AI companies focus on AI-for-AI-research

Last updated

Tracker summary: Frontier AI labs increasingly use AI systems to accelerate their own AI research and development.

At a glance

  • Assessment: Confirmed
  • Confidence in assessment: 95%
  • Predicted timing: 2025–2026
  • Primary source: Late 2025: The World’s Most Expensive AI, archived PDF p. 4

What AI 2027 Predicted

Tracker summary: The scenario describes a critical feedback loop: AI labs use their own AI systems to do AI research, creating a recursive improvement dynamic. This is portrayed as a key accelerant in the path toward superintelligence.

Scenario source locator: Late 2025: The World’s Most Expensive AI: archived PDF, p. 4.

How We Track This

We monitor:

  • Public announcements of AI-for-AI-research programs at frontier labs
  • Academic publications on automated ML research
  • New benchmarks for measuring AI R&D automation
  • Reports from lab employees about internal AI usage

Current Evidence

Public evidence documents AI use inside multiple frontier research organizations. Karpathy’s AutoResearch repository, created March 6, 2026, demonstrates a lightweight single-GPU autonomous experiment loop. PostTrainBench, first submitted March 9, 2026, was introduced to measure automated LLM post-training.

AI Discovering Novel Algorithms: Google DeepMind unveiled AlphaEvolve (May 2025), a Gemini-powered coding agent that autonomously discovers new algorithms. Key achievements: a 4×4 complex matrix multiplication method beating a 56-year-old mathematical record, and a datacenter scheduling heuristic that recovered 0.7% of Google’s global compute resources. This is a concrete, high-profile example of AI directly improving AI infrastructure: a contribution to the mechanism, without quantifying broader recursive acceleration.

Lab Automation Timelines: Sam Altman announced (Oct 2025) that OpenAI has set an internal goal of “an automated AI research intern by September 2026” running on hundreds of thousands of GPUs, with “a true automated AI researcher by March 2028.” This is the most explicit public timeline from a leading lab for automating AI research. OpenAI said in September 2026 that it had reached the intern goal; the definition and measurement are OpenAI’s own. See OpenAI’s September 6 company account and the September 29 DevDay statement reported by The Next Web.

Frontier-Lab Internal Use: METR’s 2026 Frontier Risk Report provides unusually direct cross-lab evidence that AI-for-AI-research is routine inside frontier developers. Participating companies reported broad internal AI assistance for training data, eval frameworks, algorithms, infrastructure, experiment diagnostics, model monitoring, and security work. This is stronger than product-launch evidence because it comes from an external evaluator’s structured process with multiple frontier labs.

On July 15, 2026, OpenAI disclosed GPT-Red, an internal automated red-teaming model trained through self-play against defender models. OpenAI says it directly incorporated GPT-Red into GPT-5.6 training, while precursor red-teamers supplied training data for each production model since GPT-5.3. In a replicated indirect prompt-injection arena with held-out scenarios, OpenAI reported successful attacks in 84% of scenarios for GPT-Red and 13% for human red-teamers. This is a direct AI-for-AI safety-training loop, although the result is first-party and narrower than general research automation.

On July 30, 2026, OpenAI also reported a production example of AI improving the systems used to serve AI. Within a human-led process, GPT-5.6 Sol rewrote inference kernels and ran hundreds of optimization experiments that OpenAI says reduced end-to-end serving cost by 20% and increased token-generation efficiency by more than 15%. This supports the AI-for-AI mechanism without establishing a general R&D multiplier.

On August 21, 2026, NVIDIA reported that its AVO agent explored more than 500 attention-kernel optimization directions over seven days and produced kernels that beat FlashAttention-4 by up to 10.5% on evaluated DGX B200 configurations. This is a direct AI-for-AI engineering mechanism because the agent improved low-level software used in AI computation. It remains a vendor-run systems result and does not measure a lab-wide R&D multiplier.

Anthropic’s September 17 measurement report says Claude led 26% of its measured internal AI R&D work as of August, while no measured subset was fully autonomous. It also reports about 30,000 concurrent research and engineering agents on Anthropic’s most-used internal platform. A separate report says a supervised internal research model optimized more than 30 open-source biomolecular models in just under four weeks, averaging roughly 4× faster execution with a small precision tradeoff and nearly 2× with identical outputs. These are participant measurements and bounded engineering results, not an independent estimate of a lab-wide R&D multiplier.

Sources:

Counterevidence & Limitations

  • Many documented gains concern bounded research and engineering tasks; their contribution to complete programs is harder to measure.
  • These examples do not establish how much overall capability progress was caused by AI-generated research.
  • It’s difficult to measure the actual contribution of AI tools to research productivity. Labs have strong incentives to overstate AI’s role in their research process for marketing and fundraising purposes.
  • The prediction is broad enough that almost any internal AI usage qualifies. A more demanding interpretation (that AI is making the key intellectual contributions to AI research, not just accelerating execution) would rate as “emerging” at best.

What Would Change Our Assessment

  • Maintain “confirmed”: Documented AI-for-AI research use continues across multiple organizations.
  • Upgrade confidence: Evidence that AI-generated research insights lead to capability jumps that wouldn’t have happened otherwise
  • Assess stronger claims separately: Autonomous research judgment and recursive acceleration belong to dedicated research milestones.

See Also

Update History

DateUpdate
2026-11-07Assessment reviewed. Status: Confirmed; confidence: 95%. METR’s May report documents research and engineering use across Google, Anthropic and OpenAI. This meets the use-in-AI-research claim, not autonomous discovery or a measured research-speed multiplier. Dated the primary evidence, attributed the research-intern milestone to OpenAI, and removed an unverified engineer quotation and overbroad deployment claim.
2026-09-21Anthropic reported Claude leading 26% of measured internal AI R&D work, about 30,000 concurrent agents on one internal platform, and a supervised four-week project optimizing more than 30 biomolecular models. The results strengthen the established AI-for-AI mechanism without measuring a lab-wide R&D multiplier; status and confidence unchanged.
2026-09-06Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations.
2026-08-24NVIDIA reported a seven-day AVO search that produced faster attention kernels for AI computation. This adds a direct AI-for-AI engineering mechanism, while leaving lab-wide research productivity unmeasured. Confidence remains at the 0.95 cap.
2026-08-03OpenAI reported that GPT-5.6 Sol optimized production inference kernels and experiments, reducing serving cost by 20% and increasing token-generation efficiency by more than 15%. This adds a direct AI-for-AI engineering mechanism. Confidence remains 0.95 because the result is first-party and narrower than general research automation.
2026-07-20OpenAI disclosed that GPT-Red generates attacks used to adversarially train production models, including GPT-5.6. This adds a direct AI-for-AI safety-training mechanism. Confidence remains 0.95 because the result is first-party and narrower than general research automation.
2026-05-25METR’s Frontier Risk Report reported broad AI use inside participating frontier labs for training data, eval frameworks, algorithm implementation, infrastructure code, experiment diagnostics, model monitoring, and security work. This provides direct cross-lab evidence that AI-for-AI-research is now a routine frontier-lab workflow. Confidence adjusted 0.90 → 0.95.
2026-04-20Anthropic’s “Automated Alignment Researchers” study used nine Claude Opus 4.6 copies with shared tools, code storage, a forum, and scoring infrastructure to run 800 cumulative hours of alignment research on weak-to-strong supervision. Starting from a 0.23 human-tuned baseline, the system reached a 0.97 performance-gap-recovered score on the target setup (Anthropic). The task is narrow, but it is unusually direct evidence of frontier models autonomously generating, testing, and iterating on alignment ideas.
2026-03Practice now industry-standard. Labs report AI-assisted code generation, experiment design, and paper review as routine workflows.
2025-12Anthropic, OpenAI, and Google all publicly confirm using AI systems to accelerate internal AI research and development.
2025-11METR evaluation of GPT-5.1-Codex-Max shows 8% success on hardest “AI R&D-relevant” tasks — 4x increase from GPT-5’s 2%. Trump Genesis Mission EO directs federal agencies to apply AI to 20+ science challenges. Both signal growing AI use in research, though 8% on hardest tasks shows autonomous AI research contribution remains limited.
2025-10Sam Altman announces goal of “automated AI research intern by September 2026.” Most explicit public lab automation timeline.
2025-08Anthropic study: 67% increase in merged PRs/engineer/day using Claude Code internally. AI accelerating its own creator’s R&D.
2025-05Claude Code goes GA (May 22) with enterprise customers including AI-adjacent firms. Anthropic reports internal use of Claude Code for its own model development. 5.5x revenue growth by July indicates rapid uptake. Early signal that AI coding tools are being used in the R&D pipeline that produces AI systems. Google DeepMind launches AlphaEvolve — Gemini agent discovers novel algorithms, recovers 0.7% of Google’s global compute.