Skip to content
Source · Daily Brief

AI Daily Brief — 20 July 2025

A weekend without product launches but with the highest-profile single capability claim of the year so far. OpenAI’s IMO gold ruled the discourse. Google DeepMind stayed silent. Gary Marcus pressed for disclosure. Chinese labs were pre-WAIC quiet. EU regulators counted down to August 2.

Top stories

  • OpenAI’s IMO gold reverberates through Sunday. Saturday’s announcement that an experimental reasoning LLM reached gold-medal level (35/42, 5/6 problems, no tools, two 4.5h sessions) dominated AI Twitter. Noam Brown emphasised this is a general reasoning model (not IMO-specific) and framed inference-time scaling: “o1 thought for seconds. Deep Research for minutes. This one thinks for hours.” via Noam Brown on X
  • Simon Willison publishes detailed IMO gold write-up. Aggregates Alexander Wei’s and Noam Brown’s threads, the 5/6 score, and quotes OpenAI researchers’ framing that the result came from a general-purpose model not trained specifically for IMO. One of the most-read independent takes circulating on Sunday. via Simon Willison
  • Gary Marcus presses OpenAI for disclosure. Acknowledged the achievement as “genuinely impressive” (no tools, no internet) but criticised that OpenAI “told us the result, but not how it was achieved” — flagging the lack of independent IMO validation, opacity around the model architecture and inference cost, and reports that the proofs were “very unreadable”. via Gary Marcus on X
  • Google DeepMind stays silent, holding Gemini Deep Think gold for Monday. DeepMind had also reached gold-medal performance under official IMO collaboration but agreed to wait until after the IMO closing ceremony. The 24-hour gap, during which OpenAI “largely stole the spotlight”, was the dominant meta-story of Sunday. via Google DeepMind
  • Axios: OpenAI/DeepMind IMO “gold race” is the AI story of the weekend. Frames the gold-medal level as a watershed for reasoning models, notes Google’s result was IMO-certified and OpenAI’s was self-reported, and quotes the IMO cutoff (35/42, achieved by only 72 of 630 human contestants). via Axios

Who shipped

OpenAI shipped a debate. Google DeepMind shipped silence (with a gold medal in waiting). Chinese labs shipped nothing — holding for WAIC.

Open-source pulse

Chinese ecosystem in pre-WAIC mode. The Chinese open-source story of July centred on Moonshot’s Kimi K2 alongside Alibaba’s Qwen3-Coder. Coverage pivoted to previews of Premier Li Qiang’s “World AI Cooperation Organization” proposal and Huawei’s CloudMatrix 384 demo plans.

Money, infra & hardware

Tech press primed audiences for Trump’s AI Action Plan reveal mid-week — three pillars (Accelerate Innovation, Build American AI Infrastructure, Lead in International Diplomacy and Security), companion executive orders on export of the “full-stack” AI tech package, federal permitting for data centres, and removing “woke AI” from federal procurement. via TechCrunch

Quiet corners

EU GPAI Code of Practice in implementation window after July 10 publication — conversation among compliance circles centred on the August 2 trigger date for GPAI governance rules. Through the weekend, scrutiny continued over xAI’s July 9 Grok 4 release shipping without an industry-standard system card.

By the numbers

  • 35/42 — IMO gold cutoff achieved by both OpenAI and DeepMind reasoners
  • 72 / 630 — human contestants who hit gold-medal threshold
  • Jul 28 — IMO’s requested embargo end (broken by OpenAI)
  • Jul 23 — Trump AI Action Plan scheduled reveal
  • Aug 2 — EU AI Act GPAI obligations enforcement

Compiled by AI Feed’s editor from verified web sources for 20 July 2025.