The short version
The odd thing about this lecture is the room it was given in.
On Saturday 17 October, at the IMF-World Bank Annual Meetings in Bangkok, Jack Clark, a co-founder of Anthropic, told an audience of government and finance officials: “The singularity is here.” The lecture was called Inside the Event Horizon. It ran about 28 minutes, and one number from it travelled further than the rest: US GDP of $44 trillion by 2030.
A detective’s first move with any claim is to ask what would have to be true for it to hold. So this piece does that. It checks what Clark said against the recording, works out what the $44 trillion figure is and isn’t, and then puts two of his sentences side by side. Read together, they change what the number means.
The condition that makes the number big is the same condition that makes the warning real.
Explain it like I’m ten
Imagine a school where every teacher could instantly make a thousand copies of themselves. Lessons, experiments and inventions would speed up enormously. But if the copies started improving themselves faster than anyone could check their work, the school might stop being run by its teachers. Clark says the first part is starting and the second part is the thing to prepare for.
Clue 01 · The claim
What did he actually say?
Start with the recording, because the headlines compressed it. The IMF posted the whole session on its channel. Clark’s lecture is the stretch from about 7:46 to 35:38, and the player below plays only that part.
Here is the claim in his words. Researchers and science-fiction writers have long forecast AI speeding progress until humans struggle to keep up. “They call this the singularity. And I believe that this year we’ve gained convincing evidence that it has now begun.” Moments later he says it again, flatly: “The singularity is here.”
Notice what he does not claim. He doesn’t say the world has already been transformed. He says to expect that “the whole pace of progress of not just AI but science as a whole will speed up,” and that this will bring “great wealth and potentially great peril.” Then he says he’ll concentrate on the benefits, without discounting the peril. That framing matters for everything below.
Clue 02 · The evidence
What proof did he offer?
He offered three kinds. The first was a list: AI systems proposing solutions to long-standing maths problems, drugs in trials that were designed largely by AI, and robots operating laboratory equipment on their own. The second was benchmark scores. The third was one story, which he called the latest example.
8% → 67.7%
Humanity’s Last Exam, top score at launch in January 2025 versus “this week”
2% → 94%
FrontierMath, best score at launch in November 2024 versus now
49 of 50
unpublished physics problems solved after experts corrected their own answers
The story is the most interesting of the three. Physicists around the world wrote 50 unpublished problems from their own research, so no model could have read the answers. A model first scored in the late 80s per cent. Then it insisted some of the experts’ answers were wrong, and in several cases the experts went back and found subtle errors in their own calculations. After the corrections the model reached 98%, 49 of 50. Clark quoted one of the test’s authors on this. Press coverage of the same benchmark names the model GPT-6 Astra.
His reason for caring about all this is the part to hold onto. Doing real frontier science, he argues, tests for creativity, one of the last capabilities people expected AI to lack. If AI can do that, he says, we should expect it to generalise to many fields, “including certainly AI research itself, which could have an explosive effect on innovation.” Keep that sentence in mind. It comes back.
Clue 03 · The number
Where does $44 trillion come from?
Clark introduced it carefully. Anthropic’s economists had simulated versions of the economy under different paths for the technology. If you assume models can autonomously complete a broad range of tasks and that adoption spreads, he said, “one model predicts a vast economic transformation with US GDP potentially reaching 44 trillion by 2030, which sounds very outlandish, I know, but it’s implicated by the technology continuing to get better.”
The phrase to catch is one model. The figure comes from Anthropic’s Economic Scenario Explorer, version 1.0, September 2026. It has three scenarios, and $44.4 trillion is the most extreme of them.
| Scenario | What it assumes | 2030 US GDP |
|---|---|---|
| Modest | AI’s impact is roughly like the internet’s: real gains, within the historical norm. | $34.1T · +1.6% |
| Substantial | AI does about half of knowledge work by 2030, mostly autonomously. | $36.3T · +8.3% |
| Extreme | AI beats humans at the vast majority of knowledge-work tasks. The page says it “would likely require recursively self-improving AI, adopted quickly for knowledge work.” | $44.4T · +32.4% |
Two things need untangling. Reuters reported the model as a rise of “almost 50%” by 2030 against a current level of “just over $30 trillion.” Anthropic’s own page says +32.4%. They are measuring different things. Anthropic’s percentage is against the same economy in 2030 without AI. The press figure is against today.
And the label matters. Anthropic’s page presents the scenarios as the output of a model under stated assumptions, press coverage stresses they are not forecasts, and Clark himself said “one model.” A headline saying AI could lift GDP to $44 trillion is accurate only in that conditional sense.
Clue 04 · The deduction
What has to be true for that number?
Now put two sentences side by side. The first is from Anthropic’s scenario page: the extreme case “would likely require recursively self-improving AI.” The second is Clark, discussing AI-driven AI research: it “could let us make more progress on the technology but also introduces risks and we will need to move extremely carefully here because otherwise we could lose control of the technology.”
This is the deduction, and it needs stating carefully. Clark does not say the $44 trillion scenario is the scenario in which humans lose control. What the two texts show is narrower and more useful. The biggest economic prize in the model and the biggest risk in the speech both attach to the same step: AI getting good at building better AI. You can’t quote the first without being asked about the second.
“we will need to move extremely carefully here because otherwise we could lose control of the technology.”
What the extreme case does to workers
The same scenario page spells out who pays. In the extreme case, labour’s share of GDP falls to 45.2%, from about 60% today. Knowledge-worker wages “fall by more than 10% by 2030.” Average wages can still rise in the model because other workers earn more, and total labour income stays roughly flat. Aggregate GDP can climb while knowledge-work wages fall.
~60% → 45.2%
labour’s share of GDP in the extreme case
>10%
fall in knowledge-worker wages by 2030 in that case
14.8 pts
our subtraction: 60 minus 45.2, the share moving away from labour
Clue 05 · The risks
What did he say about the downside?
More than the headlines suggest. Having said he would concentrate on the benefits, Clark spent the closing stretch of the lecture on choices. He named misuse first: being able to use AI for science “opens us up to the misuses of science as well,” and he listed “biological and cyber weapons and things we cannot yet imagine.” Then he framed three choices.
CHOICE 01
Where we direct it
An AI that can do science shouldn’t be left to do it unattended. He asked for the strictest human oversight and transparency in biological research, and said labs like Anthropic must decide how much oversight to apply to AI-driven AI research.
CHOICE 02
How much we must understand
He said humans are already spending time decoding AI advances in mathematics. Within a few years, he expects AI to invent scientific equipment built for AI minds. Do we need to understand it? “Should we?”
CHOICE 03
What we dare to imagine
Taking the progress seriously, he said, means giving ourselves “license to imagine a fundamentally different world,” one in which problems that might doom us become solvable.
His prescription is the part that rarely makes the headlines. “we’re going to need to work together to create unusual and serious rules for this technology and the companies building it.” He pairs that with a quality he calls the hardest and most noble: restraint. Build the infrastructure to benefit, he says, but also go “slowly and deliberately” on developing the technology itself.
Then comes the clock. Carl Sagan, Clark noted, thought humanity had hundreds of thousands of years to grow wiser. “we instead have a handful of years, less than a decade easily.”
A speaker who says the singularity has begun and then asks for rules on the companies building it is making one argument, not two.
Clue 06 · The doubts
Who disagrees, and what should we make of it?
Start with the speaker. Clark is a co-founder of the company whose model produced the number, and whose own system he named as the top scorer on a benchmark. That doesn’t make him wrong. It does mean the claims deserve the checking that this piece has tried to give them.
Anthropic’s own scenario page is candid about its limits. It says the model, “like every model,” is “a stark simplification of a complex reality” and that it leaves out “policy responses, business cycles, potential aggregate demand or financial market disruptions,” along with “possible catastrophic risks.” Those are large omissions for a scenario that is supposed to describe a changed economy.
Outside voices are sharper. Startup Fortune reports that critics including Gary Marcus argue the growth projections aren’t supported by current evidence. And IMF chief Kristalina Georgieva, at the same meetings, reportedly cited Amara’s Law: we tend to overestimate the short-term effects of a technology and underestimate the long-term ones. That cuts both ways, and it is a useful way to hold Clark’s number.
| Claim | Where it comes from | How firm it is |
|---|---|---|
| “The singularity is here” | Clark, spoken in the lecture | Interpretation |
| Benchmark jumps | Clark, spoken; top scorer is his employer’s model | Unchecked here |
| $44.4T GDP in 2030 | Anthropic’s scenario explorer, extreme case | A scenario, not a forecast |
| Losing control is a risk | Clark, tied to AI-driven AI research | A stated risk |
My read
Three things stand out to me. The lecture is more careful than its headline: it’s a talk about benefits that spends its closing stretch on restraint. The number is real but conditional, and the condition is the interesting part. And the most useful sentence for a policymaker may be the least quoted one, the request for “unusual and serious rules” for the companies building the technology. If you take the $44 trillion seriously, that request comes with it.
The question isn’t whether the number is right. It’s whether anyone is building the brakes at the pace they’re building the engine.
Further reading
Buy me a coffee