The Opus 4.7 tokenizer reckoning hits 594 HN pts w, Claude Design holds #2 on HN at 995 pts and 656 co, Andrej Karpathy's karpathy-skills vanishes from Gi

The Opus 4.7 tokenizer reckoning hits 594 HN pts with 415 comments — claudecodecamp.com measures 1.47× token inflation on real code and non-English content, above Anthropic's disclosed 1.0–1.35× range. At unchanged sticker prices ($5/$25 per 1M), the effective cost is +35% (≈+$105/month on a realistic workload). The first mass-scale audit of a stealth price move. Unchanged headline price is no longer trust-neutral.

Claude Design holds #2 on HN at 995 pts and 656 comments one day after launch — Figma stock closed down ~7% on Friday despite Anthropic's 'complementary, not competitive' framing. An Anthropic executive recently left Figma's board. When the second-most-discussed story on HN is about your platform and your stock drops 7%, the market has already decided which framing is true.

Andrej Karpathy's karpathy-skills vanishes from GitHub Trending top 25 — confirming yesterday's Day 4 deceleration (+7,959/day, 17% drop from Day 3) as the actual inflection point. Day 5 is off-chart. Meanwhile obra/superpowers reaccelerates to +1,713 stars/day and 158K total, up from 154.8K two days ago. The flagship curve broke; the ecosystem curve didn't. Multi-pole Claude Code ecosystem is now the empirical reality.

The self-evolving agent meme takes the top of GitHub Trending — EvoMap/evolver sustains at +737/day, lsdefine/GenericAgent at +845/day ('grows skill tree from 3.3K-line seed'), z-lab/dflash (block diffusion speculative decoding) debuts at +287/day. Every trending agent repo this week carries 'evolve,' 'genome,' or 'self-improving' branding. Marketing convergence or genuine research direction — the community hasn't yet decided.

Miguel Conner's 'I'm Coding by Hand' hits 187 HN pts with 190 comments — an ex-Aily Labs AI-agent-builder takes a 3-month Recurse Center sabbatical to hand-code a single-layer perceptron in Vim. Toby Ord's 'Are the costs of AI agents also rising exponentially?' lands same-day at 181 pts, arguing 4,000× parameter growth and 100,000× token-per-task growth over seven years. Two independent counter-narratives at the HN front page within hours of the tokenizer story. The hype hangover window has collapsed to 24 hours.

NIST officially gives up enriching most CVEs at 193 HN pts — citing a 263% surge in submissions since 2020. Going forward, only CVEs in CISA's KEV catalog, federal-used software, and EO 14028 critical software get enriched. Everything else is 'Lowest Priority — not scheduled.' The defender-of-last-resort for the open vulnerability data is officially capacity-constrained. Enterprise security posture built on NVD assumptions needs rewriting this quarter.

The iTerm2 escape-sequence RCE (164 pts) proves `cat readme.txt` is a remote code execution vector — malicious escape sequences combine with tmux integration (on by default) to inject arbitrary code. Same day: PanicLock (179 pts) reaches the front page as a MacBook Touch ID killswitch explicitly designed for 'border crossings, protests, and high-risk situations.' Fil-C (memory-safe C via capability-based pointers and concurrent GC) sustains at 159 pts. Three parallel security stories at the top — the surface is widening faster than it's hardening.

Smol Machines hits 310 HN pts — a YC-backed portable VM that boots in <200ms via Hypervisor.framework/libkrun, with a direct 'ship the machine, not the Electron bundle' thesis. Lordog/dive-into-llms (《动手学大模型》, a Chinese-language LLM tutorial book) surges to +944 stars/day and 31.7K total, Google's magika AI file-type detector at +956/day/15.6K. The week's secondary stories — Chinese dev pedagogy, lightweight isolation, native file classification — all point at the same trend: infrastructure beneath the AI layer is where the quiet compounding happens.