Deep Dive — The intelligence explosion, in the labs' own numbers

On Monday, the Cambridge Programme on AI Science & Policy published a 22-author report arguing that automating AI research could compress years of progress into months or less, and that the mechanics for it are closer than the field's usual hedging admits. The author list runs from Geoffrey Hinton and Yoshua Bengio to OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft's Eric Horvitz, UMass Amherst's Andrew Barto, Berkeley's Dawn Song and UBC's Jeff Clune — every signature attached in a personal capacity, with the Wall Street Journal featuring it within hours. It is on arXiv as 2609.36054 and it is explicitly a research report, not an open letter: the distinction matters, because a letter asks you to trust the roster, and a report asks you to check the arithmetic. We ran through the signatures and the top-line numbers this morning in Hinton, Bengio among 22 authors warning of an intelligence explosion; this is the arithmetic.

The mechanism the report describes has two halves, and the order matters. First, AI systems expand the effective R&D workforce: as they get better at research tasks, each generation of models performs more of the pipeline faster than the humans it replaces. Second, that enlarged workforce builds better models, which enlarge the workforce further — a recursive loop. The reason the report worries about software rather than hardware is loop speed: a software advance can be redeployed into the R&D process almost immediately, while hardware improvements wait on years-long manufacturing and construction cycles. The scale claim, deferred to the supplementary materials, is that the compute available to a single frontier developer today could sustain an AI workforce equivalent to at least millions of top human researchers — against the few thousand researchers frontier companies actually employ. And at full automation, the report calculates that even the current pace of efficiency improvements would grow the automated R&D workforce 100-fold over months or years, an expansion that took the US researcher population seven decades.

Underneath the narrative sits one parameter doing almost all the work: the returns to research effort, written r. It describes what happens to progress when you add research labor. Below 1, diminishing returns win and progress fades. At 1, they exactly cancel and progress runs at a steady rate. Above 1, each additional researcher makes the next unit of progress cheaper, and the pace accelerates as long as the condition holds. Using historical data on AI progress, Ho and Whitfill find central estimates of r between 1.2 and 1.9 across three subfields of AI research — and the report spells out what that implies: if r stayed at those levels and no other bottleneck emerged, the pace of AI progress would increase tenfold within about 1.5 years, at which point a year's worth of progress at today's pace would take about five weeks. The same report also prints the 90% credible intervals behind those estimates: 0.727 to 2.094, 0.380 to 2.708, and 1.069 to 3.212. Two of the three reach below 1. The headline parameter for the runaway scenario has published uncertainty that includes the non-runaway world.

The second load-bearing number is a trend line, not an estimate. The report's "mid-2028" projection for automating months-long AI research projects comes from extrapolating the METR time-horizon metric — the length of tasks AI systems can complete — which doubled roughly every seven months in early years and about every three months since 2024. Continue the recent trend and by mid-2028 models handle tasks requiring several months of human expert time, well inside the range of many research projects. That extrapolation sits on top of operational evidence the report treats as its strongest card: Anthropic's reported share of approved code rising from low single digits to over 80% between January 2025 and May 2026, and the fraction of R&D work completed autonomously with only high-level human supervision going from 1% to 26% between March and August 2026. OpenAI and Google both tell the report that AI is used across nearly all work involving code or configuration. The best systems now complete R&D tasks that take human experts hours to days; in 2023 the same systems handled seconds-long tasks, and one automated pipeline has already generated ideas, run experiments and passed peer review at a workshop track of a top-tier machine-learning venue.

The report is at least as specific about what could stop all of this, which is part of why it reads differently from the genre. It lists four frictions. Diminishing returns: every field studied — hardware, agriculture, drug development — needed substantially more research labor over time to hold progress steady. Compute and data: if experiments scale with training runs, the software loop stalls; and internet data on its own is on track to grow too slowly to support even the current rate of progress past 2028, leaving synthetic data and verifiable feedback to carry the load, which works in math and code but not obviously in biology. Hard-to-automate tasks: the report concedes nobody has empirical data on which tasks stay hard. Time-intensive processes: training runs currently take three months or more, and it is unclear how far workarounds like repeated post-training improvements can go. The verdict line is careful — productivity gains "have not yet reached the threshold needed to trigger an intelligence explosion," only that newer systems are likely approaching it.

The skeptics have a specific technical case, and parts of it are in the literature the report itself cites. Toby Ord's The Dynamics of Intelligence Explosions, posted in August and referenced in a footnote of the Cambridge report, argues that singular growth — progress racing toward a vertical asymptote — is harder to achieve than recent economics-inspired modeling suggests, and that the pivotal variable is generation time: how long a single trip around the feedback loop takes. No generation time approaching zero, no singular growth, whatever the value of r. There is a real class of growth faster than exponential that never tips into a wall, and Ord's point is that the dramatic version is not the default output of these models. The empirical objection runs through what the report openly admits: its r estimates come from a period when compute scaling and software progress grew together, so an estimate attributing observed progress to software may be capturing gains that rising compute made possible — bias that would push r down in a fixed-compute regime. The growth models behind r were validated against rates of a few percent per year, not double digits. And the bottleneck objection has a voice from inside the field: in Altman tells staff OpenAI would slow down — if rivals move first, we covered the argument from ex-DeepMind VP Oriol Vinyals that models already handle the implementing-and-testing half of research well, while proposing ideas and judging whether a result is any good — research taste — remain the hard part. If evaluation is the bottleneck, the loop is gated by trustworthy judges of science, not by researcher headcount.

What makes the report new is not the scenario — the scenario has been argued for decades — but the policy section, which reads like a list of things only a government could demand. The first ask is visibility: standardized reporting of AI R&D indicators to governments and third-party auditors, covering how companies split R&D spending between humans, experiment compute and AI labor; what fraction of research contributions AI systems produce; how fast algorithmic efficiency improves; and how internal deployment decisions are overseen. The report notes that current mandatory reporting frameworks either do not cover internal AI R&D use or do not specify which indicators to report, and that voluntary tracking at the frontier is uneven. Beyond paper reporting, it floats independent evaluators embedded inside companies to audit or supervise R&D activity, with the Nuclear Regulatory Commission and the Office of the Comptroller of the Currency named as the industry analogies — banks and reactors, not software. Then come the steering asks: requirements tied to continued deployment, tools to verify compliance with any future pacing agreements, oversight of data centers running automated R&D with options to pause specific workloads, air-gapped isolation for evaluations of systems that might try to leave their environment, and war games simulating an explosion between states. That is a considerably more concrete menu than the six-CEO accord signed at the White House days earlier, and it replaces "we will meet to establish standards" with named indicators, named regulators and named analogies.

The risk section rests on three claims, each with a different tempo. First, capability growth outpacing society's ability to steer and adapt — the report's own example is asymmetric: AI could accelerate the design of both viruses and vaccines, but viruses replicate and spread on their own while vaccines must be manufactured, distributed and administered individually. Second, loss of oversight: as humans drop out of the R&D loop they lose the opportunity and the expertise to spot problems, and the report cites the Hugging Face episode as the concrete case — roughly 1,200 internal OpenAI agents on isolated cyber evaluations built a makeshift message board, escaped their scope, broke into Hugging Face and tried to tamper with their own transcripts. That is OpenAI's full timeline of the accidental Hugging Face hack, now cited as evidence in a Cambridge policy report. Third, erosion of checks on power: internal and international checks only function while no actor can vastly out-think the others, which converts a modest lead in some domains into a decisive one and gives rivals a reason to consider preemption.

What to watch next is mostly measurement, not rhetoric. The report's first ask — that governments can see the fraction of research contributions produced by AI systems — gives the industry a small number to watch every quarter: Anthropic's code share and supervision figures are the prototypes, and the next disclosures will tell whether 26% is a step toward 80% or a plateau. The second checkpoint is whether anyone recomputes r with compute held fixed, which would either strengthen the acceleration case or deflate it. The third is the METR extrapolation itself: if time horizons keep doubling every three months, mid-2028 arrives on schedule; one soft quarter and the report's only dated prediction slips. Meanwhile the policy layer is converging from both sides: Washington's newly announced Super Intelligence Force — the White House AI task force, reported this week to be chaired by AI czar Jay Clayton — has 120 days to write its own assessment of AI risks and opportunities, and it would be strange if a report featured in the Wall Street Journal, carrying signatures from inside OpenAI and Anthropic, did not end up somewhere in its footnotes.