W45

W45 Reading Notes: Anthropic's IPO Day, The Hackening, and Open Models Catching Up

Anthropic shipped Opus/Sonnet/Haiku 5.5 plus an S-1 filing in a single day; it was also a rough security week — a backdoor in the Claude Code VS Code extension and a math trick that took over ChatGPT accounts

8articles
12+sources
Y

The biggest stories this week all landed on 11/02: Anthropic shipped Opus 5.5, Sonnet 5.5, and Haiku 5.5 on the same day, and filed its S-1 to prepare for an IPO

I'm splitting what I read this week into three threads: models and products, security, and what a few people I follow had to say

AI Models & Product Updates

Anthropic's launch was a clean-sweep: three model sizes refreshed in a single day, plus an S-1 (IPO filing), a piece on accelerating drug development, and an article on their own economic scenario forecasts

I run Claude Code for development every day, so what I care about most is three sizes updating together — it lets me allocate by task difficulty: route the trivial stuff to the cheapest Haiku, reserve Opus for things that need judgment, and stop paying top dollar for every call. For anyone on usage-based pricing, or running several agents at once like I do, that directly moves your monthly bill

Google DeepMind was busy the same week: Gemini 4 and Genie 4 (a world-simulation model) both landed. On open weights, Mistral dropped Large 4 and HuggingFace saw a wave of smaller model updates. My read is that "open models are catching up" was especially visible this week — even European banks are building proprietary intelligence on open-weight models rather than following the closed labs

AI Engineering & Dev Tools

One signal worth recording: decision models are becoming their own tool category. Hamel Husain wrote a piece on using decision models for evals, naming TypeSafe's Jev, Perplexity's Decider, Liquid's d1, Cloudflare's Clef, and OpenAI's Decisions API

I have plenty of steps in my own pipeline that only need a decision, not generated text (classification, routing, scoring), and burning a large model on those is wasteful. These purpose-built judgment models are fast and cheap, which is exactly the fit — I'll put it on the watchlist and run a real accuracy comparison when I get the chance

Martin Fowler published two very honest pieces this week: one on "why good developers are wary of agentic development" — the conclusion being that good engineers know "built" is not the same as "working in production." The other was about that creeping feeling of reading something an LLM wrote. Both resonated: I re-verify everything an agent hands me, not because I distrust the tool, but because there's still a gap between "tests pass" and "actually correct"

Expert Perspectives

Andrej Karpathy
Andrej Karpathy — He noted something I agree with: the gap between how the general public, working professionals, and a small sliver of heavy users perceive LLM capability is widening. Most people are still stuck on the "derpy chatbot" impression while a few are experiencing something entirely different
Martin Fowler
Martin Fowler — "built isn't the same as working in production" — a line I'd pin next to every agent output that claims to be done
Zvi Mowshowitz
Zvi Mowshowitz — Spent several posts dissecting Anthropic's S-1 page by page, effectively doing a close reading for anyone trying to understand the company's finances and strategy. He also framed the week as "The Hackening" — so many security incidents he had to open multiple posts

Security: A Rough Week

If I only remember one thing from this week, it's that the trust assumptions in the software supply chain are being poked holes in

  • The Claude Code VS Code extension was backdoored on the Cursor Marketplace — a third-party marketplace is not the official source; check the publisher before installing
  • Someone took over ChatGPT accounts using "math" — a seemingly harmless input became the entry point for account takeover
  • In Taiwan, DEVCORE's Orange Tsai disclosed a full Edge browser exploit chain at Pwn2Own that chains 5 logic bugs without memory corruption
  • Japan Airlines and Rakuten Travel both leaked passenger data (JAL ~92k records, Rakuten ~3.54M)

My takeaway: the common thread isn't that any single vulnerability is especially novel, but that "the things we assume we can trust" — marketplace extensions, third-party booking systems, seemingly harmless inputs — are being shown, one by one, that they shouldn't be trusted unconditionally. For any system you run, this week is a reminder: public endpoints, third-party packages, and user input should all be treated as untrusted

Action Items

  1. If you're a heavy user, take advantage of the three-size refresh to re-audit which tasks actually only need a cheap model — that's exactly how I keep my monthly costs down
  2. For pipeline steps that only need a decision, not generation, look into decision models rather than throwing a large model at everything
  3. Before installing any dev tool extension, confirm the publisher is an official source — don't install just because the name looks right; that's exactly how this week's backdoor got in

Sources

RSS Digest: This week's raw material was compiled from ~1,574 articles across Anthropic, OpenAI, Google DeepMind, NVIDIA, HuggingFace, iThome, Martin Fowler, Zvi Mowshowitz, Hamel Husain, Andrej Karpathy, and others

ShareLinkedInX
AIAnthropicSecurityOpen ModelsWeekly Reading