DeepMind CEO: Why AGI Remains No Match for Human Reasoning

February 20, 2026

TL;DR

AGI has not yet reached human-level reasoning, according to Google DeepMind CEO Demis Hassabis.
Key gaps in AI include continual learning, long-horizon planning, and consistency across tasks.
AI models are largely static and struggle to learn online from experience like humans do.
Current AI can plan short-term actions but lacks the capability for long-term strategies.
Frontier AI systems can solve complex problems but may fail at elementary questions phrased differently.
Hassabis proposed a historical stress test: discovering general relativity by 1915 after training with knowledge cut-off in 1911, which current AI would fail.
AI can be valuable as a co-researcher, potentially ushering in a new era for scientific discovery.
Practical patterns include orchestration, where broad systems delegate tasks to specialized tools (e.g., Gemini using AlphaFold for protein structures).
Technological risks exist, including misuse and societal challenges, requiring bolstered cyber and biosecurity and international standards.
Hassabis expressed cautious optimism, believing technical risks can be solved if the best minds focus on them.
The generation growing up native with AI technology is expected to achieve incredible things.

Continue reading
the original article