Former Google DeepMind VP of Research Oriol Vinyals outlined at the Agentic AI Summit 2026 why AI recursive self-improvement (RSI) is unlikely to trigger a sudden intelligence explosion. While AI will accelerate certain engineering and research tasks by tenfold or more, Vinyals argues that overall progress remains constrained by core technical and evaluation hurdles. He announced he is launching a new startup specifically focused on solving these bottlenecks.
According to Vinyals, while AI systems are making progress on implementation and experimentation, they fall short in idea generation and evaluation. Current evaluations rely heavily on indirect capability benchmarks like SWE-Bench Pro, which can lead to models exploiting scoring systems rather than making genuine scientific advances. Meaningful evaluations that test self-improvement directly are expensive, compute-heavy, and underdeveloped.
To move forward, future systems must measure research quality using criteria like originality and efficiency through reward models and reinforcement learning. Vinyals noted that teaching AI “research taste” remains an unsolved challenge, making the path to true self-improving AI a gradual process rather than a sudden leap.
Why it matters
Tempers expectations for an immediate AI intelligence explosion, pointing toward steady, incremental gains in self-improvement.
Highlights a shift toward direct RSI benchmarks and reward models that evaluate long-horizon research taste and originality.
Signals new startup opportunities focused explicitly on automated idea generation, evaluation, and research automation bottlenecks.
Source: the-decoder.com
