Skip to content
Source · Daily Brief

AI Daily Brief — 19 July 2025

A single-story Saturday. OpenAI announced an experimental reasoning LLM had achieved gold-medal performance on the 2025 IMO — a methodological milestone. The reveal broke a community embargo and ignited a process-vs-science argument across X with Google DeepMind, which had submitted for official grading and waited. ICML 2025 wrapped its final workshop day in Vancouver.

Top stories

  • OpenAI announces gold-medal performance on the 2025 IMO. Researcher Alexander Wei posted that an experimental general-purpose reasoning LLM scored 35/42 — solving P1 through P5, no tools, two 4.5h sessions, natural-language proofs graded by three former IMO medallists. OpenAI said no model with this level of math capability would ship for several months. via Alexander Wei on X
  • OpenAI official account amplifies the IMO gold thread. “We achieved gold medal-level performance on the 2025 International Mathematical Olympiad with a general-purpose reasoning LLM.” Framed as a milestone for AI + mathematics. via OpenAI on X
  • Noam Brown frames result as bigger than the headline. The IMO gold was achieved with a general reasoning LLM under the same time limits as humans, no tools — methodological generality matters more than the medal itself, and “arrived years sooner than many expected”. via Noam Brown on X
  • Simon Willison covers the OpenAI IMO claim same day. Dissects the claim — links the X thread, notes the model used “experimental general-purpose reinforcement learning and test-time compute scaling”, and that OpenAI doesn’t plan to ship anything with this math capability for several months. via Simon Willison
  • Embargo controversy: OpenAI breaks IMO press protocol. Minutes after the closing ceremony, breaking an IMO embargo that asked AI labs to wait until Jul 28. Drew criticism from the IMO community and from Google DeepMind, which had submitted Gemini Deep Think for official IMO grading and waited. IMO president Gregor Dolinar publicly noted the IMO could not validate OpenAI’s compute, human involvement, or reproducibility. via Xena Project

Who shipped

OpenAI shipped a paper-less press release. Google DeepMind shipped patience (and an official IMO submission).

Open-source pulse

Saturday weekend — quiet on releases. Mistral’s Voxtral and Moonshot’s Kimi K2 from earlier in the week continued circulating.

Money, infra & hardware

Quiet on the funding wire. Lovable’s Friday $200M at $1.8B kept ringing.

Quiet corners

ICML 2025 closed in Vancouver with its second workshop day. Conference ran Jul 13-19; Jul 18-19 were workshop days. Programs included LatinX in AI, Data-centric ML Research, and the Unifying Data Curation Frameworks workshop. Terence Tao posted commentary from Mathstodon as IMO 2025 concluded — later cautioned against drawing strong conclusions from unofficial AI claims.

By the numbers

  • 35/42 — OpenAI experimental LLM IMO score
  • 5/6 — IMO problems solved (P1-P5, missed P6)
  • 2 × 4.5h — competition sessions, same time limits as humans
  • Jul 28 — community embargo OpenAI broke
  • 3 — former IMO medallists who graded the proofs

Compiled by AI Feed’s editor from verified web sources for 19 July 2025.