This week's AI news was hot, but the more I read, the more I wanted to cool the room down
GPT-5.6 using a prompt to help close a 30-year gap in convex optimization is the kind of story that makes people excited. In the same week, we also had graphs of what AI did to Stack Overflow, "The LLM Critics Are Right. I Use LLMs Anyway," circular financing around the GPU boom, and NYC considering AI disclosure for real-estate listings
My read: AI is genuinely useful. But the more useful it becomes, the more it needs disciplined engineering, explicit risk labels, and ordinary market skepticism
AI Models & Products
GPT-5.6 being used to advance a convex optimization problem is a strong capability signal. It reminds me that large models can still surprise us in math, search, and conjecture support. But this is also the kind of story people overgeneralize. Helping an expert move a problem forward is not the same as replacing oversight in every high-risk decision
Fable 5 vs GPT-5.6 Sol on an NP-hard problem is useful as a trend indicator, but not as a direct production decision. The real question is whether your task has ground truth, can be rerun, and can be covered by negative tests
AI listing disclosure moving into city policy is important. If NYC requires landlords or brokers to disclose AI use in listings, AI-generated content has moved beyond platform policy and into regulation
AI Dev Tools & Agents
The graph of what AI did to Stack Overflow was the developer story that stayed with me. AI is not only replacing answers; it is changing the economics of knowledge communities: who asks, who answers, who gets visibility, and who still has an incentive to maintain high-quality public answers
Setting up a spare Mac for Claude Code to control shows agent workflows growing into a physical operating pattern. This is no longer just an IDE assistant. It is an AI operating an isolated machine, running longer jobs, and preserving work context. That makes environment isolation more important than before
Claude Code / OpenCode token-footprint comparisons also matter. Agent cost is not abstract: context size, pre-read strategy, and tool-call behavior all turn into latency, spend, and data exposure
Expert Takes
"The LLM Critics Are Right. I Use LLMs Anyway" captures the stance I find most useful. The critics are often right: hallucination, overconfidence, low-quality content, and governance risks are real. That does not make the tools useless. The right posture is neither belief nor rejection; it is placing AI inside a verifiable workflow
"Inviting hard questions" keeps appearing in the digest, and I read it as a vendor-maturity signal. An AI company that only wants to talk about capability, not risk, should not sit at the center of critical workflows
VC & Markets
The circular financing around Nvidia, CoreWeave, and Nebius was the week's most important market warning. AI infrastructure demand is real, but when suppliers, customers, and investors are tightly entangled, demand signals can be amplified
"AI Mania Is Eviscerating Global Decision-Making" is worth reading as a counterweight. The danger in a boom is not that people believe AI will improve; it is using that belief to skip finance, governance, risk, and basic product judgment
Action Items
- Use AI where verification exists: math, code, and data workflows can benefit, but answers need to be checkable
- Do not treat miracle cases as the baseline: breakthrough examples are signals, not procurement specs
- Isolate agent work environments: long-running jobs, remote machines, and repo access need boundaries
- Separate demand from financing structure in the GPU market: growth can be real and froth can be real at the same time
Sources
RSS Digest: see research/digests/2026-W29.md (276 articles this week, curated across AI capability, agents, and market risk)
Main sources: Anthropic, Hacker News, Meta AI, The Batch, Google Cloud Blog, Every, a16z, Reuters