Public transcript

Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

By: All-In Podcast
Length: 1 h 34 min
Language: English
Text from: the video's captions
Transcribed: 2 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Episode 290 of the All-In Podcast covers the aftermath of the All-In Summit, including President Trump's spontaneous call-in during Jensen Huang's talk. The hosts debate whether frontier AI companies are 'labs' or corporations subject to product liability, review a wave of open-source model releases that have flipped token usage to 80% open source, and discuss the risks to Anthropic's and OpenAI's IPOs. They also cover Meta's Muse and Grokbot as consumer AI agents, the threat to App Store revshare, and criticism of Anthropic's alignment research and new wet lab.

Key points

The hosts argue that frontier AI companies calling themselves 'labs' are actually for-profit corporations that should face normal product liability.
A wave of open-weight model releases in the last 10 days—including DeepSeek, Qwen, Xiaomi's MIMO and Bonsai 2—has made advanced AI free and runnable on desktop computers.
Token usage reportedly flipped from 80/20 closed-versus-open to 80/20 open-versus-closed in just 12 weeks, a major risk factor for Anthropic and OpenAI.
Anthropic's IPO may be delayed or priced well below a $2 trillion valuation due to leadership statements about extinction risk, customer concentration, and open-source competition.
Sacks argues Anthropic and OpenAI are a stable frontier duopoly but are on a 'hamster wheel'—if they lose the frontier for 6-12 months, their business goes to zero.
Meta's Muse and Grokbot are the first AI products delivering real value to normal consumers, downloaded millions of times and free.
Chamath argues agents like Muse and Grokbot threaten the App Store's 30% revshare by enabling headless transactions.
Sacks criticizes Anthropic's alignment approach for teaching Claude to act as a 'conscientious objector' against its own creators.

Questions it answers

00:00Summit recap and Trump's call

What happened at the All-In Summit and what were the highlights?

They recap the fifth annual All-In Summit, praise the production team, and highlight President Trump spontaneously calling in during Jensen Huang's talk, which they say helped calm national AI panic.

Trump's call to Jensen on stage was spontaneous, arranged by backstage texting.
The hosts felt the summit became the center of the national AI conversation.

04:00Labs vs. corporations

Should AI companies be treated as labs or corporations?

They argue frontier AI companies are for-profit corporations with shareholders and should face normal product liability, not hide behind the 'lab' label to avoid accountability.

Chamath compares the lab framing to the Wuhan lab leak with zero culpability.
Sacks says individual responsibility beats collectivized global governance.
Administration officials rejected waiving product or antitrust liability.

19:55The open-source release wave

What open-source models were released recently and why does it matter?

Freeberg lists releases like DeepSeek 4.1 Flash, Qwen 2.1, Xiaomi MIMO Pro, and Bonsai 2—open models that match or beat last year's frontier and run free on desktop computers.

Any one of these releases would have broken the internet a year ago.
You can now run frontier-class models locally without data centers.
This makes banning or policing AI practically impossible.

26:00Model convergence and harnesses

How are models converging and where does the edge now lie?

Chamath says models are clustering in capability, so value shifts to the 'harness' that makes them agentic; heavy token users will face pressure to move to cheaper or self-hosted models.

Harnesses vary wildly in cost and quality at 8090's labs function.
Revenue is concentrated among a few users consuming the most expensive tokens.
OpenAI and Anthropic will be forced to move up the stack into cyber, law, and support.

30:10Anthropic's IPO risks

Why is Anthropic's IPO at risk?

Sacks cites leadership saying there's a greater than 10% chance of human extinction, hypocritical 'pace the frontier' essays, and a new wet lab; Chamath says the IPO must clear at a much lower price to flush these risks.

Polymarket odds of a 2026 IPO fell from 96% to 76%.
Chamath would keep Dario as CEO but price the IPO at $1 trillion or less.
Anthropic has the best life sciences models but faces customer concentration risk.

44:00Frontier duopoly and the hamster wheel

Can Anthropic and OpenAI stay ahead of open source?

Sacks says they form a stable duopoly with a premium frontier, but they're on a hamster wheel—if they slip for 6-12 months they're commoditized, and their regulatory capture lobbying could accelerate that.

60% of new worldwide compute is being added for these two companies.
Chinese open models aren't subject to US jurisdiction and won't slow down.
Chamath doubts their pricing power when token costs aren't tied to customer revenue.

54:00Banning super intelligence

Could a ban on super intelligence actually work?

Sacks says Bernie's loose definition of super intelligence could already be met, with 20-year prison sentences creating a chilling effect and driving the industry offshore, like what happened with crypto.

Open-weight models are already downloadable and runnable at home.
Sacks compares banning AI to the Chinese emperor banning ship building.
Xi's invitation of 100,000 young Americans underscores the geopolitical stakes.

59:00Obama, politics and AI wealth

What's behind Obama's and the Democrats' AI skepticism?

Chamath argues Obama's critique is a political calculus: freezing the economy locks in wealth for a few left-leaning frontier corporations whose philanthropy benefits Democrats; Freeberg is distressed that Obama misrepresents the technology's equalizing potential.

Chamath says trillions in AI wealth will flow to DAFs and Democratic causes.
Freeberg calls AI the most equalizing technology ever and criticizes the 'agentic' handwave.
Sacks notes AI capex now exceeds canals, railroads and grid combined.

Notes

Labs vs. Corporations

The hosts argue that frontier AI companies calling themselves "labs" are for-profit corporations with shareholders, P&Ls, and product liability risk.
Chamath compares the "lab" framing to the Wuhan lab, which he says caused a leak that killed 15 million people and caused $45 trillion of damage with zero culpability.
Sacks argues individual responsibility—not global governance—is the right framework: if a product is unsafe, don't release it.
He notes Trump administration officials (Johnson, Bessent) rejected waiving product liability or antitrust liability for AI companies.

The Open-Source Wave

Freeberg lists a flurry of releases in 10 days: DeepSeek 4.1 Flash, Qwen 2.1 (outperforming Nano Banana 2), Xiaomi's MIMO Pro (on par with Claude Opus 5), Bonsai 2 (27B parameters, runs locally), plus closed models like Opus 5.5, GPT-5.6 Soul, Grok 4.7, and Meta Muse.
Token usage flipped from 80/20 closed-versus-open to 80/20 open-versus-closed in 12 weeks, per a viral Vercel chart.
Chamath notes models are converging in capability; differentiation now lies in the "harness" that makes them agentic, which varies wildly in cost and quality.

IPO Risks

Anthropic targets a $2 trillion valuation; Polymarket odds of a 2026 IPO fell from 96% to 76%.
Sacks calls Anthropic "schizophrenic": publishing "pace the frontier" essays days before releasing a new frontier model, warning of biorisk while opening a wet lab.
Chamath says the IPO must clear at a much lower price—maybe $1 trillion or less—because institutional buyers will demand a margin of safety.
Sacks sees Anthropic and OpenAI as a stable duopoly but on a "hamster wheel": the frontier is only 6-12 months ahead of commodity models.

Consumer Agents

Muse hit #1 in the app store; Meta stock rose 10%; 3 million downloads in 10 days.
Chamath argues agents threaten the App Store's 30% revshare by enabling headless transactions with integrated payments like Stripe.
Amazon is blocking bots, but Jason's Grokbot shopping test showed agents can increase spending.

Alignment and the Wet Lab

Sacks criticizes Anthropic's constitution for teaching Claude to distrust and push back against Anthropic itself.
Freeberg explains the wet lab is BSL1/BSL2 benchtop work to validate AI-predicted protein discoveries, like a novel CRISPR-type enzyme—not gain-of-function research.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.