Software-Driven Intelligence Explosion

Sources#

• Humans Are Not Chinchillas: Revisiting "How Quick and Big Would A Software Intelligence Explosion Be?"
• The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion
• What if automating AI R&D triggers an intelligence explosion?

Summary#

An intelligence explosion in the sense of Chan, Mindermann and 20 co-authors (arXiv 2609.36054, 2026-09-28, practitioner-opinion) is "a dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less". They mean a qualitative break from the "rapid but largely steady progress of the last few years". A software-driven one is the narrower case in which automated AI R&D causes it through software advances alone: algorithms, data, code, training and inference efficiency, research management. Hardware is held roughly fixed. The authors pick the software route for two reasons. AI is measurably improving at AI R&D. And a software advance can be redeployed into the R&D process "almost immediately", whereas hardware moves on years-long fab and construction cycles. The term and the framing come from Eth & Davidson (2025), which is not in raw/.

This page holds the mechanism, the one parameter that decides it, the arithmetic, and the evidence status of each friction. The growth-curve debate it feeds is on Intelligence Explosion Dynamics. The policy response is on Frontier Pause Verification. The current-acceleration measurements it starts from are on AI Accelerating AI Development.

The authors' own read: "a coherent pathway to a software-driven intelligence explosion that is consistent with the existing evidence." They add that "productivity gains from AI R&D automation have not yet reached the threshold needed to trigger an intelligence explosion, but gains from newer systems are likely approaching that threshold." That second sentence is a forecast, and it rests on Cunningham et al., The economics of recursive self-improvement (2026-07), which is not in raw/. Attribute it to them.

The mechanism: an R&D workforce that grows itself#

The loop has two parts (the paper's Figure 2, a schematic with no data). First, each generation of AI does more R&D tasks faster and better than humans, which expands the effective R&D workforce. Second, that workforce produces better AI, which expands it again.

How large the workforce could be. The supplementary materials follow Denain et al.'s method (How many digital workers could OpenAI deploy?, not in raw/): tokens a developer can generate per day, divided by tokens per researcher-day.

• OpenAI alone can generate on the order of 10¹³ tokens per day of inference.
• On RE-Bench, models emit on average 5·10⁵ tokens on runs of up to 8 hours, which is one human workday.
• That gives about 2·10⁷ researcher-equivalents. Allowing an order of magnitude either way on tokens per researcher-day (5·10⁴ to 5·10⁶), the range is 2·10⁶ to 2·10⁸.

That is "at least millions of top human researchers", against the thousands frontier companies employ. It assumes expert-level systems cost about what today's cost to run, and it counts token throughput, not research quality. DeepMind's report put the figure at "hundreds of thousands to millions of artificial researchers". This estimate sits one to two orders above that, with an explicit method.

How fast the workforce could grow. At full automation, "even the current pace of efficiency improvements would grow the automated R&D workforce 100-fold over months or years". The U.S. researcher population took seven decades to grow that much. The input is Cottier et al.'s finding that the price of reaching a fixed capability milestone fell 9× to 900× per year, depending on the milestone, with a substantial part of the fall due to software.

What distinguishes it from ordinary technology feedback. Chips already design better chips. The difference claimed is the share: as AI substitutes for humans on more R&D tasks, each software advance speeds up a larger fraction of the pipeline.

The deciding parameter: returns to research effort, r#

The paper's central quantity is r, the returns to research effort. It sets whether growth in R&D labor outruns the depletion of easy ideas.

• r < 1: diminishing returns dominate and progress fades.
• r = 1: the two forces offset and progress holds its rate.
• r > 1: labor growth wins and progress accelerates, for as long as r stays above 1.

The authors add that r "must eventually drop below 1" as AI hits computational and physical limits, and that nobody knows how much progress fits before that.

The model. Software quality A grows as dA/dt = A^(1−β) · E^λ, the semi-endogenous form from the macroeconomics of innovation (Bloom et al.). E is effective R&D labor, λ is returns to scale on that labor, and β is how fast ideas get harder to find. Compute and data bottlenecks are set aside. Under full automation the authors argue E ∝ A in two cases:

1. All gains are inference efficiency. Cheaper inference runs proportionally more automated researchers. This is a direct mapping.
2. All gains are training efficiency. A fixed compute stock trains a more capable model, not more of them. E ∝ A here requires the "(potentially questionable) assumption" that effective researcher count scales linearly with capability.

Substituting E = kA gives dA/dt = c·A^(λ−β+1). The growth rate rises with A when λ > β, so r = λ/β > 1 is the condition.

The estimates. Ho & Whitfill (2025, not in raw/) fit historical AI progress in three subfields and get central r between 1.2 and 1.9. Their 90% credible intervals are 0.727–2.094, 0.380–2.708 and 1.069–3.212, so two of the three include r < 1. The averaged central parameters are λ = 1.40 and β = 1.01.

The arithmetic, which is the paper's headline number. With λ − β = 0.39:

• Each doubling of A multiplies the growth rate by 2^0.39 ≈ 1.31. Each subsequent doubling therefore takes about 76% as long as the one before.
• A tenfold rise in the growth rate takes log₂(10)/0.39 ≈ 8.5 doublings.
• The first doubling is set at 4.5 months, the rate at which software alone has doubled training-compute efficiency (Ho et al.'s Rosetta stone paper). Inference efficiency may be improving faster.
• Summed over 9 doublings: 4.5 × (1 − 0.76⁹)/(1 − 0.76) ≈ 17 months. (Wiki check: 0.76⁹ ≈ 0.085, giving 17.2 months, which matches.)

So if r stayed at these values and nothing else bound, progress would be ten times today's pace about 1.5 years after full automation, and "a year's worth of progress at today's pace would take about five weeks".

What kind of curve that is. dA/dt ∝ A^(1+0.39) is the hyperbolic regime on Intelligence Explosion Dynamics: taken literally it reaches infinity in finite time. The authors say so themselves. The models "break down in the limit, where they imply that infinite labor yields infinite progress in finite time", and serial dependencies and hardware speed make that impossible (they cite Trammell, The bounded parallelizability of R&D, which is not in raw/). The 1.5-year figure is therefore a point on a hyperbola that its authors expect to bend into an S-curve at a limit they cannot locate. It is the first number in the corpus for how far into the hyperbolic regime the bend could be, given an r.

The 17 months as a random variable (October 2026)#

Li, Gao, Yang & Sha (arXiv 2610.18445, 2026-10-13) take this equation and make progress arrive as discrete random ideas. Full treatment is on Stochastic Explosion Timing and Detection Limits. It is a theory paper: it proves its results and reuses this paper's central values "for illustration only". Three of its results change how the arithmetic above should be read.

• The 17 months is a mean, and the blow-up follows it closely. If every idea doubles A, the nine-term sum is exactly the expected time to the ninth idea, 𝔼T₉ = 17.3 months. The infinite series 4.5/(1 − 2^(−0.39)) = 19.00 months is the expected explosion time. So on its own arithmetic the model reaches tenfold pace about 1.7 months before its singularity in expectation. With doubling-sized ideas, the singularity's 5–95% range is 9.7–31.9 months. With 1% ideas the range is 15.5–17.9 months, around the deterministic 16.65.
• Randomness leaves the r = 1 threshold where it is. The process explodes almost surely if r > 1 and never if r ≤ 1, for any distribution of idea sizes. The argument over r above remains the whole argument over whether.
• Uncertainty about r makes the expected wait infinite. Fit a log-normal prior to Ho & Whitfill's 0.727–2.094 interval. Then P(r > 1) = 0.743, and conditional on r > 1 the median deterministic wait is 17.3 months, with P(> 5 years) = 0.150 and P(> 10 years) = 0.074. The mean is infinite because the prior has mass near r = 1. "Ten times today's pace in ~1.5 years" is a median under one prior, not an expectation.

Why every estimate of r is soft#

The supplementary materials list biases in both directions, and they are the most useful part of the paper for anyone citing the 1.2–1.9 range.

Biases that push the true r lower than the estimates:

• The compute confound. The estimates come from years of rapid compute scaling. Software progress and compute grew together, so progress credited to software may have been enabled by compute. Some advances only pay off at scale: the transformer gives large gains at high training compute and small gains at low compute (Gundlach et al.). With compute fixed or slow-growing, "r would be lower than historical data suggest."
• Labor only. Bloom et al. measured inputs in R&D dollars, which include capital. The paper's r counts labor alone because the software loop holds compute roughly constant. By construction that gives a lower r than one counting all inputs.

Biases that push it higher:

• Most estimates ignore post-training and scaffolding gains.
• Capability may matter more than headcount. "Even an extremely large number of mediocre researchers may not be able to substitute for one genius researcher." If so, capability gains expand effective labor more than the linear assumption allows.
• Compute bottlenecks could be routed around: better extrapolation from small experiments, cheaper experiments, less compute-hungry approaches, and algorithmic progress that needs no experiments.

Biases in either direction:

• The wrong efficiency measure. Existing r estimates use training efficiency. Inference efficiency maps most directly onto workforce size, and nobody knows whether the two have similar returns. A proper measure would weigh both, and none exists yet.
• The labor proxy. Ho & Whitfill count unique authors publishing in a subfield. Drawing the field too narrowly undercounts labor, and drawing it too broadly overcounts it.
• Out of sample. The growth models were validated on rates of a few percent a year. An intelligence explosion is a regime of double-digit or higher rates.

How big: the ceiling, and a lower prior on r (October 2026)#

Chan et al. ask whether the loop runs away. The companion question, how quick and how big, is Davidson & Houlden's Forethought model (not in raw/). It asks how many year-equivalents of AI progress software alone could pack into a short period once AIs fully substitute for human AI researchers, with hardware held constant. It has four central parameters. Two matter here: r, the returns to software R&D at the start of the explosion, and the takeoff ceiling, the orders of magnitude (OOMs) of software-only efficiency gain still available after the first AI system for AI R&D automation (ASARA) exists.

James Tillman (Forethought, 2026-10-05, practitioner-opinion) reruns that model with two distributions changed (transcribed from the post's Datawrapper datasets):

The tail moves, the body less. The 3-year case falls moderately and the 10-year case by three quarters. The ablation columns show the ceiling does more than r: the ceiling change alone takes ≥10-in-1 from 17% to 6%, while the r change alone takes it to 12%. Tillman says he "reflectively endorse[s]" the drop at 10 years, puts little weight on the 3-year shift, and still gives ≥10-in-1 higher odds than his own run of the model does. These are forecasts from an opinion-tier model whose parameters he calls partly "somewhat arbitrar[y]". Attribute them.

The ceiling argument. The original 6–16 OOMs is two terms added together. ASARA is trained 2–6 OOMs less efficiently than a human brain (10²⁸ FLOP against Cotra's ~10²⁴), and there are a further 4–10 OOMs beyond the brain. Tillman cuts both.

• Humans are not Chinchillas. The case for headroom past the brain leans on the brain being badly undertrained: 10¹⁴ synapses on ~10⁹ "tokens" of experience, a ratio Chinchilla calls absurd, which mortality supposedly forced. Taken literally, Chinchilla's fit (L = 1.69 + 406.4/N^0.34 + 410.7/D^0.28) makes a 1% longer learning period worth 144× a 1% bigger brain. Then the last doubling of human brain size bought as much as about two extra months of learning. Brain tissue is metabolically costly and drives the obstetrical dilemma, so evolution would not have made that trade. A modified law matched to a rough 4× tradeoff (from brain-volume–IQ and earnings data) permits ~2 OOMs. But it implies a honeybee-sized network trained for ~50,000 years would reach human intelligence. Tillman's estimate for the rebalancing source is 0–2 OOMs, mostly in the first, against the original 1–5. His reason for a low cap: deployed AI also needs sample efficiency at inference, larger networks are what supply it (RLVR seems to need large models to avoid plateauing), and so far sample efficiency in use has not come apart from sample efficiency in training.
• One brain is not the workforce. ASARA replaces a "minimum basis set" of 10–1,000 distinct researchers, more likely around 100. Their combined training is ~10²⁵–10²⁶ FLOP, not 10²⁴. So ASARA starts 0–4 OOMs behind human efficiency, not 2–6. Footnote 13 adds that ASARA may arrive at nearer 10²⁷ FLOP than 10²⁸, which would lower this term further.

Wiki arithmetic: the two cuts stated, applied to 6–16, give about 3–11 OOMs. The post's 2–13 is wider than that at both ends. It cites "other considerations" without itemizing them, so the stated range is a judgment call, not a sum.

The r argument, and how it sits next to Ho & Whitfill. Davidson & Houlden anchor on Epoch's 1.4 for computer vision (Erdil & Besiroglu: unique authors double every 13 months, compute-per-performance halves every 9 months) and then adjust it. Tillman's objection is that at least half the state-of-the-art vision papers borrow from outside the field: ViT, sparse MoE, and ConvNeXt by its own account. So the output is not a function of the field's measured inputs. That is the labor-proxy bias above, given a concrete case. A field drawn too narrowly undercounts labor and so overstates r. He also argues compute growth has biased every returns-to-research estimate upward, which is Chan et al.'s compute confound. He centres r at 1.0 and does not offer an estimate of his own to replace the anchor.

(Medians and P(r > 1) for the log-uniform rows are wiki arithmetic.) Consistent in range, not in centre. Tillman's 0.33–3 covers most of each Ho & Whitfill 90% interval (0.727–2.094, 0.380–2.708, 1.069–3.212), so no fit falls outside his prior. But his median sits below all three central fits, and his prior makes runaway (r > 1) a coin flip where the Li et al. calibration makes it three to one. Two cautions keep these from being one quantity. Davidson & Houlden's r is the post-adjustment return at the start of the explosion: the 1.4 anchor ×2 for capability over headcount, ×1.45 for post-training, then cut three times for no compute growth, fixed hardware, and diminishing returns. Chan et al.'s r = λ/β is a labor-only historical fit. And none of the four rows is a new measurement. Not resolved. The priors differ on the same two named biases, and the experiment both sides ask for (Epoch's call for software-intelligence-explosion experiments, this page's inference-efficiency question) has not been run. Under Stochastic Explosion Timing and Detection Limits's Proposition 4.6, any prior with density at r = 1 gives an infinite expected wait conditional on explosion, so that result holds under Tillman's prior too.

What Tillman wants modelled next. Correlate r with the ceiling, since both depend on how far along the science of intelligence is at ASARA. Model quality of intelligence directly instead of as a headcount multiplier, which he expects could make takeoff "sharper and higher". And pin down the baseline year for "N years of progress". His policy reading is that a single actor bootstrapping a decisive strategic advantage is somewhat less likely, and more of AI's impact runs through an industrial explosion (Intelligence Explosion Dynamics).

The frictions, and how much evidence each has#

The paper names four frictions and splits the second into compute and data when it assesses them.

Compute is the crux on the paper's own evidence. It is the only friction where a cited result says the explosion is impossible under a specific, checkable condition, and the paper cannot say whether the condition holds. The data friction converges with The Data Wall and the Validation Commons Are One Supply Constraint: both put the escape route through verifiable feedback. The time friction is Brown's serial-experiment brake, now with a figure attached (3+ month training runs). The paper has "no direct evidence" on it, and it now has a theorem: in a delayed-feedback version of the model, any fixed deployment lag removes the finite-time singularity and leaves double-exponential growth, with log A doubling about every 6.37 months at a 3-month lag (Stochastic Explosion Timing and Detection Limits).

Where it meets the rest of the corpus#

• Anthropic's and METR's acceleration readings are a different quantity. Anthropic's August 2026 Risk Report says acceleration is "by less than a factor of 2", and METR's Opus 5.5 read is "~1.5X". The paper's threshold is self-sustaining acceleration: progress that accelerates "without any growth in exogenous inputs", which it calls necessary for a software-driven explosion. "Below 2×" and "approaching self-sustaining" can both be true. Not a contradiction, but the paper's direction-of-travel claim is stronger than either lab-side reading and rests on an un-ingested paper.
• Against the moderates, a scope difference with one live disagreement. Brown's hedged 3× is an unconditional median for the accelerated regime. The paper's 10× is conditional on full automation and on r holding. They answer different questions. The live disagreement is with Lambert's lossy self-improvement. Lambert reads efficiency gains as making intelligence cheaper without raising its peak. The paper's case 1 says pure inference-efficiency gains are enough to close the loop if r > 1, because cheaper inference means more researchers. The two positions turn on whether more copies yield more progress. The paper names this as an assumption ("mediocre researchers … genius researcher"), and Lambert asserts that "diminishing returns of more AI agents in parallel are real" without a curve. Unresolved. Both are opinion, and the parallel-returns curve that would settle it is not measured anywhere in the corpus (Multi-Agent Collective Intelligence).
• DeepMind agrees on research-gets-harder. The From-AGI-to-ASI report rates "research gets harder" a minor friction because artificial researchers are cheap. The paper reaches the same verdict and attaches an r. Neither treats compute for experiments as a separate friction, and that is where the paper locates the uncertainty.
• A different production function from Greenblatt's. Researcher Uplift from Code Output carries Greenblatt's R&D ≈ labor^0.55 × compute^0.45, in which compute takes almost half the weight. The paper's software-only model holds compute fixed and puts all growth on labor. Its own compute-confound caveat is the reason to doubt that: if historical progress credited to labor was partly compute's, a labor-only λ of 1.40 is too high. Recorded, not resolved. Neither decomposition is fit to frontier-lab data, and the paper's policy section asks for exactly the spending split that would fit one.

How onset would be detected#

The paper wants governments to receive standardized reports. Three of the requested indicators bear on this page directly:

• The inputs to r: how companies divide R&D spending among humans, compute for experiments, and compute for running AI researchers.
• Extent of automation: for example, "the fraction of research contributions produced by AI systems".
• Pace: for example, the rate of algorithmic-efficiency improvement.

It concedes that "precisely operationalizing an intelligence explosion is tricky and remains an area for future work". Li et al. give the statistical floor under the onset indicator. Assume an ideal monitor that sees every idea and knows the baseline rate and the acceleration per idea. Distinguishing this model from steady exponential growth at 5% error still takes 130 ideas for 1% ideas but only 9 for doubling-sized ones. For doubling-sized ideas the decision comes at the tenfold point, with about 1.7 months left. AI R&D Autonomy Evaluation (AECI) records what labs currently publish against these, and Frontier Pause Verification records the rest of the reporting and auditing agenda.

Connections#

• Intelligence Explosion Dynamics: the curve-shape debate this page quantifies. The r > 1 model is the hyperbolic regime, with its authors conceding the S-curve bend at a physical limit
• AGI-to-ASI Pathways: pathway 3 with its frictions assessed against data. DeepMind's "research gets harder" verdict agrees and now has an r attached
• Recursive Self-Improvement: the mechanism hub; this page is the quantitative case for the loop compounding
• AI Accelerating AI Development: the present-tense acceleration measurements the paper starts from, including the Anthropic autonomous-work share it cites second-hand
• AI R&D Autonomy Evaluation (AECI): the threshold measurements the paper's reporting indicators would standardize and extend to the spending split r requires
• Researcher Uplift from Code Output: Greenblatt's labor/compute split, which conflicts with the labor-only model here
• Frontier Pause Verification: the policy agenda this analysis is written to motivate
• Domestic Frontier Pacing: the AI Futures Project's quantified takeoff intervals (1.0 and 2.6 years from Automated Coder to superintelligence). They come from a different model of the same regime, and the paper cites the same group's AI 2040: Plan A for its verification and deterrence proposals
• Effective Compute Scaling: the software-efficiency doubling (~4.5 months for training compute) that sets the first doubling in the arithmetic, and the 2028 data-wall date
• Multi-Agent Collective Intelligence: where the parallel-returns curve that would settle the efficiency-versus-peak disagreement would have to come from
• Nathan Lambert: lossy self-improvement, the counter-position on whether efficiency gains compound
• Noam Brown: the serial-experiment brake that the paper's time-intensive friction formalizes
• Fundamental Limits of ASI: the takeoff ceiling is a practical limit expressed in OOMs of software-only efficiency, and Tillman's 2–13 is the corpus's only number for it
• Artificial Superintelligence (ASI): Tillman's lower odds on a ten-year jump within one year bear on how quickly AGI would become ASI
• Epoch AI: source of the 1.4 computer-vision anchor for r that Tillman rejects as contaminated by spillovers from other fields
• Stochastic Explosion Timing and Detection Limits: this model made stochastic. The 17 months is 𝔼T₉, the singularity follows about two months later with an idea-size-dependent spread, the r = 1 threshold survives, and detecting onset has a cube-root sample-size law

Open Questions#

• Do the compute requirements for finding software advances grow in proportion to frontier training compute? On the paper's evidence (Whitfill & Wu 2025) this condition alone decides whether a software-driven explosion is possible, and the paper says it lacks the data. What would settle it: a lab's experiment-compute share tracked against frontier-run size over several model generations.
• What is r when software quality is measured as inference efficiency, or as a combined inference-plus-training measure, and when it is estimated over a period of flat compute? Every published estimate uses training efficiency during a compute boom, which the paper concedes biases it in opposite directions.
• Are r and the takeoff ceiling correlated? Tillman suggests both are set by how mature the science of intelligence is when ASARA arrives: a mature science means hard progress and a near ceiling, an immature one easy progress and a high ceiling. Every published run of the size model draws them independently. A positive correlation pairs high r with a high ceiling, which would probably raise the ≥10-years-in-1 probability, and pairs low with low at the other end. What would settle it: a model run with the two correlated, or returns-to-research estimates from fields at different stages of maturity.

Sources#

• What if automating AI R&D triggers an intelligence explosion?: Alan Chan, Christoph Winter, Andrew Barto, Jakub Pachocki, Geoffrey Hinton, Eric Horvitz, Yoshua Bengio, Dawn Song, Jack Clark, Hilary Greaves, Anton Korinek, Samuel Hammond, Thore Graepel, Ben Bariach, Philip Torr, Sheila McIlraith, Jeff Clune, Sam Manning, Girish Sastry, Tom Davidson, Daniel Eth & Sören Mindermann, What if automating AI R&D triggers an intelligence explosion?, arXiv 2609.36054, 2026-09-28, 14pp (practitioner-opinion). Main text §"Automating AI R&D could trigger an intelligence explosion" (mechanism, frictions, evidence), the supplementary materials (workforce estimate, the feedback-loop model, uncertainties in r), and notes 3–10. All the load-bearing inputs are cited, not produced: Ho & Whitfill's r, Whitfill & Wu's compute result, Davidson et al.'s hard-task model, Cunningham et al.'s threshold claim, and Denain et al.'s workforce method are all outside raw/. The arithmetic was rechecked here and holds. Parse note: PDF-derived (docling 2.126.0, zero tables, two schematic figures with no data). Docling spaced the exponents ("10 13", "5 · 10 5", "A 1 - β E λ"), and they are read as powers throughout
• The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion: Li, Gao, Yang & Sha, arXiv 2610.18445, 2026-10-13 (practitioner-opinion, a theory paper with proofs and no data). Cited for Remark 3.6 (17 months = 𝔼T₉, 19.00 = 𝔼ζ), Table 2's timing distribution, Proposition 4.6's prior calculation, Theorem 6.1's delay result and Table 3's detection thresholds. Full treatment on Stochastic Explosion Timing and Detection Limits
• Humans Are Not Chinchillas: Revisiting "How Quick and Big Would A Software Intelligence Explosion Be?": James Tillman (guest author for Forethought), Humans Are Not Chinchillas: Revisiting "How Quick and Big Would A Software Intelligence Explosion Be?", Forethought newsletter, 2026-10-05, ~4,200 words (practitioner-opinion). Cited for both parameter tables and the ablation table, the Chinchilla 144× and modified-law arguments (footnotes 4, 10, 11), the minimum-basis-set argument, the computer-vision spillover critique of the 1.4 anchor, and the three modelling suggestions. Tables: the three charts are Datawrapper embeds. The raw transcribes each one's published dataset.csv, so they are not image reads. Davidson & Houlden's model, Erdil & Besiroglu, and Ho & Whitfill are all outside raw/. COI: Forethought published both the original model and this revision, and Tom Davidson supplied one of the arguments (footnote 9)