Current State of AI Research Automation
As of late 2026, AI heavily assists with research implementation, but agenda-setting and experiment selection remain firmly in human hands. OpenAI reports that the Decide phase ("what to work on") accounts for only 0.3% of agent output tokens, noting that "people still set our research priorities" https://openai.com/index/research-acceleration-view-inside-openai/ [https://openai.com
This was pushed back slightly, as formally stating that AI selects most experiments would likely trigger severe lab safety limits, creating strong incentives to maintain human-in-the-loop framing.
As of October 2026, no top-five AI developer has publicly attributed headcount reductions in a named research family to AI automation. Past workforce cuts at Meta Meta to Lay Off 10 Percent of Work Force in A.I. Push Mark Zuckerberg had a bold plan to replace Meta staff with ... [Meta will cut 10% o
Set alongside other governance milestones, this was pushed slightly later due to the immense PR and legal friction labs face in explicitly attributing staff reductions to AI capability.
The Gap Between Throughput and Value No top-five developer has yet stated publicly that AI has doubled its overall rate of AI research progress. The core dynamic here is the difference between coding throughput and end-to-end research value. While Anthropic reports that Claude authors over 80% of merged code and OpenAI logs 3.14 agent-workdays per researcher workday [https://www.anthropic.com/institute/recursive-self-improvement](https://www.anthropic.com/institute/recursive-self-improvement
Set against related automation milestones, this date was shifted slightly later to account for the heavy governance friction and disincentives labs face in formally declaring a 2x R&D speedup.
The estimated percentile dates for a documented 10x efficiency gain within a 12-month window are 2028-01-01 (10th), 2029-06-01 (25th), 2031-12-31 (50th), 2036-01-01 (75th), and 2042-01-01 (90th). Initial capability frameworks, historical pre-training baselines of ~3x/year AI software progress: Data & research, and current lab transparency protocols provide the relevant context. The frontier currently lacks evidence of the requisite slope change, as co
Set against related milestones, this was pushed back slightly as a 10x single-year gain represents a massive sudden leap that is structurally less likely than gradual scale-up.
Status Quo and Resolution Criteria As of October 2026, no qualifying case has been published. Near-misses like Google's AlphaEvolve finding a matrix-multiplication tiling heuristic for Gemini AlphaEvolve: A Gemini-powered coding agent for ... [AlphaEvolve: A Gemini-powered coding agent for ...](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanc
The estimate was moved slightly later to better align with the 6-to-12 month product lifecycle lag between an AI generating a novel idea and a lab successfully validating and shipping it in a frontier model.
As of mid-October 2026, no record on the modded-nanogpt leaderboard is credited solely to an AI agent. The Track 1 world-record table runs to #92 (0.665 min), and while five records credit an AI system, every single one pairs it with a human handle (e.g., "@cong_ml and AI System Recursive") https://github.com/KellerJordan/modded-nanogpt/blob/master/README.md [https://github.com/KellerJordan/modded-nanogpt/blob/master/README.
Set against related early capability signals, this distribution was slightly adjusted to reflect the time needed for social norms to accept an AI-only attribution on developer leaderboards.
The binding constraint on this resolution is the availability of a capable measuring instrument, not raw AI capability. Extrapolating the historical 128.7-day doubling time from the current 17.4-hour ceiling suggests that frontier systems might reach 160-hour capabilities by mid-to-late 2027. However, nobody currently has a published, human-baselined suite capable of verifying month-long tasks. The immense practical difficulty of establishing such a suite, compounded by the severe score degradat
The timeline was extended slightly to account for the immense practical difficulty of establishing month-long human baselines and the severe score degradation imposed by strict cheating penalties.
The Capability Gap: Implementation vs. Design Current AI R&D automation is overwhelmingly focused on task execution and implementation, not high-level design. In August 2026, the Decide and Design phases accounted for just 2.8% of OpenAI's agent output tokens, with "what to work on" making up only 0.3%. The lab affirmed that human researchers still set priorities, judge results, and decide what to pursue [https://openai.com/index/research-acceleration-view-inside-openai/](https://openai.com/
Set against related milestones, this was shifted slightly later because "principally designed" represents a profound threshold that labs are heavily disincentivized from crossing or declaring.
Current Status As of mid-October 2026, no third party has published a periodic assessment containing a research speedup rate it measured itself. The first edition of METR's Frontier Risk Report (2026-05-19) relayed what companies self-reported rather than measuring the speedup directly https://metr.org/blog/2026-05-19-frontier-risk-report/. The closest precedent is a preliminary estimate of "~1.5X overall acceleration in capabilities
Set against related milestones, this was shifted slightly later to account for the methodological difficulty of isolating and publishing an independently measured rate, even with embedded access.
As of October 2026, no jurisdiction holds binding authority to set a minimum share of a frontier developer’s compute for non-capability purposes. Current governance efforts rely on voluntary frameworks or alternative regulatory levers. The White House Accord on Super Intelligence, signed in September 2026 by major developers, is strictly voluntary and contains no compute-allocation provisions, only noting that codifying its steps "may make sense" over time [White House Accord on Super Intelligen
This distribution remains positioned in the deep long tail, reflecting the extreme unlikelihood of governments adopting this highly specific, intrusive regulatory mechanism.
Current Status of the Proposal A commitment to a "negative internal-public gap"—deploying a model externally before using it for internal AI research—remains purely an advocacy proposal. The concept was formalized by the AI Futures Project in August 2026 How to pace the US frontier How to pace the US frontier, expanding on earlier suggestions that public release before inte
This was pushed further into the tail, as reversing the deployment gap forces developers to surrender massive strategic advantages, an outcome likely only under extreme regulatory duress.
I estimate roughly a 20% chance that a pacing letter reaches 5,000 frontier-lab employee signatories before the end of 2032. I place the 10th-percentile date at January 1, 2029 and the 25th-percentile date at June 1, 2033, while the median and upper percentiles fall much later: January 1, 2038 at p50, January 1, 2048 at p75, and January 1, 2065 at p90. Any plausible path to reaching the threshold within the next six years relies heavily on sudden, shock-driven surges rather than steady, incremen
Following the complete stalling of the current open letter, this was shifted later to reflect that a successful petition now requires a massive external shock to mobilize a saturated employee base.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited