Skip to content
Source · Daily Brief

AI Daily Brief — 10 June 2025

Three big AI events on one day. Mistral entered the reasoning-model wave with Magistral — including the first Western 24B-class Apache 2.0 reasoning model. Altman published his most ambitious AI-futures essay to date. And ChatGPT had one of its worst infrastructure days of 2025, going down globally for 10+ hours the same day its CEO published ‘The Gentle Singularity.’

Top stories

  • Mistral launches Magistral reasoning family. Magistral Small (24B, Apache 2.0 on Hugging Face) plus Magistral Medium (proprietary, via Le Chat, Mistral API, partner clouds). Chain-of-thought reasoning across English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese. Medium scored 73.6% on AIME2024 (90% with majority voting @64); Small scored 70.7% / 83.3%. Powers a new ‘Think mode’ in Le Chat plus ‘Flash Answers’ (~1,000 words/sec via Cerebras hardware, ~10x typical speed). via Mistral
  • Magistral analysis hits press the same day. TechCrunch, VentureBeat, Axios, InfoWorld and Simon Willison all published. Mistral’s RL training stack built ground-up (no traces from prior models). The arXiv paper (2506.10910) introduces an RLVR — Learning from Verifiable Rewards — pipeline delivering a 50% gain on AIME-24 pass@1 over the Mistral Medium 3 base. via TechCrunch
  • Sam Altman publishes ‘The Gentle Singularity’ essay. Sketches a near-term AI timeline: 2025 = agents doing real cognitive work (especially code); 2026 = systems generating novel insights; 2027 = useful real-world robots; 2030s = wildly abundant intelligence and energy. Frames the singularity as transformation so gradual people barely notice — ‘wonders become routine, and then table stakes’ — while flagging safety, governance and broad access to superintelligence as the open work. via Sam Altman
  • ChatGPT and Sora suffer 10+ hour global outage. Took ChatGPT (free and paid) and Sora offline globally starting around 02:45 ET / 07:45 BST. Downdetector showed high concentrations across North America, Europe, India and Australia. Post-mortems attributed it to an API surge — reportedly ~4.2M requests/min, 37% above prior peak — jamming load balancers and cascading across microservices. Significantly more severe than the March 2025 blackout. via OpenAI
  • WWDC25 Day 2: Foundation Models framework deep-dive for developers. Day 2 sessions put Apple’s new framework front and center. API gives third-party apps direct, free, on-device access to the ~3B-parameter LLM at the core of Apple Intelligence — including guided generation and tool calling — with as few as three lines of Swift. Pitched as a power move for offline, privacy-preserving features.
  • Apple Intelligence shipped features confirmed in developer sessions. Live Translation across Messages, FaceTime and phone calls; upgraded Visual Intelligence that can act on whatever is on the iPhone screen; Image Playground able to call ChatGPT for new styles; Genmoji upgrades that let users mash two Unicode emoji into a new character.
  • Apple publicly admits the personalized Siri AI upgrade is delayed. Federighi acknowledged the more-personalized Siri Apple Intelligence features ‘need more time to reach our quality bar’ and would not ship until at least 2026. Forrester’s Dipanjan Chatterjee called ‘the silence surrounding Siri … deafening.’ via Fortune

Who shipped

Mistral ran the model day. OpenAI ran the essay day — and the outage day. Apple ran the developer day at WWDC25.

Open-source pulse

Magistral Small (24B, Apache 2.0) is the first Western reasoning model released with permissive commercial weights at this scale. Built on Mistral Small 3.2 (2506) with SFT from Magistral Medium traces plus RL. Community quantizations and llama.cpp / vLLM serving guides followed within hours. via Simon Willison

Money, infra & hardware

The Meta-Scale AI mega-deal was actively being negotiated ahead of formal disclosure Jun 12-13: Meta investing $14.3B for a 49% non-voting stake (valuing Scale >$29B) and hiring founder Alexandr Wang and several top staffers into Meta’s AI org. Scale CSO Jason Droege lined up as interim CEO.

Quiet corners

No major release from Anthropic, Google DeepMind, xAI, DeepSeek or Alibaba/Qwen on the date. No standalone NVIDIA or AMD chip news anchored to Jun 10; AMD’s Advancing AI 2025 event with Instinct MI350 was two days out.

By the numbers

  • 24B / Apache 2.0 — Magistral Small size + license
  • 73.6% / 90% — Magistral Medium AIME2024 / with majority voting @64
  • ~1,000 words/sec — Flash Answers speed via Cerebras
  • 10+ hours — ChatGPT global outage duration
  • ~4.2M req/min / +37% — reported API surge / above prior peak
  • 2025 / 2026 / 2027 / 2030s — Altman’s agents / novel insights / robots / abundance milestones
  • Most-mentioned company: Mistral

Compiled by AI Feed’s editor from verified web sources for 10 June 2025.