DeepMind's New AI Just Changed Science Forever

TL;DR

DeepMind's Aletheia AI autonomously solves open math problems and contributes core content to peer-reviewed research papers, marking a first in AI-driven scientific discovery.

Key Points

• 1.Aletheia uses a separated thinking-verification architecture to prevent self-deception. The messy reasoning chain is hidden from the verifier so the AI cannot blindly agree with its own flawed logic — a key breakthrough over prior systems.
• 2.A 100x compute efficiency gain was achieved without sacrificing intelligence. A stronger base model was trained so the current Aletheia matches a 6-month-old version's capability while using 100 times less compute, boosting math olympiad scores from ~65% to ~95%.
• 3.Heavy tool-use training stopped the AI from hallucinating fake citations and authors. Enabling web search and training the AI to synthesize dozens of cutting-edge papers was what finally grounded its outputs in real, verifiable knowledge.
• 4.Aletheia autonomously solved 4 open Erdős math problems and contributed core content to 5 research papers. Topics include constants in arithmetic geometry and new limits for interacting particles; independent math experts verified the work for both correctness and novelty.
• 5.The system reaches Level 2 autonomous publishable research, with Levels 3–4 (groundbreaking work) still out of reach — but the host questions for how long. Given the pace of progress, he suggests those levels could fall within months.