Daily AI Brief — Wednesday, February 18, 2026

Generated: 2026-02-18 05:00 Items: 43 new stories


Claude Sonnet 4.6 — Anthropic releases its latest flagship model with enhanced capabilities across reasoning, coding, and multimodal tasks. Anthropic

Research

GLM-5: from Vibe Coding to Agentic Engineering — Research paper detailing GLM-5's development and its performance in agent-based coding tasks. arXiv

The Geometry of Alignment Collapse — Study reveals how fine-tuning can break AI safety measures through geometric analysis of alignment failures. arXiv

Perceptive Humanoid Parkour — New method chains dynamic human skills via motion matching for advanced humanoid robot movement. arXiv

CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing — Novel approach for editing large language models without destroying existing capabilities. arXiv

Tools

Async/Await on the GPU — Technical deep dive into implementing asynchronous programming patterns for GPU computation. VectorWare

Alibaba's Qwen3.5-397B-A17B reaches #3 in open weights rankings — New massive open-source model climbs to third place in the Artificial Analysis Intelligence Index. Reddit Image

Industry

Thousands of CEOs admit AI had no impact on employment or productivity — Major CEO study reveals AI's promised productivity gains have yet to materialize in the workplace. Fortune

Anthropic deploys $20M to support AI regulation ahead of 2026 elections — Major AI company invests heavily in regulatory advocacy as election cycle approaches. CNBC

OpenAI, US government, and Persona built identity surveillance machine — Investigation reveals collaboration on identity verification system with surveillance implications. VMFunc

Community

LLM food truck survival experiment — Community test gives 12 language models $2,000 and a virtual food truck business, with only 4 managing to stay profitable. Reddit

GLM-5 and DeepSeek top Game Agent Coding League — Both models secure positions in top 6 across five different gaming environments in agent performance testing. Reddit

llms.txt standard for AI crawlers — New proposed standard helps websites communicate with LLM crawlers about their content and preferences. Anna's Archive


Local Sources

@@tom_doerr: Locks computer when Bluetooth device leaves range... twitter-bookmarks

@@tom_doerr: Resources for systematic trading and quantitative finance... twitter-bookmarks

@@steipete: how it started:... twitter-bookmarks

@@blader: used claude code to make a little claude code skill that learns new claude code skills as you use cl... twitter-bookmarks

@@shotovim: Skills が多くなり管理に困っていたんですが、Obsidian の Bases を使ったらきれいに整理できました。... twitter-bookmarks

@@melvynxdev: urgent pro tips with Claude Code :... twitter-bookmarks

@@iruletheworldmo: i never want to read any other way again. https://t.co/hc4edKpDcJ... twitter-bookmarks

@@trq212: https://t.co/X2iu8WdIb8... twitter-bookmarks

@@openclaw: 🦞 Clawdbot v2026.1.12... twitter-bookmarks

@@googleaidevs: Announcing MedGemma 1.5 🩺... twitter-bookmarks

Frontier Model News (Perplexity Research) ai-landscaping

Claude Sonnet 4.6 (Anthropic) ai-landscaping

GPT-5.3-Codex-Spark (OpenAI) ai-landscaping

Frontier Model News (Perplexity Research) ai-landscaping

GPT-5.3-Codex-Spark (OpenAI) ai-landscaping

Qwen3.5 (Alibaba) ai-landscaping