Intelligence Explosion Dynamics

Sources#

• From AGI to ASI
• How to pace the US frontier
• Idea Search: Guiding Tree Search with Ideas to Explore Diverse Scientific Methods
• Noam Brown – Agent swarms, alignment, & recursive self-improvement
• Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
• Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown

Summary#

Where Recursive Self-Improvement names the mechanism (AI building its successor) and the pathways report places it as pathway 3, this page is about the growth dynamics — the shape of the curve. The "From AGI to ASI" report distinguishes three regimes and stresses that which one obtains is poorly understood and has no historic precedent to fit:

• Exponential — constant multiplicative growth rate (e.g. effective compute at ~10×/yr).
• Hyperbolic (super-exponential) — the growth rate itself increases with the quantity that grows; the characteristic property is infinite growth in finite time — a singularity. First raised for AI self-improvement by Solomonoff (1985); the basis of Kurzweil's "Singularity" and many fast-takeoff scenarios (Good 1965's "intelligence explosion", Bostrom, Chalmers, Davidson).
• S-curve — in natural finite systems, frictions and boundary conditions bend growth down before a singularity. For AI-research automation, "the point at which these frictions kick in is unknown."

The feedback loop#

If AI speeds up AI research, that progress yields faster/more-capable/more-numerous AI, which speeds research further — a positive feedback loop. The most dramatic dynamics (potentially hyperbolic) arise if AI research is fully automated, but weaker recursive effects are already in play (e.g. "thinking" models curating better training data; neural architecture search; AI-assisted hardware design). The report's working judgment: it is a strong assumption that hundreds of thousands to millions of artificial researchers would have negligible impact — so unless progress halts first, recursive improvement is likely at least a significant accelerant.

Four mechanisms (mapped to human evolution)#

The report maps recursive-improvement flavors onto the engines of human capability growth — AI may run each far faster:

1. Genetic (genotypic RSI) — improving the "blueprints": code (architectures, optimizers, harnesses) and hardware designs. Slow for humans; potentially very rapid for self-modifying AI.
2. Cultural (memetic RSI) — improving intellectual artifacts: automated dataset curation, synthetic data, AlphaZero-style recursive distillation, tool formation. Human cultural evolution drove the last 50,000 years; AI's could be much faster given how fast artifacts are produced/shared/consumed.
3. Cooperative (sociogenic RSI) — division of labor: specialization frees resources → larger collectives → further specialization. Requires cooperation; its weight for AI collectives is unclear (today's models specialize instantly via prompting). See Multi-Agent Collective Intelligence.
4. Data — AI curating/generating/simulating higher-quality datasets for the next generation (the AlphaZero distill-search-back-into-prior loop). The economic pressure to harvest test-time-compute returns from billions of users makes this concrete.

Formal barriers & what dampens the explosion#

• Schmidhuber's Gödel machines formalize provably-optimal self-modification but require complete self-knowledge and are limited by Gödel's incompleteness (see Fundamental Limits of ASI).
• Christiano's iterated amplification — capability bootstrapping while preserving alignment by recursive task decomposition.
• Physical & real-time brakes — even superhuman-speed digital researchers must run experiments and wait for outcomes; anything requiring physical manipulation (better chips, wet-lab validation) can't be sped up arbitrarily. This is the embodied bottleneck and the report's main argument against an unbounded singularity. Iterated recursion also tends to plateau (diminishing returns, cf. AlphaZero) or degenerate (training on self-generated data).

The report's overall low-confidence read: a hard plateau exactly at AGI is unlikely; more probable is either a pre-AGI plateau or a relatively smooth AGI→weak-ASI transition — unless recursive improvement produces dramatic acceleration, which "cannot be ruled out" and would make the transition rapid.

A practitioner's brake: test-time compute as the pacing constraint#

Noam Brown (OpenAI, practitioner-opinion) supplies an independent, mechanism-level argument against the hyperbolic singularity — from inside a frontier lab rather than from theory. Because a model's greatest capability is unlocked only by large-scale test-time compute — runs that can take weeks or months — time itself becomes the binding constraint. "Things can only go so fast because the models need to run for long enough to actually do something really, really powerful." An overnight intelligence explosion (a breakthrough that instantly makes the model smarter, cascading in moments) is ruled out not by a compute ceiling but by the wall-clock the compute has to burn. Brown's name for the realistic shape is a "gradual takeoff": RSI accelerates the parts that parallelize and gets bottlenecked on the parts that don't — the same Amdahl's-law structure as Recursive Self-Improvement's futures — so "over time the things we're getting bottlenecked on are going to shrink" without ever going instantaneous.

This converges with the report's own embodied bottleneck (The Abstraction Barrier) from a different direction: DeepMind argues physical experiments can't be sped arbitrarily; Brown argues that even pure-inference research can't, because inference has its own irreducible duration at the frontier of capability. Both land on "time-paced, not instant."

The same practitioner, three months later: same verdict, different brake, and a number (September 2026)#

Brown restates the anti-hyperbolic conclusion in September 2026 (practitioner-opinion) and the mechanism underneath it has changed — worth marking, because the June argument is the one this page has been carrying as his distinctive contribution.

The inference-duration brake above is not mentioned. What is in its place is the ordinary experimental one:

"In mathematics, you're purely bottlenecked by thinking… When you look at things like RSI, you do have to run experiments. It's not enough to just be extremely smart. … It's running experiments serially, because they take a while to either train new models or to get the results. It's having the GPUs to run those experiments."

So inference wall-clock is Brown's binding constraint (not superseded in fact — he never retracts it — but displaced as of 2026-09-17: asked the same question after the Navier-Stokes result, he reaches for serial experiment latency and GPU supply instead, and the argument that made his June position distinctive from DeepMind's embodied bottleneck does not appear.) The practical effect is that his position has moved toward the report's, not away from it: what was a novel pure-inference brake is now the standard compute-and-experiments brake, and the convergence noted above is closer to an identity.

Two things are genuinely new. First, the counterfactual he uses to price compute against intelligence, which is a compact statement of where he thinks the binding input is: "if you had 100x less compute and all the most brilliant people in the world working at OpenAI, how much progress would you be making relative to having the amount of compute that we have now with the amount of people we have? I suspect it would be less progress, actually." Pressed on whether that means 100× less, he declines — "No, not 100x less. But… a lot less." Compute is the scarcer input in his reading, and the elasticity is unstated.

Second, a number where June had only a shape, and it is the sharpest thing he says about this page's subject:

"We do see a speedup, and we see a significant speedup. But I don't think it's an overnight intelligence explosion where we go 100x faster… If you put a gun to my head and ask me for a number, I could see things going 3x faster. That is huge… But there's a big difference between that and 100x faster."

With the error bars he volunteers: "It could be that things only go 50% faster. I think it's unlikely, but it's possible that things go 10x faster. There's a lot of uncertainty around this." And the framing that keeps 3× from reading as deflationary: "considering how fast things are going now on an exponential, if that exponential is 3x faster, that is massive."

Where to file 3×. It is a personal median from one researcher under interview pressure, hedged three ways, with no model behind it and no definition of the denominator (faster at what — AI R&D, model releases, benchmark progress?). Its value is as a comparand: the corpus's other quantified acceleration estimates are Kwa's ~2.5× serial researcher uplift back-solved from Anthropic's 8×-code figure, and the AI Futures Project's 1.0-to-2.6-year AC-to-superintelligence takeoff intervals. Brown's 3× and Kwa's 2.5× are measuring different things — a forecast of the accelerated regime versus an estimate of the current one — but they are the same order of magnitude, and both sit two orders below the hyperbolic case this page's third regime describes. That is the first time a frontier-lab insider's number and an outside estimate have agreed about the order of magnitude in this corpus, and it is the strongest available argument that the near-term regime is exponential-with-a-faster-constant rather than hyperbolic.

He also supplies the epistemic caveat himself, and it is the honest frame for all of the above: "I don't know what the world looks like in 2030. That's the truth."

The measurement problem underneath every acceleration number (September 2026)#

The same interview contains the clearest statement in the corpus of why the acceleration figures this page depends on are so hard to pin down, and it is a definitional problem rather than an instrumentation one. Brown names three confounds in sequence:

• Attribution. "If it's the human directing the AIs to do the work, how much do you attribute to the human? How much do you attribute to the AI?"
• Jaggedness plus substitution. The models are "exceptionally good at looking over data sets and checking every single data point" so "you can disproportionately use the AIs for those things" — and "of course, if something is suddenly 100x faster and 100x better, you're going to do more of that thing." The task mix moves in response to the capability, so a before/after comparison is not comparing the same work.
• And the direction of the counterfactual changes the answer. "Are you comparing it to a speedup of three years ago? Is the question more, 'Given what we were doing three years ago, how much faster are we able to do it now?' versus 'Given what we're doing now, how much slower would it have been three years ago?' Those are actually two very different questions."

That last distinction is the one this corpus has not been making. Every uplift figure it holds — the 8×-code number, Kwa's back-solve, the internal-acceleration curves — implicitly picks one of the two counterfactuals and none of them says which. Brown's own usable datum in the same passage is an input measure rather than an output one, which sidesteps the problem by not claiming to measure productivity at all: OpenAI's internal-acceleration post reports its top 1% of researchers spending $7,000–8,000 a day on Codex as of early August, "on an exponential." A spend curve is checkable-in-principle and says nothing about what the spend bought — which is exactly the activity-versus-capability confusion this page records elsewhere, arriving this time in a form its author does not mistake for a capability claim.

Takeoff length as a policy variable (August 2026)#

The sources above argue about the shape of the curve. The AI Futures Project's pacing proposals (2026-08-05, practitioner-opinion) are the first in the corpus to put numbers on the takeoff interval and then treat it as something a regulation moves. Their model reports Automated Coder arrival and the AC-to-superintelligence takeoff under two forecasters' median parameters — Daniel Kokotajlo's and Eli Lifland's — for a no-intervention baseline and for each proposed intervention.

Scenario	Kokotajlo medians: AC / takeoff	Lifland medians: AC / takeoff
No intervention	2028.3 / 1.0 yr	2029.5 / 2.6 yr
Compute cut to one fifth, applied at the no-cut AC date	2028.3 / 2.3 yr	2029.5 / 5.0 yr
Compute cut applied immediately (2026.75)	2029.4 / 1.6 yr	2031.1 / 3.7 yr
9-month AI-R&D capability lag from 2026.75	2028.4 / 2.4 yr	2029.7 / 4.1 yr

Two results matter here and neither is stated in the post's prose — both are read off the figure legends.

• Takeoff extension is non-monotonic in intervention timing, and peaks when the intervention lands on Automated Coder. Cutting compute earlier or later than AC extends takeoff less, on both parameter sets. Adding AC arrival to takeoff length (wiki arithmetic, not the authors') gives the year superintelligence arrives, and on that metric the ranking inverts — the earliest cut wins on both. So "maximize takeoff length" and "maximize total delay" are different objectives with different optima, and the proposal's preference for the former rests entirely on the claim that time late in the explosion is worth more than time early, because "we care most about slowing down late into the intelligence explosion… you have smarter AIs, so the time is much more valuable."
• A capability lag buys takeoff without buying delay. Forbidding AI R&D on models younger than nine months, applied immediately, roughly doubles-to-quadruples takeoff length while moving Automated Coder by 0.1-0.2 years. That is the cleanest instance in the corpus of an intervention that acts on the curve's slope in the explosive region rather than on its start date.

Treat the numbers as calibrated forecasts, not measurements: the model, the priors and the proposal share authorship, and the whole exercise sits at practitioner-opinion. What it contributes is that the exponential-vs-hyperbolic-vs-S-curve question above has a policy dual — the interval from AC to ASI is the quantity every proposal on this trajectory is actually trying to buy.

The first longitudinal self-modification dataset measures the wrong axis (August 2026)#

Everything above argues about curve shape from theory, forecast, or an unrun extrapolation. Ouroboros/Hope (Razzhigaev et al., arXiv 2608.08311, case-study) is the corpus's first source that could have supplied an actual early-onset self-improvement series: 161 continuous days of a deployed agent rewriting its own core, 1,085 self-modification commits at 94.2% agent-authored, monthly datapoints published throughout. That is the data shape this page's first open question asks for.

It is measured on activity, not capability. Figure 6 — the paper's only time series — plots four cumulative deployment metrics at monthly endpoints Feb→Aug 2026: model spend ($110.6K), tokens (79.7B), published code size (175,755 LOC), memory artifacts (227 MB). Spend and tokens grow steadily with mild deceleration and flatten in the final month; the two artifact series are step-shaped, and published LOC declines slightly at the end. There is no benchmark re-run, no task-success rate, no capability metric of any kind over the deployment. Worse for the extrapolation: the paper's benchmark scores were produced on frozen seeds with self-evolution disabled, so the document contains no experiment anywhere linking self-development to a score.

Two things this contributes, and they are both about what such a dataset has to contain rather than about the curve:

• Growth in scale is not growth in capability, and the two are trivially confusable when only one is instrumented. A deceleration in spend or a plateau in published LOC says nothing about whether the loop's returns are diminishing — it may equally mean the system got more efficient, or that its remaining problems got harder. Without a capability curve beside the activity curve, every shape in Figure 6 is compatible with exponential, S-curve, and no improvement at all.
• One suggestive counter-signal, hedged. Lifetime, 1,085 commits from 1,522 reviewed self-edit attempts implies ~71% acceptance, against a stated recent review block rate of 63.5% (~37% acceptance). If that gap is real rather than an artifact of two undefined denominators, the fraction of self-edits that survive review fell substantially over the deployment — which is the direction diminishing returns would predict, and is the closest thing in the document to a bend.

Weight it as case-study with total author COI: one lineage, no control population (the paper's own Limitations say so), and counters self-reported by the system under study.

Connections#

• Continuous Self-Modification Under Review — the first deployed, unbriefed, continuously self-modifying agent in the corpus and the closest thing to an early-onset RSI dataset: 161 days, 1,085 self-modification commits, monthly published datapoints — all of them measuring spend, tokens, code volume and memory artifacts rather than capability
• Domestic Frontier Pacing — takeoff length as a policy variable: quantified AC-to-superintelligence intervals (1.0 and 2.6 years at full speed) and the amounts different pacing interventions extend them, plus the non-monotonic result that takeoff extension peaks when the intervention lands on Automated Coder
• Recursive Self-Improvement — the mechanism; this page is its growth-dynamics complement (Anthropic's When AI builds itself is the sibling source)
• AGI-to-ASI Pathways — recursive improvement is pathway 3; the report treats its dynamics as the largest forecasting uncertainty
• Effective Compute Scaling — exponential compute growth is the substrate a recursive loop bends upward; data-as-RSI overlaps the data-wall counters
• Multi-Agent Collective Intelligence — cooperative/sociogenic RSI is collective specialization compounding
• Advantages of Digital Intelligence — high-bandwidth memetic/cultural evolution (lossless experience-sharing) is one engine of super-exponential self-improvement
• The Abstraction Barrier — the embodied bottleneck converts a would-be hyperbolic loop into one paced by empirical science
• Fundamental Limits of ASI — Gödel/physics/real-time limits are what bound the explosion
• AI-Driven Formal Proof Search — FunSearch/AlphaEvolve are concrete algorithmic self-improvement (LLM-guided program search finding novel constructions)
• Task Time-Horizon Scaling — the trendline that makes "is the curve bending?" empirically checkable
• AI R&D Autonomy Evaluation (AECI) — measuring the extent of AI-R&D automation (Chan et al. 2026; Anthropic's AECI) is the empirical input a "recursive improvement scaling law" would need
• Artificial Superintelligence (ASI) — whether the AGI→ASI transition is smooth or explosive is what these growth regimes decide
• The Data Wall and the Validation Commons Are One Supply Constraint — mechanism 4 (data) costed against the corpus's measured side, and the answer narrows it: AI curating and generating the next generation's data is cheap in FLOPs and rationed by verifier existence, verifier latency, and generator diversity collapse. So the data engine is not a compute-paced accelerant at all — it runs at the speed of the verification substrate, which puts it on the same side of the ledger as the embodied bottleneck rather than opposite it
• Large-Scale Test-Time Compute — Brown's mechanism for a time-bottlenecked (gradual, not hyperbolic) takeoff: peak capability requires long inference runs, so wall-clock binds even for pure-inference research
• Noam Brown — the practitioner source for the test-time-compute pacing argument, and for its September 2026 displacement: same anti-hyperbolic verdict, serial-experiment-and-GPU brake instead of inference wall-clock, and a hedged 3× as his median for the accelerated regime
• AI Accelerating AI Development — where Brown's three confounds bite: every uplift figure in the corpus implicitly picks one of his two counterfactuals ("how much faster now" versus "how much slower then") and none of them says which, while the task mix shifts toward whatever the models got good at
• Evaluation Horizon Versus Release Cadence — the same wall-clock argument relocated from the researcher to the evaluator: if long runs are where capability lives, the party who runs out of calendar first may be the one checking the model rather than the one improving it
• Researcher Uplift from Code Output — supplies the labor-side coefficient for the growth question: Kwa's ~2.5× serial researcher uplift plugged into Greenblatt's R&D ≈ labor^0.55 × compute^0.45 gives ~1.77× R&D speedup today, and with compute tripling yearly both inputs grow exponentially — the semi-endogenous case for sustained (not plateauing) research-output gains
• Balance-of-Power Superintelligence — the explosion scenario stated by a builder: Zuckerberg's manifesto concedes a self-improving system could "squeeze 100x or more intelligence out of each gigawatt" and command more effective compute than everyone else combined, then answers it with a compute-allocation ratio rather than a capability limit
• Transformative Creativity — the same fixed-model-plus-search datum read as a ceiling on what can be conceived rather than on what can be scored: an automated discovery loop searches a conceptual space it is handed, so its reach is bounded by that space's contents, and the ideas an LLM adds to one are all pre-existing techniques

Open Questions#

• Can "recursive improvement scaling laws" be formulated — predicting self-improvement curves (and their plateau point) from early-onset datapoints? Partially answered, in the negative (2026-08-12): Ouroboros/Hope supplies the corpus's first early-onset dataset from a real self-modifying deployment — 161 days, 1,085 self-modification commits, monthly published datapoints — and every series in it is deployment activity (spend, tokens, published LOC, memory artifacts), not capability, while its benchmark scores are single frozen-seed snapshots taken with self-evolution disabled. So the question is not advanced on the fitting problem; what moves is the specification, which now has a concrete instance: a usable dataset needs the capability curve beside the activity curve, and no published deployment has produced one.
• How far can a fixed model's performance be pushed with test-time search alone, and under what conditions does recursive distillation degenerate vs. compound? Partially answered on the first half (2026-08-12): Idea Search (empirical) holds the model fixed (Gemini 2.5 Pro) and varies only the search, over 2,000 nodes on scRNA-seq batch integration. Unguided Tree Search saturates early — ~300 nodes at 0.678 ± 0.011 — and the ceiling moves when the search's input space is enlarged rather than when the model changes: seeding it with an explicit bank of ideas decomposed from ten expert methods pushes saturation to ~500 nodes and the mean to 0.697 (best single solution 0.728). Two qualifications make it a small datum rather than a result: the ~0.02 gain is on the order of the 0.008–0.018 trial-to-trial spread, and enlarging the space is not sufficient by itself — adding 49 LLM-brainstormed ideas helped the bandit sampler (0.692 → 0.703) but hurt uniform sampling (0.712 → 0.698), and raising the sampler's own exploration coefficient α from 1 to 4 lowered the mean (0.712 ± 0.012 → 0.703 ± 0.008). So a fixed model's search ceiling is set jointly by the space and the selector, and turning the exploration dial up naively lowers it. Nothing here bears on the recursive-distillation half.
• Which binds first — algorithmic ceilings, the embodied bottleneck, or compute/energy supply — determining exponential vs. hyperbolic vs. S-curve? Partially answered: RSI Growth Curves: Which Friction Binds First? — both this report and Anthropic's locate the binding constraint outside cognition (the slowest un-acceleratable step coupling the loop to reality); the embodied bottleneck re-paces rather than halts, data-wall/research-harder demote into compute, and the abstraction barrier is the one candidate fundamental blocker. Retagged #oq/now → #oq/source 2026-08-10: the candidate is named but unranked, and ranking it needs external evidence rather than further synthesis.

Sources#

• From AGI to ASI — Section 2 ("Is the Singularity near?"), Section 5.3 (recursive self-improvement, four mechanisms), Section 7.1 (research agenda item 4)
• Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown — Noam Brown (No Priors, 2026-06-26), practitioner-opinion: test-time-compute dependence makes time the takeoff bottleneck; "gradual takeoff," no overnight explosion
• Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution — Razzhigaev, Gritsaev, Kaznacheev, Dragunov, Yampolskiy & Kuznetsov (MSU / Skoltech / Joi Lab / AIRI, arXiv 2608.08311, 2026-08-08, case-study — downgraded from empirical at compile): Figure 6 (the four monthly activity series, read from the page image per the two-pass rule — its shapes and the final-month LOC decline appear nowhere in the text), Table 4's deployment counters, and Appendix C's evolution off scaffold disclosures. Total author COI; one lineage, no control population. Parse warnings and full treatment on Continuous Self-Modification Under Review
• Noam Brown – Agent swarms, alignment, & recursive self-improvement — Noam Brown (OpenAI) with Dwarkesh Patel, Dwarkesh Podcast, 2026-09-17 (practitioner-opinion). §00:22:02 "What math progress tells us about recursive self improvement": the serial-experiment and GPU brake, the 100×-less-compute counterfactual, the 3× median with its 50%-to-10× range, the two-counterfactuals distinction, and the Codex internal-spend figure. First-party throughout and hedged by the speaker — "I could totally be wrong. I admit that", "if you put a gun to my head and ask me for a number" — with no model, no denominator for "faster", and no published version of the internal-acceleration post in raw/. Full treatment of the June→September delta on Noam Brown
• Idea Search: Guiding Tree Search with Ideas to Explore Diverse Scientific Methods — Wang, Cui, Brenner & Venugopalan (Caltech / Google Research / Harvard), arXiv 2608.08958 (2026-08-09, empirical): §5.1–5.3, cited here only for the fixed-model search-ceiling datum annotated on the second open question. Zero tables in the source; all figures quoted from prose, Figure 5 verified at 4× under the two-pass rule. Effect sizes sit within the trial-to-trial spread — carried as direction, not magnitude. Full treatment on Transformative Creativity
• How to pace the US frontier — AI Futures Project, 2026-08-05 (practitioner-opinion): the AC/takeoff figures from §"Effect of compute allocation minimums", §"Limit capabilities of models used for AI R&D" and §"When to start pacing the frontier". All values are transcribed from figure legends under the image two-pass rule, not from prose; the non-monotonicity and the AC-plus-takeoff sums are wiki-derived. Same-authorship COI: the AI Futures Model and one of the two parameter sets are the proposers' own. Full treatment at Domestic Frontier Pacing