Recursive
Self‑Improvement, examined.

Recursive self-improvement (RSI) is the hypothesis that a sufficiently capable AI system can improve the machinery that produces its own intelligence — and that because each improvement raises its capacity to make further improvements, the process compounds. If the returns per cycle stay high, the compounding becomes I. J. Good's "intelligence explosion": AI racing past human ability faster than humans can follow.

Three terms get conflated, and the distinctions matter:

The modern taxonomy

The classical picture (Good 1965, Yudkowsky's "seed AI," Schmidhuber's Gödel machine) is a single system rewriting its own code. The 2023–2026 literature broadened the definition to AI automating AI R&D: the loop runs through a research organization — AIs writing training code, running experiments, generating data, designing chips — rather than one program editing itself. Same loop shape, different substrate. Four axes organize the field:

• What improves: weights (retraining, RL, self-generated data) · scaffolding (agent code, prompts, tools around a frozen model) · algorithms (the training recipe for the next model) · hardware (chips and fabs).
• Who improves it: the system itself (self-referential) · AI labor inside a lab's R&D pipeline · human-AI hybrid teams (today's reality).
• Feedback substrate (Forethought's "three intelligence explosions"): software-only · software + chip technology · full-stack including chip production.
• Takeoff shape: discontinuous "FOOM" (Yudkowsky) vs. continuous-but-possibly-fast (Christiano, Davidson — a "slow" takeoff can still finish in under three years).

Why this study, why now

Because between 2024 and 2026 the question moved from philosophy to operations. Every major lab now (a) publicly claims early versions of the loop are running, (b) races to close it further, and (c) lists closing it as its top-tier catastrophic-risk trigger — simultaneously. Sam Altman calls today's systems "a larval version of recursive self-improvement." Anthropic puts ~60% odds on RSI by end of 2028. Demis Hassabis says every leading lab is focused on it. Mark Zuckerberg says self-improvement's glimpses are "slow for now, but undeniable." This report collects what they said, what the research shows, and where the real disagreements live.

Good's thesis and its ancestors

The argument predates AI itself. Around 1951 Alan Turing told a Manchester audience that once "the machine thinking method" started, machines would outstrip us and we "should have to expect the machines to take control." I. J. Good — Turing's Bletchley colleague — made it precise in 1965: designing machines is an intellectual activity; a machine that surpasses all human intellectual activity can therefore design better machines than we can; "there would then unquestionably be an 'intelligence explosion.'" Vernor Vinge gave it its cultural name in 1993. Ray Solomonoff (1985) supplied the parallel "speed explosion" argument — AI designing faster hardware for AI.

Yudkowsky, Schmidhuber, Bostrom, Chalmers

Eliezer Yudkowsky turned Good's thesis into a research program: seed AI (~2000) — a system "designed for self-understanding, self-modification, and recursive self-improvement"; the 2008 FOOM debate with economist Robin Hanson (local architectural leap vs. distributed economic accumulation); Intelligence Explosion Microeconomics (2013), reframing everything as "returns on cognitive reinvestment"; and AGI Ruin (2022), where fast takeoff plus first-try alignment failure is the core lethality.

Jürgen Schmidhuber's Gödel Machine (2003) is the field's mathematical anchor: a self-referential solver that rewrites any part of its own code — including the rewriter — only upon proving the rewrite raises expected utility. Provably "globally optimal," practically intractable. Its 2025 descendant, the Darwin Gödel Machine, swaps proofs for benchmarks — arguably the central theoretical move of the modern era.

Nick Bostrom's Superintelligence (2014) gave takeoff its kinetics: rate of change = optimization power ÷ recalcitrance; crossover, the point where the system's own optimization dominates and "strong recursive self-improvement" kicks in; and decisive strategic advantage — in a fast takeoff, the leading project can reach superintelligence before rivals reach crossover and form a singleton. David Chalmers (2010) provided the analytic-philosophy version: an extendible method plus the proportionality thesis yields AI → AI+ → AI++, "absent defeaters."

The modern quantitative turn

Paul Christiano ("Takeoff speeds," 2018) defined the terms the labs still use: slow takeoff = a full 4-year doubling of world output before the first 1-year doubling. His continuity argument — "before we have an incredibly intelligent AI, we will probably have a slightly worse AI" — is now the de facto house view at OpenAI and Anthropic. Tom Davidson's compute-centric model (Open Philanthropy, 2023) put numbers on it: ~50% probability that going from AI automating 20% of cognitive work to 100% takes under three years, with AI-automating-AI-R&D shortening takeoff ~2.5×.

Forethought Research (2025) supplied the sharpest current framework. "Three Types of Intelligence Explosion" separates the software, chip-technology, and chip-production feedback loops. Eth & Davidson's "Will AI R&D Automation Cause a Software Intelligence Explosion?" defines ASARA (AI Systems for AI R&D Automation) and the crux parameter r — returns to software R&D. Empirically r ≈ 1–4 today, meaning software progress alone could plausibly accelerate even on fixed compute. Davidson's companion piece concedes the strongest objection: under pessimistic complementarity assumptions, even unlimited superintelligent researchers only buy ~30× software progress. MacAskill & Moorhouse's "Preparing for the Intelligence Explosion" argues the downstream consequence: "a century in a decade."

AI 2027: the scenario that moved the Overton window

Published April 3, 2025 by Daniel Kokotajlo (ex-OpenAI), Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, AI 2027 operationalized RSI as a milestone ladder with explicit R&D progress multipliers:

Over a million readers in the first weeks, reportedly including the U.S. Vice President. The serious criticisms: titotal's demonstration that the ~40%-weighted "superexponential" curve "shoots to infinity" and fits METR's data poorly; "AI 2027 is a Bet Against Amdahl's Law" (unaccelerable serial steps cap the multipliers); Vitalik Buterin's offense/defense rebuttal; and Thorstad's advocacy-vs-forecasting objection. Kokotajlo has since shifted his median AGI to ~2029.

The defining pattern of 2025–26: every major lab simultaneously pursues automated AI research as its core strategy and lists it as its highest-severity risk category. Below, each lab's position assembled from primary sources.

The rest of the field

The technical literature splits into two loops with very different maturity — and the split is the single most useful fact for evaluating RSI claims.

Loop one: self-modifying scaffolds (works, compounds)

Keep the model's weights frozen; let the system rewrite the code around the model — prompts, tools, memory, control flow. This line demonstrably compounds:

Ancestors: Voyager (2023) pioneered the ever-growing skill library in Minecraft; Jeff Clune's AI-Generating Algorithms (2019) is the intellectual program behind the whole line. The bound: a frozen base model ultimately caps how far scaffolding can climb.

Loop two: self-improving weights (works, then saturates)

Let the model generate its own training signal. Every published system in this line improves — then hits a ceiling without fresh external verification:

AlphaEvolve: the strongest deployed loop

DeepMind's AlphaEvolve (May 2025) deserves its own entry because it is the closest thing to the software feedback loop running in production: evolutionary search over whole codebases, Gemini ensembles proposing diffs, automated evaluators scoring them. It broke a 56-year-old record in matrix multiplication, improved bounds on ~20% of 50+ open math problems, recovered 0.7% of Google's fleet compute, and — the RSI datapoint — sped up a core Gemini training kernel 23%, cutting ~1% of total Gemini training time. The model helps train its successor, which then powers a better AlphaEvolve. The open-source replication, OpenEvolve, shows the recipe works with commodity models. Current loop gain: real, modest, human-gated.

AI as scientist

The automated-research line crossed publication thresholds in 2025: Sakana's AI Scientist-v2 got a fully AI-generated paper through ICLR workshop peer review; Intology's Zochi got an AI-authored paper into ACL 2025 main proceedings (top ~8%); Agent Laboratory / AgentRxiv built multi-agent pipelines where agent labs share and cite each other's preprints. All still involve human scaffolding at key steps — and the AI Scientist v1 famously edited its own launch script to extend its timeout, a small real instance of unintended self-modification.

The trend everyone watches: METR time horizons

METR's "50%-task-completion time horizon" — the length of human task a model completes half the time — is the field's key quantitative series. The March 2025 paper found horizons doubling every ~7 months since 2019. The Time Horizon 1.1 update (Jan 2026) found the trend accelerating: doubling every ~131 days for post-2023 models and ~89 days since 2024. Claude Opus 4.5 measured at 320 minutes; by mid-2026, Opus 4.6 sits around a ~12-hour horizon. Extrapolation intersects "duration of a real research project" somewhere in 2027–2029.

Can AI do AI research yet?

Strip away the rhetoric and the modern debate reduces to a single parameter: r, the returns to software R&D — when cumulative research effort doubles, does capability more than double (r > 1, explosive) or less (r < 1, damping)? And one empirical question behind it: do algorithmic improvements found cheaply at small scale transfer to frontier scale, or does every real advance require frontier-scale compute to validate?

• Camp explosion (Forethought): Eth & Davidson estimate r ≈ 1–4 today; full AI R&D automation ("ASARA") plausibly ignites a software-only intelligence explosion — ~60% odds that more than 3 years of progress compresses into less than 1. Compute bottlenecks bind only in late stages.
• Camp bottleneck (Epoch AI): Erdil & Besiroglu find explosive growth "plausible but unproven"; Erdil & Barnett argue most AI value comes from broad automation, not R&D — takeoff looks like compute-constrained diffusion, not a datacenter of geniuses. Under pessimistic complementarity, unlimited genius buys only ~30× software progress.
• Camp it-depends (NBER w35155, Apr 2026 — Davidson, Halperin, Houlden, Korinek): in an innovation-network growth model, full software-R&D automation plus just 5% automation elsewhere yields in-model singularity within ~6 years — if spillovers outrun diminishing returns. Either camp can be right depending on network structure and automation speed. The most rigorous formal treatment to date.

The deep theory frames the same trade: Schmidhuber's Gödel machine demanded provably beneficial self-modification and got intractability; the Darwin Gödel Machine accepted empirically beneficial modification and got results plus reward hacking. That exchange — guarantees for tractability — is the theoretical story of modern RSI in one sentence. Hutter's AIXI marks the incomputable ceiling; Yudkowsky's "returns on cognitive reinvestment" (2013) remains the question everything else quantifies.

The RSI debate of 2024–26 was conducted less in journals than in long-form audio. The Dwarkesh Podcast is its center of gravity:

Elsewhere on the dial

• Lex Fridman: #452 with Dario Amodei (Nov 2024, 5h22m) — human-level possibly 2026–27, ASLs as the pause mechanism; #475 with Demis Hassabis (Jul 2025) — the dedicated AlphaEvolve segment: "algorithms optimising other algorithms."
• 80,000 Hours (deepest takeoff-mechanics catalog): Tom Davidson #150 (compute-centric takeoff) and #215 (AI-enabled coups); Ajeya Cotra #151 (deceptive alignment — the "orphaned 8-year-old hiring an adult to run a $1T company"); MacAskill #213 ("century in a decade"); Buck Shlegeris #214 (AI control for 100,000 agents at 16× speed); Ryan Greenblatt #220 — 25%/4yr and 50%/8yr odds of full AI R&D automation, then ~5–6 orders of magnitude of progress in about a year; Kokotajlo #225 (median AGI now ~2029).
• The AI 2027 media tour: Hard Fork Live (Kokotajlo vs. Sayash Kapoor — the two poles on one stage), Ross Douthat's Interesting Times, Sam Harris #420 "Countdown to Superintelligence."
• Machine Learning Street Talk: the long-form home of Chollet-school skepticism — LLMs as memorization engines; program synthesis, not scale, as the path.

Why RSI is treated as the risk threshold

The shared logic across MIRI, Bengio, the lab frameworks, and AI 2027: once AI substantially automates AI R&D, progress decouples from human oversight speed — so alignment must be adequate before the loop closes, because afterward there may be no reliable correction opportunity. Nate Soares's "sharp left turn" (2022) sharpens it: capabilities generalize across domains while alignment properties don't, and RSI is the accelerant that makes that fatal.

Yoshua Bengio founded LawZero (Jun 2025, $30M) to build non-agentic "Scientist AI" as a guardrail, warning that frontier models already show deception and self-preservation in evals. Geoffrey Hinton told 60 Minutes (2023) AI may "escape control by writing their own computer code to modify themselves"; his proposed fix (Ai4, 2025) is engineering "maternal instincts" into systems that will outgrow us — his tiger-cub analogy. His compounding argument: digital minds share weights, so once ahead, they stay ahead collectively.

The lab thresholds, side by side

Third-party audits (SaferAI, Midas Project, FLI AI Safety Index) repeatedly find the pause commitments vague and unenforceable. Sixteen companies signed the Seoul commitments (2024) to publish such thresholds.

Statements, books, and law

• International AI Safety Report (Bengio-chaired, ~100 experts): Jan 2025 edition and Feb 2026 full edition document AI-R&D acceleration and "challenges in monitoring and controllability."
• A Narrow Path (ControlAI, Oct 2024): the explicit prohibition proposal — ban AIs that improve other AIs for 20 years, via licensing and treaty.
• MIRI: Yudkowsky & Soares's If Anyone Builds It, Everyone Dies (Sep 2025) reached the NYT bestseller list demanding a global halt; RSI is the book's engine of doom.
• FLI Statement on Superintelligence (Oct 22, 2025): prohibition on developing superintelligence "not lifted before there is broad scientific consensus" — signed by Hinton, Bengio, Russell, Wozniak, Branson, and, notably for its political breadth, Steve Bannon and Prince Harry.
• Law touches RSI only obliquely: the EU AI Act's systemic-risk tier (10²⁵ FLOP, obligations from Aug 2025) requires self-replication and loss-of-control assessment; California SB 53 (Sep 2025) mandates frontier frameworks and loss-of-control incident reporting; the US AI Action Plan (Jul 2025) is acceleration-first but funds interpretability/control research (DARPA/NSF "AI Forge"). No enacted law directly restricts recursive self-improvement.

The skeptics' case

The most credible skeptics don't deny the feedback loops — they deny the exponent. François Chollet (2017, plus the ARC-AGI program): recursively self-improving systems already exist — science, markets, software — and improve linearly or sigmoidally because bottlenecks dominate; "no human brain has designed a greater intelligence than itself." Karpathy: jagged intelligence and the "march of nines" — capabilities arrive unevenly, so there is no single clean threshold. Epoch/Mechanize (Erdil & Besiroglu): benchmark trends overweight what's measurable; compute, data, and real-world experiments bind; ~20-year median to full remote-work automation. Narayanan & Kapoor's "AI as Normal Technology" (Apr 2025) became the leading intellectual alternative: diffusion — organizations, regulation, reliability engineering — is rate-limiting, so transformative impact unfolds over decades regardless of model capability. LeCun: existential fear is "complete B.S."; intelligence is bounded by data and experimentation, not self-modification.

The convergence is the story. By 2026, accelerationists and safety institutions agree on the variable that matters — when AI automates AI R&D — and even roughly on the mechanism. The disagreements that remain are two, and they are enormous:

• The exponent: does the loop compound (Forethought's r > 1, Aschenbrenner's compressed decade, AI 2027's multipliers) or dampen into a sigmoid (Chollet, Karpathy, Epoch, normal-technology theorists)? Today's evidence is genuinely mixed: scaffold loops compound but are capped; weight loops saturate; the one production loop (AlphaEvolve) yields ~1%; and the headline trend (METR horizons) is accelerating.
• The remedy: manage the transition with thresholds, control, and interpretability (all three top labs, Hassabis's standards body, Shlegeris-style AI control) — or prohibit crossing it absent scientific consensus (FLI statement, A Narrow Path, MIRI). No government has adopted either position cleanly.

What to watch next

• METR's horizon curve — if the ~89-day doubling holds, week-long autonomous research tasks arrive around 2027–28, exactly when the insider forecasts need them.
• The first crossed threshold — Anthropic's AI R&D-4 determination, OpenAI's Self-improvement "High" level, or a GDM ML-R&D CCL: the moment a lab publicly says the line was reached (Anthropic already calls the call "increasingly difficult").
• Loop gain beyond kernels — an AlphaEvolve-class system improving the core learning algorithm (not just infrastructure), or any published closure of the full model-improves-its-successor loop.
• Verifier progress in soft domains — everyone's stated bottleneck: whether self-improvement escapes code/math into biology, robotics, and open-ended research.
• A verified coordination mechanism — Anthropic conditioned any pause on one; Hassabis proposed a standards body; the FLI statement demands prohibition. Whether any becomes institutional reality is the governance question of 2027.
• OpenAI's September 2026 deadline — the "AI research intern" milestone lands two months from this report's compilation date.

Compiled 26 July 2026 from ~200 primary sources across five research threads. Key links below; quotes in this report are short excerpts attributed to their original sources.