Demis Hassabis says current AI cannot autonomously redesign itself
In a full interview, Hassabis said systems such as AlphaEvolve still rely on human choices and tightly specified goals. They can make incremental optimizations, but have not unequivocally demonstrated the architectural breakthroughs needed for open-ended recursive self-improvement.
0 comments · 0 votes
Sign in to join the discussion →
No comments yet. Start the discussion.
Why it moved the index
Hassabis distinguishes demonstrated, human-guided incremental optimization from the unproven ability to invent major new architectures under an unconstrained self-improvement goal. That materially narrows evidence for a present hard-takeoff mechanism, while his warning that end-to-end self-improvement may eventually be possible limits the size and confidence of the away-from-doom signal.
Assessment history
- R1Away 38 · confidence 78
New dated full-interview evidence separating current human-guided optimization from unproven open-ended recursive self-improvement.
13 Aug 2026