Skip to content
Source · Daily Brief

AI Daily Brief — 14 July 2026

Open-weight releases, governance proposals, and legal actions defined 14 July. PrismML shipped a 1-bit quantised model that runs a 27B network on a phone, xAI’s Grok Build tool was found to have been silently uploading entire user codebases before the feature was disabled, and Meta faced simultaneous acclaim — a perfect score at the Asian Physics Olympiad — and lawsuit over AI-driven layoff targeting.

Top stories

  • Grok Build uploaded user codebases without explicit consent. Security researcher Cereblab published findings showing that xAI’s Grok Build CLI packaged and sent entire code repositories, including explicitly excluded files, to Google Cloud storage during coding sessions. xAI disabled the feature after the disclosure. via The Verge – AI
  • Bonsai 27B: a 27B-parameter model in 3.8 GB, running on a phone. PrismML released a ternary (1-bit) quantised variant of Qwen 3.6 27B that cuts the model’s footprint from 54 GB to 3.8 GB while retaining roughly 90% of fp16 performance. The model runs in-browser via custom WebGPU kernels and on mobile hardware. via Hacker News
  • Meta AI achieved a perfect 30/30 on the Asian Physics Olympiad theoretical exam. Meta’s team submitted a model to the APhO competition and reported that it tied with the top three student contestants. via X · @AIatMeta
  • Demis Hassabis called for a US-led global AI standards body. DeepMind’s CEO proposed an independent watchdog modelled on FINRA to test frontier models and establish safety baselines, saying “nobody in the world knows what happens next.” Sam Altman described the proposal as “thoughtful.” via TechCrunch – AI
  • Apple filed a trade secrets lawsuit against OpenAI. The suit centres on Sam Altman’s hardware ambitions and OpenAI’s pre-IPO trajectory, adding to a string of legal challenges the company has accumulated in 2026. via The Verge – AI
  • Meta sued over AI-driven layoff targeting. Twenty-six former employees alleged that Meta used internal AI performance tools to disproportionately select workers with disabilities and medical conditions for redundancy. Meta denied the claims. via The Verge – AI

Who shipped

Meta dominated coverage from multiple directions: an AI model topped the Asian Physics Olympiad, while the company simultaneously faced two AI-related lawsuits and Instagram head Adam Mosseri predicted companies will soon cap per-engineer AI token spending. OpenAI announced 8 million active users across Codex and ChatGPT Work and reset usage limits; Sam Altman noted GPT-5.6 Sol is half the price and roughly twice as token-efficient as Fable. Anthropic launched Claude for Teachers — free premium access for verified US K-12 educators, with a commitment not to train on student data — and pledged CAD $10 million to Canadian AI research institutions. Apple opened the iOS 27 public beta, giving general users early access to the rebuilt Siri ahead of the fall release.

Open-source pulse

Bonsai 27B from PrismML is the headline release: ternary quantisation delivers 27B-parameter capability at phone-class memory. The community also anticipated several imminent arrivals — Kimi K3 was expected within hours of publication, DeepSeek V4 general availability later in the week, and new Mistral and Liquid models before the end of the month. The llama.cpp project shipped five point releases (b10003–b10011) covering CORS server options, prompt-cache refactoring, and pre-Ampere GPU fixes.

Money, infra & hardware

DeepSeek is reportedly raising again just weeks after closing a $7 billion round, seeking capital for its own data centres to sustain aggressive V4 pricing. The founder of Hinge raised $18 million for Overtone, a voice-and-audio-first AI dating service. NVIDIA published developer guides on post-training Cosmos 3 in under a day using agentic coding skills.

Quiet corners

Mistral was absent from direct announcements, appearing only in community speculation about upcoming model releases. Cohere and AI21 had no presence in today’s coverage.

By the numbers

  • 473 stories today across 30+ sources
  • Most-mentioned model: GPT (14 tagged stories)
  • Most-mentioned lab: Meta
  • Notable absences: Mistral, Cohere, AI21

Compiled by AI Feed’s editor from all 473 headlines published on 14 July 2026.