Where is the explosion?

Two clocks

One forecast for 2027 and one diagnosis for today. They do not have to contradict each other.

At the turn of September and October 2026, two answers were given to the question of when artificial intelligence will start improving itself faster than humans do.

Rayan Krishnan, head of Vals AI (independent benchmarks), said on 3 October 2026 that according to their index Anthropic will reach full RSI (recursive self-improvement) in July or August 2027. That is a forecast: an extrapolation from a limited set of tasks. That full automation of R&D follows from it is something the benchmark assumes rather than tests.Vals methodology

Ramez Naam, an investor and futurist who described the solar revolution before it became obvious, published a diagnosis at Noah Smith's: under his assumptions (chapter 6), today's loop is 5–10 times too weak to sustain itself.

Five meanings of self-improvement

"RSI" is five different things. Most disputes start by confusing the rungs.

Naam proposes a taxonomy that puts the discussion in order. Types 1–4 run into diminishing returns. Only type 5, the take-off to superintelligence, requires increasing returns.

Krishnan's forecast concerns type 4 for a single profession; the explosion is type 5. Naam judges that without a major conceptual breakthrough type 5 is unlikely; that is an opinion, not a measurement.

Benchmark versus laboratory

A benchmark, lab work and a forecast measure three different things. Before you compare the numbers, check the unit.

The METR task horizon is the length of a task, counted in hours of expert work, at which a model reaches 50% or 80% success. The last published estimate in the series shown concerns an early version of Claude Mythos Preview from April 2026: 17.4 h at 50% (8.5–55 h) and 3.1 h at 80% (1.6–6.6 h). Above 16 hours the METR suite stops measuring reliably.METR definition

For newer models (GPT-5.6 Sol, GPT-6 Astra, Claude Fable 5.1, Opus 5.5) there are no comparable points in the cited METR horizon series. METR itself deemed its estimates for Sol unreliable.METR on Sol

OpenAI (report of 6 September 2026) measures something else: tasks that researchers actually delegate to agents, with success meaning no human intervention. Naam read off the chart a horizon of about 15 minutes at 80% (January–July 2026). The same report announces an "automated research intern" that, under human direction, performs tasks lasting several days.OpenAI report

These numbers cannot be divided by one another: different models, period, tasks and definition of success. The conclusion: a good benchmark result is not enough to declare readiness for autonomous research. Autonomy in the lab requires a separate measurement.

And the forecast? The authors of AI 2027 assessed on 16 August 2026 that reality is running at 70–90% of the pace of their scenario, depending on the choice of indicators. The path in Exhibit 3 illustrates one variant, not the whole model.AI Futures, 16 Aug 2026

A chain with its second half missing

Token consumption is growing much faster than the number of experiments. The cited data do not allow us to judge how the value of discoveries has changed.

The OpenAI report shows big numbers: 124 times more tokens per researcher, about 7 times more code per engineer. These are activity measures. Closest to an outcome is the number of experiments: in August 2026 researchers were running 1.6 times more of them per person than the 2025 average.

Further down the chain breaks: how many experiments produced discoveries, and how many discoveries produced better models, the report does not say. Naam estimates, calling it speculation himself, that under typical diminishing returns 60% more experiments amount to roughly 10% faster progress.

Anthropic looks from the other side. In a voluntary survey of 130 employees, the average self-reported productivity gain is about 4×. According to the system card, doubling the pace of progress would require a gain "roughly an order of magnitude larger", i.e. around 40×. That is a self-assessment set against a model-based estimate whose arithmetic the card does not show.

"Productivity gains do not translate one-to-one into capability progress."Anthropic, Claude Mythos Preview system card (as cited by R. Naam)

In its evaluation of Opus 5.5 (22 September 2026) METR relays a preliminary estimate: about 1.5× acceleration of capability progress thanks to AI inside Anthropic, without access to the data and without a clear period. We do not set it against Naam's estimate for OpenAI.METR, 22 Sep 2026

The strongest counter-argument, which Naam himself concedes: a better researcher may run fewer experiments, but better-aimed ones. That is why the last two links in the exhibit are empty, not zero.

More fuel, faster, but not explosively

Inputs are growing exponentially. The public capability index accelerated with reasoning models and has grown in a straight line since.

The world's total AI chip compute, including inference and with a projection at the end of the period, grew according to Epoch roughly 127 times in a little over three years (February 2023–2026, H100 equivalents). That is infrastructure context, not the input to a single model and not a return on research.Naam, chapter 5

With reasoning models the trend of the Epoch capability index (ECI) accelerated: about 14 points a year since o1 (September 2024), against about 6 before. That alone does not settle how much of it is recursive self-improvement and how much a change of recipe (more inference compute, more RL). Since the end of 2024 there is no further upward bend.Epoch, 1 Sep 2026

"We believe that internal use of the latest AI models has been a key factor in maintaining the current pace of progress, but we do not yet see clear signs of dramatic acceleration beyond that pace."Anthropic, Claude Fable 5.1 & Mythos 5.1 system card (italics Anthropic's; as cited by R. Naam)

The loop

The loop exists. The question is whether each turn yields more than the previous one, or less.

Better AI speeds up research, research yields better AI. Tom Cunningham and co-authors calculated how much each additional ECI point must deliver for the loop to sustain itself: about 15% more research productivity. Naam, after a calibration correction, gets about 19%. These are two calibration variants; Naam himself advises not to read much into the difference.Cunningham et al., pp. 28–29

The current state: from the Anthropic survey Cunningham et al. get about 9%, from the OpenAI experiment logs Naam gets about 3%. Hence his 5–10× gap. At 9% the gap is only 1.7–2.1×: the choice of input data weighs more than the choice of threshold.

Below the threshold growth continues, driven by chips, data centres and capital, but the software loop itself dies out. Above the threshold it picks up speed. Move the slider.

Two thresholds on one map

Krishnan measures capability. Naam measures amplification. There is a scenario in which both are right.

Capability (threshold A) is the question of whether a model can do an ML engineer's job; the Vals index asks exactly that. Amplification (threshold B) is more than acceleration: AI can speed up research relative to work without AI (the 1.5× estimate from chapter 4) and still be below the threshold. Amplification here means whether an improvement in AI triggers feedback strong enough for successive turns to sustain or increase the pace of progress at constant external inputs.

A scenario consistent with both: in 2027 a model does an ML engineer's job, and the capability curve keeps rising in a straight line. Automating a known recipe is climbing a known mountain.

Why clever steps go sideways

Today's models are excellent executors and weaker discoverers. A fresh study suggests how training can restrict exploration, and offers a partial remedy.

That is an observation from one workshop. New sources allow it to be taken apart.

The Opus 5.5 system card. Anthropic writes about its own newest model:

"Like previous models, it is weaker at open-ended research: internal users report that it mostly tests incremental ideas and prefers less ambitious hypotheses, and in a human-led biology exercise it stuck to the published literature and had difficulty developing new ideas."Anthropic, Claude Opus 5.5 system card, section 2.3.3 (as cited by R. Naam)

On CoBench, a benchmark of historical R&D problems in AI, Opus 5.5 gains only 2.6 percentage points over Opus 5, within the margin of error. METR supposes, with great uncertainty, that full automation of R&D will require major progress in prediction, in building one's own feedback loops and in the researcher's "judgment".METR, 22 Sep 2026

Sharpening Tax. Researchers at Meta Superintelligence Labs (arXiv, 1 October 2026) describe a matching mechanism. RL training "sharpens" the model: first-attempt accuracy (pass@1) rises, the benefit of further attempts falls. In an observational analysis of 42 combinations of commercial models and benchmarks the "tax" at 128 attempts was positive in 36, which does not always mean a drop in pass@128.paper, §4

In a controlled experiment on Qwen2.5-7B-Instruct, standard RL (PPO) on Sokoban more than doubles pass@1, and pass@128 falls by 28% relative. The remedy, PTGS, is a sampling temperature matched to task difficulty during RL training: it raises both measures at once and recovers about 68% of the pass@128 loss on Sokoban and 39% on FrozenLake.paper, §5, tab. 5

The analysis of commercial models does not establish the cause of Opus's weakness in discovery. It shows a mechanism consistent with the observations and that it is not inevitable.

Where the loop works. Where verification is cheap and automatic. Poteto (Cursor, now SpaceX AI) says that in a month he "shipped" 2,000–2,500 PRs because his agents run the application and check their own work themselves, and a human reviews samples in the morning (recording, 39:12; an anecdote, not a measurement, a large part being code clean-up). Kent Beck on attempts without a human in the loop: "it always goes off the rails" (Prodacity 2026, 31:00). The direction of research has no cheap verifier.

What would change the assessment

Let us not ask who is right. Let us ask which observation would reset which clock.

The intuition is sound, with two corrections. The same premise holds up Krishnan's clock, which extrapolates today's recipe for building models. And for Naam a breakthrough does not invalidate the analysis, it only forces a recalibration; stronger feedback could also come from better model judgment, memory, data or coupled loops. Besides, diminishing returns did not arrive with the Transformer: scaling curves with diminishing returns were being fitted as early as 1993.

A map for the investor

Do not ask when the explosion will come. Ask how quickly and how reliably the result can be checked, and which variable AI really moves.

The analysis in this text gives that opinion a criterion. Autonomy is favoured by fast and reliable verification, especially when the cost of error is bounded and the result can be checked before use. Naam expects narrow superintelligence where machines generate the data and check the result themselves: in chess and go, in formal mathematics, in parts of programming and cybersecurity. The further from such verification, the more a mixed model pays off: AI executes, an expert directs and evaluates.

This concerns tasks, not fields: programming includes both a change with a full test suite and an architecture decision, biotech both in-silico docking and a clinical trial endpoint. In biotech, moreover, Eroom's law applies: between 1950 and 2010 the real R&D cost per newly approved drug doubled roughly every 9 years. AI will shorten the verifiable stages in software fastest, but it also creates value through lower cost and better selection (Exhibit 10b).Scannell et al., 2012