Vals AI measures how close the frontier models are to doing AI research alone, and it now puts that point at August 2027. The date comes from the company doing the measuring rather than from a lab with a release to promote, which is the reason to read it carefully.
Rayan Krishnan, co-founder and CEO of Vals AI, told Bloomberg Tech that models will pass the researchers who build them by August 2027. Vals AI is an independent evaluator: it scores frontier models on real tasks and publishes the results, and one of those scores is aimed at exactly the capability he is forecasting.
What is being forecast
The claim is not about benchmark scores. Krishnan is forecasting full recursive self-improvement, RSI for short: a model that autonomously develops its own next version, with no researcher taking part in the decisions. 🔁
The illustration he uses is plain. GPT-6 builds GPT-7 on its own, and Claude builds its own successor.
What full RSI takes away from people
today AI takes part in building new models, people make the key decisions
full RSI hypotheses -> experiments -> training the next model, all run by the model
example GPT-6 builds GPT-7, Claude builds its own successor
forecast August 2027, Vals AI's projection, not a confirmed date
AI is already inside the process at every large lab. It writes and runs a good share of the work, and a person still makes the decisions that matter: which hypothesis gets tested, which experiment counts, which run becomes the successor. Full RSI is the name for the state where all three of those decisions belong to the model.
That is why a higher score on a test does not qualify. A score measures output. The forecast is about who decides.
How firm the date is
Not firm. The month is Vals AI's projection, made by Krishnan's team, and no lab has committed to it. The press version of the quote adds "or sooner", which hedges in the other direction and still leaves it a forecast.
Where Vals AI's own index puts things
Vals AI keeps an RSI Index for this capability. Models run autonomous research tasks in language-model development, under fixed compute and time, and each task is scored on a log scale: 0 is the task's starting baseline, 0.5 a published, human or frontier-model reference, 1 the theoretical best.
The index leader as it stands is Claude Fable 5.1 at 35.03%. Vals' reading of the results is that the models are strongest at running experiments and correcting misleading measurements, and still lack the judgment to identify higher-leverage directions. 🧪
Read those two Vals statements together. The leaderboard says the missing part is judgment about direction. The forecast says the missing part shows up within about a year.
Who this is for
Researchers, the people who plan headcount for research, and anyone who has filed recursive self-improvement under distant. The month is the least reliable fact here and the definition is the most reliable one.
Does the gap between a 35.03% index score and a model choosing its own research direction close in a year, or is the month doing more work than the measurement supports?
The interview: https://www.youtube.com/watch?v=MpYAoufS588
The RSI Index: https://www.vals.ai/benchmarks/rsi_index
Follow me for more on AI, LLMs, and Software Development:
@khasky — LinkedIn / Patreon / GitHub / Bluesky / Mastodon
@khaskydev — X / Threads / Instagram / Pinterest / Facebook
Top comments (0)