Demis Hassabis and Sergey Brin decode the next leaps: thinking systems, test-time compute, and what “AGI” should actually mean.

TL;DR

• Thinking/Reasoning models at test time can add hundreds of “ELO” in games; similar gains are expected beyond games as world models improve.0430
• Scale still matters — but algorithmic advances plus tool-use and multi-agent “thinking” are key for the next jumps.
• AGI should be consistent across tasks at best-human levels (Einstein/Mozart/Curie), not easily poked with trivial failures.

Key Claims

• Reinforcement-learning style “thinking” on top of language models can boost capability far beyond first-try outputs; test-time compute matters.
• Serving and long “thinking” workloads will create huge inference compute demand, not just training.
• AGI should be consistent across tasks at best-human levels (Einstein/Mozart/Curie) rather than easy to break.
• Self-improvement may accelerate progress, but the real world is messier than games.

Why It Matters

Enterprises should plan beyond “bigger models” to reasoning budgets (time/steps), tool-use, and multi-agent planning — the levers that move reliability and ROI.

Action Steps

• Run A/B tests on test-time compute (short vs. long thinking) for your highest-value tasks.
• Add a tools layer (retrieval-augmented generation, code runner, browser) to your agent graph.
• Define consistency metrics (pass@k with tool-use) to move from demo to production.

Tools Box

• Agent Frameworks
• Code Copilots & Test-Bots
• Research Copilots
• Ops Bots
• Creatives (Design & Video)

FAQs

Will AI replace most programmers in one year?

Some high-profile voices predict rapid displacement; many expect role redesign over pure replacement. Prepare for both: double down on specific

ation, architecture, oversight, testing, and product judgment.

What are infinite context windows?

Models that can read and retain very long inputs (documents, repos), enabling step-by-step planning wi

What are AI agents?

Systems that observe, plan, and act with memory and tools—moving from assistance to semi-autonomy with human approval gates.