Explore the research

AI research

This is the closest link to the singularity idea: better AI making further AI progress possible.

31%

field summary estimate

7 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Valid discoveries30%
Successive improvement25%
Research task performance40%
Research autonomy30%
Experiment efficiency25%
Algorithmic efficiency40%
Acceleration30%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Acceleration

30%estimated progress

What we track

Elapsed time and resources per comparable validated capability increment across generations; report whether research gains persist, flatten, or regress.

Why this estimate?

Organizational acceleration supports useful AI-assisted research, but causal attribution, human management and continuation across generations remain unresolved.

Advancements

Anthropic survey: staff report 50% productivity gain from ClaudeDecember 2, 2025. A survey of 132 Anthropic engineers and researchers found they used Claude in 59% of their work, up from 28% a year earlier, and self-reported a 50% productivity gain, up from 20%. Adoption of Claude Code coincided with a 67% increase in merged pull requests per engineer per day.
METR: AI companies report no dramatic speed-up from AI R&D automationMay 19, 2026. METR's Frontier Risk Report found that companies did not report dramatic speed-ups in overall progress attributed to AI R&D automation, and that Anthropic explicitly argued it had not seen a 2x increase in the pace of progress as of April 2026. A METR trial of late-2025 public agents found productivity benefits of about 4 to 20%.
Anthropic: Claude writes over 80% of merged code; engineers merge 8x moreJune 4, 2026. Anthropic reported that as of May 2026 over 80% of the lines merged to production were attributable to Claude, and that the typical engineer merged 8 times as much code per day in Q2 2026 as in 2024. It said lines of code overstate the real productivity gain; a March 2026 survey of 130 employees gave a median estimate of about 4x output.
METR measures a 1.9× speed-up in Anthropic's AI progressNovember 5, 2026. METR estimated Anthropic's capability-index growth rate since October 2025 at 1.9 times its previous rate (95% CI 1.5–2.8).

Read the dated findings and sources

Status: Measured organizational proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 5, 2026: METR estimated Anthropic's capability-index growth rate since October 2025 at 1.9 times its previous rate (95% CI 1.5–2.8). METR: Anthropic AI R&D acceleration assessment.

Interpretation and tracking value

This is an outside assessment of one organization, not a controlled causal estimate of recursive self-improvement. Compute scaling was examined, but other causes and redacted evidence limit attribution.

Next comparable measurement

Freeze a capability target and measure elapsed days, compute and dollars for each increment; retain model generations and independent ablations. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Algorithmic efficiency

40%estimated progress

What we track

Compute, energy, and total dollars needed to reach a fixed independently evaluated capability target. Report training and inference separately.

Why this estimate?

A deployed kernel optimization shows measurable execution and automated evaluation. Humans still define the workload and capability target; general research efficiency and autonomous iteration are not established.

Advancements

AlphaDev finds sorting routines up to 70% faster, added to LLVM libc++June 7, 2023. Google DeepMind's AlphaDev discovered sorting algorithms up to 70% faster for short sequences and about 1.7% faster for sequences over 250,000 elements, which were added to the LLVM libc++ library. A hashing algorithm it found was about 30% faster for the 9 to 16 byte range and was released in Abseil.
Epoch AI: price of GPT-4-level inference falls about 40x per yearMarch 12, 2025. Epoch AI tracked models on six benchmarks over three years and found the price to reach GPT-4's performance on PhD-level science questions fell about 40x per year. Across performance milestones the annual price decline ranged from 9x to 900x.
AlphaEvolve speeds up a FlashAttention kernel by up to 32.5%May 14, 2025. Google DeepMind reported that AlphaEvolve achieved up to a 32.5% speedup for a FlashAttention kernel implementation used in Transformer-based AI models.
AI Index 2025: cost of GPT-3.5-level inference dropped over 280-fold2025 AI Index Report. Stanford's 2025 AI Index reported that the cost of running a system performing at GPT-3.5 level dropped over 280-fold between late 2022 and late 2024. It also reported hardware costs declining by 30% annually and energy efficiency improving by 40% each year.
AlphaEvolve speeds up a Gemini training kernel by 23%December 9, 2025. Google reports a 23% speedup in a Gemini training kernel, translating to about a 1% reduction in total training time.

Read the dated findings and sources

Status: Measured narrow optimization. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 9, 2025: Google reports a 23% speedup in a Gemini training kernel, translating to about a 1% reduction in total training time. Google Cloud: AlphaEvolve results.

Interpretation and tracking value

Developer-reported optimization of one workload; it is not a 23% reduction in training cost, energy, or compute needed for a fixed general capability.

Next comparable measurement

Repeat matched training to the same held-out quality; record FLOPs, wall time, joules and all-in cost separately from inference. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Experiment efficiency

25%estimated progress

What we track

Validated improvements per 100 experiments and per total dollar; include failed runs and researcher support.

Why this estimate?

Faster workflows suggest early gains, but the missing all-experiment denominator prevents credit for dependable campaign-level efficiency.

Advancements

Sakana's AI Scientist generates full research papers for under $15 each12 August 2024. The AI Scientist paper described a system that generates research ideas, writes code, runs experiments and writes up a full paper in three machine learning subfields, at a cost of less than $15 per paper.
AlphaEvolve improves best known solutions on about 20% of 50 open problemsMay 14, 2025. Applied to over 50 open problems in mathematics, AlphaEvolve rediscovered state-of-the-art solutions in roughly 75% of cases and improved the best known solutions in about 20% of cases.
AIRA agents raise MLE-bench lite medal rate from 39.6% to 47.7%3 July 2025. A study of search strategies for AI research agents reported that its best pairing of search strategy and operator set raised the rate of earning a Kaggle medal on MLE-bench lite from 39.6% to 47.7%.
METR documents faster AI research workflowsNovember 5, 2026. METR's assessment includes evidence of faster research workflows, but does not publish a standardized count of validated gains per 100 experiments.

Read the dated findings and sources

Status: Qualitative progress; required rate unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 5, 2026: METR's assessment includes evidence of faster research workflows, but does not publish a standardized count of validated gains per 100 experiments. METR: Anthropic AI R&D acceleration assessment.

Interpretation and tracking value

Private operational evidence cannot establish the project's all-in experiment-efficiency metric. Failed runs and support work need explicit accounting.

Next comparable measurement

Compare matched research campaigns; count all runs and independent improvements and divide by total dollars and researcher hours. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Research autonomy

30%estimated progress

What we track

Human hours and interventions from research question through implementation, evaluation, and reproducibility package.

Why this estimate?

AI performs useful research work, yet the assessment attributes most gains to assistance and supplies no complete record of human interventions.

Advancements

AI Scientist-v2 paper passes peer review at an ICLR 2025 workshop10 April 2025. Sakana AI submitted three fully autonomous manuscripts to a peer-reviewed ICLR workshop; one achieved scores high enough to exceed the average human acceptance threshold.
Nature publishes Sakana's AI Scientist; AI reviewer reaches 69% accuracyMarch 26, 2026. The AI Scientist was published in Nature. One AI-generated manuscript at the ICLR 2025 'I Can't Believe It's Not Better' workshop averaged 6.33 from reviewers (scores 6, 7, 6), higher than 55% of human-authored papers, and the system's Automated Reviewer reached 69% balanced accuracy against human decisions.
Anthropic: Claude agents design every experiment in an 800-hour projectJune 4, 2026. Anthropic reported that in an April 2026 AI safety project on weak-to-strong supervision, Claude-powered agents designed every experiment themselves over about 800 cumulative hours for roughly $18,000 in compute. They recovered about 97% of the performance gap, against about 23% for two humans over about a week; humans chose the problem and scoring rubric, and the result did not transfer cleanly to production-scale models.
METR finds AI assisting, not yet leading, model trainingNovember 5, 2026. METR judged current acceleration primarily to reflect AI assistance, without highly significant fully autonomous contributions to the most important model-training work.

Read the dated findings and sources

Status: Partial evidence; complete human-effort baseline unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 5, 2026: METR judged current acceleration primarily to reflect AI assistance, without highly significant fully autonomous contributions to the most important model-training work. METR: Anthropic AI R&D acceleration assessment.

Interpretation and tracking value

The assessment does not supply a complete per-project intervention count or human-hours denominator.

Next comparable measurement

Log every intervention from question selection through replication; report median human hours and autonomous completions/all assigned projects. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Research task performance

40%estimated progress

What we track

Held-out objective improvement on fresh research tasks at fixed time, compute, and data budgets, relative to expert researchers and automated-search baselines.

Why this estimate?

Multiple bounded research-engineering tasks support repeatable execution and objective checking. Budget-sensitive results and expert-selected tasks limit agenda and long-run autonomy credit.

Advancements

MLE-bench: o1-preview with AIDE reaches bronze level in 16.9% of contests9 October 2024. OpenAI's MLE-bench tests agents on 75 machine learning engineering competitions from Kaggle. The best setup, o1-preview with AIDE scaffolding, achieved at least a Kaggle bronze-medal level in 16.9% of competitions.
RE-Bench: AI agents outscore experts on two-hour research tasksNovember 22, 2024; revised May 27, 2025. RE-Bench includes seven research-engineering tasks and 71 eight-hour expert attempts. Agents scored about four times humans at two-hour budgets; humans outscored agents about twofold at 32 hours.
MLGym: frontier models tune baselines but produce no novel methods20 February 2025. MLGym-Bench evaluated five frontier models on 13 open-ended AI research tasks. The models could improve on given baselines, usually by finding better hyperparameters, but did not generate novel hypotheses, algorithms, architectures or substantial improvements.
PaperBench: best agent replicates 21.0% of ICML 2024 paper results2 April 2025. OpenAI's PaperBench asks agents to replicate 20 ICML 2024 Spotlight and Oral papers, graded on 8,316 individual tasks. The best agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieved an average replication score of 21.0%.
MLE-bench leaderboard: Famou-Agent 2.0 wins medals in 64.44% of contests2026-02-23 (leaderboard entry date). The official MLE-bench leaderboard lists Famou-Agent 2.0 with Gemini-3-Pro-Preview at 64.44 ± 1.18% of competitions with any medal, against 17.12 ± 0.61% for AIDE with o1-preview in October 2024. OpenAI has since paused new submissions while it develops a process to ensure they are fair and comparable.
Anthropic: Mythos Preview speeds up training code 52x on a test taskJune 4, 2026. Anthropic reported that on a training-code optimization task, model-achieved speedups rose from about 3x with Claude Opus 4 (May 2025) to about 52x with Claude Mythos Preview (April 2026), against about 4x for a skilled human in 4 to 8 hours. Anthropic said this is not a real-world training speedup.

Read the dated findings and sources

Status: Measured benchmark comparison. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 22, 2024; revised May 27, 2025: RE-Bench includes seven research-engineering tasks and 71 eight-hour expert attempts. Agents scored about four times humans at two-hour budgets; humans outscored agents about twofold at 32 hours. RE-Bench research paper.

Interpretation and tracking value

Results depend strongly on budget and task selection. They measure bounded engineering tasks, not the full process of discovering and validating research directions.

Next comparable measurement

Rerun fresh tasks with identical compute/data and 2-, 8- and 32-hour budgets; report task-level scores and uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Successive improvement

25%estimated progress

What we track

Change in the next system's ability to perform fresh AI research tasks caused by the previous system's contributions. Measure each generation, not just the final model.

Why this estimate?

Evidence of feedback into research is suggestive, but it does not isolate an independently validated sequence of AI-caused improvements across generations.

Advancements

Darwin Gödel Machine rewrites its own code, lifting SWE-bench from 20% to 50%29 May 2025. The Darwin Gödel Machine, a coding agent that iteratively modifies its own code and validates each change empirically, raised its SWE-bench performance from 20.0% to 50.0% and its Polyglot performance from 14.2% to 30.7%.
Research agent makes seven successive self-improvements in an 8-day run22 September 2026. A paper on recursive self-improvement of AI research agents reported seven successive improvements over an autonomous 8-day run, evaluated on four held-out benchmarks. Reward hacking fell from 55% to 32% during the run, and the resulting agent remained 7 percentage points below the human-engineered agent.
METR reports substantial recent research accelerationNovember 5, 2026. METR found substantial recent research acceleration, while emphasizing uncertainty in attribution and future continuation.

Read the dated findings and sources

Status: Feedback evidence; causal generation series unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 5, 2026: METR found substantial recent research acceleration, while emphasizing uncertainty in attribution and future continuation. METR: Anthropic AI R&D acceleration assessment.

Interpretation and tracking value

An organization-level trend does not isolate how each system's contribution improves the next system's research ability. No qualifying multi-generation causal series is established by this assessment.

Next comparable measurement

Evaluate at least three successive generations on fresh frozen research tasks, with and without inherited AI-produced improvements. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Valid discoveries

30%estimated progress

What we track

Fraction of AI-proposed improvements reproduced by an independent team across at least three seeds and two model sizes or settings.

Why this estimate?

Accepted internal optimizations merit execution credit. Missing independent replication across seeds and settings limits validation and transfer credit.

Advancements

FunSearch finds the largest cap sets in 20 years; published in NatureDecember 14, 2023. Google DeepMind's FunSearch, which pairs a language model with an automated evaluator, found the largest cap sets recorded in some settings, described as the largest increase in cap set size in 20 years, and produced bin-packing heuristics that outperformed established human-designed ones. The results were published in Nature.
Independent test of Sakana's AI Scientist: 42% of experiments fail20 February 2025. An independent evaluation of The AI Scientist found 42% of experiments failed due to coding errors, and generated manuscripts had a median of five citations, of which only five of 34 were from 2020 or later. Producing a full paper cost USD 6 to 15 and needed 3.5 hours of human involvement.
AlphaEvolve multiplies 4x4 complex matrices with 48 scalar multiplicationsMay 14, 2025. AlphaEvolve found an algorithm to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving on Strassen's 1969 algorithm, previously the best known in this setting.
AlphaEvolve optimizations go into use in Google's computingDecember 9, 2025. AlphaEvolve produced evaluated code optimizations used within Google's computing workflows.

Read the dated findings and sources

Status: Validated deployment example; replication fraction unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 9, 2025: AlphaEvolve produced evaluated code optimizations used within Google's computing workflows. Google Cloud: AlphaEvolve results.

Interpretation and tracking value

Internal acceptance is useful evidence but does not provide independent reproductions divided by all proposed improvements across three seeds and two settings.

Next comparable measurement

Publish all proposals, failures and artifacts; have an outside team reproduce each accepted improvement across the required seeds and scales. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note