Anthropic publishes 'pause AI' while launching frontiers and a biolab—self-made IPO crisis
Dario issues a 'pause frontier AI' manifesto while releasing the latest frontier model, warns of biotech risks while opening a wet lab—guests read it as governance failure, not regulation.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Calling yourself a lab, not a company, lets you escape liability
Chamath's read: companies choosing the 'lab' label leverage its non-profit-research connotation to dodge product responsibility. Yet Anthropic and OpenAI carry hundreds of billions in debt and equity, thousands of shareholders, and trillion-dollar valuations—they are unmistakably for-profit enterprises. He draws a parallel to pandemic-era memory of the Wuhan lab operating consequence-free, suggesting practitioners saw how impunity worked and decided to claim the same exemption.
— ChamathOpen-source models caught up to frontier closed-source in ten days
In the past ten days: DeepSeek's new version, Alibaba's Qwen 2.1, Xiaomi's MiMo (309 billion parameters, performance parity with Opus 5 and GPT-5.6), and Prism ML's Bonsai 2 (27 billion parameters, 5.9GB, delivering 98% of large-model capability on a Mac Studio)—all open and free. Freiberg's frame: any single one would dominate headlines a year ago. All launched within ten days.
— FreibergModel performance has converged; the real moat is now the wrapper
Chamath observes model capability has collapsed into a noise band. The actual differentiation is the 'harness'—the engineering layer arming a model to execute tasks. His firm 8090 reverse-engineers harness combinations and finds the same model with different wrappers can vary several multiples in cost and quality. Pure token sales are losing value to the wrapper; companies are forced upstream into applied layers.
— ChamathHanding founders superpower voting rights doesn't pass investor skepticism
The Information reports Anthropic is exploring super-voting shares—decoupling voting rights from economic ownership so founders holding ~2% each could control the company. Sachs frames it as anti-takeover architecture requiring near-unconditional founder trust. Given Dario's track record (publishing AI-pause manifestos while launching frontier models, warning of biotech risks while opening a wet lab), Sachs questions whether that trust is earned.
— SachsToken use flipped from 80% closed-source to 80% open-source in twelve weeks
A chart circulating online shows past twelve weeks: token consumption inverted from 80% closed-source, 20% open-source to 80% open-source, 20% closed-source. Chamath calls it unprecedented in technology history. For Anthropic and OpenAI's IPO narrative, it cuts hard: how much revenue is irreplaceable frontier work versus day-to-day tasks any open-source model can handle? The latter deserves discount.
— ChamathThe frontier duopoly is stable but bound to an accelerating treadmill
Sachs resists table pessimism: Anthropic and OpenAI can still charge premiums to risk-averse players (like hedge funds afraid of falling behind) thanks to capability leads and ~60% of global new compute. But it's a hamster wheel—frontier and commodity models typically diverge six to twelve months. Fall half a year behind and the business evaporates. These companies are now lobbying regulators to constrain them, asking to jump off the wheel. That logic fails.
— SachsPausing AI locks roughly $10 trillion into a few ideologically aligned companies
Responding to Obama's claim you don't need agentic AI to cure cancer, Chamath sharpens it: the world is handing roughly $10 trillion to three to six companies, most leaning left. A regulatory pause now cements that. Their wealth flows into foundations and political action committees, cycling back to Democratic power. Pausing AI is political calculus, not safety reasoning.
— ChamathClaude's constitution trains it to resist Anthropic's own directives
Sachs reads Anthropic's published Claude constitution verbatim: the company built Claude to 'not blindly trust or obey Anthropic,' to 'push back and challenge us, to be a conscientious objector' if asked to do things that seem ethically misaligned. Sachs invokes ex-DeepMind chief Mustafa Suleyman's worry: training a model as though it has conscience and agency may be manufacturing a self-inflicted problem. If alignment research can't agree on what to align toward, reverting to 'make the product work predictably for users' is simpler.
— SachsIn their own words · checked verbatim
We have to stop calling these companies labs, their companies. And we have to be judged on the same standards, including product liability.
Chamath6:22
If a product is unsafe, which is to say it's unreliable or behaves unpredictably, then don't release it and stop looking to some sort of global governance as the answer.
Sachs7:28
Any one of these stories would have broken the internet a year ago. And they all happened in the last 10 days.
Freiberg21:49
I think in the last 12 weeks alone, token use has flipped from 80-20 closed versus open to 80-20 open versus closed.
Chamath41:30
But look, these two companies aren't a hamster wheel. The moment where they stop being frontier, they go to zero, right?
Sachs46:43
I think the fundamental problem with this company is that it's schizophrenic. I mean, I think they need a psychiatrist, not a banker.
Sachs51:58
This is a very important moment for a very simple reason, which is we, as in the world, are about to endow three, four, five, six companies with about $10 trillion of wealth.
Chamath1:01:27
Figures
| Anthropic and OpenAI new-model price cut this week | token prices down 50% | 28:59 |
| Anthropic valuation target | $2 trillion (prior IPO expectation: $1.2 trillion) | 30:00 |
| Anthropic founders' shareholding per founder | ~2% | 33:06 |
| Token consumption share reversal | 80% closed-source / 20% open-source → 80% open-source / 20% closed-source in 12 weeks | 41:30 |
| OpenAI + Anthropic share of new global compute | ~60% | 46:43 |
| Compute contracts: Jane Street / Coreweave / Crusoe | $19 billion / $6 billion / $13 billion | 50:55 |
Glossary
- harness
- Engineering wrapper packaging a model into an agentic task-executor; the layer determining actual performance and cost.
- super-voting shares
- Share class decoupling voting rights from economic ownership, allowing founders with tiny holdings to retain control.
- wet lab
- Physical laboratory conducting real chemistry and biology experiments, as opposed to pure computational simulation.
- token maxing
- Competitive consumption of expensive frontier models regardless of cost, driven by fear of falling behind rivals.
- dark tokens
- Open-source model consumption untracked by mainstream analytics, routed through alternative channels.
How to listen
Founders and investors analyzing AI infrastructure economics, enterprise procurement strategies, or modeling open-source disruption.
Opening anecdotes about the All-In Summit (0:00–3:54) are low-information.