Will AI fully automate AI research and engineering (achieve 'full recursive self-improvement', full RSI) by the end of the given year?

rationale:

The Resolution Bar and Status Quo As of September 2026, frontier labs are deeply into partial automation but have demonstrably not closed the AI R&D loop. The resolution criteria are unusually strict and conjunctive: top-human-level novel ideation, plus end-to-end execution, plus no meaningful human contribution sustaining algorithmic progress. While OpenAI reached an "automated research intern" milestone and Anthropic reports >80% of merged code is authored by Claude, human oversight remains heavily required 2 sources. More than half of OpenAI's successful 4–8 hour agent tasks still require human intervention takeoff.watch, and Google’s Co-Scientist is explicitly framed as an assistive partner deepmind.google. Frontier AI meaningfully accelerates internal work, but a fully autonomous loop remains unrealized.

The Binding Bottleneck of Research Taste The primary obstacle is not raw coding ability, which has seen rapid exponential gains, but rather "research taste" and open-endedness. Current frontier models suffer a pronounced capability gap in choosing viable research goals, identifying dead ends, and evaluating results anthropic.com. Empirical benchmarks support this: Anthropic’s TASTE benchmark shows the best models at 60% versus 77% for human researchers alignment.anthropic.com, and a Princeton-led study on NeurIPS-caliber work found agents were "rote, formulaic," and unable to backtrack from unpromising approaches technologyreview.com. Furthermore, Epoch AI's EBR-bench playthroughs reveal a lack of improvement across iterations, casting doubt on the assumptions required for a rapid software-intelligence explosion epoch.ai.

Rapid Trends vs. Governance and Supervision Drag Bullish indicators rely on extremely fast subcapability trajectories—such as METR's calculated 89- to 131-day doubling times for task-completion horizons metr.org—and aggressive internal lab targets. However, even OpenAI’s target for March 2028 explicitly focuses on an automated researcher under human supervision, a critical rung below this question’s "no meaningful human contribution" bar openai.com. Crucially, safety frameworks and recent governance moves present deliberate friction. In mid-2026, calls from top executives (Amodei, Altman, Hassabis) and the "Pacing the Frontier" letter signed by over 1,300 lab employees signaled an intent to deliberately slow capability deployment and keep humans in the loop to ensure alignment 3 sources. This drastically reduces the likelihood that labs will openly deploy and verify a purely autonomous loop soon, even if latent capabilities emerge.

Horizon Synthesis (2030 vs. 2035) By the end of 2030 (24%), full RSI is plausible but remains a distinct minority outcome. While "automated coder" capabilities will likely be achieved, completely removing human agenda-setting and bridging the research-taste gap in four years requires navigating severe technical and institutional headwinds. Structured expert elicitations, such as FutureSearch, assign only a ~5% chance of a top developer claiming full automation before 2032 futuresearch.ai. However, by 2035 (55%), the probability tips into slightly-likely territory. Five additional years of compute build-out, data accumulation, and next-generation architectural shifts (e.g., algorithm-discovery systems like AlphaEvolve) make bridging the open-endedness gap highly credible 2 sources. Still, it is far from certain; diminishing returns on scaling, severe energy constraints, or a persistent, gradual transition where humans permanently retain a supervisory role keep the 2035 likelihood tightly bounded around a coin flip epoch.ai.