Part of: this research — view all 12 rows
Current Status and the Resolution Bar
As of November 14, 2026, no qualifying case exists. The resolution criteria demand explicit public attribution by a top-five developer that an AI originated a training-method or architecture change that shipped in a released frontier model. Previous high-profile examples of AI-assisted engineering do not meet this bar. Google's AlphaEvolve 2 sources and its subsequent deployments 2 sources represent kernel optimization and hyperparameter tuning within a human-designed search space. Publicly, labs still frame AI strictly as a collaborator. OpenAI maintains that "people still set our research priorities" openai.com, while Anthropic recently affirmed that its models' advantages stem mostly from judgment rather than generating novel research ideas 2 sources.
Capability Trajectory
Despite this current lack of attribution, the underlying capabilities are advancing rapidly. Anthropic's internal capability index shows growth at 1.88× its earlier rate, and an independent assessment estimates a ~2× acceleration with roughly 80% attributed to AI metr.org. Importantly, the capability for an AI to discover novel methods already exists in narrower settings, as demonstrated by Google’s Lion optimizer arxiv.org and DeepMind’s DiscoRL nature.com. Furthermore, Anthropic announced on November 10 that it expects to imminently meet its Responsible Scaling Policy’s AI R&D threshold based on projected capability growth. Because the resolution requires only one incremental training-method or architecture change—not full end-to-end research autonomy—the remaining leap is primarily one of transfer to frontier scale and subsequent disclosure.
Pipeline Lag and the Attribution Bottleneck
Estimates must account for the structural lag between discovery and public release. Even if a model like Claude Fable 5.2 or Gemini 4 Carbon originates a qualifying idea today, that idea must be adopted, tested, integrated into a frontier training run, and eventually released with a published technical report. This pipeline typically takes 6 to 12 months. Furthermore, explicit public attribution is a formidable bottleneck. Labs heavily guard architecture and training details for competitive reasons. They also face safety and pacing incentives to downplay autonomous AI capabilities, preferring to describe improvements as achieved "with human guidance" to avoid triggering regulatory scrutiny or internal safety thresholds sec.gov.
Estimate Rationale
The early percentiles (February to July 2027) account for the possibility that a modest, qualifying post-training innovation is already in the pipeline, perhaps soon to be publicized via Anthropic's promised acceleration reports or its IPO narrative regarding "compute multipliers" simonwillison.net. The median of February 2028 aligns with the expected pipeline lag for current frontier models to generate adoptable ideas, and closely tracks OpenAI's stated goal of an automated researcher by March 2028. However, the long right tail (extending to mid-2031) reflects the substantial risk that labs persistently refuse to explicitly attribute core architectural changes to AI due to commercial secrecy, ambiguous human-AI provenance, or deliberate pacing pauses.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited