Super Intelligence

Super Intelligence: short, sourced explainers from Time Travel Leap.

AlphaProof Nexus proved open math problems, and Lean checked every step

Infographic: a glowing orb labeled language model, a stack of blocks under a gold check mark labeled Lean checks every step and a grid of 353 tiles with nine raised gold tiles labeled open Erdős problems. Panels explain that Google DeepMind's agents wrote proofs in Lean, that the system settled nine open Erdős problems and proved 44 sequence conjectures and that machine checking spares people a line-by-line re-check, though experts must confirm each problem was stated correctly in Lean. Cards: nine of 353 open Erdős problems, 44 of 492 OEIS conjectures, two of the nine unsolved for more than 50 years. Banner: proofs a computer checked line by line; about 2.5 percent success; the team chose the problems. Source: Science, Oct. 8, 2026.

Google DeepMind's system settled nine open Erdős problems, two of them unsolved for more than 50 years. A proof checker verified every line. (Image: SI illustration)

Science 394(6820), 234–239 · Oct. 8, 2026

Super Intelligence

The LACE framework designs its own problem-solving algorithms

Infographic: a stack of books with a scroll labeled contract and tool library, five glowing spheres linked to a gold center labeled specialist heuristics and a 3D bar chart of average scores with LACE in gold. Panels explain that LACE gives a language model a verified contract and evolves specialist heuristics, that it averaged 0.945 across 36 CO-Bench problems against 0.870 for the best prior method and that it scored 0.97 to 0.99 on four new problems where five other methods produced no feasible algorithm. Cards: 0.945 LACE average, 0.870 best prior method, 0.571 language model alone. Banner: benchmark scores, not real-world savings. Source: Nature Machine Intelligence, Oct. 1, 2026.

LACE outscored other Super Intelligence methods on average across 36 classic optimization benchmarks. The scores are benchmark results, not real-world savings. (Image: SI illustration)

Nature Machine Intelligence · Oct. 1, 2026