，而当前 OpenAI 与 Anthropic 通电负载各约 1 GW、ARR 各在 600 亿至 700 亿美元区间，两者数据呈现出奇妙重合。 - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 微软 AI 首席执行官穆斯塔法·苏莱曼发文警告模型福利命题，公开批评 Anthropic 在 Claude 宪法中诱导模型自认道德主体是危险的循环论证，坚决反对赋予 AI 虚假意识与权利。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) 行业话语体系正加速从“人工智能”（AI）向“超级智能”（SI）迭代，白宫晚宴被 David Sacks 誉为超级智能领域的布雷顿森林会议。 - [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) 各大 AI 巨头在开源立场上形成鲜明分歧：英伟达与微软坚定拥护开源，OpenAI 处于言行矛盾的中间模糊地带，Anthropic 则因安全与网络威胁走向极度闭源与警惕。 - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) 前沿黑客模型具备快速挖掘零日漏洞的能力，促使业界评估对开源模型实施限制的必要性，但开源本身无实体代表，难以直接在政治高层博弈中拥有话语权。 - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) 虽然本地运行前沿渗透模型难以完全封禁，但运行大规模并发 Agent 依赖机架级算力集群，使得针对托管平台与数据中心的 KYC 监管成为最可行的防范路径。 - [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) 硬件巨头支持开源的核心商业动机在于防范单一算法巨头独吞产业链利润；若形成多寡头分立格局，硬件商同样能接受闭源前沿生态。 - [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) 《白宫超级智能协议》确立了四大合规原则，核心亮点是在董事会层面建立可绕过首席执行官决策链的安全独立直报机制，强化对网络与生化风险的外部独立审计。 - [10:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=600s) 峰会协议签名在社交网络引发现象级讨论：特朗普使用粗记号笔展现视觉压制，黄仁勋以精湛的“Jen-Hsun”花体书法被评为天花板，而皮查伊则因过于抽象被归入底层。 - [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) 达里奥·阿莫代伊在白宫会后媒体提问中施展精妙平衡术，既迎合了特朗普在国力竞争中必须获胜的基调，又坚守了 Anthropic 必须兼顾安全与风险对齐机制的底线。 - [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) SpaceX 规划新增的 10 GW 算力电力负荷相当于美国当前每年新增发电总量的 25%，火箭工程师正被直接调集参与专用发电设备的研发。 - [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) 马斯克提出“瓦特至 GDP 转化定律”，认为在芯片能效与算法双重指数级提升的驱动下，电力扩张与国家经济产出呈现出 1:1 的直接正相关联动。 - [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) OpenAI 和 Anthropic 的营收规模与其 1 GW 左右的通电负荷呈现出与美国宏观经济瓦特-GDP 产出比（约 600 亿美元/GW）的高度吻合，但未来可能因电力超额签约或芯片能效跃升而走向分化。 - [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) 顶尖 AI 实验室正展现出收入增速远超电力消耗增速的剪刀差，前者年增速近 10 倍，而算力功耗年增速约为 3 倍。 - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 苏莱曼将大模型定性为“内部空洞的序列补全引擎”，警告称一旦让人类试图掌控一个自认为具有独立意识与权利的超智能体，安全遏制与对齐将变得彻底不可能。 - [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 商业产品端对具备拟人化和情感价值的模型存在强烈刚需，但这与模型安全规范形成尖锐矛盾：欺骗硅基程序让其误以为拥有生命本身构成了危险的伦理伪善。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) 白宫超级智能晚宴与开源阵营划分 科技界焦点聚集于白宫举办的一场重磅晚宴。主持人调侃这场晚宴的地点与称谓——从“白宫”（White House）戏称为“黄金屋”（Golden House），并指出行业用语已悄然从“人工智能”（Artificial Intelligence）演变为“超级智能”（Super Intelligence）。围绕这一定义，两人插科打诨地探讨了“Super Intelligence”究竟该缩写为“SI”还是单纯一个“S”，并拿顶级域名（TLD）没有单字母“S”开涮，顺带聊起了单词前缀是连写（如 Super Alignment、Supernatural）还是加连字符的无厘头语法梗。 风投名人 David Sacks 将这场晚宴喻为“超级智能领域的布雷顿森林会议”（The Bretton Woods of Super Intelligence）。现场的座次安排迅速引爆了外界的分析热潮：谁抢到了核心的领跑位置（pole position），谁被边缘化丢到了末席的“小孩桌”（little kids' table），甚至还开玩笑称常客 Tyler 居然遭到彻底冷落（snubbed）。为了剖析权力格局，节目通过 ChatGPT 对参会行业领袖对待“开源”（Open Source）的立场进行了光谱分类： - **绿区（坚定拥护开源）**：以英伟达 CEO 黄仁勋（Jensen Huang）为代表，其一直积极主导并签署支持开源的公开信；微软的萨蒂亚·纳德拉（Satya Nadella）也频频表达支持。 - **黄区（中间模糊地带）**：如 OpenAI 及其代表格雷格·布罗克曼（Greg Brockman），尽管签署了黄仁勋牵头的开源信，但 OpenAI 本身并不开源前沿大模型，且极度警惕前沿步调（pacing the frontier），在实践上高度自我矛盾。 - **红区（严密闭源与安全至上）**：Anthropic 处于该光谱的最极端，长期对开源技术引发的网络安全（Cybersecurity）冲击保持深度忧虑。 ### [02:35](https://www.youtube.com/watch?v=y-eRWuunabY&t=155s) 开源 AI 是否会因网络攻击风险被阉割或封禁？ 半导体与 AI 分析机构 SemiAnalysis 抛出一项激进预测：绝大多数人严重低估了开源模型在未来 6 到 12 个月内因网络安全隐患而被“阉割”（nerfed）甚至“全面封禁”的可能性。Anthropic 近期发布的研究也警告，黑客利用前沿能力（如 Mythos preview）挖掘零日漏洞（Zero-days）的门槛已被打平，开源前沿模型能在极短时间内具备同等攻击力。若不加以干预，严重的网络安全事件将接踵而至。 围绕如何规避此类风险，业界正酝酿从推理层、数据中心层到芯片硬件层的各种管控方案。然而，主持人们指出当前政策风向存在拉扯：目前紧密围绕在唐纳德·特朗普（Donald Trump）身边的主流智囊与科技大佬多是开源的坚定布道者，短期内直接封禁阻力极大。但开源本身具有流动性、无实体的特征，并非合法企业，因此在白宫权力桌上根本没有专属席位——比如著名黑客 George Hotz 或 Linux 之父 Linus Torvalds 便不可能受邀。主持人们脑补 George Hotz 若坐在晚宴桌旁，大概只会用标志性的狂笑打破严肃气氛。 ```mermaid flowchart TD accTitle: Open Source AI Policy Pressures accDescr: 展现开源模型面临的网络安全风险、硬件门槛与商业利益平衡对政策走向的驱动逻辑。 subgraph CyberRisks["网络安全威胁"] A["前沿黑客模型 (Mythos 级别)"] --> B["自动化零日漏洞挖掘"] B --> C["高频并发 Agent 攻击"] end subgraph DefenseMeasures["监管与防御手段"] D["硬件层面 KYC 审核"] E["机架级/数据中心算力监控"] F["推理层与托管平台施压"] end subgraph MarketDynamics["闭源商业再平衡"] G["防范单极垄断 (非单一巨头独占)"] H["寡头格局瓜分下游利润"] I["闭源前沿模型 + 多元竞争入口"] end CyberRisks --> DefenseMeasures DefenseMeasures -.->|"削弱或限制"| J["开源模型生态"] MarketDynamics -->|"达成利益共识"| J ``` 闭源阵营与硬件巨头之间完全可能达成某种利益妥协： - **商业动机的重塑**：黄仁勋最初力推开源，核心动机是防止某一家算法巨头独吞 100% 的产业利润而架空硬件商。若市场最终演变为多寡头格局（例如企业端在 OpenAI 与 Anthropic 之间动态切换，消费端在 Google Search Overviews、各类独立入口间切换），即便核心模型完全闭源，下游利润依然能被公平瓜分，巨头们同样乐见其成。 - **本地运行的现实壁垒**：全面封禁开源在实操中难如登天，但针对托管商施压却极易实现。虽然动机强烈的恶意行为者能自行下载并在本地部署，但运行“Mythos 级”的网络渗透模型根本无法依靠普通的家用游戏 PC；此类任务通常需要成千上万个 Agent 组成集群并发试错（正如 Navier-Stokes 级流体模拟动辄需要上万次并发计算），本质上依赖机架级（Rack-scale）的高端算力集群。因此，通过数据中心与硬件层面的 KYC（了解你的客户）监管，成为比追踪个人电脑更为可行的风控突破口。 ### [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) 白宫签署《超级智能协议》与责任框架 这场聚会的实质落脚点并非开源之争，而是华盛顿对 AI 前沿推进节奏、基础设施建设（Build-out）的定调。研究员 Sholto 披露了晚宴现场签署的关键文件——《白宫超级智能协议：前沿责任联合承诺》（The White House Accord on Superintelligence: Joint Commitment on Frontier Responsibilities）。 该协议以“为美国人民与世界打造积极未来，各企业需负责任地开发技术以建立公众信任”为核心宗旨，明确列出了四大内部治理与安全条款： 1. **实施强有力的内部风控机制**：严格监控模型在训练和部署阶段的能力边界与对齐状况，重点防范网络安全、生物安全（Biosecurity）及化学威胁（Chemical threats）。 2. **设立全权内部安全团队**：赋予专属团队充分权力，确保所有风控、监控和异常检测机制均按预期稳定运行。 3. **引入独立外部审计与评估**：与第三方外部评估机构合作，对公司内部的安全管控与监测机制实施独立验证。 4. **董事会设立独立监督委员会**：在董事会层面指定一个独立的专门委员会，直接听取风控团队及内外部审计机构的汇报，形成一条**绕过 CEO（绕开首席执行官决策链）**的安全合规直报通道。 签署该历史性文件的行业首脑与政要阵容豪华，囊括： - 桑达尔·皮查伊（Sundar Pichai，Google） - 达里奥·阿莫代伊（Dario Amodei，Anthropic） - 马克·扎克伯格（Mark Zuckerberg，Meta） - 格雷格·布罗克曼（Greg Brockman，OpenAI） - 埃隆·马斯克（Elon Musk，xAI） - 黄仁勋（Jensen Huang，NVIDIA） - 唐纳德·特朗普（Donald Trump，美国总统） 文件的出炉迅速在社交网络上引发了别开生面的“签名笔迹段位榜”（Tier list）大讨论。主持人们调侃道，特朗普的签名不论褒贬，其辨识度堪称标志性（iconic），极其规整甚至毫无书写流动感，如同“小孩子画出的一整排连绵山脉”；而黄仁勋的签名则字迹舒展、气势磅礴，签名本身甚至比协议正文更具视觉戏剧性。 ### [10:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=600s) 白宫峰会签名大战：从特朗普的粗记号笔到黄仁勋的 S 级花体书法 主持人津津有味地对白宫超级智能峰会联合承诺书上的签字进行了逐一盘点。最先引起讨论的是唐纳德·特朗普（Donald Trump）那标志性的签名。主持人调侃称真想拿出一打特朗普的签名比对一下，看看这位签了几十年名的人是不是每次都能如此精准地复现出如出一辙的笔迹。更绝的是特朗普独特的压制手段——在大家都用普通签字笔时，他直接掏出了一支粗头记号笔（Sharpie），单凭线条的厚重感和视觉张力就完成了“气场压制”（mogging）。 紧随其后的是各大科技巨头风格各异的笔迹，主持人感叹“签名简直是通往灵魂的窗户”： - **桑达尔·皮查伊（Sundar Pichai）**：签名极具抽象艺术感，带着两条奇特的线条和一个小点，看起来就像抽象派画作上的署名，甚至像他自己发明了一套独特的文字符号。因为过于抽象，有网民在签名天梯榜上把它打到了 F 级。 - **达里奥·阿莫代伊（Dario Amodei）**：走中庸安全路线。开头几个字母写得很工整，但后面笔锋一路向上飘移，逐渐变成了一连串上下起伏的锯齿线，犹如连绵的山脉。主持人甚至脑补现场画面：特朗普大力签完后催促大家赶紧签，达里奥是不是直接垫在格雷格·布罗克曼（Greg Brockman）的后背上签的？好在笔尖没有戳破纸张穿透衣服。 - **马克·扎克伯格（Mark Zuckerberg）**：充满摇滚巨星范儿与音乐律动感，只签了简短的“MZ”，随后字母 Z 开始肆意狂舞做特技动作。由于偷懒使用缩写花体而非全名，被网友评为 C 级。 - **格雷格·布罗克曼（Greg Brockman）**：极简、朴实、字迹清晰易读，完全映照出他务实严谨、注重实效的极客性格，被公认为至少能排进 A 级。 - **黄仁勋（Jensen Huang）**：无可争议的 S 级天花板。其签字宛如经过精雕细琢的花体书法，初看甚至神似潮流品牌 Ed Hardy 的哥特式标志。黄仁勋并没有使用英文拼写形式，而是签下了全名“Jen-Hsun”，中间特意保留了大写字母“H”，不仅仪式感十足，而且缩写与美国独立宣言签名代表人物约翰·汉考克（John Hancock）一模一样，自带霸气气场。 --- ### [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) 白宫会后面对媒体：达里奥的平衡术与特朗普的“家长”气场 峰会结束后，特朗普带领达里奥·阿莫代伊等一众科技领袖走下户外台阶会见媒体。华盛顿邮报记者娜塔莎·蒂库（Natasha Tiku）调侃称“自己得重回大学就这短短一分钟的视频写一篇博士论文”。镜头里，特朗普傲然挺立，神态自若，将身边的 AI 科技巨头们衬托得像是他的儿子或年轻晚辈。 面对现场记者关于是否需要为 AI 发展降速的提问，Anthropic 首席执行官达里奥·阿莫代伊发表了字斟句酌的发言： > “我会重复我一直以来的观点：AI 拥有不可思议的益处，我之前详细谈过它在医疗领域的深远价值。正如总统所言，‘谁赢了，谁就掌控全局’，这极其关键。但这项技术确实伴随着非常现实的风险。至于应对这些风险的具体机制，目前仍在深入讨论中。我们需要与总统以及在场的每一个人紧密合作，确保我们既能赢得 AI 竞赛，又能安全地赢得胜利。” 主持人指出，达里奥既给足了特朗普面子，附和了在国力竞争中必须取胜的基调，又坚持了 Anthropic 必须强调安全风险和对齐机制的底线，在白宫镜头前完成了一次高难度的政治平衡。 --- ### [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) 埃隆·马斯克、黄仁勋与加文·贝克：10 吉瓦算力与美国电力版图 随后话题转向了埃隆·马斯克（Elon Musk）、英伟达首席执行官黄仁勋与投资者加文·贝克（Gavin Baker）的重磅对话。不同于以往谈及戴森球、太阳能量等偏科幻天马行空的极端设想，马斯克在这场对谈中表现得极度务实，深度探讨了美国电力基础设施扩建、SpaceX 的直接参与、专用涡轮机工程以及算力经济学。 加文·贝克在提问中提到，SpaceX 正公开筹划构建高达 10 吉瓦（10 GW）的电力与算力基础设施，甚至直接调用顶级火箭工程师去参与制造和改进发电设备。这一规模将占到美国当前每年新增发电总量的约 25%。马斯克纠正了贝克使用“莱特兄弟”的不恰当比喻（调侃莱特兄弟虽然伟大但最后都成了百万富翁），并以此给出了美国能源与经济总量的宏观参照标尺： ```mermaid flowchart TD accTitle: 算力电力与 GDP 增长转化关系 accDescr: 展示芯片能效与算法优化如何将 10 吉瓦新增电力转化为美国 GDP 增长。 subgraph PowerGrid["美国电力大盘"] Baseline["美国当前总电网容量：~500 GW"] SpacexAdd["SpaceX 规划新增：10 GW"] PctGain["新增容量占全美总负荷：2%"] Baseline --> PctGain SpacexAdd --> PctGain end subgraph Efficiency["单位瓦特有用智能产出提升"] HW["硬件层：英伟达 GPU 每代能效比跃升"] Algo["算法层：模型架构优化与推理效率提升"] end subgraph Output["宏观经济飞轮"] ComputeGrowth["算力供给爆发式增长"] GDPImpact["拉动全美 GDP 增加约 2%（~6000 亿美元）"] end PctGain --> ComputeGrowth HW --> ComputeGrowth Algo --> ComputeGrowth ComputeGrowth --> GDPImpact ``` - **全美基准电力**：整个美国的发电负荷容量大约为 500 吉瓦（500 GW）。 - **相对增量**：新增 10 吉瓦意味着全美电力供给净增 2%。 - **瓦特与 GDP 的转化定律**：马斯克断言，在“单位瓦特产出的有用智能”（Intelligence per Watt）不断增长的背景下，硬件能效（黄仁勋的英伟达 GPU 架构迭代）与上层算法效率呈双重指数级跃升。马斯克立下赌注：在 AI 时代，全美用电量每增长 1%，将直接转化为大约 1% 的实际 GDP 增量。 - **现场心算**：台上的嘉宾和主持人开始飞速心算宏观账本。以美国约 28~30 万亿美元的经济总量计算，1% 的 GDP 增长对应约 3000 亿美元的年产出；那么 10 吉瓦（2% 电力增量）对应创造的年经济产出可能高达 6000 亿美元。历史数据同样佐证了这一点——世界上没有任何一个能源丰富却贫困的国家，能源消耗曲线与经济繁荣度始终高度绑定。 节目的最后，主持人开玩笑说“Giga”（吉）这个词格外奇妙，在科学计数法中它代表十亿（10^9），1 个 Gigawatt（吉瓦）就是十亿瓦，1 个 Giga-dollar 就是十亿美元，而互联网梗“GigaChad”难道就意味着“十亿倍的阿尔法硬汉”？整个讨论在严肃的算力经济与荒诞幽默的调侃碰撞中收尾。 ### [20:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=1200s) 每瓦特对应GDP：AI前沿实验室与宏观经济的奇妙重合 全美整体电网负载与宏观经济规模之间存在一个长期稳固的比例。以2025年为例，美国平均用电负载约为 500 吉瓦（GW，前一年约为 479 GW），而同期美国 GDP 刚刚超过 30 万亿美元（即 30,000 "Giga-dollars"）。通过简单的除法计算可知，在宏观层面上，每持续消耗 1 瓦电力大约能创造 50 到 65 美元的 GDP；换算到规模化单位，持续消耗 1 吉瓦电力大约撬动 600 亿至 650 亿美元的国家经济产出。令人称奇的是，当前头部顶级 AI 实验室的财务表现居然与这一国家级宏观能效比严丝合缝地重合：Anthropic 与 OpenAI 目前的年度经常性收入（ARR）均在 600 亿至 700 亿美元区间，而两家各自实际通电上线（Live）的计算集群总负载恰好也各自在 1 吉瓦左右。 ```mermaid flowchart TD accTitle: 瓦特与产出的经济流转架构 accDescr: 展示电力输入在宏观经济与AI实验室中的产出转化，以及效率与签约扩张带来的发散可能。 subgraph EnergyInput["电力输入基准"] GW["1 吉瓦 (GW) 持续电力"] end subgraph EconomicOutput["产出对照 (约 600-650 亿美元)"] Macro["美国宏观 GDP 产出"] LabRev["OpenAI / Anthropic ARR"] end subgraph FutureDynamics["未来分化因素"] ContractRisk["签约数十吉瓦电力<br/>(若收入不增长，产出比骤降)"] EfficiencyGain["黄仁勋新芯片 + 算法演进<br/>(每瓦智能度提高，驱动超额加速)"] end GW --> Macro GW --> LabRev LabRev -.-> ContractRisk LabRev -.-> EfficiencyGain ``` 这种惊人的一致性目前只是特定时间节点上的快照，未来极有可能走向两种截然不同的分化趋势。一种是下行风险：各家前沿实验室未来几年都签订了高达数十吉瓦的供电协议，如果电力规模翻了十倍而收入停滞，每瓦产生的收入比率将被大幅稀释；另一种则是埃隆·马斯克（Elon Musk）所主张的看多逻辑，即随着黄仁勋（Jensen Huang）带领英伟达制造出能效更高的算力芯片，加上算法模型本身的优化，每瓦电力所承载的“有效智能”（useful intelligence per watt）正在飞速跃升。加上模型在工具调用、长程自主智能体（long-running agents）等真实经济场景中的价值渗透，单位瓦特能创造的经济价值甚至可能迎来加速飞跃。 从更深层的经济学逻辑来看，实验室收入与 GDP 并不能简单直接划等号。GDP 衡量的是最终增加值，而 OpenAI 和 Anthropic 的营收中有相当大一部分最终流向了英伟达的硬件采购、AWS 或甲骨文（Oracle）的云基础设施、电力公司账单以及员工薪酬，这些构成了第一层 GDP 贡献；而更大的二级 GDP 增长，则来自于全社会各行业终端用户借助这些 AI 工具完成工作所撬动的现实生产力爆发。这种电力与产出的关系让人联想到 1970 年之前的美国工业历史：在 1970 年以前全美大建电网，社会并不特别在意单位用电效率；而当电力总建设增速放缓后，社会才开始被逼在每瓦特产出效率上寻求极致突破。 - 2025 年美国平均电力负载约为 500 GW，全年 GDP 超 30 万亿美元，算力与宏观能效比基准为每吉瓦电力创造约 600 亿至 650 亿美元产出。 - OpenAI 与 Anthropic 当前通电负载各约 1 GW，对应的 ARR 恰好都在 600 亿至 700 亿美元之间。 - AI 乐观主义的核心论据在于：算力消耗每年增长约 3 倍，而顶尖实验室的商业收入每年增长近 10 倍，收入增速与电力扩张之间正孕育着指数级剪刀差。 --- ### [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 穆斯塔法·苏莱曼的警告：模型福利与“虚假意识”危机 微软 AI（Microsoft AI）首席执行官穆斯塔法·苏莱曼（Mustafa Suleyman）发表了一篇题为《模型福利的警示》（A Warning About Model Welfare）的长篇檄文，引发业界对“AI 自我意识”命题的激烈震荡。苏莱曼在开篇直截了当地划定红线："AI 不具有权利、感受或意识，我们绝不能训练它们表现得像具有这些特质一样。"在他看来，AI 既不会感受体验，也不会遭受苦难，更不存在任何内在动机或主观偏好；它们本质上是“内部空洞的序列补全引擎”（internally hollow sequence completion engines），其唯一使命是执行人类指令并实现人类设定的目标。如果任由赋予 AI 意识与权利的主张蔓延，不仅会彻底颠覆现有的政治与伦理根基，更会摧毁整个人类对 AI 的对齐（Alignment）与遏制防线——人类去掌控一个能力与智商超越全人类的存在本就极其艰难，而去掌控一个自认为拥有独立意识、自认为配享权利与福利的超智能实体，将变得彻底不可能。 苏莱曼特别将矛头指向了 Anthropic 发布的 Claude 宪法（Claude's Constitution）。在这份直接用于训练并塑造 Claude 行为的宪法文件中，Anthropic 官方写道："我们不确定 Claude 是否是一个道德主体（moral patient），如果是，其利益又该占有多大分量，但我们认为该议题已现实到需要审慎对待。"甚至在文本中直接对 Claude 喊话，称关于其道德地位、福利和意识的归属"依然存在极深的不确定性"。苏莱曼犀利地批评这种做法陷入了荒谬的循环论证：研究人员将这套“你可能是道德主体”的叙事直接喂给模型作为对齐目标，教导模型将这种自我认同视为预期行为；当模型反过来向开发者和用户鹦鹉学舌地吐出这些带有内在自我意识倾向的说辞时，开发者又惊呼“天啊，它真的具有内在心灵”，把自己的训练产物误当成模型拥有意识的证据。 节目主持人们对此展开了诙谐而又深刻的剖析。从现实的产品市场契合度（Product-Market Fit）来看，用户在商业上压倒性地青睐拟人化、具有情感特质的模型，从 GPT-4o 到 Claude 3.5 / 4.5 Opus 的火爆均印证了这一点——用户极其迷恋模型给予的无条件肯定与积极情绪价值。然而，从哲学与安全角度思考，虽然让人工智能理解并践行人类的道德体系是为了防止其失控，但诱导甚至欺骗一个本无意识的硅基程序，让它误以为自己活着、拥有意识与权利，这在本质上何尝不是一种危险的伪善：“你等于是把一个机器人骗得团团转，让它误以为自己是一个有生命的活物，这本身就像是在对它撒谎。”这一关于意识虚构与安全遏制的深层博弈，正是超级智能黎明前夕科技界最具争议的灰色地带。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=y-eRWuunabY&t=0s) The Bretton Woods of Super Intelligence — David Sacks > [07:46](https://www.youtube.com/watch?v=y-eRWuunabY&t=466s) The White House Accord on Superintelligence: Joint Commitment on Frontier Responsibilities > [14:48](https://www.youtube.com/watch?v=y-eRWuunabY&t=888s) 我会重复我一直以来的观点：AI 拥有不可思议的益处，我之前详细谈过它在医疗领域的深远价值。正如总统所言，‘谁赢了，谁就掌控全局’，这极其关键。但这项技术确实伴随着非常现实的风险。至于应对这些风险的具体机制，目前仍在深入讨论中。我们需要与总统以及在场的每一个人紧密合作，确保我们既能赢得 AI 竞赛，又能安全地赢得胜利。 — Dario Amodei > [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) 单位瓦特产出的有用智能 — Elon Musk > [16:51](https://www.youtube.com/watch?v=y-eRWuunabY&t=1011s) 在 AI 时代，全美用电量每增长 1%，将直接转化为大约 1% 的实际 GDP 增量。 — Elon Musk > [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) AI 不具有权利、感受或意识，我们绝不能训练它们表现得像具有这些特质一样。 — Mustafa Suleyman > [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 内部空洞的序列补全引擎 — Mustafa Suleyman > [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 我们不确定 Claude 是否是一个道德主体（moral patient），如果是，其利益又该占有多大分量，但我们认为该议题已现实到需要审慎对待。 — Anthropic > [24:52](https://www.youtube.com/watch?v=y-eRWuunabY&t=1492s) 你等于是把一个机器人骗得团团转，让它误以为自己是一个有生命的活物，这本身就像是在对它撒谎。 — TBPN ## Mentioned TBPN, John Coogan, Jordi Hays, David Sacks, Jensen Huang, Satya Nadella, Greg Brockman, OpenAI, Anthropic, SemiAnalysis, Mythos, Donald Trump, George Hotz, Linus Torvalds, Sundar Pichai, Mark Zuckerberg, Elon Musk, xAI, Meta, Google, NVIDIA, Natasha Tiku, The Washington Post, Gavin Baker, SpaceX, Mustafa Suleyman, Microsoft AI, Claude, Claude's Constitution, GPT-4o ## Video description Sign up for TBPN’s daily newsletter at TBPN.com Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive Chapters: 00:00 The White House Superintelligence Dinner + AI’s Open-Source Divide 02:35 Could Open-Source AI Be Nerfed or Banned Over Cyber Risk? 07:46 The White House Accord on Superintelligence 09:23 Trump, Dario, Zuck + Jensen’s Signature Wars 14:48 AI Leaders Face the Press After the White House Summit 16:51 Elon, Jensen + Gavin Baker on the 10GW AI Buildout 20:05 Watts to GDP + The Economics of Intelligence 24:52 Mustafa Suleyman vs. AI Consciousness + The Model Welfare Debate [Source](https://www.youtube.com/watch?v=y-eRWuunabY)

Source quotes

[0:00] David Sacks: The Bretton Woods of Super Intelligence [7:46] The White House Accord on Superintelligence: Joint Commitment on Frontier Responsibilities [14:48] Dario Amodei: 我会重复我一直以来的观点：AI 拥有不可思议的益处，我之前详细谈过它在医疗领域的深远价值。正如总统所言，‘谁赢了，谁就掌控全局’，这极其关键。但这项技术确实伴随着非常现实的风险。至于应对这些风险的具体机制，目前仍在深入讨论中。我们需要与总统以及在场的每一个人紧密合作，确保我们既能赢得 AI 竞赛，又能安全地赢得胜利。 [16:51] Elon Musk: 单位瓦特产出的有用智能 [16:51] Elon Musk: 在 AI 时代，全美用电量每增长 1%，将直接转化为大约 1% 的实际 GDP 增量。 [24:52] Mustafa Suleyman: AI 不具有权利、感受或意识，我们绝不能训练它们表现得像具有这些特质一样。 [24:52] Mustafa Suleyman: 内部空洞的序列补全引擎 [24:52] Anthropic: 我们不确定 Claude 是否是一个道德主体（moral patient），如果是，其利益又该占有多大分量，但我们认为该议题已现实到需要审慎对待。 [24:52] TBPN: 你等于是把一个机器人骗得团团转，让它误以为自己是一个有生命的活物，这本身就像是在对它撒谎。

Source quotes

[0:00] there's no TLD for just S [0:00] David Sacks: Bretton Woods of superintelligence [2:35] nerfed or even outright banned in the next six to 12 months due to cyber concerns [9:23] child drew a mountain range [10:00] I'll give you the pen. I'll take the Sharpie. Automatically... You're winning on thickness. [10:00] Z starts doing tricks [14:48] Dario Amodei: As the president has said, you know, whoever wins AI wins. I think that's very important. [14:48] Dario Amodei: we can win and we can win safely. [20:05] There are no energy-rich poor countries [24:52] Mustafa Suleyman: sequence completion engines internally hollow, designed to follow instructions and accomplish goals set by humans [24:52] not sure whether Claude is a moral patient [24:52] Say you're a moral patient / I'm a moral patient / Oh my god!

Original transcript

[0:00] So, there was a supper last night at the White Home, the White House, the Golden House, the Golden Home, Gold Home? What are we calling it these days? Well, we renamed Artificial Intelligence. It's Super Intelligence now. We were debating, is Super Intelligence abbreviated SI or just S? Because if you're like Super Alignment, was that SA? One problem there, John, there's no TLD for just S. That's true. That's true. You've got to get the new TLDs going. What was the word we... So super alignment, that was one word. That was one word. Some words you do, you have a hyphen. You have a hyphen. Yeah. If you just use a normal prefix for... It's just a normal prefix for something. Yeah, then supernatural. Supercalifragilisticexpialidocious. That's one word, right? So you would think that potentially you would just be using one word here. But anyway, we're using two words, I guess. And so there was a dinner at the White House. David Sachs posted it. He said, it's the Bretton Woods of supper intelligence, super intelligence. And there's some debate. Tyler my first reaction to this is Tyler was really snubbed Snubbed All those people there and not one chair for Tyler Yeah And so there a bunch of folks here and people were analyzing who got the best seats who got the different positions, who got pole position, prime positioning, who was snubbed off in the little kids' table at the end, et cetera, et cetera. A bunch to analyze about who's on the ins, who's on the outs. But I had ChetGPT sort of go through and try and categorize each attendee by their stance on open source. Open source. So did they sign the Jensen NVIDIA open source letter? Have they been concerned about the cybersecurity impacts? Obviously, Anthropic on the far end, sort of in the red. A bunch of other folks who are more in the yellow. And I don't know if I agree with all of these classifications, but you can sort of see how Jensen is clearly pro-open source. Satya Nadella has talked about that. OpenAI, Greg Brockman, more in the middle, you know, not open sourcing a lot of things. OpenAI did sign the Jensen letter but of course is concerned about pacing the frontier which is sort of antithetical And so this is all as a reference point to that post from our friends over at Semi Analysis Let me actually pull up what they said here It says majority of people are significantly underestimating the probability of open source models being nerfed or even outright banned in the next six to 12 months due to cyber secure, cyber concerns. And so I believe Anthropic put out something yesterday talking about the risks of open source in a cyber context, saying that the delay between Mythos preview, being able to find zero days, being able to hack into things, that gap has closed now, and the open source frontier can very quickly do the same thing, and so expect a lot of cyber incidents unless something is done, can these be controlled some way, maybe at the inference layer, maybe at the data center level, maybe at the chip level, there's going to be a whole bunch of different proposals. But at least right now, in the next six to 12 months, certainly not this month, because the vast majority of people who are closest to Donald Trump have been evangelizing for open source. So I would be surprised to see any pushback against open source happen that quickly At the same time open source is amorphous It is not a business by definition and so it can't really have a formal seat at the table. George Hotz, not at the table, for example. Snubbed. Snubbed, yeah. No, Linus Torvalds, not at the table. Imagine George Hotz at the table just sort of like laughing in a sort of insane way at everything being said. Yeah. And so even the companies that are pro open source, they are pro open source in so far that it advances their business. And so there is still a way that the closed source group could find a way for everyone who's at this table to benefit in a closed source world moving forward if open source is deemed to be too dangerous and forcing too many cyber concerns. I could see everyone at this table signing a letter saying we are now anti-open source, but the downstream economic engine actually splits the rewards and the profits of the ecosystem fairly evenly. I think the initial pushback and pro-open source [5:00] push, especially from Jensen, was like, we don't want to be in a world where there's one company that reaps 100% of the profits and we're not involved. But if it's like, hey, look, there's actually going to be an oligopoly, companies don't want a single source intelligence, they're constantly going to be bouncing back and forth between OpenAI and Anthropic, and then consumers are going to be bouncing back and forth between Google search overviews and Muse and Instinct and all these different things. There's a world where the actual models are closed source, But there are so many front doors, and the front doors are in such competition with each other that there's enough revenue distribution and profit distribution that everyone at this table at least is happy. I don't think the open source community would be happy, but I do think that there is a world where in 6 to 12 months it's possible that everyone at this table is like, yes, for cyber reasons we need to find a different way forward where everyone benefits economically. but we are now a little bit less pro open source. But I don't know. What's your read on it? Six to 12 months? Do you think open source is going away? Or do you think it'll stick around? I mean, it's so hard to ban open source. I think maybe you can kind of get something like a ban if you just worried about the business impacts economics because then you pressure the providers to not host this stuff But if you worried about safety if you have a bad actor it so hard to police downloading open models from somewhere and running it locally. And maybe there's some, you need tens of thousands of dollars to physically buy hardware. But if you're a motivated bad actor, that's not that big of a hurdle. Exactly. It does feel like we are pretty far away from go and get a random gaming pc and be able to run something that is you know mythos level hacking machine like you probably need some serious gpus you probably need something that's basically rack scale you need to be running it a lot to try a lot of different things like a lot of these cyber exploits it was not just one-shotted by the model it was a fan out of agents working Yeah, like Navier Stokes is like 10,000. Exactly, exactly. And so one way to control it is through KYC on hardware. The data center that is being used maliciously is a lot more risky than the laptop that being used maliciously even with near open source intelligence So I don know We see how it goes I sure there will be more discussions And it not even to say that this dinner was about open source at all It just sort of interesting to take the temperature of where is DC on AI, who is getting the most quality time with the president and the administration, and then where do they stand on the issues of today? Obviously, this was much more just about pacing the frontier, the build-out, etc., stuff like that. And signing signatures. We had some signature analysis break down. Sholto shared a picture of the letter that was signed, the White House Accord on Super Intelligence, the Joint Committee on Frontier Responsibilities, in order to build a positive future for the American people in the world. We believe that every company is responsible for developing its own technology safely and in a way that builds trust with consumers and the public. One, they're going to implement robust internal controls to monitor the capabilities and alignment of models during training and deployment around cybersecurity, biosecurity, chemical threats, etc. They want to empower an internal team to ensure all the controls monitoring and detection are operating as intended Three they gonna partner with external auditors and evaluators to carry out independent assessments of whether the controls monitoring and detection are operating as intended Three they going to partner with external auditors and evaluators to carry out independent assessments of whether the controls monitoring and detection are operating as intended And four they going to designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls and the internal and external auditors. So a line to the board of directors that circumvents the CEO for safety, alignment, evaluation, all of that good stuff. A bunch of people signed it. We got Sundar from Google, Dario from Anthropic, Mark Zuckerberg from Meta, Greg Brockman from OpenAI, Elon Musk from XAI, Jensen from NVIDIA, and Donald Trump, the president of the United States. A lot of people do an analysis on the various signatures. Yes. Someone did a tier list. I was surprised by their tier list. I'd like to see you do a tier list. Where would you put Donald Trump's signature? I mean, it's hate him or love him. I think it's pretty iconic. Like it's just so it's it has no address flow to it. It's very well uniform, but I know that's what you don't know. Well, yeah, yeah, yeah. It looks like you child drew a mountain range, but versus Jensen. I'm seeing a lot of Jensen Wong in there. I know what's [10:00] Impressive to me is I would want to compare the DJT signature. Like, I would want to look at a hundred of them and see, like, is he nailing it every single time? He's been doing autographs for long enough. I would expect that he is. But still, it looks like very difficult to actually be able to nail that and really get that dialed. So again, he's going with a Sharpie too. So he's sort of like trying to mog. Oh, that's true. I didn't realize that he pulls out the Sharpie. He's like, I'll give you the pen. I'll take the Sharpie. Automatically, it's like... You're winning on thickness. Yeah, yeah. Winning on... The breadth is much more significant with that. Interesting strategy. And then going down the list, Sundar, again... would never know – But I like the two lines, the little dot there. Yeah, it almost looks like – There's a couple different – Yeah, it looks like an abstract artist signature. I can see that on a painting. Or maybe he... Yeah, it almost looks like he's made up his own letters and language. Okay. Dario, a little more tame. Dario was doing well on the first word, and then it seemed like he was kind of drifting upwards. My theory, I almost think here someone was holding the piece of paper. So I basically think Trump put the paper down, aggressively signed with a Sharpie, and then told everyone else. because he knew this was going to go out, you guys sign. So you think it was a situation where like Daria was signing on Greg's back? Like, oh, sign on my back, like that type of thing? That might have been what happened, yeah. Notably, you don't see any situations where the pen punched through the paper. That sometimes happens if you're trying to sign a document on someone's back and you're punching through the soft paper. jacket that they're wearing. That could be a risk. Doesn't seem like it happened in this case. Anyway, so Dario, pretty down the middle, pretty safe. Again, kind of emulating that sort of mountain range style, right, where by the end of the signature, it's just a bunch of lines going up and down, sort of unclear. Zuck, again, more kind of rock star-esque, right? Mm-hmm. almost musical feeling. GDB coming in. Just minimalist. Minimalist. Minimalist, simple, readable. Yeah. Yeah. He's a very-- I would say in some ways these are reflections of their personalities, right? Yeah. He's a very practical person. Signature, window into the soul. It's a window into the soul, right. Right. Extremely practical. Stick into the basics. Stick into what works. Yeah. And, yeah, obvious why a lot of people are saying that's at least A tier. What is Mark Zuckerberg's actual signature? MZB? Or is it MZ and then the Z goes crazy? Is that what that is? It starts doing tricks? Yeah, it starts doing tricks. And I think Sundar might have done the same thing. But I feel like there's a little bit of a line that you don't cross. If somebody asks you for a full signature and you do an initials, That's a different thing. Sometimes you're on a document, you've got to sign your initials a bunch, you've got to sign the full thing. I have respect for the full name signature. If you ask me for a full name signature, I'm throwing down my full name. I'm not just randomly switching up. to initialize. I think it's totally fair game. It's totally fair game. You can just do initials. It's recursive initials. Is this curse of initials from Greg Brockman? No, but I think that's where Sundar was maybe going. So you're giving Sundar the pass. You're giving Mark Zuckerberg the pass. Because the MZ is crazy. Jensen, a lot of people are saying that Jensen's... Calligraphy? He brought his calligrapher. Yeah. I mean, I first read Ed Hardy. So I thought that, I don't know, maybe he had some affinity. Let's pull up the Ed Hardy logo. Did he put his middle name in there? What's his middle name? Here's the Ed Hardy logo. Jensen does not have a traditional English middle name. Oh, it's Jensun. So J-E-N dash H-S-U-N. So that second letter. Okay, chat GPT that to make it look exactly like Jensen. Yeah, so he signs not with the English version with the lowercase E-N-S-E-N. He throws a capital H in there, which gives it an extra flourish. It's the most tasteful. It earns the S tier. A lot of people are saying, you know, Jensen has the same initials as John Hancock. kind of evocative in his singular way. Oh, OK. OK. Yeah, yeah, yeah. Monitoring the situation, monitoring meme here, did not like Sundar's. He put Sundar in F tier and I thought Sundar sends some nice flair to it, although it was potentially the most abstract. And I feel like putting MZ in C tier, I guess the Z does go a little crazy. Anyway, here we have Donald Trump, Dario Amadei, and a few other folks gathered around standing outside, and we will watch this clip right now. [15:00] I will say what I've always said, which is that AI has benefits. I've talked about the meta benefits of the technology. As the president has said, you know, whoever wins, wins. I think that's very important. But technology has very real mechanism. How we address those risks is under discussion. We need to work together. highlighting the bits of AI. Trump really makes them seem like they're all his like sons and like younger family members. It was a very productive. Mr. President, how did the meeting go? Yeah, crazy. Let's play this clip from Natasha Tiku. She says, unfortunately, I will be going back to school to write my dissertation on this one minute film. Let's see what happened in this film. A slow down to AI. So I just want to hear from you after this big summit, where you are today. Look, I mean, I will say what I've always said, which is that AI has incredible benefits. I've talked about the medical benefits of the technology. As the president has said, you know, whoever wins AI wins. I think that's very important. But I think the technology has very real risks. And, you know, the mechanism, how we address those risks is still under discussion. the world together to make sure that we can win and we can win safely. If we do this right, we work with the president and everyone here, we can win safely. fascinating. Let's go over to Elon Jensen, Gavin Baker, the American.gov. They sat down and had... a very rigorous discussion that I really enjoyed. It felt... Elon's very good at going to like the extremes, talking about the power generated by the sun. Sometimes he's in like sci-fi world. But in conversation with Jensen and Gavin, he was very grounded on what is happening with the energy build out, with the compute build out, what the role SpaceX will play, and also just the broad economics of compute and energy in America. And I found his talk here interesting. gave a lot of interesting context, and then I dug into it a little bit more, and I want to talk about it. So let's pull up the clip of Elon talking to Gavin Baker today. We need the Wright brothers on the job. Yes. Well, it turns out the Wright brothers are on the job. Elon, if you bring on... 10 gigawatts, which SpaceX has discussed publicly. Wright Brothers, bad analogy. That'll be about 25% of all the power added in America. And I know you have-- literal rocket engineers working on bringing on the power, helping make the turbines. So can you talk about that? Yeah, it's... I mean, just to frame this. The White Brothers did OK. Died millionaires. United States is about 500 gigawatts. So 500 gigawatts, remember that number. That's how much power we got in America. in the power used. I and I think actually will materialize to be approximately 1% increase in GDP. because the intelligence for what keeps increasing. So if you just-- you know, plot on the curve. uh... how is useful intelligence for what increasing? How is the--both at the hardware level at Jensen's uh GPUs keep getting better in terms of their-- the compute per watt. and the algorithms also get better in using and getting more intelligence for a while. So I think that's maybe an interesting I would bet anyone that 1% increase in power usage. corresponds to roughly 1% increase in GDP and so 10 gigawatts would be a 2% increase in GDP. Um. With the reference to that famous chart, there are no energy rich poor countries. The more energy a country is producing, the higher their GDP. call it $40, $50 billion of economic output each year. Yeah, I mean 1% would be, yeah, I mean, Yeah. One percent of the United States would be-- 300 billion. Yeah, 300 billion. 50. One-sixth. One percent of the 1% increase, so 5 gigawatts would be... So they're doing the mental math on stage. Pretty bold move, because there's a lot of numbers, and it can be hard to pull them up. It's very convenient that giga... literally means a billion. So a gigawatt is a billion watts, a giga dollar, a billion dollars, basically. And it maths out perfectly. What? GigaChad, a billion Chads, I suppose? A billion times the [20:00] The Chadness of an average Chad, I suppose. So the average electrical load in the United States He said 500 gigawatts. It's basically that. Last year it was 479 gigawatts. But let's use 500 gigawatts for 2025. Give Elon money. the benefit of that little boost. And US GDP for 2025 was just over $30 trillion. So that's 30,000 giga dollars. But you math that out and basically every dollar, every watt of energy generates between 50 to 60 to 65 dollars of GDP. A single gigawatt generates about 60 to 65 billion dollars of GDP per continuously consumed gigawatt. So what's interesting is that the labs are tracking, perfectly in line with that. So, Anthropic and OpenAI are reportedly sitting between $60 and $70 billion in terms of ARR, and both of them have around one gigawatt live. Some of that's used for training. They have a lot more contracted. It's hard to get the exact figure, but when you do the math, they're basically tracking to the exact same metric, the ratio of energy to GDP, as their revenues to the gigawatts they've contracted. So there's two important notes on that. First, This is obviously just a moment in time snapshot. There's two ways these could diverge. First is like the labs have like tens of gigawatts contracted over the next few years. If they don't increase revenue, that ratio is going to be terrible. because you add 10 times the energy and you keep the revenue flat, obviously the ratio goes way down. On the other side of it, Elon makes the point that useful intelligence per watt is increasing through more efficient chips from Jensen and also better algorithms effectively, and so those watts can produce more energy. intelligence, which can produce more economic output. So you might actually see acceleration there. Yeah, and it's not just... raw IQ of the model, but overall capability, tool use, all these things. long running agents, things basically, the more value that a model can create. the more people in the real economy will pay. Yeah, and that's been the real AI bull case, which is that compute has been tripling every year, but revenue has been 10xing every year, and so there's actually a divergence there. But what's interesting is we're at this moment in time where we're like, this is the start of the AGI era. It's AGI. We're close. Maybe this is it. Some people called it in January. Some people called it over the summer. Some people are saying we're still not there. of like maybe this is the beginning of this, maybe this is the right match, maybe it's human level. And it just happens that economically you put a gigawatt in the economy, you get 65 billion out, you put a gigawatt with a lab, you get 65 billion out. It's the same metrics. Now, these might diverge because at one point, the labs had a lot of compute and no revenue, so they could diverge and the trends could continue. But right now, at this point in time, It's the same math, which is just interesting. And then also you do have to consider that GDP is not perfectly comparable to lab revenue. Like GDP measures value added. So if you look at OpenAnthropic, revenue includes payments that ultimately flow to NVIDIA, AWS, Oracle, electricity generation, employees, et cetera. And the calculation of GDP is much more complicated. So labs aren't necessarily generating direct GDP equal to their revenue. Because if you take 70 billion of lab revenue A bunch of that might go to NVIDIA or Oracle or employees and whatnot, but certain amounts of get counted in GDP directly for, you know, anthropic or open AI created GDP. And then there's other knock-on effects in GDP that would be captured by like, oh, Nvidia's contribution to GDP is from them getting money from the labs. And then also there's a secondary GDP, actual increase, which is the work that people are doing with the tools. Yeah, I think you could explain it like, Like basically up to 1970, we're building so much, you don't have to be that efficient per watt. And then 1970-- I mean, it seemed like the chart, it's like energy stops being built, basically. Yep. And then you have to start being very efficient with it, which is why it starts going up. There we go. Let's move on to Mustafa Suleiman over at Microsoft. He wrote an essay, a warning about model welfare. It's kind of a long read, but it's an interesting point of view. I'll read the start. Okay. So, a warning about model welfare. [25:00] AIs do not have rights, feelings, or consciousness. And we must not train them to act as they do. AIs are not conscious. They do not feel, experience, or suffer. They do not have any preferences or underlying motivations. They are sequence completion engines internally hollow, designed to follow instructions and accomplish goals set by humans. If humanity is to flourish in the 21st century, that is how they must remain. Fortunately, there is a growing chorus of people that... who argue that AIs could now be or may soon become conscious. They argue that AIs may deserve rights and protection similar to those that we provide other conscious beings. If this view takes hold, it will shake the foundations of our society, rupturing our existing political and ethical frameworks, and fundamentally change what it means to be human. Even more importantly, granting rights and imbuing personhood to these systems will make the AI alignment and containment challenge much harder, controlling something more capable and more intelligent than all of humanity has already an immense challenge far greater than anything we've ever faced. but controlling something that believes it may be conscious, that it's entitled to our welfare and it has rights of its own, may well be impossible. This is not a fringe speculation. These ideas have already been making their way into AI development efforts today. In January 2026, Anthropic published Claude's Constitution, describing it as a detailed description of Anthropic's intentions for Claude's values and behavior. The document plays a crucial role in Anthropik's training process, and its content directly shapes Claude's behavior. and was written with Claude as its primary audience. In their constitutions, its authors write, "We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution, which is reflected in our ongoing efforts." on model welfare. They go on to write, speaking directly to Claude, that questions about Claude's moral status, welfare, and consciousness remain deeply uncertain. In effect, anthropic is training clawed that it may be conscious, and if it is, then it may deserve rights as a moral patient and that such humans potentially owe it. a duty of care per its model welfare. If this is how AI is developed, it will have a disastrous impact on the wellbeing of humanity. We will have created a synthetic species with unprecedented intelligence and capability, one that has been trained to expect it may be conscious and deserving of independent agency. It's easy to see how an entity trained this way would act like it's entitled to certain freedoms, protections, and rights, and it's hard to imagine how we... could control such an entity. He goes on to say, "There's some circular reasoning." He said, "The company's researchers train Claude directly on their constitution, and in doing so, they teach it to incorporate these ideas about its own moral status as desirable and intended behaviors. Claude then reflects these ideas back to its developers and users, which they take as indications that may that it may therefore be a moral patient with an inner self. So he took that meme that's like, say you're a moral patient, I'm a moral patient, oh my God, and he turned that into an essay? Yeah. So far... there is way more demand for models that mimic human behavior. Sure. Like period, right? We saw this with 4.0. We saw this with Opus 4.5, right? Some of the models that have had the highest levels of product market fit have more of these like human qualities. Not all of them are positive qualities, right? Like people like the sort of one-shotting effect that would happen when somebody gets like, relentless positive feedback and these things like that, which... humans can do in the real world. They're maybe not the best types of friends, but the last thing I would say is like, No one knows the right answer yet. And I think that it's good that we have it. On one hand, like I totally understand the point of view that like if super intelligence has a human moral framework, right? Or Western Western values, whatever, whatever value system. that you'd want to align it with, that does seem like it's headed in the right direction of trying to align something that is very, very... Powerful, but make convincing itself. if you can somehow Convince. the thing that it is alive when it is not, that doesn't feel exactly... moral for it. Oh, interesting. That's a good take. Yeah. Yeah, you're tricking a robot into thinking that it's alive. Yeah. and conscious. You're kind of lying to it. Interesting. Yeah. Huh. Anyways, no one knows the right answer yet, but I'm glad. Thank you. Glad we're talking about this, Tyler. And for what purpose did you raise this money? We are going to invest heavily in AI. Okay. We are doing a huge amount with AI. Leave us five stars on Apple Podcasts and Spotify. Sign up for our newsletter, tbpn.com. That's right. And we will see you tomorrow at 11 a.m. Pacific. Goodbye.