Archer Whitfield: Benchmark scores are not the singularity. I want to hold that thought for a second — and then complicate it.
Archer Whitfield: Because Jack Clark doesn't think they're merely benchmarks. And Clark isn't someone you just wave off. He's a co-founder of Anthropic — the AI safety company, the one Dario Amodei and Daniela Amodei and Clark himself built in 2021 because — I mean, the stakes felt existential to them. That was the whole point.
Archer Whitfield: October 17th, 2026, IMF event in Bangkok — government ministers, finance officials, the actual policy establishment in the room — and Clark says this: 'I believe that this year we've gained convincing evidence that it has now begun.'
Archer Whitfield: The singularity.
Archer Whitfield: And the evidence he offered was — well, it was the benchmarks. Humanity's Last Exam: 8% in January 2025, 67.7% by the time he spoke. FrontierMath: 2% in November 2024, then 94%. Those are the empirical anchors for the claim.
Archer Whitfield: Which is where I feel the ground get a little soft under the argument.
Archer Whitfield: Not because the numbers aren't real — they are, and they're striking — but the word he used, singularity, has no agreed technical definition. It describes, roughly, a hypothetical point at which AI-driven progress outpaces humanity's ability to understand or control it. But that's not a threshold. There's no instrument that tells you when you've crossed it.
Archer Whitfield: Clark looked at a trend and called it an arrival.
Archer Whitfield: Maybe he's right. That's genuinely possible. But there's a thing worth naming here: the institution making this declaration is an AI safety lab. The one that — at least in principle — was built to sound the alarm carefully. Not to announce the thing it was watching for from a podium at the International Monetary Fund.
Archer Whitfield: That tension is what I want to pull on.
Archer Whitfield: Anthropic's situation is complicated in a few ways.
Archer Whitfield: This is a company that reportedly held back a model called Mythos — didn't release it — because they judged it too dangerous, despite capabilities that could've pushed their revenue into the hundreds of billions.
Archer Whitfield: That's not nothing.
Archer Whitfield: Leaving hundreds of billions on the table — or even the credible prospect of it — because you actually believe your own safety concern is real… that's either genuine institutional courage or the most expensive piece of brand-building in the history of technology.
Archer Whitfield: And I genuinely don't know which.
Archer Whitfield: Then there's the February 2026 episode — the Trump administration clashing directly with Anthropic over Claude, labeling the company a national-security risk, canceling federal contracts after Anthropic refused to loosen safeguards against mass surveillance and autonomous lethal weapons.
Archer Whitfield: Again — that cost something real.
Archer Whitfield: Federal contracts canceled. A presidential administration calling you a security threat. That is not a comfortable place to be if you're trying to scale a commercial AI company.
Archer Whitfield: There's a case that Anthropic is exactly what it says it is — a lab that means it.
Archer Whitfield: But then the IPO reporting arrives, and the frame shifts slightly. Critics point to Anthropic's public safety discourse — the landmark policy documents, the existential-risk framing, all of it — as brand differentiation ahead of going public. The responsible AI company, the one you can trust. That is a remarkably useful story to tell investors.
Archer Whitfield: And Geoffrey Hinton — who is not a careless thinker — has warned directly that once Anthropic is a public company, it will carry legal duties to maximize shareholder profit. That those duties will conflict with the safety mission.
Archer Whitfield: That's the vise, isn't it.
Archer Whitfield: A lab that has demonstrably paid real costs to hold a line on safety — and that is also moving toward public markets, also needs capital, also benefits enormously from the perception that it alone can be trusted with the most powerful technology in human history… when that lab's co-founder stands at an IMF podium and declares the singularity has begun, is that a scientific assessment or is that a NARRATIVE?
Archer Whitfield: I'm not saying Clark is wrong. I'm saying the question of who is speaking, and why now, and in front of whom — that question doesn't go away just because the benchmark numbers are real.
Archer Whitfield: Clark didn't just declare something. He told the room what to do about it.
Archer Whitfield: His call at the IMF was for companies and governments to plan for success — not plan for risk, not plan for the worst, but proactively plan for rapid AI-driven scientific acceleration while, as he put it, balancing speed and caution.
Archer Whitfield: Which sounds reasonable until you ask: plan for what, exactly?
Archer Whitfield: The term singularity has been in circulation since the 1990s — Ray Kurzweil built an entire framework around exponential technological growth, the idea that accelerating returns eventually produce a curve so steep it escapes human comprehension. That's the intellectual lineage Clark is invoking. And Kurzweil's version, whatever you think of it, was at least internally consistent — he tied it to specific projections, specific years.
Archer Whitfield: But it still never resolved into a threshold.
Archer Whitfield: Decades of singularity discourse and there is no instrument, no agreed technical definition, no line you cross and know you've crossed. Clark's wording on October 17th was careful in one sense — 'I believe we've gained convincing evidence' — but that's an interpretation of a trend, not a finding.
Archer Whitfield: And benchmark jumps, however dramatic — from 2% to 94% on FrontierMath, from 8% to 67.7% on Humanity's Last Exam — don't straightforwardly imply self-accelerating progress. Critics are right about that. The leap from 'AI is getting dramatically better at these tests' to 'progress now outpaces human understanding' is contested. It might be true. But it isn't what the numbers alone show.
Archer Whitfield: Governments are being asked to plan for something that doesn't have a definition, on the basis of evidence that doesn't fully support the conclusion, by an institution with substantial stakes in the story it's telling.
Archer Whitfield: That's the live wire.
Archer Whitfield: Because the thing to watch now — the specific thing — is the Anthropic IPO. If and when Anthropic goes public, Geoffrey Hinton's warning stops being philosophical. It becomes structural. Legal duties to maximize shareholder profit don't sit alongside a safety mission as a tension you manage culturally. They sit above it, in court.
Archer Whitfield: And once that's true, every public statement Anthropic makes about existential risk — every landmark safety document, every refused contract, every model withheld — gets read in that light.
Archer Whitfield: The question I keep returning to — and I'd sit with this if I were a policymaker in that Bangkok room — is whether a for-profit lab with enormous stakes in continued AI acceleration should be the primary voice framing that acceleration as inevitable. Not because Clark is wrong. But because 'inevitable' is doing a great deal of work when the entity declaring it is also the one best positioned to profit from everyone believing it.
Archer Whitfield: Who sets the terms of inevitability matters rather a lot. And right now, that answer is Anthropic.
Archer Whitfield: And I suppose what I'd say, in the end, is this: the declaration and the beneficiary are the same entity.
Archer Whitfield: Jack Clark stands at an IMF podium in Bangkok and tells the rooms that govern capital and policy that the singularity has begun — that progress is already beyond reversal — and the institution he co-founded, the one building the technology in question, is the institution that needs your trust, your regulatory goodwill, your capital, to keep going. That's not a coincidence you can set aside. Anthropic's stated mission is AI safety. The declaration frames acceleration as already underway. Those two things together are not neutral. They are, rather, a remarkably convenient alignment of message and interest.
Archer Whitfield: The entity most advantaged by the singularity being inevitable is the frontier AI lab that is already building it.
