Morning brief

Tuesday, July 14

3 people · 0 companies · 5 topics

TL;DR Claude Code shipped v2.1.208+v2.1.209 today with screen-reader support and a wave of agent reliability fixes. Gemini 3.5 Pro is targeting Thursday July 17 after a full architectural rebuild, but no official Google confirmation exists yet.

Tools & dev

  • claude-code v2.1.208 (July 14, 01:10 UTC, 45 changes): screen reader mode (--ax-screen-reader or CLAUDE_AX_SCREEN_READER=1), vimInsertModeRemaps setting for jj-to-Esc sequences, CLAUDE_CODE_PROCESS_WRAPPER env var for corporate launcher support, mouse-click support for multi-select menus in fullscreen, fixed fast mode not auto-restoring after model switches, fixed background agent replies lost on delivery failure, fixed context window briefly resetting after CLI auto-updates, fixed HTTP/2 in supervised/background sessions, paginated 200+ row markdown tables, fixed several Bedrock streaming and OAuth MCP server bugs. v2.1.209 hotfix followed six hours later, unblocking /model and other dialogs that were incorrectly blocked in background sessions. GitHub releases

Models & research

  • model-releases Gemini 3.5 Pro is targeting July 17 GA per third-party reporting — Google rebuilt the architecture from scratch rather than shipping what it had, adding a claimed 2M-token context window and a reasoning layer to compete with Fable 5 and GPT-5.6. As of July 13, no official model card, pricing page, or Gemini API listing exists; every specific claim comes from unnamed sources. DeepSeek V4 also goes from preview to stable on July 17, and the legacy deepseek-chat/deepseek-reasoner API aliases hard-retire July 24 — requests using old names will error after that. TechTimes (July 13) · DEV.to DeepSeek migration guide

  • cactus-compute/needle — 26M-param function-call model designed to run on phones, watches, and glasses; distilled from Gemini using an MLP-free “Simple Attention Network” architecture (MLPs dropped entirely, relying on external tool lists instead). Achieves 6,000 tokens/sec prefill and outperforms FunctionGemma-270m on single-shot function calling. Weights fully open-source. GitHub

People

  • sama vs. elonmusk on compute: Altman publicly dismissed SpaceX’s orbital data center pitch on July 13, saying it “is not going to be a serious business anytime soon” and accusing Musk of “selling public market investors on short-term space datacenters.” Context: space compute is a load-bearing pillar of SpaceX’s $2T valuation. TechCrunch

  • emollick pushed back on OpenRouter token-share graphs as a measure of real AI usage: “I think OpenRouter is not a good measure of actual model usage in a world of agentic tools… We really need better data indicators for AI!” — pointing out that usage is shifting to Codex/Code/Cowork without surfacing in token-share metrics. X

OSS

  • Dicklesworthstone/destructive_command_guard — Rust PreToolUse hook that intercepts shell commands before AI coding agents execute them; blocks git reset --hard, rm -rf, database drops, and cloud resource deletions across Claude Code, Codex, Gemini CLI, Cursor, Copilot Chat, and others. Uses SIMD-accelerated regex for sub-millisecond latency; 50+ modular security packs (git, filesystem, Postgres, Docker, K8s, AWS/GCP/Azure). 3.2K total stars, 481 new today. GitHub

  • github/spec-kit — Toolkit for Spec-Driven Development, trending today with 807 new stars. GitHub

  • cactus-compute/needle — see Models above; 113 new stars today in Python trending. GitHub