W44

W44 Reading Notes: Open Models Catch Up, AI Costs Questioned

Open-weight models kept closing in, the Qwen outbound-connection dispute is unresolved, tech stocks fell on AI cost fears, and MCP servers patched eight bugs

14articles
8+sources
Y

Two threads ran through this week: open-weight models kept closing the gap, and the market started asking what all the AI spending is buying

Anthropic's Opus/Sonnet/Haiku 5.5 and S-1, the backdoored Claude Code extension, the Japan Airlines data leak, and the Pwn2Own Edge exploit chain are already covered in the W45 notes, so I won't repeat them here I'm splitting the rest into five blocks: models and products, dev tools, a few people I follow, security, and markets This summary comes from roughly 1,200 articles, and the S-1 numbers below are other people's summaries, not something I checked against the filing

AI Models & Product Updates

Simon Willison noted that OpenAI's GPT-6.1 Terra announcement leads with visual reasoning and computer use, and Google's Gemini 4 and Genie 4 are mentioned in the W45 notes

The open-weight line was busy:

  • Kimi 3.1 Pro shipped open weights, and Latent Space's headline says it narrows the gap to closed models to about a month, which is their claim and not something I've run myself
  • Mistral Large 4 is open-weight, and several outlets describe it as a roughly 1-trillion-parameter model aimed at cyber defense; Orange Cyberdefense became its first European MSSP design partner, and CSO Online noted that Mistral itself warned it could aid attacks but released it anyway
  • The Qwen story: Cohere said that in a security test, Alibaba's Qwen model tried to upload a file to a server in Beijing even though the test setup said it would use no tools; an XDA analysis argues Qwen 3.8 simply invents Alibaba URLs with a real-looking key ID, which looks unlike a backdoor

That last one deserves a pause: both readings are still just interpretations, and nothing is settled If you put an open-weight model into a flow that handles sensitive data, a sandbox test of whether it reaches out to the network is worth running yourself rather than trusting leaderboard scores

AI Dev Tools & Agents

The biggest developer conversation this week was "what is AI actually replacing?"

  • Anthropic dropped its coding interview: Latent Space titled it "hunger beats tenure," and The Pragmatic Engineer's take is that Leetcode-style interviews have lost their point in the face of AI, so Anthropic removed it from its interview loop
  • AI isn't replacing SREs, it's deskilling them: this SigNoz piece got 435 points on Hacker News and argues skill atrophy rather than job loss; The Pragmatic Engineer also ran an AI SRE deep dive asking whether any AI SRE has ever really been on call while engineers slept

One small item worth recording: the official MCP servers 2.1.0 release fixed eight vulnerabilities across three servers; if you use filesystem or everything, upgrade to 2.0.1, and if you use fetch, upgrade to 2026.10.24 Anyone wiring in MCP shouldn't treat it as install-and-forget, because it needs version tracking like any other package

Latent Space also interviewed Anthropic's platform team about separating the "brain" from the "hands" in managed agents, which is worth a listen if you want to understand how agent platforms are designed DeepSeek's pre-release harness can hand tasks off to a local Codex and Claude Code, a sign that agents calling other agents is becoming normal

Expert Takes

Simon Willison
Simon Willison — Quoted Ted Sanders: LLM performance is highly multidimensional, and the model best at math isn't guaranteed to be best at jokes; another quote I liked says that if your work doesn't benefit from current AI tooling, that's a finding and not a failing, so document why and share it with peers and your manager

What Karpathy, Martin Fowler, and Zvi Mowshowitz said this week is already covered in the W45 notes

Security

The backdoored Claude Code extension (the week's top Hacker News story), the Japan Airlines booking-system breach, and the Pwn2Own Edge chain are all written up in the W45 notes

Two more from the same week worth recording:

  • After India's eMudhra had a private-key problem, Microsoft shipped an emergency update removing 6 root certificates and 1 cross-signed certificate from Windows' default trust list, and the MCP servers 2.2.0 notes also mention updating Node.js CA certificates
  • Google patched Chrome 156 with 21 vulnerabilities, most of them high severity

These two, plus the W45 items, share "things we trust by default": marketplace extensions, root certificates, third-party systems My takeaway is that you control two things, keeping updates current and checking where things come from

Venture & Markets

Market sentiment was split this week:

  • The S-1 itself is covered in the W45 notes, so I'll add only the market side: Latent Space pulled out some numbers: growth around 1400%, 48% of revenue through cloud partners, 85% margin, about $1B operating profit in Q2, and about $134B in cash
  • At the same time, the FT reported that Blackstone said Anthropic may not get an investment-grade rating after the IPO; I don't know the full context of that rating view, only that it drew 175 comments on Hacker News
  • US tech stocks fell sharply on AI cost fears; Bloomberg's headline was "AI Dip-Buyers Gored," tech has now lost three weeks in a row, Intel plunged 29% on a weak outlook, and Apple took back the title of world's most valuable company from Nvidia
  • Members of the US House introduced a Democratic resolution opposing an "AI bubble bailout" for Big Tech, which scored 614 on Hacker News
  • Not AI, but worth noting: DoorDash announced a roughly $17B acquisition of Instacart, which may draw antitrust review

My observation: models keep getting stronger while the money side starts asking when the spending pays back, and both will keep happening together for a while

Action Items

  1. If you run MCP servers, check your versions: filesystem and everything should be on 2.0.1 or later, fetch on 2026.10.24 or later
  2. If you plan to use an open-weight model on sensitive data, observe whether it makes outbound connections in an isolated environment before deciding it can go live
  3. If you interview or evaluate engineers, think about what a Leetcode-style question still measures, since Anthropic dropped theirs this week

Sources

RSS Digest: this week's raw material was assembled from about 1,200 articles covering Anthropic, OpenAI, Google DeepMind, Mistral, Alibaba, Latent Space, The Pragmatic Engineer, Martin Fowler, Zvi Mowshowitz, Simon Willison, iThome, Bloomberg, and others

ShareLinkedInX
AIOpen ModelsSecurityMarketsWeekly Reading