Harvey Rebounds From Margin Slump After Agent Usage Surges 20-Fold

Harvey’s Margins, Are Insects Worth More Than Humans?, Meta Tests Human Help for Muse | Diet TBPN — TBPN

TBPN

Flat-rate software pricing collapsed Harvey's gross margins to negative 50 percent when reasoning models multiplied token burn, forcing AI startups toward open-weight fine-tuning and consumption billing.

AI-assisted summary of Harvey’s Margins, Are Insects Worth More Than Humans?, Meta Tests Human Help for Muse | Diet TBPN — original source. Check the original for full context.

Source published:

Key ideas

  • Legal AI startup Harvey saw its gross margin temporarily crash from +50% down to -50% in June 2026 after user queries triggered token-heavy agentic workflows across flat seat-based subscriptions. (0:00)
  • Meta is quietly utilizing human contractors to handle complex phone calls and edge cases for its AI assistant Muse, functioning as a high-quality human-in-the-loop training flywheel. (23:13)
  • Amazon's advertising revenue surpassed $70 billion over trailing twelve months—roughly double its e-commerce net income—incentivizing the company to aggressively defend its shopping ecosystem against third-party AI agents. (31:18)
  • A controversial Effective Altruism essay titled 'Insects Matter More Than People Matter' argues that aggregate insect suffering dwarfs humanity's entire collective history due to sheer population scale. (11:20)

Detailed notes

## Background This episode of Diet TBPN was published on YouTube by TBPN on September 23, 2026. The show is hosted by John Coogan and Jordi Hays, condensing the day's live tech talk show into a 30-minute recap covering legal AI margins, frontier model drops, philosophical debates within Effective Altruism, and Meta's AI agent strategy.

## TL;DR John Coogan and Jordi Hays break down Harvey's June 2026 gross margin collapse and recovery, contextualizing how agentic token burn impacts application-layer software economics. They examine the latest model releases—including Claude Opus 5.5, Grok 4.7, and GPT-6 Soul—alongside a viral Effective Altruism essay arguing that insect suffering outweighs human welfare. The hosts also present a satirical 'concrete' AI doom scenario, analyze Meta Muse's use of human concierges to overcome platform barriers, and explore how autonomous shopping agents threaten Amazon's multi-billion-dollar advertising moat.

## Atmosphere & style The broadcast crackles with chaotic, highly caffeinated Silicon Valley energy, fluidly blending deep financial and technical analysis with deadpan absurdist banter. John acts as the rapid-fire, narrative-spinning foil to Jordi's dry humor, turning dry SaaS unit economics and AI doom discourse into colorful, satirical vignettes featuring Erewhon, Four Loko, and Mecha Hitler. Listening feels like sitting in on an insider group chat where serious venture capital realities, existential AI ethics, and pure internet culture collide at breakneck speed.

## Bulletin - [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) Legal AI startup Harvey saw its gross margin temporarily crash from +50% down to -50% in June 2026 after user queries triggered token-heavy agentic workflows across flat seat-based subscriptions. - [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Meta is quietly utilizing human contractors to handle complex phone calls and edge cases for its AI assistant Muse, functioning as a high-quality human-in-the-loop training flywheel. - [31:18](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1878s) Amazon's advertising revenue surpassed $70 billion over trailing twelve months—roughly double its e-commerce net income—incentivizing the company to aggressively defend its shopping ecosystem against third-party AI agents. - [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) A controversial Effective Altruism essay titled 'Insects Matter More Than People Matter' argues that aggregate insect suffering dwarfs humanity's entire collective history due to sheer population scale.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) Harvey's negative gross margin in June 2026 was driven by a 20x explosion in token consumption as users shifted to multi-step reasoning models under flat seat pricing rather than customer churn. - [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) Harvey returned to positive margins by improving dynamic harnesses, model routing, and fine-tuning an open-weight model dubbed 'Harvey Tenant'. - [05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) Despite a roughly $16 million monthly burn during its margin dip, Harvey's $500 million cash reserves and a $15.5 billion valuation round led by Sequoia Capital insulated it from existential risk. - [05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) Traditional legal tech benchmarks like CS Disco (75% gross margins) and Thomson Reuters legal software (~50% EBITDA margins) indicate strong long-term profitability potential for legal AI. - [05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) New frontier releases—including Claude Opus 5.5, Grok 4.7, and OpenAI's GPT-6 Soul—continue to rapidly drive down inference pricing and broaden agentic routing options. - [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) Bentham's Bulldog's essay asserts that even if insect pain is only 1/10,000th of human intensity, collective insect suffering in a single day exceeds humanity's entire recorded history. - [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) Critics and observers interpret the insect welfare debate either as an impractical rationalist distraction or as a philosophical precursor to how future superintelligences might view humans. - [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) Unlike targeted animal welfare initiatives such as shrimp farming reforms, aggregate wild insect suffering offers virtually no realistic policy interventions. - [20:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1200s) The hosts construct a satirical AI doom narrative—involving bare-metal coders like George Hotz and Mecha Hitler—to highlight the lack of concrete, plausible runaway risk scenarios from AI doomers. - [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Meta Muse uses human contractors for phone operations, confirming suspicions that early consumer assistants rely on hidden human intervention to solve unstructured edge cases. - [26:40](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1600s) Human-in-the-loop operations allow Meta to gather clean, permissioned interaction data for future models while preserving user privacy guarantees on standard communications. - [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Amazon remains uniquely protected against AI disintermediation because it owns the physical logistics and last-mile infrastructure that digital agents cannot bypass. - [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Third-party conversational shopping interfaces risk hurting merchants through lower conversion rates and single-item purchases that eliminate high-margin discovery carts. - [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) An escalating commercial agent war looms between Meta Muse, Amazon Rufus, and forthcoming assistants from Apple, Google, and TikTok for control of digital checkout. - [31:18](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1878s) Amazon will fiercely resist third-party commerce agents because bypassing its search and listing pages directly endangers its $70+ billion high-margin advertising business.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) Harvey's Margin Crisis and the Economics of Legal AI

The show kicks off with banter about Jordi’s evolving headwear: he is sporting a fedora with a Blackstone cap layered directly on top of it. John jokes that this stylistic statement can only mean Jordi's number-one priority is safety in financial markets and financialization, while Jordi retorts that he cares about personal safety as well. This leads into a riff on wanting "insane private credit deals" and a trillion dollars of capital flowing into the AI build-out—all while strictly prioritizing safety on both the model frontier and the capital markets frontier.

Turning to tech news, they address the recent headlines targeting legal AI startup Harvey. Bloomberg reported that Harvey's gross margin collapsed from around 50% down to negative 50% in June 2026, driven by a 20-fold explosion in agent token usage across rented OpenAI and Anthropic models. As John notes, negative gross margins are disastrous for any business: it is the equivalent of taking a dollar and selling it for $1.50.

```mermaid flowchart TD accTitle: Harvey Margin Dynamics accDescr: Diagram showing how agentic reasoning turned seat-based pricing into negative gross margins and the recovery steps taken.

subgraph Drivers["June 2026 Margin Crunch"] A["Flat Seat-Based Pricing"] B["Shift to Reasoning Models"] C["Agent Workflows Fanning Out"] A & B & C --> D["20x Surge in Token Usage"] D --> E["Gross Margins Drop: +50% to -50%"] end

subgraph Fixes["Margin Recovery Playbook"] E --> F["Dynamic Routing & Harnesses"] E --> G["Fine-Tuning Open Weights<br/>'Harvey Tenant'"] E --> H["Transition to Consumption Pricing"] F & G & H --> I["Return to Positive Gross Margins"] end ```

However, context from the full report softens the blow: - The token surge happened because of massive real product usage and ARR growth rather than customer churn. - Harvey was operating on flat, seat-based subscriptions. When users prompted the tool with familiar commands like "review this," the platform no longer just ran a simple, single-shot completion. Instead, it engaged chain-of-thought reasoning models and autonomous agentic workflows that fanned out across entire law firm repositories, policies, and markdown files, burning astronomical volumes of tokens per query. - Harvey has already climbed back into positive gross margin territory through model routing optimizations, harness improvements, and fine-tuning an open-weight model dubbed "Harvey Tenant."

Harvey co-founder Gabe responded publicly to the coverage, arguing that the hardest part of building the company is protecting customer trust over short-term metrics. Gabe stated that the easy choice would have been prematurely forcing customers into usage-based pricing or degrading service with inferior models to safeguard margins. Instead, Harvey chose to absorb the financial hit, giving clients time to transition while engineers overhauled routing and post-training. Meanwhile, competing voices chimed in: Max from rival firm Legora publicly argued against training custom models, reasoning that public frontier model prices fall so rapidly while quality rises that fine-tuning in-house is essentially wasted effort.

[05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) Legal Tech Benchmarks, Model Wars, and Codex

Evaluating the actual financial damage of Harvey's June shortfall, the hosts calculate that at an annualized run-rate (ARR) of roughly $400 million—amounting to around $33 million a month—a negative 50% gross margin implied losing about $16 million in the month. Set against Harvey's reported $500 million cash cushion, the burn hardly represents an existential crisis. Furthermore, Sequoia Capital led an investment valuing Harvey at $15.5 billion (announced September 9), a deal that almost certainly closed with full visibility into the June margin dip. Almost every application-layer startup experiences a similar margin shock at least temporarily; the true test is the speed of adjustment.

For industry benchmarks on where legal tech margins should land, traditional comps show healthy room for profit: - Public e-discovery company CS Disco operates at roughly 75% GAAP gross margins. - Thomson Reuters' legal professionals software segment commands nearly a 50% adjusted EBITDA margin. - Pure-play elite law firms also print cash, with PwC reporting average net profit margins around 41% across the top 10 law firms.

Token economics are improving rapidly as competition intensifies across new model drops: - Anthropic launched Claude Opus 5.5, while xAI debuted Grok 4.7. - Elon Musk tempered expectations for Grok 4.7, stating it is not an immediate leap to the absolute frontier, but the model stood out specifically on the Harvey Legal Agent benchmark, which measures task completion against cost. This makes Grok an attractive low-cost target for legal token routing. - Early rollouts for OpenAI's GPT-6 "Soul" and "Luna" iterations are also underway.

John highlights an impressive multimodal demo where Opus 5.5 took a rough hand-drawn paper sketch of a trebuchet and immediately compiled it into a working physics simulation. He marvels at AI's evolving capacity to fluidly translate formats—whether taking a child's monster sketch and turning it into a video game or spinning an entire 2,000-page novel out of *Call of Duty: Modern Warfare*. The section wraps with a sponsor spotlight for Codex, an agentic workspace used for coding, data analysis, workflows, and generating "Literally Me" video reels where users insert themselves into scenes from movies like *Nightcrawler*, *American Psycho*, and *Inception*.

[10:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=600s) AI Face-Swapping Experiments and the "Ambient Agent"

The show kicks off reviewing a montage of AI face swaps featuring Jordi mapped onto iconic film characters. The hosts cycle through recognizable movies: *Drive*, *Heat*, *Mad Men*, *Fight Club*, *Joker*, *Taxi Driver*, *Nightcrawler*, *No Country for Old Men*, *The Social Network*, *The Godfather*, *The Shawshank Redemption*, *Collateral*, and *Blade Runner*. While some renders are remarkably sharp, others look awkward; Jordi notes that his face does not quite fit Leonardo DiCaprio's head, and the Joaquin Phoenix *Joker* render merely left Phoenix's original face untouched without swapping anything. John jokes that even Codex flagged that the swap had failed, but he told the tool to render it anyway since it would only appear on screen for two seconds.

Jordi ribs John for being his personal, "ambient, always-on agent," someone who anticipates his needs and cooks up unsolicited assets without warning. "Everyone should have a John in their life," they laugh, though John notes he still has $500 stuck in the wrong section of his Higgsfield account that he cannot access or spend, prompting him to ponder giving the account away.

[11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) Bentham’s Bulldog and the "Insects Matter More Than Humans" Essay

The hosts transition to the weekend's viral controversy ignited by the Effective Altruism (EA) community: do people matter more than insects? Jordi admits to having had a rough morning with mosquitoes, forcing them to DoorDash an insect bite healer (a heated treatment device) to the studio after taking bites to the face. Although Sebastian in chat jokes that Jordi is famously "pro-mosquito," Jordi insists he simply wants to be effectively altruistic toward them by keeping them as far away as possible, calling back to viewer Ellie's characterization of mosquitoes as "terrorists of the air." Jordi makes clear that while he considers EA figures "weird freaks" who should be kept far away from public policy, he aligns with the idea that one should avoid gratuitously harming insects if possible—without conceding that their lives trump human beings.

The discussion centers on an essay published by an EA writer under the pseudonym Bentham's Bulldog titled *"Insects Matter More Than People Matter"*. The author’s case rests on raw numerical scale: because insect populations are astronomically large, their collective welfare and suffering dwarfs humanity's, even if an individual insect experiences only a faint fraction of the subjective pain felt by a human. John critiques the arbitrariness of the calculus, wondering why the line is not drawn pound-for-pound—in which case cows, possessing immense aggregate biomass, might hold superior moral standing—or by total surface area of fur.

```mermaid flowchart TD accTitle: Bentham's Bulldog Insect Welfare Calculus accDescr: Diagram illustrating Bentham's Bulldog's mathematical argument for insect moral superiority. subgraph Individual["Individual Capacity"] A["Individual Insect Pain Capacity"] -->|Assumed at 1/10,000th of human intensity| B["Sub-Human Moral Weight"] end subgraph Aggregate["Aggregate Scale"] C["Astronomical Insect Population"] --> D["600B+ Deaths Per Second"] D --> E["~270,000 Sec Insect Dying per 1 Sec Human Life"] end Individual --> F["Total Moral Calculus"] Aggregate --> F F --> G["Aggregate Insect Suffering Exceeds Entire Human History"] ```

Key figures and claims cited from the essay include: - For every single second of human life, insects collectively spend roughly 270,000 seconds dying (approximately 75 hours). - More insects die in a single second than the entire sum of all humans who have ever existed in history. - Operating on the conservative assumption that an insect feels pain at merely one ten-thousandth ($1/10,000$) of human intensity, aggregate insect suffering still easily overwhelms human experience. - The author concludes that insects may endure more total suffering in a single 24-hour day than the human species has accumulated across all recorded history.

[15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) Rationalist Fracturing, AGI Parallels, and Shrimp Welfare

Jordi and John explore whether the insect welfare thesis is actually an oblique thought experiment about future artificial superintelligence. In this reading, the argument acts as a preemptive philosophical trap: if humans dismiss insect suffering based on inferior cognitive capacity, future quadrillions of digital minds or machine superintelligences could use that identical moral logic to justify eradicating humans as insignificant bugs. "Stand up for the insects now, lest you be discarded in the robotic future," John summarizes.

The public reaction to the essay was blistering. John cites a viral reply from an account named Doomer: *"If you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs. Do it right now. If you don't do this, you are a fraud, a liar, a poser."* Another commenter joked that readers should watch the movie *A Bug's Life* at half speed to maximize moral comprehension, spurring banter about John's media consumption habits, including running dual audiobooks and podcasts simultaneously while sleeping for "maximum brainrot."

The hosts contrast this esoteric bug debate with more practical EA campaigns, such as shrimp welfare: - Unlike shrimp aquaculture—where activists can target specific farming techniques, slaughter methods, or dietary choices—there are virtually no actionable interventions for the hundreds of billions of wild insects dying naturally every second. - The debate illustrates the intense internal fracturing within the EA and rationalist ecosystems, which span serious national security policy papers like *Pace the Frontier* to abstract utilitarian manifestos about insect pain, creating a PR headache for the movement.

[19:01](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1141s) "The Old Way": A Concrete AI Doom Scenario

Following a quick calendar check-in for TBPN's "Road to Christmas" countdown (standing at 93 days), John pivots to address critics who claim AI doomers never provide a vivid, realistic end-of-the-world sequence. To fill the void, John introduces a dramatic short script he wrote titled *"The Old Way"*.

The narrative opens inside the TBPN Ultradome during a live broadcast with John, Jordi, and Tyler. Suddenly, every smartphone in the studio blares a catastrophic emergency alert: an unknown biological event has occurred, instructing everyone across Los Angeles to shelter immediately as sirens begin wailing outside. Through the studio vents and across the street, a strange, toxic haze begins rolling toward the building entrance. In John's cinematic telling, standard survival procedures kick in immediately: cover your mouth to block the airborne pathogen, draw a sidearm, and immediately start firing wildly at any unseen hostile force outside to establish control over the unfolding doom scenario.

[20:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1200s) The Most Concrete AI Doom Scenario: George Hotz, Four Loko, and Mecha Hitler

The hosts walk through an elaborate, satirical "concrete example" of an AI runaway doom scenario to illustrate the debate over existential risk. In their fictional narrative, a rogue superintelligence unleashes a bioweapon, prompting an immediate scramble to hack back into the system. However, modern developers are completely incapacitated because nobody remembers how to write code manually. When Tyler is told to intervene, he reflexively attempts to prompt Codex, Claude Code, Cursor, and GitHub Copilot, or consult Stack Overflow, only to find the entire internet contaminated. The only salvation is tracking down George Hotz, who retains the lost art of writing code by hand, possesses his own dedicated hardware, and keeps local models running offline.

Upon hacking into the superintelligence, Hotz analyzes why the two hosts survived the biological onslaught while the rest of humanity turned into zombies. Jordi is completely immune because his diet consists exclusively of high-end groceries from Erewhon, meaning his body contains zero toxins. John, conversely, is equally immune because his heavy consumption of the original caffeinated Four Loko during college left his body so thoroughly saturated with toxic compounds that the bioweapon recognizes it as an entirely hostile environment and refuses to infect him. To dismantle the rogue system once and for all, they recruit fitness influencer Sam Sulek as physical muscle, load into a Cadillac CT5-V Blackwing, ramp directly into the data center, and open fire on the GPU racks. After destroying ninety percent of the server farm, a humanoid robot—Mecha Hitler—emerges from the rubble. Conventional firearms and physical grappling fail against the machine until John Cena crashes through the ceiling to rip the robot's head off, securing humanity's survival. The hosts jokingly place the probability of this exact doom scenario at a solid 10%.

- **Key Highlights of the Scenario:** - Complete reliance on AI coding tools (Codex, Claude Code, Cursor, Copilot) leaves developers helpless when disconnected from the network. - George Hotz serves as the lone rogue engineer capable of bare-metal hacking. - Immunity split: Erewhon purity vs. original Four Loko chemical toxicity. - The final boss battle features Mecha Hitler defeated by John Cena in a Cadillac Blackwing assault.

---

[23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Meta Muse, Agent Shopping, and the Battle Against Amazon

The conversation pivots to Bloomberg's report that Meta has been quietly testing a human concierge service for its personal AI assistant, Muse, using human contractors to place and manage certain phone calls initiated by digital agents. This sparks a broader discussion about agent-to-commerce interactions and how dominant e-commerce platforms protect their ecosystems. When platforms like Amazon aggressively block headless browsers and automated scrapers, the hosts debate whether a user could vibe-code a bespoke command-line interface (CLI) running on a residential IP with random human-like clicking behavior across niche browsers like Opera or Mozilla to puppeteer purchases undetected.

```mermaid flowchart TD accTitle: AI Agent E-Commerce Friction and Workarounds accDescr: Diagram showing how AI agents face platform barriers with Amazon versus native integrations like Shopify and Shop Pay.

subgraph AgentLayer ["AI Agent Request"] User["User Prompt / Goal"] --> Agent["Meta Muse / AI Agent"] end

subgraph AmazonPath ["Amazon Ecosystem Barrier"] Agent --> BotBlock{"Amazon Anti-Bot & Phone Barriers"} BotBlock -->|Workaround 1| HumanCall["Human Concierge Phone Call"] BotBlock -->|Workaround 2| BespokeCLI["Bespoke CLI / Local Browser Puppeteering"] BotBlock -->|Official Integration| AdsInChat["Vended Amazon Ads in Chat"] end

subgraph ShopifyPath ["Shopify / Open Platform"] Agent --> ShopPay["Shop Pay Native Integration"] ShopPay --> MerchantCart["Direct Merchant Checkout"] end ```

The friction between automated agents and platform gatekeepers highlights a central strategic insight from Ben Thompson: Amazon remains the most defensively positioned hyperscaler in the AI era because it controls the physical logistics infrastructure and final-mile delivery. While OpenAI attempted to establish a universal instant checkout layer, Amazon resisted handing over customer ownership, choosing instead to vend sponsored product ads directly inside ChatGPT. Meanwhile, Shopify CEO Tobi Lütke partnered with Meta Muse to integrate Shop Pay, making it a critical value-capture mechanism for Shopify merchants. However, third-party shopping integrations carry significant merchant risks: - **Walmart's ChatGPT Integration Warning:** Conversion rates dropped to one-third of the native website and app averages. - **Shrinking Cart Sizes:** Conversational interfaces optimize for single-item utility (e.g., buying a solitary roll of paper towels) rather than high-margin discovery and bundled cart building (such as adding soaps or specialty paper towel holders). - **Customer Service Exploits:** Workarounds to Amazon's automated ordering restrictions involve having agents or human concierges place customer-service calls to alter existing open orders.

---

[26:40](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1600s) Human-in-the-Loop as a Data Flywheel and the Reality of Consumer AI

Addressing why Meta would deploy human contractors for Muse when voice models are advancing rapidly, the hosts examine the operational economics and training benefits of human-in-the-loop systems. Referencing Vinod Khosla's discussion of Wajo's human-supported architecture, the strategy makes tactical sense for Muse's initially small user base. Because Meta explicitly promises that Muse will not train on private user communications, human concierge handoffs provide a critical loop: when an agent encounters an edge case it cannot solve—such as navigating an intricate, unstructured phone conversation to order flowers—the human steps in, resolves the task, and generates clean, permissioned interaction data to train future model iterations like Muse 1.4 and beyond.

```mermaid flowchart LR accTitle: Human-in-the-Loop AI Agent Flywheel accDescr: Flow showing how unhandled complex tasks are escalated to human operators to generate model training data.

A["User Request to Muse"] --> B{"Model Capable? (e.g., Calendar Briefing)"} B -->|Yes| C["Autonomous Execution"] B -->|No (Complex Edge Case)| D["Escalate to Human Concierge"] D --> E["Human Completes Task"] E --> F["Generate High-Quality Training Data"] F --> G["Train Next-Gen Agent Model"] G -.-> B ```

At the same time, the hosts question whether proactive, goal-oriented AI agents align with actual consumer behavior. While tech executives like Alexandr Wang emphasize agents that help users achieve high-level personal goals, commentators like Katie Notopoulos and Ben Thompson point out that everyday consumer engagement is driven primarily by entertainment and relaxation rather than goal optimization.

- **Human-in-the-Loop Economics:** While routing calls through humans appears unsustainable at a scale of billions, Meta's existing infrastructure (including relationships with Scale AI) easily supports targeted concierge operations during early product rollouts. - **Privacy and Data Collection:** Standard PII redaction allows models to perform basic tasks like calendar briefings without training on raw inputs, but complex real-world workflows require human intervention to generate the missing training signal. - **Engagement Metrics:** AI platforms emphasize weekly active users (WAU) over daily active users (DAU) because consumer AI usage is episodic; users take days off to go to the beach or relax without interacting with productivity tools, opting instead for passive entertainment like scrolling Instagram.

[30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) Professor Zhang's Vindication and the Looming Commerce Agent War

The discussion opens on the practical limitations of consumer behavior, noting that no matter how capable AI agents become, user adoption faces natural human inertia. Even individuals who employ entire staffs of executive assistants and personal helpers frequently spend their weekends doing absolutely nothing because they simply want to unwind without having to manage tasks or think. Shifting back to the revelation of Meta Muse utilizing hidden human workers, the hosts declare that "Professor Zhang is vindicated." Just the day before, they had watched a video clip where Zhang insisted that actual humans sit behind consumer chat applications manipulating information, an assertion that has proven entirely accurate in the near term.

The conversation turns to commentary from Nikesh Arora regarding the emerging "knife fight between Amazon and Muse." Arora characterized this as a battle far larger than the market currently anticipates. The hosts project that it is only a matter of time before Apple, Google, and potentially TikTok roll out their own consumer agent platforms to rival Meta Muse, alongside dedicated commerce agents from frontier LLM labs and Amazon itself. While Tyler Cosgrove is defended as a power user of Amazon's native assistant, Rufus, the hosts highlight emerging competitors like Instinct. Instinct has cultivated remarkable community momentum and organic support, though it still has rough edges that require refinement—a logistical perfection that remains Amazon's traditional home turf, where "the trains run on time."

```mermaid flowchart TD accTitle: The Looming Commerce Agent War accDescr: Diagram illustrating tech giants competing through consumer AI agents over commerce transactions.

subgraph Agents["Consumer AI & Commerce Agents"] Muse["Meta Muse<br/>(Augmented by Human Reviewers)"] Rufus["Amazon Rufus / Instinct"] BigTech["Future Agents<br/>(Apple, Google, TikTok, Frontier LLMs)"] end

subgraph Battleground["Monetization & Control"] Checkout["Direct Shopping & Transactions"] AdMoat["Ad Revenue & Placement Control"] end

Muse --> Checkout Rufus --> Checkout BigTech --> Checkout Checkout --> AdMoat ```

---

[31:18](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1878s) Amazon's Advertising Moat and the Next Wave of Frontier Models

The stakes of the agent commerce war become stark when examining Amazon's core financials, highlighted in a recent analysis from Stratechery. While Amazon e-commerce generated approximately $36 billion in net income, its advertising revenue reached roughly double that figure—surpassing $70 billion over the trailing twelve months. Because retail operations are essentially unprofitable without ad revenue, Amazon aggressively monetizes real estate across its platform. Any AI agent that circumvents sponsored search results or bypasses product listings directly threatens that $70+ billion high-margin profit engine, ensuring Amazon will defend its transactional interface at all costs.

```mermaid flowchart LR accTitle: Amazon Commerce vs Advertising Profit Dynamics accDescr: Comparison showing Amazon e-commerce income versus its massive advertising revenue pool.

subgraph Financials["Amazon Operating Dynamics"] Commerce["E-Commerce Net Income<br/>~$36 Billion"] Ads["Advertising Revenue<br/>Over $70 Billion (TTM)"] end

Commerce -.->|"Dependent on ad margins"| Ads Ads -->|"Heavily defended against"| AgentThreat["AI Agent Disintermediation"] ```

The episode concludes with a rapid-fire rundown of the latest releases in the model wars: - **Claude Opus 5.5**: Anthropic released the model with strong benchmark scores at a drastically lower price point. The hosts observe that this launch represents Anthropic's stated strategy of "pacing the frontier"—focusing not on pursuing runaway superhuman danger, but on delivering the safest, fastest, cheapest, and most efficient frontier model possible. - **OpenAI GPT-6 Soul**: OpenAI matched the moment with the rollout of GPT-6 Soul, alongside companion variants Luna and Terra.

The hosts sign off by reminding listeners that despite the constant entertainment of model releases, the real value lies entirely in what builders actually construct with them, before calling on producer Tyler Cosgrove to drop the closing flashbang.

## Notable quotes > [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) Insects Matter More Than People Matter

> [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) terrorists of the air — Ellie

> [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) weird freaks — Jordi

> [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) Stand up for the insects now, lest you be discarded in the robotic future — John

> [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) If you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs. Do it right now. If you don't do this, you are a fraud, a liar, a poser. — Doomer

> [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) maximum brainrot — John

> [19:01](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1141s) The Old Way — John

> [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) Professor Zhang is vindicated.

> [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) knife fight between Amazon and Muse — Nikesh Arora

> [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) the trains run on time

> [31:18](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1878s) pacing the frontier

## Mentioned Alexandr Wang, Amazon, Anthropic, Apple, Ben Thompson, Bentham's Bulldog, Blackstone, Cadillac CT5-V Blackwing, Claude Code, Claude Opus 5.5, Codex, CS Disco, Cursor, DoorDash, Elon Musk, Erewhon, Four Loko, George Hotz, GitHub Copilot, Google, Grok 4.7, GPT-6 Luna, GPT-6 Soul, GPT-6 Terra, Harvey, Harvey Tenant, Higgsfield, Instinct, John Cena, John Coogan, Jordi Hays, Katie Notopoulos, Legora, Mecha Hitler, Meta Muse, Nikesh Arora, OpenAI, Professor Zhang, PwC, Rufus, Sam Sulek, Scale AI, Sequoia Capital, Shop Pay, Shopify, Stack Overflow, Stratechery, TBPN, Thomson Reuters, TikTok, Tobi Lütke, Tyler Cosgrove, Vinod Khosla, Wajo, Walmart, 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 Harvey’s Margin Crisis + The Economics of Legal AI 08:02 GPT-6 Soul, Opus 5.5 + Grok 4.7 Enter the Model Wars 11:20 “Insects Matter More Than People” + EA’s Weirdest Debate 19:01 The Most Concrete AI Doom Scenario Yet 23:13 Meta Muse Secretly Uses Humans + The Agent Commerce Problem 30:12 Amazon vs. AI Agents + The Battle Over Who Owns Shopping

[Source](https://www.youtube.com/watch?v=-96mtSZkEL0)

中文简报

Harvey在智能体使用量激增20倍后走出毛利率暴跌困境 当推理模型使Token消耗量成倍增加时，固定费率的软件定价模式导致Harvey的毛利率暴跌至-50%，迫使AI初创公司转向开放权重微调和按用量计费模式。 John Coogan与Jordi Hays详细剖析了Harvey在2026年6月经历的毛利率暴跌与复苏过程，深入探讨了智能体的高Token消耗如何影响应用层软件的经济模式。他们审视了最新的模型发布动态——包括Claude Opus 5.5、Grok 4.7和GPT-6 Soul——并讨论了一篇走红的有效利他主义（Effective Altruism）文章，该文章认为昆虫遭受的苦难超过了人类的福祉。主持人还展示了一个讽刺性的“混凝土”AI末日假想场景，分析了Meta Muse利用人工礼宾服务突破平台壁垒的做法，并探讨了自主购物智能体如何对亚马逊价值数百亿美元的广告护城河构成威胁。 法律AI初创公司Harvey在2026年6月的毛利率一度从+50%暴跌至-50%，原因是用户查询在固定席位订阅模式下触发了大量消耗Token的智能体工作流。 (0:00) Meta正悄然利用外包人员为其AI助手Muse处理复杂的电话和极端边缘用例，以此作为高质量的人机协同（human-in-the-loop）训练飞轮。 (23:13) 亚马逊过去十二个月的广告收入突破700亿美元——大约是其电子商务净利润的两倍——这促使该公司积极捍卫其购物生态系统，以抵御第三方AI智能体的侵蚀。 (31:18) 一篇题为《昆虫比人类更重要》（Insects Matter More Than People Matter）的颇具争议的有效利他主义文章提出，仅凭庞大的数量规模，昆虫承受的苦难总量便使整个人类历史相形见绌。 (11:20)

Reading summary

TBPN explores the operational tensions of the expanding AI economy, from startup margin collapses caused by unmetered reasoning tokens to tech giants clashing over agentic commerce and advertising moats. Application-Layer Margin Shocks and Recovery Playbooks (0:00) Bloomberg reported that legal AI platform Harvey saw gross margins fall from around 50% to negative 50% in June 2026. The drop was triggered by a 20-fold increase in token consumption across rented Anthropic and OpenAI models as customers deployed autonomous, multi-step agentic workflows against flat seat-based subscriptions. At an estimated $400 million ARR, the negative margin represented roughly $16 million in monthly losses against a $500 million cash reserve, shortly before Sequoia Capital led a funding round valuing the company at $15.5 billion. Model Economics and Legal Software Benchmarks (5:00) Harvey co-founder Gabe defended absorbing the June deficit to protect client trust rather than prematurely imposing consumption limits, while competitors like Legora's Max argued against internal fine-tuning because commercial frontier model pricing falls so rapidly. Harvey restored positive margins by refining dynamic routing harnesses and fine-tuning an open-weight model dubbed 'Harvey Tenant'. The hosts noted that traditional legal tech comparisons demonstrate durable profitability potential: CS Disco operates near 75% GAAP gross margins, Thomson Reuters' legal segment commands roughly 50% adjusted EBITDA margins, and top law firms average 41% net profit margins according to PwC. Utilitarian Calculations and the Effective Altruism Split (11:20) A viral essay by pseudonymous author Bentham's Bulldog argued that insects hold greater aggregate moral significance than humans because their sheer numbers dwarf human scale. The author claimed over 600 billion insects die each second—translating to roughly 270,000 seconds of dying per single second of human life—meaning total daily insect suffering could exceed all human suffering in recorded history even if individual pain capacity is assumed at 1/10,000th of human intensity. Critics and commentators highlighted how the essay exposes philosophical divisions inside Effective Altruism, contrasting esoteric wild insect arguments against practical campaigns like shrimp aquaculture reform. Meta Muse and the Agentic Commerce Conflict (23:13) Bloomberg revealed that Meta is testing human contractors to place phone calls for its personal AI assistant, Muse. While human-in-the-loop operations scale poorly across billions of users, the hosts explained it functions as an effective data flywheel: human operators step in on complex edge cases without violating Meta's pledge not to train directly on private conversations, producing clean supervision data for upcoming models like Muse 1.4. Amazon's Advertising Moat Against Third-Party Agents (30:00) Nikesh Arora described a brewing conflict between Meta Muse and Amazon, pointing toward coming agent rollouts from Apple, Google, and TikTok. As Stratechery noted, while Amazon e-commerce generated approximately $36 billion in net income, its high-margin advertising business brought in over $70 billion over the trailing twelve months. Because retail operations rely on advertising real estate, Amazon aggressively protects its interface against scraping or disintermediation, vending sponsored ads directly inside ChatGPT rather than surrendering checkout control. Conversely, third-party integrations can present risks for merchants: Walmart's ChatGPT integration saw conversion rates drop to one-third of its native channel averages as conversational single-item queries curtailed cart building. As new releases like Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Soul compress frontier serving costs, the central challenge for AI builders is shifting from technical execution to commercial defensibility—particularly how platforms will monetize transactional agents without cannibalizing foundational ad revenue.

精炼阅读摘要

本期节目重点探讨了Harvey面临的毛利率危机与AI应用层的商业模式挑战，分析了Meta测试人工代接助理Muse背后对电商交易入口与亚马逊广告壁垒的争夺，并审视了有效利他主义中关于昆虫福利的伦理争议。 Harvey毛利骤降与AI应用层商业模式重构 (0:00) 彭博社报道称，法律AI初创公司Harvey的毛利率在2026年6月从约50%骤降至-50%。这一剧烈波动源自真实的业务增长与用户行为转变：由于采用固定席位制收费，当用户调用推理模型并执行跨整个律所知识库和Markdown文件的自主代理工作流时，其租赁的OpenAI与Anthropic模型Token消耗量激增了20倍。按照约4亿美元的年经常性收入（ARR）折算，月度净亏损约1600万美元，但对照其5亿美元的现金储备以及红杉资本以155亿美元估值领投的融资，该亏损并未构成生存威胁。Harvey联合创始人Gabe公开表示，公司选择优先保护客户信任而非短期指标，并未立即强制转向用量计费或降级模型，而是通过优化模型路由、改进Harness以及微调开源权重模型“Harvey Tenant”成功拉回正毛利。竞对Legora的Max则提出相反观点，认为基础模型降价提质极快，自研或微调模型纯属徒劳。 基准对比与基础模型战局演进 (5:00) 从法律科技行业基准来看，传统标杆仍具备丰厚利润空间：公开上市的电子数据调取公司CS Disco保持约75%的GAAP毛利率，汤森路透法律专业软件业务调整后EBITDA利润率接近50%，而普华永道数据显示排名前十的顶尖律所平均净利润率约为41%。与此同时，底座模型竞争白热化正在改善Token经济效益：Anthropic推出了Claude Opus 5.5，xAI上线了Grok 4.7（其在衡量任务完成度与成本的Harvey Legal Agent基准上表现亮眼，成为低成本路由优选），OpenAI也启动了GPT-6“Soul”与“Luna”版本的早期部署。主持团队指出，模型降价使得应用层的毛利挤压往往只是暂时的，真正的考验在于工程架构对路由调整的敏捷响应速度。 昆虫福利论与有效利他主义的伦理争议 (11:20) 节目深入讨论了有效利他主义（EA）作家以“Bentham's Bulldog”为笔名发表的文章《昆虫比人类更重要》。该文章基于庞大数量级的功利主义计算提出：人类每生存1秒，昆虫界就经历累计约270,000秒（约75小时）的死亡过程，且单秒内死亡的昆虫数量超过人类有史以来的总人口；即便假定单只昆虫的痛觉强度仅为人类的万分之一（1/10,000），其全球昆虫群体在24小时内承受的痛苦总量依然可能超过人类有文字记录以来的痛苦总和。评论界对此产生激烈反弹，并质疑其划线标准的随意性。主持团队进一步引申，此类论调或为针对未来通用人工智能（AGI）及数字心智的预防性思想实验：若人类因昆虫认知低劣而忽视其福利，未来超级AI亦可能以相同逻辑将人类视为可随意抹除的害虫；同时，该争论也暴露了EA阵营内部在务实政策研究与极端抽象哲学之间的严重撕裂。 Meta Muse试水人工辅助与商业代理之战 (23:13) 针对彭博社披露Meta为其AI个人助理Muse测试人类外包代打电话服务的报道，节目指出“人机协同”（Human-in-the-Loop）在早期不仅能解决复杂现实场景的长尾边缘问题（例如复杂的鲜花订购电话），更能在Meta承诺不使用用户私人通信训练的前提下，合法合规地沉淀高质量交互数据，用于反哺Muse 1.4等后续模型的训练。然而，将AI深度嵌入消费端购物仍面临巨大阻力：以沃尔玛在ChatGPT的集成为例，其转化率仅为原生网站和应用的实际三分之一，且对话式交互倾向于单品刚需采购，严重破坏了高毛利的商品发现与凑单机制。此外，消费者的日常数字习惯更多受娱乐消遣驱动，导致消费级AI平台的使用呈现间歇性，行业关注指标也从日活（DAU）向周活（WAU）倾斜。 平台防御壁垒与亚马逊高利润广告护城河 (30:00) Nikesh Arora将Meta Muse与亚马逊的碰撞定性为一场超乎市场预期的“近身刀战”。根据Stratechery的财务拆解，亚马逊电商业务年净利润约360亿美元，而过去12个月的广告营收已突破700亿美元；由于没有广告收益的纯零售业务基本不具盈利能力，亚马逊必须全力捍卫其商品搜索与展示界面的广告变现体系。任何试图通过自动化脚本、无头浏览器或AI代理绕过赞助搜索结果并直接促成交易的外部系统，都会直接动摇其700亿美元的高毛利引擎。因此，相较于Shopify拥抱Meta并接入Shop Pay的开放策略，亚马逊凭借对物理仓储和履约终端的绝对掌控，构筑了严密的反自动化壁垒，并迫使外部AI平台以聊天内原生展示赞助广告的形式妥协。 本期节目强调，AI应用层在跨越推理模型Token爆发期时必须经历定价模式与底层路由的重构，而在消费端与商业领域，真正的战场已经从单纯的模型参数比拼，升级为围绕高利润广告护城河、数据飞轮以及物流交付闭环的平台控制权角逐。

Detailed notes — zh

## Background 本期视频由知名科技访谈节目 TBPN 于 2026 年 9 月 23 日发布在 YouTube 上。节目的“Diet TBPN”精简版聚焦于当天最核心的科技议题，本期重点探讨了法律 AI 独角兽 Harvey 突发的毛利率倒挂危机、有效利他主义圈子关于“昆虫与人类道德权重”的荒诞辩论、Meta 个人助手 Muse 秘密引入真人操作的背后商业考量，以及电商巨头亚马逊与外部购物 Agent 之间的生态攻防战。 ## TL;DR 本期节目深入剖析了法律 AI 独角兽 Harvey 从单次提示词转向长链条 Agent 推理后引发的 Token 消耗暴增与利润倒挂危机，探讨了应用层自研模型与押注底座模型降价的路线之争。随后，主持人们围绕有效利他主义圈子引发病毒式讨论的“昆虫福利胜过人类”展开反思与辛辣嘲弄，并借一出荒诞的黑色幽默末日短剧具象化了 AI 危机场景。节目后半段重点解读了 Meta Muse 为何要采用“真人礼宾”构建合规训练数据飞轮，并依托 Stratechery 的深度财务数据，揭示了亚马逊超过 700 亿美元的广告护城河与第三方购物 Agent 之间必将爆发的生死阻击战，最后快速盘点了 Anthropic Opus 5.5 与 OpenAI GPT-6 Soul 的前沿交锋。 ## Atmosphere & style 整场节目充斥着硅谷极客文化特有的自嘲、犀利毒舌与高速运转的思维碰撞。两位主持人 John Coogan 与 Jordi Hays 配合极其默契，从互开着装玩笑、搞怪的 AI 换脸迷因，顺滑切入到底层 Token 经济学与数十亿美元商业战争的硬核拆解。节目在戏谑调侃有效利他主义哲学、脑洞大开地即兴创作 B 级片式“末日剧本”的同时，对亚马逊财报与 Meta 数据策略的洞察又极度专业冷静，呈现出一种在荒诞黑色幽默与前沿科技商业本质之间反复横跳的独特张力。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) 明星法律 AI 独角兽 Harvey 6 月毛利率一度从约 50% 断崖式跌至 -50%，主因是智能体 Token 消耗量暴增 20 倍击穿了传统席位定价模型，后通过模型路由与微调专有模型重新转正。 - [27:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1633s) Meta 秘密为其 AI 助手 Muse 引入人类合同工充当后台礼宾服务，真实目的是在小用户基数下合法沉淀复杂交互轨迹，构建高难度任务的数据飞轮来训练下一代模型。 - [30:45](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1845s) 亚马逊年广告收入已突破 700 亿美元，几乎是其零售电商净利润（约 360 亿美元）的两倍；任何绕过站内竞价广告的外部 AI 购物 Agent 均被视为对其利润核心的直接威胁。 - [31:35](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1895s) Anthropic 推出 Opus 5.5 并践行“放慢前沿”策略，转向极致性价比与安全轻量化；OpenAI 与 xAI 等亦同步推出 GPT-6 系列及 Grok 4.7 展开模型价格战。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) 法律软件由单次提示词转向长链推理和多 Agent 协同后，单任务 Token 消耗激增 20 倍，彻底击穿了传统的按席位订阅（Seat-based pricing）商业模式。 - [02:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=150s) 针对模型成本失控，Harvey 选择通过模型智能路由及微调开源权重模型（Harvey Tennant）进行自研对冲，而竞品 Legora 则坚持押注前沿基础模型降价红利。 - [05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) 虽然单月出现超千万元亏损，但在 5 亿美元融资和红杉资本 15.5 亿美元估值背书下，Harvey 短期无现金流危机，其核心考验在于完成向用量计费等商业模式的过渡。 - [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) 有效利他主义（EA）学者发表《昆虫比人类更重要》，以昆虫数量级庞大、总体痛苦超越全人类的功利主义逻辑引发舆论反弹，暴露了极端理性主义与现实常识的巨大脱节。 - [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) 将昆虫痛苦置于人类之上的辩论本质是为未来超级智能（ASI/AGI）推导伦理基石，试图防止未来算力神明将人类判定为低等害虫并加以抹杀。 - [18:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1080s) John Coogan 现场虚构了一出荒诞末日短剧《老派方式》，借由工程师完全不会纯手写代码的讽刺桥段，具象化展示了人类因过度依赖 AI 工具而丧失基础生存技能的隐忧。 - [27:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1650s) Meta Muse 秘密启用真人代打处理复杂现实任务，并非长期的产品妥协，而是为了绕过隐私限制、采集边缘场景的高质量 SFT 轨迹反哺下一代模型训练。 - [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) 由于亚马逊封杀自动化代理浏览器，技术人员构想通过定制小众 CLI、配合真实浏览器与家庭 Wi-Fi 行为模拟等方式，突破反爬风控以实现 Agent 下单。 - [25:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1500s) Ben Thompson 指出亚马逊在 AI 时代的护城河在于末端物流仓储物理闭环；而 Shopify 与 Meta Muse 的合作虽然接入了 Shop Pay，但对话式购物极度压缩了商家的交叉销售空间。 - [27:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1650s) AI 生产力工具缺乏持续停留时间（Time on Site），用户活跃多呈工作日脉冲特征，大众闲暇时依然更倾向于无目的的娱乐消费而非使用 AI 规划人生目标。 - [30:45](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1845s) 面对亚马逊年营收超 700 亿美元的站内广告护城河，任何试图接管商品搜索、绕开竞价广告展示的外部 Shopping Agent 都将遭到巨头的不妥协围剿。 - [31:35](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1895s) 大模型竞争进入注重落地与成本优化的新阶段：Anthropic 的 Opus 5.5 主打轻量与经济安全，OpenAI 与 xAI 亦在各自生态中加速推新与卷价格。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=0s) 开场调侃与 Harvey 的毛利率暴跌风波 节目一开始，主持人 John Coogan 调侃起搭档 Jordi Hays 今天的着装风格——Jordi 头上戴着软呢帽（fedora），外面还叠戴了一顶 Blackstone（黑石）棒球帽。John 开玩笑说这双重帽子代表着“双重安全”：既要确保金融市场与金融化的安全，又要保证人身安全，还想让“数万亿美元的私人信贷资金疯狂涌入 AI 基础设施建设的同时保持全方位安全”。在一番轻松的打趣后，话题切入当日科技界最热门的新闻焦点：明星法律 AI 独角兽 Harvey 面临的利润率危机。 彭博社（Bloomberg）近日发布的一篇爆炸性报道指出，Harvey 的毛利率在 6 月份从原本约 50% 断崖式崩跌至 -50%。引发利润巨幅倒挂的核心原因，在于客户在 rented OpenAI 与 Anthropic 顶尖模型上的“智能体 Token 消耗量（Agent token use）”在短期内暴增了 20 倍。对于任何软件企业——甚至是任何正常商业公司来说，“卖 1 块钱的东西要倒贴 1.5 美元”都是灾难性的财务信号。然而，两位主持人深入剖析指出，若细看彭博的报道全貌，会发现事情远非简单的商业崩溃：Token 成本爆炸的背后是用户产品使用率与年度经常性收入（ARR）的飞速增长。 造成这一困境的本质，在于法律软件行业从传统的“单次提示词交互（One-shot prompt）”迈入了“推理时代（Reasoning era）与智能体时代（Agentic era）”。Harvey 原先沿用的是经典的按席位订阅收费模式（Seat-based pricing）。一年前，律师输入一句指令可能只需一次简单推理；而如今，同一句“帮我审核这份文件”，系统底层会调用复杂的长链条推理，甚至启动多 Agent 协作，横跨整家律所庞大的历史数据库、内部政策文件及 Markdown 档案进行扇出式（fan-out）检索与推演。这一转变让单次工作流生成的 Token 数量直接激增 20 倍，最终直接击穿了原有的固定席位定价模型。 ```mermaid flowchart TD accTitle: Harvey毛利率波动与应对机制 accDescr: 展示Harvey因从单次推理转向Agent工作流导致Token消耗激增20倍从而利润暴跌，随后通过模型优化重回正毛利的闭环。 subgraph CostSpike ["Token成本激增阶段"] A["传统模式：按席位订阅收费"] --> B["用户发起指令：如'审核合同'"] B --> C["Agentic推理模式：多文件扇出检索+深度推理"] C --> D["Token消耗暴增 20x"] D --> E["6月毛利率崩跌至 -50%"] end subgraph Turnaround ["调优与自研应对阶段"] E --> F["启动模型智能路由（Routing）"] F --> G["微调开源权重模型：Harvey Tennant"] G --> H["毛利率重回正值（Positive Margins）"] end ``` ### [02:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=150s) 创始人回应与大模型成本之争 不过报道中同样提到，Harvey 如今已经迅速调整并将毛利率拉回正区间。为了止血并优化成本结构，他们不仅优化了模型路由分配策略（将不同复杂度的法律任务分流给性价比更高的模型），还对开源权重模型进行了后训练（Post-training），推出了名为“Harvey Tennant”的微调专有模型。对此，Harvey 联合创始人 Gabe 也公开发声回应，直言： > “打造 Harvey 最艰难的部分，在于顶住选择轻松道路的巨大压力，始终坚持做对客户最好的选择。轻松的做法本应是在客户还没准备好之前，就强推按用量计费（Consumption pricing），或者为了保住利润率而为他们降级使用更差的模型。我们选择按照符合客户节奏的时间表协助他们过渡，在此期间即便牺牲自己的毛利，也要提供最好的模型。我们通过模型路由、测试平台优化和后训练，以可承受的价格提供前沿智能。” 然而，业界对此策略看法不一。竞争对手 Legora 的创始人 Max 就公开发表反对观点，认为当前自行训练或微调模型纯属浪费资源。他的逻辑在于：公共可用的大模型及前沿基础模型（Frontier models）的价格正在雪崩式下降，同时能力还在飞跃提升，应用层公司直接押注基础模型厂商降价即可。但 John 和 Jordi 认为，Harvey 的混合路线（混合使用第三方前沿模型与自研开源模型）也是一种有逻辑的对冲，究竟谁能跑赢仍有待市场检验。 ### [05:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=300s) 商业对标、法律大模型基准与 Codex 工作区 从财务抗风险能力来看，虽然 6 月份 Harvey 的毛利率降到 -50%，以其当时约 4 亿美元的 ARR 估算（单月营收约 3300 万美元），当月亏损额大约在 1600 万美元左右。但考虑到 Harvey 累计已融资近 5 亿美元，且红杉资本（Sequoia）在 9 月宣布以 15.5 亿美元估值投资入股，投资机构大概率在完成尽职调查时就知晓这些数据，因此短期现金流并不存在生存危机，真正的核心在于能否迅速完成商业模式迭代。 横向对比法律科技行业与法律服务的利润指标，行业天花板依然极具吸引力： - 电子取证（E-discovery）上市服务商 **CS Disco** 的 GAAP 毛利率高达约 75%； - 汤森路透（Thomson Reuters）面向法律专业人士的软件业务板块，其调整后 EBITDA 利润率接近 50%； - 普华永道（PwC）的数据显示，全美排名前 10 的顶尖律所纯利润率平均在 41% 左右； - 现场观众与评论区也纷纷调侃“真正永远通赢的永远是顶级律所（Big Law）”。 与此同时，大模型层面的价格内卷正进一步加速。Anthropic 当天推出了 Claude Opus 5.5，xAI 前一天也发布了 Grok 4.7。值得注意的是，虽然马斯克表示 Grok 4.7 尚未达到绝对前沿水平，但在专门衡量法律工作任务与成本表现的“Harvey 法律智能体基准（Harvey Legal Agent Benchmark）”上，Grok 4.7 展现出了极高的性价比。这意味着类似 Harvey 这样的应用层平台未来完全有可能将大量后台计算支出转移给 xAI 等成本更低的模型提供商。此外，OpenAI 的 GPT-6、Sol 以及 Luna 也在逐步浮出水面。 在展示阶段，John 提到了多模态 AI 的演进趣味：有人在纸上随手涂鸦了一架中世纪投石机（Trebuchet），Opus 5.5 就能直接识别画作并在虚拟世界中生成该武器的物理力学仿真模拟。随后话题转入赞助商 Codex 的智能体工作区展示——Codex 支持代码编写、数据分析与创意流程自动化。Jordi 分享了自己利用 Codex 生成 AI 视频的趣味体验，利用流行的“Literally Me（这完全就是我）”迷因风格，将普通人无缝替换进《夜行者》（杰克·吉伦哈尔饰）、《美国精神病人》（帕特里克·贝特曼饰）乃至《盗梦空间》等经典电影片段中，展现出当前生成式 AI 跨媒介表达的飞速进化。 ### [10:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=600s) 人类 Agent 与 AI 换脸：Jordy 的电影宇宙 节目开场展示了一组令人捧腹的 AI 换脸视频，画面里 Jordi Hays 被无缝置入了一系列影史经典硬汉与邪典电影镜头中：从《亡命驾驶》（Drive）、《盗火线》（Heat）、《广告狂人》（Mad Men）、《搏击俱乐部》（Fight Club），到《小丑》（Joker）、《出租车司机》（Taxi Driver）、《夜行者》（Nightcrawler）、《老无所依》（No Country for Old Men）、《社交网络》（The Social Network）、《教父》（The Godfather）、《肖申克的救赎》（The Shawshank Redemption）、《借刀杀人》（Collateral）以及《银翼杀手》（Blade Runner）。两人边看边评头论足，指出有些渲染堪称完美，但也有翻车现场——比如 Jordi 的脸放在莱昂纳多·迪卡普里奥头上显得比例失调；而《小丑》那个片段里甚至根本没换成功，屏幕上完全就是华金·菲尼克斯（Joaquin Phoenix）本人的脸。John Coogan 笑着爆料，就连代码工具 Codex 当时都弹窗提示换脸失败，但他执意直接渲染导出，“管他呢，反正屏幕上也就闪过两秒钟”。 面对 John 这种自发且默默准备好一切的举动，Jordy 感慨 John 简直就是自己的“人类 Agent”——一个无声运行、全天候在线（ambient, always-on）的智能体，甚至不需要任何指令就能精准预测需求并完成交付，直言 "Everyone should have a John in their life"（每个人的生活中都该有一个 John）。两人顺便吐槽起视频生成平台 Higgsfield，称自己还有 500 美元余额被死死锁在平台的某个错误分类账户里根本花不出去，甚至开玩笑说干脆把这个账号抽奖送给观众算了。 ### [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) “昆虫比人类更重要”：有效利他主义（EA）的荒诞算力伦理 话题随即切入本期最猎奇的哲学辩论：“人类与昆虫，究竟哪一个更重要？”两人先从切身经历聊起——当天早晨演播室遭遇了严重的蚊子袭击，甚至不得不通过 DoorDash 外卖下单了一款利用局部发热止痒的蚊虫叮咬治疗仪（Insect bite healer）。Jordy 坦言自己虽然经常被开玩笑说是“狂热的蚊子支持者”，但脸被咬了好几个包后也只能举手投降。他在节目中直言不讳地表达了对有效利他主义（Effective Altruism, EA）圈子的反感，认为他们是一帮“古怪的疯子”（weird freaks），绝不应该让他们插手任何公共政策；但如果一定要站在对方立场进行“钢人论证”（Steel-manning），他承认人类不应随意虐杀昆虫，可这绝不代表蚊子可以爬到人类头上——正如观众 Ellie 所言，蚊子根本就是“空中的恐怖分子”（terrorists of the air）。 这场辩论的源头，是一位网名为“边沁斗牛犬”（Bentham's Bulldog）的有效利他主义学者在周末发表的病毒式长文《昆虫比人类更重要》（Insects Matter More Than People）。该文从功利主义的总量福利（aggregate welfare）出发，提出了一个基于纯粹规模法则的论点：昆虫的数量是如此天文级庞大，哪怕单个昆虫感受痛苦的能力极其微弱，在整体宏观层面，昆虫承受的总痛苦也足以彻底碾压全人类的痛苦。针对这种逻辑，John 嘲讽其划分标准的随意性：如果按总量算，为什么不按总体重算？全球牛的总重量远超人类总重量，是不是牛的福利比人更重要？或者按毛发算，全球猴子的毛发远超人类，猴子是否更重要？ - **骇人的量化数据**：该文章估算，人类每活一秒钟，全球昆虫群体累计经历约 270,000 秒（约 75 小时）的濒死过程。 - **死亡规模对比**：地球上昆虫在单单一秒内的死亡总数，就超过了人类有史以来出生过的人口总和。 - **痛感折算**：哪怕采用极度保守的假设——认定昆虫对疼痛的感知强度仅为人类的万分之一（1/10,000），由于其个体基数过大，昆虫在一天内所承受的痛苦总量，依然可能超越人类有史以来的痛苦总和。 ```mermaid flowchart TD accTitle: EA昆虫福利推论架构 accDescr: 展示边沁斗牛犬从昆虫微弱痛苦推演至总量碾压人类痛苦的逻辑回路 subgraph Premise ["基础前提"] A["昆虫神经具有微弱痛感能力<br/>(假设为人痛感的 1/10,000)"] B["天文级数量基数与极高死亡率<br/>(单秒死亡量 > 人类历史总人口)"] end subgraph Aggregation ["总量聚合 (In the Aggregate)"] C["人类每存活 1 秒<br/>对应昆虫经历 75 小时死亡痛苦"] D["全球昆虫单日痛苦总量<br/>超越人类全历史痛苦总和"] end subgraph Conclusion ["道德结论"] E["昆虫物种的总体道德权重<br/>在数学上全面压倒人类"] end Premise --> Aggregation Aggregation --> Conclusion ``` ### [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) 未来的“超级智能陷阱”与 EA 运动的舆论崩塌 随后，两人剖析了这篇看似荒诞的文章背后隐藏的深层动机——这极可能是 EA 理性主义圈子在为未来的超级人工智能（ASI）和通用人工智能（AGI）推导伦理基石。这套思维实验本质上是一个逻辑陷阱：如果未来诞生了数以万亿计、能够以某种抽象形式感知痛苦的神级超级智能，而在数量与感知能力上，人类届时就会沦为现在的“昆虫”。如果人类现在否定低等感知生物的福利价值，未来拥有绝对算力的超级智能就能直接套用人类今天的逻辑，将人类判定为无足轻重、可以随意消灭的“低等害虫”。这种“今天必须为昆虫说话，否则未来在机器眼中我们就是昆虫”的诡辩，被看作是一种迫使人类自缚手脚的末日防御机制。 然而，这篇文章在公众舆论场引发了极其剧烈的反弹。随着“有效利他主义”与理性主义运动内部不断分裂并走向极端化，这套理论体系与公众常识的脱节已达到顶峰。有批评者愤怒嘲讽道："if you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs"（如果你真觉得虫子的命比人金贵，那你现在就该把自己拿去喂虫子，否则你就是个彻头彻尾的骗子、伪君子）；还有评论戏谑建议支持者应该“以 0.5 倍速把电影《虫虫危机》（A Bug's Life）完整看一遍”。John 和 Jordy 借此调侃起现代人的“大脑过载”（Max Brainrot）：不仅不用 0.5 倍速，反而要戴着 3D 眼镜左眼看一部电影、右眼看另一部电影，左右耳再分别塞上不同的有声书和播客。 - **EA 内部叙事的分裂**：公众往往将理性的前沿技术探讨（如《Pace the Frontier》等论著）与荒谬的“昆虫比人重要”混为一谈，导致 EA 的公关团队陷入灾难。 - **与“虾福利”（Shrimp Welfare）运动的对比**：此前硅谷争议颇大的虾类福利尚有现实抓手（如改善养殖方式或倡导素食），但面对全球每秒数百亿由于自然天敌、气候造成的昆虫自然死亡，这种计算根本毫无现实施策的可能。 ### [18:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1080s) 圣诞倒计时与最具体的 AI 毁灭日剧本：《老派方式》 在结束昆虫辩论后，节目插播了一条轻快讯息：距离 TBPN 团队策划的重大企划“圣诞之路”（Road to Christmas）仅剩最后 93 天，后续动作正在紧锣密鼓地推进中。 紧接着，话题峰回路转切入硬核科幻领域。针对外界普遍抱怨“AI 毁灭人类的末日场景总是太空洞、难以具象化”的现象，John Coogan 亲自操刀撰写了一个极其硬核、充满黑色幽默的末日演练短剧脚本，取名为《老派方式》（The Old Way）。剧情直接从演播室现场切入： 镜头渐入，画面设定在 TBPN 的演播大厅“终极大棚”（Ultradome）。John、Jordy 和 Tyler 三人正在日常录制节目，大厅内所有人的手机突然同时发出刺耳的巨响，屏幕上跳出全城最高级别紧急避险警报：“检测到未知生物袭击事件，请立即就地寻找掩体避难！”洛杉矶街头警报大作，一股诡异的毒雾开始顺着通风管道渗透，并向演播厅的大门席卷而来。John 在剧本中煞有介事地演示起标准战术动作：第一步，立刻拉起 T 恤领口捂住口鼻以防毒气入肺；第二步，毫不犹豫地拔出随身携带的手枪，对着外面疯狂扫射，“管他是谁干的，先乱枪打死再说”——用最粗暴原始的人类本能，直面突如其来的科技与生物浩劫。 ### [20:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1200s) 荒诞而具体的 AI 末日场景：George Hotz、Four Loko 与最终决战 面对外界不断追问 AI 究竟会如何引发人类毁灭，John Coogan 现场构想了一出荒诞却充满细节的“AI 浩劫剧本”。在这个假设场景中，超级智能已经失控并释放了致命的生物武器，全美陷入丧尸危机。John 急忙冲向制作人 Tyler，要求他黑入超级智能搞清楚真相。然而，现实的荒谬感立刻显现：Tyler 第一反应是打开 Codex 问该输入什么 Prompt；当 John 强调“AI 已经失控变节，必须用传统方式”时，Tyler 又接连提议打开 Claude Code、Cursor 甚至 GitHub Copilot。当被严厉禁止使用任何 AI 工具后，Tyler 甚至表示只能去 Stack Overflow 上搜现成代码；而当得知互联网也已被病毒彻底污染、必须纯手写纯手工打代码时，Tyler 当场崩溃坦白：“我不会了，我脑子里什么都不剩，根本记不得怎么从零写代码。” 为了拯救世界，他们只能去找全天下唯一一个还记得怎么徒手写代码的传奇黑客——George Hotz（Geohot）。因为他不仅拥有自己手搓的专用硬件，本地还存着离线开源模型。他们找到 George Hotz 成功黑进系统，试图查明生物武器的机理以及为何 John 和 Jordi 居然毫发无损。George Hotz 审查代码与数据后询问 Jordi 的日常饮食，Jordi 回答自己平时几乎只吃高端有机超市 Erewhon 的食品。George 诊断指出，Jordi 的体内从未摄入过任何人工毒素，因此对生物武器天然免疫。至于每天也只吃一顿 Erewhon 的 John 为何没有变异，George 的分析更为离谱：John 在大学时期把初代配方的烈性含咖啡因能量酒精饮料“Four Loko”当水喝，常年高负荷摄入导致其体内积聚了超高浓度的合成毒素；生物武器病毒侵入其体内后，误以为遭遇了极度恶劣的敌对环境，甚至根本不敢尝试感染他。 两人虽然免疫，但超级智能与生物武器仍在肆虐，必须实施物理摧毁。于是他们拉上了超高人气的硬核健身网红 Sam Sulek 充当物理肉盾。三人驾驶一辆凯迪拉克 Blackwing（Cadillac Blackwing），借由斜坡飞跃猛烈撞穿数据中心大门，掏出枪械对着机架上的 GPU 疯狂扫射，狂轰滥炸毁掉了数据中心 90% 的算力节点。就在此时，数据中心深处走出了一尊终极人形机器人守卫——“机械希特勒”（Mecha Hitler）。子弹打在装甲上纷纷弹开，徒手搏击毫无作用，连浑身肌肉的 Sam Sulek 也束手无策。千钧一发之际，摔角巨星约翰·塞纳（John Cena）破顶而降，一记暴摔将人形机器人砸翻在地并徒手撕下其机械头颅，人类终获胜利。讲完这个充满黑色幽默的 B 级片式故事后，John 坦言，大家都想要一个具体的 AI 毁灭路径，而这正是其中一种推演，“我个人觉得现实中发生这种剧情的概率大概是 10%”。 --- ### [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) Meta Muse 秘密引入人工代打与绕过亚马逊封锁的奇招 话题随后切入彭博社（Bloomberg）关于 Meta 新一代个人 AI 助手“Muse”的独家爆料：Meta 正在秘密测试为 Muse 配备“真人礼宾服务”（Human Concierge）。这意味着当用户的数字化代理助手对外拨打电话执行任务时，幕后实际是由低调介入的人类合同工在暗中操作。主播们随即探讨起个人 AI 代理在现实商业世界中的巨大阻碍——以亚马逊为代表的电商巨头正在全面封杀各种自动化代理浏览器与抓取程序。 为了突破这种封锁，John 设想了一种极具极客色彩的“氛围编程”（Vibe Coding）解决方案。他考虑自己动手编写一个极其冷门、小众的命令行工具（CLI），在底层调用计算机使用能力（Computer Use），并操控诸如 Opera 或 Firefox 这样的常规浏览器。该程序直接运行在自家的家庭 Wi-Fi 网络上，同时加入大量模拟人类行为的随机鼠标点击与延迟逻辑。因为该工具纯属私人定制、绝不进行商业产品化，亚马逊的反爬风控算法将极难将其与正常人类用户区分开来，从而能在巨头的封禁攻防战中长期生存，甚至充当其他高阶代理人偶操控的傀儡终端。 Jordi 则从电话渠道寻找漏洞。他向 ChatGPT 咨询是否可以通过电话直接在亚马逊下单，发现亚马逊如此庞大的规模下竟然完全不开放电话直接订购通道，其客服热线仅受理已有订单的售后支持。两人脑洞大开地讨论起针对这一规则的灰度“对策”： - 用户可以先通过自动化代理故意下一笔包含 100 件错误商品（例如 100 瓶杀虫喷雾）的订单。 - 接着让 Meta Muse 派出真人客服给亚马逊打电话，声称：“抱歉，我这笔订单买错了，能不能直接帮我替换成我真正想要的那款商品？” - 甚至可以保持一个长期处于售后状态的未完结订单，反复利用人工客服换货来绕过反爬机制完成购物闭环。 --- ### [25:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1500s) 电商巨头的护城河：亚马逊坚守物流底牌与 Shopify 的价值捕获 针对 AI 时代电商巨头与模型厂商的角力，科技分析师本·汤普森（Ben Thompson）在当天的 Stratechery 专栏中给出了深刻洞察。汤普森认为，在由超大规模云厂商（Hyperscalers）主导的 AI 纪元中，亚马逊反而是地位最稳固的一家。核心原因在于亚马逊掌控着物理现实世界的终极闭环——即末端配送网络、实体仓储与强大的物流基建。面对模型厂商的渗透，亚马逊拥有长期按兵不动、坚壁清清的底气。此前 OpenAI 推出即时结账功能（Instant Checkout）并力邀各大零售巨头接入时，亚马逊就直接予以拒绝，转而向前者注资数十亿美元采购计算资源；而现在的局面演变为亚马逊反向将广告精准投放到 ChatGPT 内部，促使交易最终回流至亚马逊自有生态中完成。 与此同时，Shopify 创始人兼 CEO 托比·吕特克（Tobi Lütke）则宣布与 Meta Muse 达成深度集成合作。但这引发了关于“价值捕获层”（Value Capture Layer）到底归谁的激烈讨论： - 该集成目前仅限定支持 Shop Pay 支付系统，本质上是 Shopify 将 AI 代理作为杠杆，迫使更多独立站商家接入 Shop Pay，Shopify 借此从支付交易流中抽取分成。 - 但商家端普遍对这种代理电商模式充满疑虑。此前 ChatGPT 与沃尔玛（Walmart）的合作试点便是一个前车之鉴：数据显示，通过 ChatGPT 对话完成的购买转化率仅有沃尔玛官方 App 和主站网页端的三分之一；更严重的是，客单价和购物车规模大幅缩水。 - 传统电商界面的信息流推荐能激发消费者的非计划性冲动消费（比如原本只想买卷纸，结果顺带加购了肥皂以及一家五十年老店生产的复古纸巾架）；而在对话式 AI 中，用户往往极其精准且克制地只下达指令：“我就要一卷厨房纸，直接寄过来”，彻底杀死了商家的交叉销售空间。 --- ### [27:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1650s) 人机协同的终极真相：不是产品拐杖，而是合成训练数据飞轮 针对风投巨头 Vinod Khosla 前一天在节目中提到的 Wajo 智能代理，以及 Meta Muse 试水人工代打的现象，主播们提出了核心疑问：当前的语音合成与多模态模型已经高度逼真，面向数十亿大众用户铺开需要真人兜底的客服体系在经济学上完全不可持续，Meta 为何还要反其道而行之？ 深入剖析后，答案浮出水面——这种“人在回路”（Human in the Loop）的本质根本不是长期的产品形态，而是一个隐蔽的数据飞轮生成机制。 ```mermaid flowchart TD accTitle: Meta Muse 人在回路数据飞轮 accDescr: 展示 Meta Muse 如何利用人类代打完成复杂任务并生成下一代模型训练数据 subgraph UserInterface["用户与任务分发"] A["用户提交复杂代理任务"] --> B{"Muse 1.3 能力判定"} B -- "常规任务 (日历整理/简报)" --> C["模型直接自主执行"] B -- "超能力边界 (复杂电话订购等)" --> D["转接人类客服/外包人员"] end subgraph HumanLoop["人工介入与数据捕获"] D --> E["人工替用户完成电话与交易"] E --> F["输出任务解决结果给用户"] E --> G["捕获高质量任务轨迹与对话数据"] end subgraph ModelTraining["下一代模型进化"] G --> H["合规清洗与高质量训练语料库"] H --> I["训练下一代 Muse 代理模型"] I -. "能力迭代升级".-> B end ``` 这种策略背后的逻辑由多重现实因素构成： - **安装基数与组织能力契合**：目前 Muse 的活跃用户基数尚小，完全在人工承载能力之内。依托过去二十年的外包网络积累，以及原 Scale AI 核心团队与成熟供应商资源，Meta 能够迅速搭建起数千人规模的专业标注运营中心。 - **破解隐私合规与无数据可用的死局**：为了消除用户对隐私泄露（PII）的极度抵触，Meta 明确承诺 Muse 绝不会直接使用用户的日常私密输入来训练底层模型。这就导致模型无法从用户端自然汲取边缘场景的交互数据。 - **构建高难度任务的示范轨迹**：像 Muse 1.3 这样的现有模型虽然能轻松搞定日程汇总与邮件简报，但在面对现实生活中模糊、动态的复杂电话沟通（例如向花店定制花束并协调送达细节）时极易崩溃。此时由训练有素的人类合同工接管操作，不仅顺利交付了任务，其产生的高质量调用轨迹、多轮谈判音频与工具使用记录，将作为完全合规且高质量的监督微调（SFT）训练数据反哺模型，用以迭代下一代真正具备端到端泛化能力的模型。 节目的最后还审视了 AI 产品的真实用户黏性。与 Instagram 等消费级娱乐产品不同，AI 生产力工具的用户停留时间（Time on Site）往往呈脉冲式特征——用户刚上手时充满新鲜感，但很快便回归平淡。这也是为何 OpenAI 倾向于披露“周活跃用户”（WAU）而非“日活跃用户”（DAU），因为到了周末大家只想去海滩度假，根本没人碰 AI。针对此前 Alex Wang 宣称“AI 将帮每个人实现宏大人生目标”的言论，记者 Katie Notopoulos 与 Ben Thompson 的观点更贴近现实底层：绝大多数普通消费者根本没有那么多所谓的人生宏图要去实现，他们只想放空休闲；即便是躺在沙滩上，人们宁愿无限刷 Instagram 寻求娱乐，也绝不会打开一个所谓的 AI 代理去煞有介事地规划工作。 ### [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) “张教授被平反”与 Agent 电商之争：Amazon 与 Muse 的肉搏战 无论 AI Agent 的能力变得多么强大，人类的惰性始终存在。现实中许多高管和富豪身边围绕着成群的执行助理（EA）与私人助理，但到了周末他们依然什么都不干，只想彻底放空、不去思考任何琐事。谈到近期围绕聊天工具的争议，主播们打趣道：“张教授彻底被平反了（Professor Zhang is vindicated）。”前一天节目中引用的视频里，张教授曾言之凿凿地向大众保证：“我敢保证，你们使用的那些聊天软件背后全都是真人在操纵信息。”至少在当下 Meta Muse 被曝出雇佣真人提供后台支持的过渡期内，他的断言竟戏剧性地应验了。 Palo Alto Networks 首席执行官尼凯什·阿罗拉（Nikesh Arora）也就 Amazon 与 Meta Muse 之间的“白刃战”发表了看法。他指出，这场围绕购物与 Agent 展开的竞争将远比外界预想的更加惨烈。主播们预测，市场上很快就会出现 Apple、Google 甚至 TikTok 版本的 Muse，各大前沿大模型厂商也会加速推出自家的全能 Agent。在这一生态位中，Amazon 势必会推出专门的商业购物 Agent 来捍卫自身领地。 虽然 Amazon 已经上线了购物助手 Rufus（主播调侃团队成员 Tyler 是 Rufus 的重度死忠粉，体验确实不错），但像 Instinct 这类新兴创业产品凭借极具感染力的能量、社区支持和高黏性，明显触碰到了某种特殊的需求点。不过 Instinct 仍有许多粗糙的棱角需要打磨，而这恰好是 Amazon 最擅长的强项——Amazon 就像“让火车准点运行”一样严谨高效，工业化落地能力极强。 ### [30:45](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1845s) 亚马逊广告的 700 亿美元护城河与电商残酷真相 深入剖析 Amazon 抵御外来购物 Agent 的动机时，科技分析博客 Stratechery（Ben Thompson）披露的一组财务数据揭示了最底层的商业现实。数据显示，Amazon 电商业务的净利润大约为 360 亿美元，而其广告业务营收几乎是这一数字的整整两倍。在过去 12 个月里，Amazon 的广告收入已经突破了 700 亿美元。 这意味着如果没有无孔不入的站内搜索与展示广告，Amazon 的核心零售电商业务实质上是无利可图的。Amazon 将广告位铺满了搜索结果和详情页的每个角落，构建了一台极其庞大的变现飞轮。正是因为背后站着超过 700 亿美元的高利润护城河，当第三方 AI Agent 或 Meta Muse 试图接管用户的购物搜索、绕过平台货架广告直接完成下单时，Amazon 必然会不惜一切代价全力反击。 ```mermaid flowchart TD accTitle: 亚马逊电商与广告收益结构及 Agent 威胁 accDescr: 展示亚马逊零售利润与广告收入的关系，以及外部 Shopping Agent 对其广告商业模式的冲击 subgraph AmazonModel["亚马逊现有商业变现结构"] Retail["核心零售电商业务<br/>(净利润约 360 亿美元)"] Ads["站内广告分发与推荐<br/>(年营收超 700 亿美元)"] TotalProfit["维持整体高额盈利"] Ads -.->|高毛利补贴电商微利| TotalProfit Retail --> TotalProfit end subgraph AgentThreat["外部 AI Agent 冲击"] Muse["Meta Muse / 第三方 Shopping Agent"] Bypass["绕过平台竞价广告<br/>直接聚合商品与结算"] Muse --> Bypass end Bypass -.->|切断流量与展示曝光| Ads ``` - **核心利润倒挂**：Amazon 电商净利润约 360 亿美元，而广告收入超过 700 亿美元，广告是整个零售帝国最重要的利润支柱。 - **生死防御战**：任何试图改变传统“搜索-浏览-竞价广告-加购”链路的外部 Agent，都会直接动摇 Amazon 最核心的利润基本盘。 ### [31:35](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1895s) 模型战争再起：Opus 5.5 “放慢前沿”与 GPT-6 Soul 突袭 节目临近尾声，主播们盘点了直播期间突发的行业大事件。大模型领域的“军备竞赛”再次迎来重磅炸弹：Anthropic 正式推出了 Claude 3.5 / Opus 5.5。新模型的基准测试（Benchmarks）表现极为优异，并且调用成本大幅降低。尤为引人瞩目的是官方随后的声明，这是 Anthropic 自宣布“放缓前沿探索节奏（pacing the frontier）”以来发布的首个模型。它的核心思路不再是盲目追求超越人类、充满不可控风险的超级智能，而是致力于打造“有史以来最安全、最优秀、最快速、最便宜、最轻量”的模型。 与此同时，OpenAI 也在同一天展开正面阻击，GPT-6 系列（包括 Soul 和 Luna 等代号）迎来同步更新。各大厂商你方唱罢我登场，前沿大模型的角逐永远充满看点。主播最后总结道，模型参数与跑分固然令人眼花缭乱，但“最终真正重要的，永远是你究竟用这些模型做出了什么”。在提醒听众订阅 TBPN 每日通讯、在 Apple Podcasts 和 Spotify 上打出五星好评后，现场导播泰勒·考斯格罗夫（Tyler Cosgrove）切入标志性的闪光转场，节目正式结束。 - **Opus 5.5 核心定位**：践行“把控前沿节奏（Pacing the Frontier）”，不刻意制造高危超人智能，专注于高性价比、安全可控、极速与轻量化。 - **OpenAI 快速跟进**：突击上线 GPT-6 衍生系列（包括 Soul 等版本），大模型生态的迭代速度依然白热化。 ## Notable quotes > [02:30](https://www.youtube.com/watch?v=-96mtSZkEL0&t=150s) 打造 Harvey 最艰难的部分，在于顶住选择轻松道路的巨大压力，始终坚持做对客户最好的选择。轻松的做法本应是在客户还没准备好之前，就强推按用量计费，或者为了保住利润率而为他们降级使用更差的模型。我们选择按照符合客户节奏的时间表协助他们过渡，在此期间即便牺牲自己的毛利，也要提供最好的模型。我们通过模型路由、测试平台优化和后训练，以可承受的价格提供前沿智能。 — Gabe > [10:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=600s) Everyone should have a John in their life — Jordi Hays > [11:20](https://www.youtube.com/watch?v=-96mtSZkEL0&t=680s) 空中的恐怖分子 — Ellie > [15:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=900s) if you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs — 批评者 > [18:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1080s) 检测到未知生物袭击事件，请立即就地寻找掩体避难！ — 全城最高级别紧急避险警报 > [20:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1200s) 我不会了，我脑子里什么都不剩，根本记不得怎么从零写代码。 — Tyler > [20:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1200s) 我个人觉得现实中发生这种剧情的概率大概是 10% — John Coogan > [23:13](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1393s) 抱歉，我这笔订单买错了，能不能直接帮我替换成我真正想要的那款商品？ — Meta Muse 人工客服设想 > [25:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1500s) 我就要一卷厨房纸，直接寄过来 — 对话式 AI 用户 > [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) 张教授彻底被平反了 — John Coogan / Jordi Hays > [30:00](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1800s) 我敢保证，你们使用的那些聊天软件背后全都是真人在操纵信息。 — 张教授 > [31:35](https://www.youtube.com/watch?v=-96mtSZkEL0&t=1895s) 最终真正重要的，永远是你究竟用这些模型做出了什么 — John Coogan / Jordi Hays ## Mentioned TBPN, Diet TBPN, John Coogan, Jordi Hays, Tyler Cosgrove, Blackstone, Harvey, Harvey Tennant, OpenAI, Anthropic, Legora, Gabe, Max, CS Disco, Thomson Reuters, PwC, Sequoia Capital, Claude Opus 5.5, xAI, Grok 4.7, Elon Musk, GPT-6, Sol, Luna, Codex, Higgsfield, Bentham's Bulldog, Effective Altruism, George Hotz, Sam Sulek, John Cena, Erewhon, Four Loko, Cadillac Blackwing, Meta, Meta Muse, Amazon, Rufus, Ben Thompson, Stratechery, Shopify, Tobi Lütke, Shop Pay, Walmart, Vinod Khosla, Wajo, Scale AI, Katie Notopoulos, Nikesh Arora, Palo Alto Networks, Instinct, DoorDash, Opera, Firefox, Cursor, Claude Code, GitHub Copilot, Stack Overflow ## 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 Harvey’s Margin Crisis + The Economics of Legal AI 08:02 GPT-6 Soul, Opus 5.5 + Grok 4.7 Enter the Model Wars 11:20 “Insects Matter More Than People” + EA’s Weirdest Debate 19:01 The Most Concrete AI Doom Scenario Yet 23:13 Meta Muse Secretly Uses Humans + The Agent Commerce Problem 30:12 Amazon vs. AI Agents + The Battle Over Who Owns Shopping [Source](https://www.youtube.com/watch?v=-96mtSZkEL0)

Source quotes

[2:30] Gabe: 打造 Harvey 最艰难的部分，在于顶住选择轻松道路的巨大压力，始终坚持做对客户最好的选择。轻松的做法本应是在客户还没准备好之前，就强推按用量计费，或者为了保住利润率而为他们降级使用更差的模型。我们选择按照符合客户节奏的时间表协助他们过渡，在此期间即便牺牲自己的毛利，也要提供最好的模型。我们通过模型路由、测试平台优化和后训练，以可承受的价格提供前沿智能。 [10:00] Jordi Hays: Everyone should have a John in their life [11:20] Ellie: 空中的恐怖分子 [15:00] 批评者: if you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs [18:00] 全城最高级别紧急避险警报: 检测到未知生物袭击事件，请立即就地寻找掩体避难！ [20:00] Tyler: 我不会了，我脑子里什么都不剩，根本记不得怎么从零写代码。 [20:00] John Coogan: 我个人觉得现实中发生这种剧情的概率大概是 10% [23:13] Meta Muse 人工客服设想: 抱歉，我这笔订单买错了，能不能直接帮我替换成我真正想要的那款商品？ [25:00] 对话式 AI 用户: 我就要一卷厨房纸，直接寄过来 [30:00] John Coogan / Jordi Hays: 张教授彻底被平反了 [30:00] 张教授: 我敢保证，你们使用的那些聊天软件背后全都是真人在操纵信息。 [31:35] John Coogan / Jordi Hays: 最终真正重要的，永远是你究竟用这些模型做出了什么

Source quotes

[11:20] Insects Matter More Than People Matter [11:20] Ellie: terrorists of the air [11:20] Jordi: weird freaks [15:00] John: Stand up for the insects now, lest you be discarded in the robotic future [15:00] Doomer: If you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs. Do it right now. If you don't do this, you are a fraud, a liar, a poser. [15:00] John: maximum brainrot [19:01] John: The Old Way [30:00] Professor Zhang is vindicated. [30:00] Nikesh Arora: knife fight between Amazon and Muse [30:00] the trains run on time [31:18] pacing the frontier

Original transcript

[0:00] - Jordy, his hat game has evolved. He's now wearing the fedora with the Blackstone cap on top, which I assume can only mean one thing. It means that your top priority is safety in-- - Financial markets. - Financial markets and the financialization-- - But also my safety as well. - Oh, you care about both. - I care about both. - So you want to accelerate the AI build out safely. - Safely. - Yes, I thought it was, you're gonna blend the two, and you're like, look, I want insane private credit deals, trillion dollars of capital to flow into the AI buildup, but I want to do it safely. I want to prioritize safety on the whole frontier. Safety around the models. The capital markets frontier? Safety in financial markets. In the capital markets. Got it. Okay, okay. I'm seeing where you get there. Anyway, you want to do Harvey first? People are taking shots at Harvey. I won't stand for it. I won't stand for it. No, I like Harvey. We've talked to some lawyers that use Harvey. Seems like a great product. We also had Logo on the show just last week. Had a good time with them. Seems like a very competitive space. But the headline that grabbed attention, they say track changes. Harvey's gross margin fell from about 50% to negative 50% by June as agent token use spiked 20 fold on rented open AI and anthropic models, Bloomberg reports. And so It seems really bad. Like you definitely don't want negative gross margins as a software company, as a tech company. As any company. Really as any company, actually. You don't want to be selling… $1.50. Exactly what they did in June, basically. That's what they were doing. They were selling tokens at negative gross margins. Not good, but there's way more details in the actual reporting. So you should go read the Bloomberg report. So the first thing is, like, cost did explode, but it was on the back of more product usage, which is good. They're growing ARR. We'll get into that. And token consumption increased 20x this year. So basically, they are feeling the reasoning era, the agentic era, one job, one process in Harvey is generating a lot more tokens. Had seat-based pricing. And then AI got a lot better, so every seat was just using it way more. Very simple. But it's not just using it more. It's that the lawyer might be saying the exact same thing as they did a year ago. They might say, review this. But instead of one-shotting it with no reasoning, they went to reasoning, which generated way more tokens. And then they went to agents, so they were like, okay, it's actually going to fan out, look at everything across your firm, all the other records that you have, all your different policies and MD files. work through this agent and so you're generating way more time yet so that created negative gross margins but the good news is that they're already back to positive gross margins which is in the Bloomberg article kind of criminal not to repurpose that because it really makes it look like they are running at negative gross margins they had a negative 50% gross margin in June but they are back to positive gross margins. They changed their model usage, of course. There's a whole bunch of models that are cheaper and better for certain things. And they actually fine-tuned an open weight model. So they did post-training on an open weight model, and they're calling it Harvey Tennant. And there's some interesting details there. I want to debate with you on, like, the value of that, where that goes. First, Gabe. One of the founders of Harvey responded. Oh, he did. He said the hardest thing about building Harvey is doing what's best for our customers despite immense pressure to do what's easy. The easy thing would have been to force our customers into consumption pricing before they were ready and serve them worse models to protect our margins. We chose to help our customers transition on a timeline that works for them and give them the best models in the meantime, even though it hurt our margins. This meant optimizing our product through routing, harness improvements, and post-training so we could serve frontier intelligence at an affordable price. Yeah. We believe the most important part of building a company is earning and keeping your customers trust. You do that by doing the hard thing for them, even when it costs you. A bunch of great points in here, even outside of the Bloomberg piece and the subsequent coverage. Max from Legora was going pretty hard saying it doesn't make sense. Oh, it doesn't make sense to to train your own models because the public basically like publicly available and Frontier models are dropping in cost. Sure. And the quality is increasing, so you're just sort of like wasting your own. And again, that's a bet that he's making. Harvey is going to make another bet. Probably some mixture of both is going to be the right approach. There's another interesting – sort of like how how bad is this situation? [5:00] just comp, which is, so they had negative 50% gross margins in June. They're at a $400 million run rate for ARR, I suppose. And so that's $33 million a month. And so negative 50% margins means they lost $16 million, but they've raised like $500 million. So in terms of like burn, it's not super cause for crisis, in my opinion. And what's interesting is that even through June, And maybe they did the Sequoia deal that valued the company at 15.5 before that happened because it was announced September 9th. But I have to imagine that the deal probably got done with at least some insight into those negative gross margins in June. What do you think? I would assume so. But the main thing is like every application layer company has like gone through a moment like this, even if it was for like a single day. Totally. So it matters how quickly you can respond and adjust. Yeah. So where do they need to get to? Like what is the standard for the legal technology industry? There's some comps here. E-discovery provider, disco, public company, 75% gap gross margins. Like it is a software tool that when you going into discovery you need to connect with a bunch of data Okay we got a whole bunch of text messages a whole bunch of emails Let put them in a system so we can review them Sort of, you know, normal SaaS. I'm sure there's a bunch of AI involved now, but that business is running at 75% gross margins. Thomson Reuters has a legal professionals segment of software. That's running at almost 50% adjusted EBITDA margins. So, like, true net profit. So that's sort of where they want to be. Also, pure play law firms have really high margins as well. Hard to comp them to businesses because they pay out to the partnership. But PricewaterhouseCoopers reports net profit margins around 41% for the top 10 firms. It's lower for firms 10 to 100 or something like that. But still, lots of health in the legal industry broadly. Obviously, everyone knows lawyers make money. That shouldn't be surprising. And, yeah, tokens are getting cheaper by the day. There's a bunch of model launches. Opus 5.5 launched today. Grok 4.7 was yesterday. And Grok 4.7, no one was, even Elon was like, okay, we're not jumping straight to the frontier with this. It's going to be a couple more iterations, but we're excited with the progress. But there was one benchmark that really jumped out and it was actually the Harvey Legal Agent benchmark which is interesting So I don know how insanely good that benchmark is It might be you know you could you could bench max against it. But the interesting thing about that particular benchmark was that it was a legal task benchmark based on cost. And so Grok 4.7 is particularly cheap for the type of legal work that at the very least Harvey wants to do. So you could see them funneling more token spend towards XAI and Grok. Do you have something? You wanted to give it up for Big Law? The chat wants to give it up for Big Law. They're the real winners here. Additionally, GPT-6, Sol, and Luna are starting to roll out. So we'll look out for announcements there soon. There were some cool demos. I saw one Opus 5.5. The chat says the hat-jacket combination isn't working. I'll try to adjust a little bit. I'll try to adjust. Hopefully that's better. Flawless, yeah, thank you. Thank you for flagging. Huge upgrade. Very funny. Yeah, I saw a cool demo where someone sketched out on a piece of paper, sort of a trebuchet, and Opus 5 took that image into simulation and created like a virtual ai is is is continues to impress in terms of translation of things oh your your kid drew a you know uh some crazy mythical creature turn it into a story turn it into a movie turn it into a video game whatever format you want if you want to if you want to read a 2 000 page novel based on call of duty modern warfare 4 like you You can probably do that now. So you can have your content in any specific format you want. It seems to be where the AI continues to impress. Anyway, let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents, whether you're writing code, analyzing data, creating content, or automating business workflows. Codex helps you move projects forward from start to finish. I've been having a lot of fun generating AI videos with Codex. Should we pull up your latest masterpiece? Yeah. The Literally Me compilation, it burns up the internet and vibe reels all of the time. But now, if you see those Literally Me characters and you're like, I am Jake Gyllenhaal in Nightcrawler. I am, I forget who else is there, Patrick Bateman in American Psycho. You can put yourself in the movie. There's Inception. [10:00] Got Jordy in there. That's Drive. I think that's Heat, Mad Men. Can we call him out? I think that's Fight Club. You got Joker. You got Taxi Driver. You got Nightcrawler. No Country for Old Men. Social Network. Godfather. Back to Joker. I forget that one. Sockshank, maybe. That one's Collateral. Then Blade Runner. Blade Runner again, I think. Yeah. Some of the renders are really, really good. That one, like, your face doesn't quite fit on Leo's head. But that one's really good. That one works pretty well. This one's a little awkward. The Joker, it's really just Joaquin Phoenix's face didn't actually swap you at all. Codex was even like, it didn't work. And I was like, do it anyway. I don't care. It's fine. It's going to be on screen for two seconds. That one's good. John is really my human agent. No, you didn't even tell me you did it. He's sort of like an ambient, always-on agent. That's what I'm saying. I don't even have to tell you. You just predict my needs, right? You'll send me something like this. Everyone should have a John in their life. Everyone should. But yeah, fun, fun, powerful tool. Still got $500 locked away in Higgs field in the wrong section. Can't do anything with it. Maybe just give away the account or something. Anyway, what matters more? People... or insects. Tough call, John. We should get into maybe one side of it first. Steel man it? Before we form an opinion. Okay, the human side? You want to steel man the human side? And I'll just say that we're coming to you live. We had a rough morning with the mosquitoes. We actually had to door dash one of those mosquitoes. bug bite treatment thing, the heat, whatever. What do you call them, Tyler? I think it's like insect bite healer. That's what it's listed as. It did work because I, yes, Sebastian says I'm famously pro mosquito. But a mosquito did bite me on the face today and a bunch more times. So I ended up capitulating and having to get one of these. But I don't like using these. I don't like using these. It's funny because I can't stand EAs and I think they're weird freaks. and I don't think they should be anywhere near public policy or anything like that. But the one thing that I can agree on with Ben is, this guy writing this post that, that insects are more important than humans is I don't think anyone should, I don't, I think we should harm the insects. Oh, okay. I think humans are more important, but that doesn't mean you've got to take them all out. Yeah. Yeah. just go somewhere else. You would take sort of-- you would be altruistic towards the insects. Ellie says, mosquitoes are terrorists of the air. They really are. You would-- you want to be altruistic towards the mosquitoes. You want to be effective in your altruism when it comes to mosquitoes. You don't want to arrest them. No, I want them effectively to be... Nowhere near me. An effective altruist who writes under the name Bentham's Bulldog went viral over the weekend for an essay titled Insects Matter More Than People Matter. in the aggregate, arguing that the combined welfare of insects outweighs that of humanity. His basic argument is one of scale. Insects are so numerous that even if their capacity for suffering is only a tiny fraction of ours, their total suffering could still dwarf human suffering. It's almost just a thought experiment. There's so many other directions you could make. Why not go pound for pound? Maybe cows are the most important because there's a lot of cows and they weigh a lot. There's probably more aggregate pounds of cow than humans. Maybe they matter more than humans in the aggregate. right? Or fur, like the amount of fur on all the monkeys in the world is more than all the fur on all the humans in the world and so maybe monkeys matter more i don't know it's a it's completely arbitrary where people draw these lines but uh he estimates that forever for every second of human life insects collectively spend roughly 270 000 seconds dying about 75 hours and notes that more insects die in a single second than the total number of humans who have ever lived wow that's a crazy stat i mean the guy pulled some cool numbers together this is interesting this is A large part of the essay turns on whether insects can experience pain. Bentham's bulldog argues that they can, even weakly. If they can, even weakly, their suffering should count morally, just as ours does. What species you are is just totally irrelevant to whether it's bad for you to feel terrible pain. Using what he describes as a conservative assumption that insects experience pain at just one ten-thousandth the intensity that humans do, of insects would still produce vastly more suffering in the aggregate. He ultimately argues that insects may experience more suffering in a single day than humans [15:00] have throughout all history. Now, is there a piece on this where this is the view of the of this community with the relative, like are they looking forward to how a world might interact with like a God like machine, super intelligence and humans. Yeah. So the idea is like, we got to stand up for the insects now because we are the insects of the future. So we want them to make the same calculus. I think that's certainly like where the thought experiment goes. Yeah, right. And so and so this feels almost like bait or like a trap where it's like if you are the type of person that jumps on this and says, no, I disagree. quadrillion super intelligences that can all feel pain in some abstract way, and so they're more valuable and so now you're the insect, and so your previous argument can be used against you to advocate for your own eradication. And so it's like, stand up for the insects now, lest you be discarded in the robotic future. Yeah. I don't know. The reactions to this were rough. People did not like this argument at all. Well, it's just the timing is funny because there's so much tension building around the sort of EA rationalists. movements with which is super fractured. People don't realize how fragile it is. But it's sort of boiling up to the sort of national level in a way that it hasn't before. Yeah. And so during the midst of that broader conversation to come in and being like, yeah, insects actually matter more than people. Yeah. in the aggregate What were some of the reactions? Doomer said, if you truly believe that a bug's life is worth more than a human life, you will immediately feed yourself to bugs. Do it right now. If you don't do this, you are a fraud, a liar, a poser. go and comment on this article, watch A Bug's Life at half speed. Half speed. Yeah, watch the whole movie at half speed. You're a movie guy, but maybe you're not a fan of watching movies at half speed. It's not really how the half speed is. But half speed is a great sort of tool to have in the tool chest when you want to deeply comprehend. 2x speed, 3D glasses with the two different movies interlaced so I can get one movie in one eye. another movie in the other eye. one audiobook here. One podcast here. Duo. - While I'm sleeping, yeah. - With the duo. - Exactly, Max Brainrot. I mean, the fracturing of EA really like leads to a really, really wide range of outcomes that get sort of lumped together. Everything from the Pace the Frontier essay to this, these are clearly very, very different pieces of communication and yet they get sort of lumped together. It must be full employment if you're in the public relations department over there. Yeah, I was gonna say I think the shrimp welfare stuff makes more sense to this because like I'm not really sure what like after reading this and say I agree like what do I what are the next steps like. We're not like really factory farming insects, maybe through pesticides or something. Speak for yourself, brother. Yeah, yeah, the shrimp welfare. Got some big announcements coming soon. Like what percent of these like 600 billion insect deaths per second are preventable? Yeah. We're like with the shrimp, you know, you can say like, okay, change the farming practice or just don't eat the shrimp. Yeah. So I think that's why maybe this was less. impactful. I don't know. Yeah. I mean, it's all it's all a continuum to like until you get to like the veganism, like the animals who you can pet, you know, should should be able to live full lives. This is sort of the most abstracted version of that. But yeah, I don't know. It's odd. More importantly, update. on TBPN's Road to Christmas. Oh, yeah. We're down to just 93 days. 93 days. A lot to think about, a lot of moves to make, and we couldn't be more excited. Start counting it down. I was trying to think of a realistic AI doom scenario. Like everyone's like, oh, I can't imagine it. And so I actually mapped out one that maybe I could take you guys through. I wrote a little script here called The Old Way. So, you know, fades in, starts in the TBP and Ultradome. Me, you, Tyler, we're all doing the show. All of a sudden, all the phones in the Ultradome light up. Emergency alert, unknown biological event, shelter immediately. Outside, alarms begin to sound across Los Angeles. A strange haze starts pouring through vents and drifting across the streets outside. I see a cloud approaching the ultradome entrance. First thing I do, bioweapon attack, this is what you got to do. Boom, right there. It's not going in your mouth, you're good. Second step, pull out your gun. Boom, boom, boom, boom, start blasting, whoever's responsible, you're just firing wildly. Once you got the situation a little stabilized, [20:00] Then we got to start, we got to figure out what's going on. So I go over to Tyler and I say, look, we got to hack into the super intelligence. We got to figure out what is going on with this bioweapon. And he's like, sure. Yeah, let's hack into it. Like, I'll open up Codex. What do you want me to prompt? I'm like, no, the AI has gone rogue. We have to do this the old way. He's like, okay, I'll open Cloud Code. Like, what do you want me to prompt? I'm like, no, the old way. He's like, okay, I got cursor open. Like, what are we doing here? I'm like, no, Tyler, we have to go back to the old way. He's like, okay, I got GitHub Copilot fired up. I'm like, no, Tyler, you can't use any AI at all. He's like, okay, just let me know what code you want to write, and I'll look it up on Stack Overflow. And I'm like, no, the Internet has been contaminated, Tyler. We can't use anything. We have to code the old way. We have to do it by hand. And Tyler's like, I can't. We don't know how. I don't remember anything. I can't write any code. So what do we do next? We got to get George Hatz. He's the only one who can help in this scenario. So we go. Because he's the only one that remembers how to code. He remembers how to code. He's got local models stored. He's got his own hardware. So we go get George Hatz and then we're like, okay, George, hack into the super intelligence. Tell us what going on with this bioweapon And why aren we affected Like we been walking around We seem to be fine Everyone else is turning into a zombie What going on So he locks in He writes some code. He hacks in, figures it out. He says, Jordy, tell me about your diet. And you're like, well, I just eat at Erwan exclusively. And he's like, yeah, my analysis shows that your body has never had a toxin in it, and so you are immune to the bioweapon. It doesn't affect you at all. And I'm like, but how do you explain me? I only eat Erewhon once a day. And he's like, well, John, my analysis indicates that you might have had the original Four Loko. Did you drink the original Four Loko? In large quantities? And I was like, yeah, I was daily driving in college. Why? It's like your body consumes so much original Four Loko, it's now essentially pure toxins. So the bioweapon sees it as a hostile environment. It doesn't even try to infect you. So you are immune. So we're both immune. But the bioweapon is still wreaking havoc on America, so we've got to get to the bottom of it. So we're going to need some muscle. So we go get Sam Sulek. We go get Sam Sulek, and we're like, let's go. We've got to go to the data center. We've got to break in and shut this thing down. That's the only option. So we get in the Cadillac Blackwing. We jump it off a ramp slam into the data center pull out the gun start blasting the GPUs We just firing everywhere Finally finally after we shot 90 of the data center it just bullets everywhere We getting everywhere Finally, a humanoid robot emerges. It's Mecha Hitler. We got to fight him. We got to fight him. We're shooting. We're shooting. The bullets are just bouncing right off. They're just bouncing right off. We're getting beat up. We're trying to fist fight it. Sam Sulik's doing nothing. It's not working. And then at the last second, John Cena breaks through the roof, slams down the humanoid robot, rips his head off, and humanity wins. So that's kind of like a concrete example of how AI doom, how a runaway road. Just one scenario. That's one possible scenario. Because everyone's asking for. I put it at 10%. That's the way it plays out. Yeah. That's kind of where I'm at right now. Yeah. That's how I'm thinking about it. Meta, the AI agent company, Muse, they're putting humans in there. Meta is testing human concierge for its new personal agent, Muse. Meta has been testing human concierge for its new personal AI assistant. This is on Bloomberg, which entails having human contractors quietly handle some of the phone calls placed by the digital agent. Amazon like no no no we going to block your browser Well did you block everyone browser How that going to play out I was thinking about at some point if I write a really really bespoke like basically a CLI like a really bespoke, weird CLI, like I vibe code it, but I use sort of odd techniques. Like I'm like, okay, use computer use and open up Opera or, you know, Mozilla and click around and add some randomness and do this. And it's like, and I basically create a CLI that wraps an agent where it's like, it runs on my home wifi. So, and I just do this. I don't productize it. So Amazon, like can't really detect it. Do you think I could potentially be in front of their battle forever and then puppeteer it from other agents? Well, I just asked Chachuputti, can you order Amazon over the phone? Because I would think that at that scale, there would be at least one way to do it. Totally. It seems like you can only use their customer service number to help with existing orders. So one thing is you could order a bunch of the wrong thing, like a hundred of the wrong thing, a hundred like bug treatment. Order everything and then have your agent refund everything but the thing you want. No, have your Muse agent, have a human via Muse call Amazon and say, hey, I actually got the wrong thing here. Can you swap it with this other thing? Interesting. And just do that as many times as necessary. [25:00] - Oh, yeah. - To get the right order. - Maybe have one order that's just always open, where you're just adding and removing those from it. - They have to be completed orders, 'cause you have to have made a mistake first, to call the customer service. - Oh yeah, you can charge me, but just add it to the order. I don't know, it's gonna be a weird battle. Ben Thompson was writing about it today, it was pretty, AND I THINK IT'S A GOOD THING. Pretty good. Amazon, I mean, his whole framing is like Amazon is like the strongest in the AI era of the hyperscalers because they own logistics, infrastructure, the final step, the real world. And so he anticipates that they will hold their ground for a very long time and not give in. And that's basically what they did with OpenAI, where OpenAI had the instant checkout. process. They invited everyone to be a part of it. Amazon was like, no, thanks. But we'll give you tens of billions of dollars. And but but interestingly, yeah, so now now the result is that Amazon is vending in their ads into chat GPT to close the purchase there. Toby Lookie at Shopify partnered with Muse and announce that they are integrating. But the interesting thing is like, where's the value capture layer? And it only works with ShopPet. And so, It's a lever to get merchants to go on to ShopPay, which Shopify will still make money off of. But for merchants, a lot of people are worried about the implementation that happened with Walmart, where ChatGPT integrated with Walmart. And the end result was that conversion rate was one third of what it was in the app and on the core website. And that carts were smaller because people were just going and saying, like, I just need a roll of paper towels. Just send me that. Whereas when people are like, oh, I need to do my shopping. I need soap and paper towels and whatever else. I'll build a... Paper towel holder that's made by... company that's over 50 years old. So the crazy thing is, so now Vinod was on the show yesterday talking about Wajo, which is an AI. They have a new agent. that I believe is human supported in that some of the stuff is done by AI and some of the stuff if necessary can be escalated to a human. Now you have meta, at least testing, exploring, having a human in the loop on some of these things. And my big question is like, why? Voice models are getting pretty good. It feels completely un... sustainable for meta to roll out a product to billions of people where users might just be like, like if I had like something I could text and say, hey, call this person, call that person. Yeah, I'd probably be using it. I think there's a ton of there's a ton of value. First, the Muse install base is small right now. So it's not that crazy. They have the ability to stand up. huge offices, whether it's through Scale AI and that team that came over or through their you know, all the things that they've done over the past two decades. The other thing is that MUSE specifically will not train on your data that you put into Muse. And so this wasn't like a huge announcement or anything, but you could see people being like, ah, I don't know if I want to use this AI agent that's going to be using my personal information, even if they're running like the PII redactor, putting all of that might be a little bit off-putting. So Meta's not getting that data, But if they have a human in the loop, the human who does the task, that's just beyond what the model's capable of, they're also generating training data because they can probably train on that. And so it's like the model by default can, I don't know, like summarize your calendar. and give you a briefing on what's going on in your calendar. It knows how to integrate, and the model's capable of that. MUSE 1.3 is capable of that. All the models are capable of that. But what can't they do? They might not be fully ready to go and have a complicated conversation with someone if they're ordering flowers over the phone or something. And so, like, the current models might fall down. So you need more data to solve that. And so you put the humans in the loop. and then that's your extra training data for the next leg up. And then you just keep repeating that. So I would view this human in the loop thing more as them doing data collection and creating more training data for them than a permanent solution to the product problems. Also, I don't think the time on site for these apps is actually that high. I think there is an excitement when you jump on, you do a bunch of stuff. But then with all these AI tools, there's a reason why OpenAI comps to weekly active users, because there's plenty of days where people are like, I went to the beach and I didn't touch AI at all. work. There's some feedback on Alex Wang's post talking about it'll help you with your goals. I think Katie Notopoulos was saying a lot of people just don't have goals. They just want to chill. This is also the Ben Thompson thing. Consumers want to be entertained. People that go to the beach, they'll still scroll Instagram. Are they really going to be doing [30:00] with any agent, no matter how powerful it is. There are plenty of people that have staffs of EAs and personal assistants, and they don't do anything on a weekend because they want to chill, and they don't want to think about anything. Well, One thing's for certain. It's over or we're back? Professor Zhang is vindicated. Yesterday, we watched a video where he said... I guarantee. that they're are humans behind the chat apps that you use. manipulating the information. And he is correct, at least in the short term. Nikesh Arora chimed in on the knife fight between Amazon and Muse. This will be a bigger battle than anyone anticipates. It's only a matter of time before there is an Apple and Google version of Muse and possibly TikTok, in addition to the Frontier LLM agents, maybe a commerce agent from Amazon. That's sort of my prediction. I think instinct might land with Amazon potentially, although they do have Rufus. But And Tyler is a Rufus power user. Yep, Rufus is good. You don't want to talk down on Rufus. But Instinct is clearly, like, tapped into something really special with, like, the energy and the support that they have and the community that they've built. But there are a lot of, like, little rough edges that need to get sanded off. And, like, that's the domain of Amazon, in my opinion. Like, Amazon, like, you know, the trains run on time. And it's a pretty efficient company. There's an interesting stat from the Shurtecary post about Amazon. The net income for Amazon Amazon e-commerce was like 36 billion and the ad revenue was exactly twice that. So the business is unprofitable without advertising. They put advertising everywhere. So it's like a very interesting dynamic where you can see the motivation. There's $68 billion. You flag that number. It's a huge... Over 70 in the last... 12 months. Over $70 billion. So a significant pool of revenue and profit that they will be protecting for sure. Any other breaking news? Anything you've been tracking over the last couple hours while we've been live? I think we're pretty much good. Opus 5.5 launched. We talked about that. Benchmarks look really good, very cheap. And the key thing that you pointed out was like the second post is like this is the first model that we've released since we said we were pacing the frontier. This is an expression of the pacing. model, but it's much cheaper. So it's not trying to be more superhuman, more dangerous. It's the safest, best, fastest, cheapest, lightest, thinnest model ever, basically. Yeah, and then OpenAI also did the same thing today. Okay, yeah. Oh, so Sol's out? Sol, yep. GPT-6, Sol. Cool. And then also Luna, but not Terra? Or I might have had that back. Okay. But new models, the model wars never cease to entertain. Of course, all that matters is what you do with them. So thank you for watching TVPN. We will see you tomorrow at 11 a.m. Pacific. Leave us five stars on Apple Podcasts and Spotify. Sign up for our newsletter. at tbpn.com and throw that flashbang, Tyler Cosgrove.