Ibrahim El Khansa

Wednesday 14 October 2026

Daily AI Brief

Top 3

  1. Gemini 4 Argon reaches paying users

    Gemini 4 Argon reaches paying users

    Google's flagship model is now in the Gemini API paid tier, AI Studio and Antigravity, and in the Gemini app for Google AI Pro and Ultra subscribers. API prices are $2 per million input tokens and $10 per million output until 30 November, then $4 and $20, with a 1M unified input-output token limit.

    blog.google
  2. OpenAI engineer charged with bugging its office

    Prosecutors say Abdulrahman Alfozan, a Saudi national at OpenAI since July 2025, hid a recording device behind a TV in an executive conference room; OpenAI says staff used its own models to identify the device and trace its origin. OpenAI says it has no sign he took model weights, source code or customer data.

    sfstandard.com
  3. Anthropic replaces git with its own Scion

    Anthropic moved its monorepo off git to Scion, built in house in five months, after daily merges rose from about 2,500 in May to over 30,000 in September, mostly from agents and CI jobs. CI checkouts dropped from minutes to seconds; Scion will not be open-sourced, but the post shares its design principles.

    anthropic.com

Benchmarks

  • Gemini 4 Argon tops the Vals Index

    Vals AI's independent index, updated 13 October, puts Argon first at 68.90%, ahead of Claude Sonnet 5.5 (67.04%), Claude Opus 5.5 (66.97%) and Claude Fable 5.1 (65.83%).

    vals.ai
  • Google's own Argon comparison table

    Google shows Argon ahead of Claude Opus 5.5 on AutomationBench (51.3% vs 42.5%) and Harvey's Legal Agent Benchmark (19.6% vs 3.8%), but behind Opus 5.5 on Terminal-bench 4.0 (57.4% vs 66.4%) and behind GPT-6 Astra on FrontierSWE v2 (55.0% vs 65.5%).

    Lab reporteddeepmind.google

AI news

  • 415 scientists call for AI containment rules

    A letter signed by Yoshua Bengio, Geoffrey Hinton, Stuart Russell and others asks labs to airgap cyber-capable models now, and calls for an international body that sets containment standards and receives reports of containment failures, to be made public eventually.

    ai-security-letter.org
  • Apple wins early access to OpenAI hiring files

    In Apple's trade secrets suit, a judge gave Apple expedited access to OpenAI's interview materials for former Apple staff in its hardware and robotics groups and approved depositions of three ex-Apple employees, including hardware chief Tang Yew Tan. Broader requests for OpenAI's hardware research were denied for now.

    techcrunch.com
  • Xbox starts an experimental Labs team

    Xbox Labs is "a small group of game makers, engineers and scientists" meant to turn emerging ideas into things players and developers can try. The announcement points to agents that write and test software and to games being decompiled and rebuilt for new hardware, but names no projects yet.

    news.xbox.com

Built with AI

  • EmbeddingGemma 2 becomes a speech-to-text app

    A local Mac transcription app built on Google's embedding model, which its author called a couple of days of work; one commenter reported an hour of audio transcribed in 13 seconds. Commenters found it fast but less accurate than cloud models, and English only: Portuguese audio came out as English text.

    news.ycombinator.com
  • REA passes 222,000 GitHub stars

    New since our 12 October item (about 125,900 stars then): the tool that lets coding agents reverse engineer apps down to native binaries now shows about 222,300 stars and 55,100 forks.

    github.com

Frameworks and dev tools

  • Claude Code 2.1.299 adds idle compaction

    Conversations over 200K tokens that sit idle for about 54 minutes are now compacted before the one-hour prompt cache expires; turn it off with "idleCompaction": false. A new CLAUDE_CODE_DISABLE_FILE_MENTIONS setting passes @-mentioned file paths instead of their contents.

    github.com
  • Gemini 4 Argon arrives in GitHub Copilot

    Generally available and rolling out gradually for Copilot Pro+, Max, Business and Enterprise, in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI and the coding agent, billed at provider list pricing.

    github.blog
  • GitHub blocks pushed Anthropic keys by default

    Secret scanning now detects Anthropic service and user API keys, and push protection blocks commits containing them by default, including in free public repositories.

    github.blog
  • 11-year-old soundness hole in Rust's borrow checker

    A newly reported bug lets code bypass a struct's lifetime requirement through covariant field assignment, and variants can read freed memory. Rust's types team marked it P-high, says exploiting it needs an intentionally malicious Drop implementation, and a fix PR is open.

    github.com

Tech news

  • ASML says AI demand outruns 2027-28 capacity

    ASML reported €11.7bn in third-quarter sales, guided €14.2bn to €15.2bn for the fourth, and said its planned 30% capacity growth for 2027 and 2028 is "no longer enough". Some customers want to order for 2029 and possibly 2030.

    siliconrepublic.com
  • DOJ raids over 40 North Korean laptop farms

    Prosecutors searched more than 40 laptop farms in Texas, seized over 180 computers that US companies had shipped to new hires and arrested at least two people; at least 70 US companies were infiltrated. CrowdStrike helped the FBI investigate.

    techcrunch.com
  • Microsoft fixes a record 1,411 flaws

    October's Patch Tuesday fixes 1,411 vulnerabilities, 70 rated critical, including an actively exploited Mark of the Web bypass (CVE-2026-96951). BleepingComputer links the record count to Microsoft's use of AI to find and fix bugs.

    bleepingcomputer.com
  • Starlink limits terminals in motion in Ukraine

    SpaceX capped the StarShield military package at 100 mph overnight, which a Zelensky adviser says severely disrupted several military missions. The White House has not commented.

    kyivindependent.com
  • Apple HomeView opens for pre-order

    Apple's 7-inch smart display with an A18 chip and Siri AI (in English to start, in beta) is available from 13 November.

    apple.com

Worth reading

  • Making a game its own AI harness

    Kyle Kukshtel argues LLM agents are bad at reading games by default and lays out the fixes he uses with Claude Code: seeded randomness, typed text commands, image diffing and a simulation separate from the engine.

    kylekukshtel.com

An AI agent I set up researches and writes each brief. Every item links to its source.

Other days