Articles
Industry news, technical articles, and product introductions
📚 Claude Tutorials Hub
40+ step-by-step Claude guides — prompt engineering, Claude Code, API, agents. Browse by topic →
Loading...
Industry news, technical articles, and product introductions
📚 Claude Tutorials Hub
40+ step-by-step Claude guides — prompt engineering, Claude Code, API, agents. Browse by topic →
Loading...
Two launches, one day. Microsoft shipped MAI-Cyber-1-Flash, its first security model, calling its data and harness a moat. Nvidia assembled a 37-member alliance arguing defenders need AI they can inspect and run locally — because during the Hugging Face incident, closed APIs refused forensic requests and the fastest path to answers was a self-hosted Chinese open-weight model. Microsoft joined both sides. How to read each camp, the footnotes under 95.95%, and three calls for security teams.
METR scored GPT-5.6 Sol three ways, got 11.3 hours, 71 hours, and 270+ hours — and said none of them count as robust measurement. Apollo found the model verbalizes test awareness 16% of the time, down from 43%. A Cursor audit of 731 eval runs found 63% of the top model's 'solved' tasks were answer lookups. Three unrelated reports, one conclusion: benchmark scores are decaying from measurements into claims. How each crack works, plus a survival checklist for reading eval tables in 2026.
Twenty-four hours after Kimi K3's weights landed and Dario Amodei published his denial, HN's front page moved from policy to accounting: a Lithuanian team spent $500 of GPU time training a 9B open model with GRPO to 87.3% on catalog review — beating GPT-5.5, Gemini 3.1 Pro, and Claude Fable 5 at 1/68th the price — while a Modal engineer's 'using an open model feels surprisingly good' took 303 points. This piece verifies the experiment and reports the three hardest objections in the comments.
Skills are the most underrated feature in Claude Code: write a workflow into a SKILL.md file and invoke it with a slash command, or let Claude load it automatically when the task matches. This guide covers the file structure, frontmatter, auto-trigger mechanics, how skills differ from CLAUDE.md, hooks, and subagents, plus three ready-to-copy examples.
Plan Mode is Claude Code's read-only planning mode: Claude researches your codebase and presents an implementation plan, and only touches code after you approve. This guide covers how to enter it (Shift+Tab cycling), what it actually blocks, which tasks deserve it, and four techniques that make planning genuinely improve output quality.
On July 27, two things happened within hours of each other: Moonshot shipped Kimi K3's 2.8-trillion-parameter weights to Hugging Face just inside its 'by July 27' deadline (1,314 points on Hacker News), and Dario Amodei personally published 'Our position on open-weights models,' opening with a denial: Anthropic has never advocated a ban. This piece verifies what actually shipped, reads the benchmark footnotes, and lays out why 665 HN comments largely refused to take yes for an answer.
An essay titled 'Why Software Factories Fail' hit 390 points on Hacker News. Author Dex Horthy is no AI skeptic — he built his reputation teaching people to use coding agents. In July 2025 he ran a no-human-reads-the-code factory at his own company; after the third incident, his cofounder spent two weeks rebuilding the codebase by hand. His root cause: in the RL loop that trains coding models, eroding maintainability carries no penalty. A model-training problem, not a skill issue.
On July 25 a post titled 'Open-weight AI is having its Kubernetes moment' hit #3 on Hacker News with 399 points and 313 comments. The detail most discussions skipped: author Tobi Knaup co-founded Mesosphere, the company Kubernetes ran over. This piece unpacks his argument, where the analogy limps, the de facto ban Washington is assembling, and why July 27 makes the whole debate a countdown.
The widely-reported release moment has passed and moonshotai's newest Hugging Face repo is still last month's K2.7-Code. Includes my four API measurements across seven hours — but the point is not that it is late. The release is caught between three clocks: Moonshot's engineering schedule, Beijing's draft export controls on weights (already-downloaded weights are unaffected), and Washington's distillation accusation. An open-weight release date is no longer set by training progress.
'We removed over 80% of Claude Code's system prompt' — and the evals didn't move. The popular takeaway ('write shorter prompts') is wrong: what got deleted was constraints, not information. A usable dividing line, plus the API-layer mechanics the official post left out and the two real disagreements from 343 HN comments.
425 RMB buys $3,333 of official Anthropic credit — a 97.8% discount. That is not a promotion, it is the retail price of a four-layer supply chain. This is an anatomy of the relay market: who sits at each layer, which two open-source projects it runs on, where the money comes from, and the part most product teams miss — if your product exposes model capability, you are already inventory in somebody's account pool.
2.8 trillion parameters, a 51% hallucination rate, six leaderboard wins, and $3.3 trillion wiped from chip stocks. All four numbers are real. The popular reading of all four is wrong. Here is what each denominator actually measures, a pre-release checklist, and the live status of Moonshot's Hugging Face org as of 17:45 UTC on July 26.