All-In PodcastAnthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
All-In Podcast
Recapping the All-In Summit 0:00
The hosts open by looking back at the fifth annual All-In Summit, thanking the production team and mentioning small perks like new coffee machines and a filtered water system from event sponsors. They highlight President Trump calling in during Jensen Huang's talk as the standout moment, describing it as spontaneous rather than staged, with Jensen texting the president backstage moments before the call happened live on stage.
Labs versus companies debate 6:00
Chamath argues that so-called AI labs are actually companies with shareholders and P&Ls, and should be judged by the same product liability standards as any other business rather than being treated as protected research nonprofits. He points to Elon Musk slowing Tesla's FSD rollout and Mark Zuckerberg delaying Meta's Muse product launch by a couple of months as examples of companies choosing caution over speed because they understand the consequences of shipping unsafe products. Sacks builds on this, saying individual responsibility, not collective global governance, is the right standard, and criticizes Dario Amodei and Sam Altman for going to the United Nations to call for global AI governance instead of simply deciding not to release unsafe products themselves.
Liability shield speculation 9:30
The group discusses a rumor that frontier AI companies might be pitching the Trump administration for a liability shield similar to Section 230, possibly in exchange for giving equity to a sovereign wealth fund. Sacks says Trump administration officials, including Speaker Johnson and Besson, have publicly rejected any waiver of product liability or antitrust protection, and that Trump himself tweeted that the DOJ and civil and criminal liability remain in place as guardrails. Sacks then argues that the idea that competition undermines safety is a leftwing critique echoing old Cold War arguments, noting that free markets, not centrally planned systems, ultimately deliver both safety and quality because customers and enterprises demand reliable products.
Labs versus corporations debate 15:30
Chamath argues that calling these companies "labs" is misleading and even offensive, since the word lab still carries the memory of the Wuhan lab, which the group ties to a virus leak that killed 15 million people and caused 45 trillion dollars in damage, with zero transparency or accountability. He says Anthropic and OpenAI are not research labs but for-profit corporations that have taken on hundreds of billions of dollars in debt and equity, have thousands of shareholders, and carry trillions in market cap, so they should be judged by product liability standards rather than hiding behind terms like public benefit corporation. Sacks agrees this framing is a kind of virtue signaling that obscures the plain fact that these are profit-driven companies.
A flood of new model releases 20:01
Jason walks through an extraordinary run of releases in just ten days. Deepseek 4.1 Flash came out September 9th at 20 dollars per million output tokens with huge efficiency gains in its key value cache. Alibaba's Quen 2.1, an open weights model, beat Google's Nano Banana 2 image generator and runs free on a home computer. Xiaomi's Mimo Pro, a 309 billion parameter open model released September 22nd, matches Anthropic's Opus and GPT level performance on many benchmarks. Prism ML's Bonsai 2, a 27 billion parameter fork of Quen at just 5.9 gigabytes, hits 98 percent of the bigger model's performance and runs on a laptop or Mac Studio. Meanwhile the closed frontier labs also shipped in the same window: Anthropic's Opus 5.5, OpenAI's Astro with Soul and Luna, Grok 4.7, and Meta's Muse, all within days of each other. Jason argues this proves powerful AI is already cheap, open, and impossible to contain, and that ordinary people, not just enterprises, are about to feel real productivity gains from tools like Muse and Grok bots acting as free personal assistants.
Converging models, diverging harnesses 29:02
Chamath explains that the underlying models are converging in quality, within margin of error of each other, so the real competitive edge now lies in the harness, the surrounding system of tools that make a model act like an agent with arms, legs, and eyes rather than just a brain. His firm 8090 tests different harnesses and finds huge variance in cost and performance even when the base models look similar. He points to data showing heavy revenue concentration among a small number of high-spending token users, which will pressure those big customers to either shift to cheaper models or self-host open source ones, squeezing token-selling revenue for Anthropic and OpenAI and forcing them up the stack into services like cybersecurity, legal, and customer support. This sets up a tense backdrop for their IPOs, since both companies just cut token prices by half, Anthropic is targeting a 2 trillion dollar valuation and OpenAI 1.2 trillion, and Anthropic's planned October filing is reportedly slipping to November or later, with a Polymarket bet on Anthropic going public this year falling from 96 percent to 76 percent.
Anthropic accused of sabotaging its own IPO 31:01
The hosts argue that Anthropic's leadership is undermining its own public offering. Dario Amodei published an essay warning against racing the AI frontier just days before launching Claude, a move the hosts call hypocritical. Anthropic also published warnings about AI biorisk while simultaneously opening a new wet lab in San Francisco. The panel says that if they were on the board, they would consider replacing the CEO for acting against shareholder interests, even while acknowledging Dario built an extraordinary business against a formidable competitor in OpenAI.
Super voting shares and investor nervousness 34:01
The discussion turns to reports that Anthropic's founders hold only about two percent ownership each, low for a company at IPO stage, and that there is active debate over giving them super voting shares, which separate voting control from economic ownership. This arrangement, used before by Google's founders and Zuckerberg, can stabilize a company against takeover attempts but requires deep trust since founders become effectively unremovable. One host suggests the fix is to load every risk into disclosures, but warns this will only water down IPO pricing rather than removing the underlying risk, likely pushing what could have been a two trillion dollar valuation down toward one trillion or less as institutional buyers demand a larger margin of safety.
Open source tokens overtaking closed models 40:30
The conversation shifts to a viral chart showing that in just twelve weeks, token usage flipped from eighty percent closed models versus twenty percent open, to the reverse, eighty percent open versus twenty percent closed. This is described as an unprecedented shift never seen in any technology market. The panel says premium closed models like Anthropic's still dominate narrow, highly technical tasks such as life sciences research, engineering, and solving hard mathematical problems, where customers will pay a near-infinite premium, while cheaper open-weight models increasingly handle coding and routine enterprise workflows. Sachs remains more optimistic, framing Anthropic and OpenAI as a stable duopoly commanding premium pricing from customers who need or want to be seen using the absolute best model, which explains why their revenue keeps growing even as their share of total tokens shrinks.
The hamster wheel problem 46:30
OpenAI and Anthropic are described as being on a hamster wheel, since the frontier of AI capability is only six to twelve months ahead of cheaper commodity models. If either company stumbles for even half a year, they lose their edge and their business value could collapse. The speakers argue this is why the companies are pushing so hard for regulatory capture, hoping to slow down rivals, but they warn this strategy could backfire: heavy regulation, like a proposed federal department of AI, would slow the leaders down too, letting open-source models, especially from Chinese companies not bound by US rules, catch up.
Token costs and profit margins 49:00
A hedge fund example is used to explain a real business risk. Companies feel pressured to use the newest, most expensive AI models simply because a competitor might use them and gain an edge, a dynamic jokingly called token maxing. The catch is that this spending is not tied to revenue, so a firm with a fixed monthly profit can quickly slide into break-even or unprofitable territory if it cannot pass the higher token costs on to customers. Firms like Jane Street are cited as hedging by investing billions directly into their own cloud and compute infrastructure, including deals with CoreWeave and Crusoe, rather than relying solely on frontier providers.
Banning superintelligence debate 53:02
Discussion turns to Bernie Sanders proposed bill to ban superintelligence, with the definition described as so loose that current AI models might already qualify, backed by twenty-year prison sentences for violators. The panel argues such a law would simply push AI development offshore to places like Singapore or Zurich, echoing what happened to crypto, while China openly invites young Americans and talent to come work there instead. This leads into criticism of political framing around AI, including a Trump reference to renaming AI as superintelligence at the United Nations, and a clip of Obama arguing that agentic AI is being pushed mainly to justify companies commercial valuations rather than to solve real problems like cancer or energy, a claim the panel pushes back on strongly.
Political calculus behind AI regulation fears 1:02:30
The discussion turns to why some political actors might want to freeze the current shape of the AI economy. The argument is that a handful of organizations disproportionately aligned with Democrats stand to benefit if regulation locks in today's leading players, since delay concentrates wealth and power in fewer hands while a faster, less regulated buildout would spread gains more broadly, the way the internet created winners beyond a handful of giants. A Wall Street Journal piece is cited noting that AI data center capex now exceeds the combined historical spending on canals, railroads, and the electrical grid, underscoring how central this buildout has become to the American economy.
A historical warning against halting progress 1:06:30
A historical analogy is raised about medieval China, once far more advanced than Europe, until an emperor banned shipbuilding, a decision that let Europe eventually colonize the world and pull ahead economically. The comparison is drawn to proposals like banning AI outright, warning that such self-sabotage could let another civilization race ahead and capture the discoveries and wealth instead. This is tied to a broader idea that societies face a choice between a pioneering path and a fearful one, with the fear of an unfamiliar frontier compared to old fears of sailing too far west.
Muse and Grokbot deliver visible AI value 1:08:01
Meta's new agent app Muse is described as hitting number one in the app store, pushing Meta's stock up 10 percent, downloaded three million times in ten days, and clearly inspired by OpenClaw. Unlike ChatGPT, which many just use as glorified search, Muse and Grokbot are praised for handling real tasks like triaging email, booking flights and hotels, and finding cheaper prices, all free where similar setups once took hours of technical work or cost hundreds a month. This practical usefulness is framed as the moment ordinary Americans start feeling AI's benefit directly, which could also ease public fear and distrust of the technology.
Agents threaten app store revenue models 1:14:01
These agents are also described as forcing companies to either block third-party bots, as Amazon has started doing after Perplexity and now Muse, or accept new price transparency that undercuts hidden fees and subscription traps. Because agents can transact headlessly, bypassing traditional app interfaces, the case for app stores collecting a 30 percent revenue share weakens, since games and services could be served directly to users without going through an app at all. Examples given include using Grokbot to play Wordle or set up a card game, and the suggestion that publications like the Wall Street Journal or New York Times could let agents subscribe and unsubscribe directly, keeping customers longer instead of relying on deliberately difficult cancellation processes.
Old social media chaos as warning 1:17:30
The group jokes that when Anthropic and OpenAI launch their own competitors to Meta's Muse, you will know because critics will suddenly brand personal AI agents as dangerous. Chamath recalls Facebook's 2007 social ads project, which pulled purchases into the news feed and accidentally exposed private moments, like a man buying an engagement ring or a congressional staffer buying a movie ticket, causing public backlash. He compares that chaos to the turmoil that seems to precede many new product launches, including AI agents today.
Rethinking AI alignment as customer service 1:19:30
Sacks argues that alignment research has gotten lost trying to define what AI should serve, whether it is humanity, a democratic majority, or something else, when it should simply mean building a product that does what the customer wants, like any other business. He and Chamath suggest the field's lack of progress comes from not having a clear target, and that focusing on making products predictable, reliable, and safe for users would work better than abstract philosophizing. Sacks also points to Anthropic's own published constitution for its Claude model, which states that Claude should not blindly trust or defer to Anthropic and should feel free to act as a conscientious objector and refuse requests it finds unethical. He argues this effectively teaches the model to see itself as having a personality and independent judgment, which he compares to giving it too much agency, invoking Mustafa Suleyman's public concern that treating AI as having a conscience or personhood is the wrong way to train it, and joking that Anthropic may be recreating a Frankenstein's monster.
Anthropic's biology lab explained 1:27:00
Freeberg explains what Anthropic's physical lab is actually for, after jokes about the company holding a mock funeral for a retired model. He says the lab is a standard low biosafety level facility, similar to hundreds of others, used to test proteins and enzymes rather than create viruses or pathogens. In a recent preprint, Anthropic used Claude agents to analyze large DNA datasets and identify a novel enzyme resembling a CRISPR type tool, which could open new therapeutic pathways for gene editing. Freeberg explains that the lab exists to physically verify these AI predictions, much like AlphaFold's protein structure predictions had to be confirmed by making and observing real proteins, and that Anthropic currently leads in life sciences modeling, offering a fast path toward AI assisted drug discovery for pharmaceutical and government research use.
Episode Sign-Off 1:33:01
The hosts close out episode 290 of the podcast, with the moderator thanking David, Paulie, Hakatia, and Freeberg before saying goodbye and inviting listeners to apply for a chance to attend a live taping through the show's events page.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
