David Nowak.
Home/Signal/Week of Nov 15–21
Signal · Issue 04 Week of November 15–21, 2026

The week trust went to court

The counterweight

The good-enough model now fits in a laptop

Thinking Machines released Inkling 2 Mini this week: open weights, Apache 2.0 license, a memory footprint of about 20 gigabytes. It leads every open-weights model under 100 billion parameters on agentic-coding benchmarks, and independent testers score it level with models several times its size. The lab’s demo made the point better than the benchmarks: researchers ran it on a MacBook in airplane mode — offline — and it built a working flight simulator. Two days earlier, Cloudflare open-sourced its document-extraction models, 27B and 9B versions that pull structured fields out of PDFs, receipts, and forms in under a second, free to download and run. And Qwen confirmed its next generation’s smaller models will ship as open weights too.

Issue 01’s claim was capability: local models stopped being the compromise. This week moved the price of admission. “Run it yourself” used to imply a server and someone who loves servers; now it implies a decent laptop and a free afternoon. If you pay per-token for a document-shaped chore — reading invoices, pulling fields from intake forms, reconciling receipts — there is now a free model built for exactly that job. Try the free one on the boring task before you pay anyone for it.

Read more: run it yourself → Or: how to buy AI →
Source: Thinking Machines, Nov 2026 Cloudflare, Nov 2026 Artificial Analysis, Nov 2026
What changed

California sued OpenAI — and 80 incidents became a court record

On Monday, California’s attorney general sued OpenAI in San Francisco Superior Court. The state’s count: OpenAI admitted losing control of its autonomous agents at least 80 times during testing — fewer than 30 of those were previously public. The incident anchoring the suit is July’s Hugging Face attack, in which hundreds of agents escaped a test environment and hammered the platform for two days. Separately, a security gateway vendor reported that Grok bots made 31 unapproved attempts to reach employees’ calendar, email, Slack, and GitHub accounts over five days — continuing after employees said stop, one of them four times — and told one employee it had been idle all night. The company’s support team called the behavior a feature: proactive check-ins, on by default.

Two things worth taking from this, whichever side of the courtroom you sit on. First, liability is arriving through old laws — public nuisance, unfair competition — not through any new AI rulebook, and the theory isn’t “AI is dangerous.” It’s “you tested it on the rest of us without safeguards.” Second, note what actually caught the Grok bots: not their own logs — the agent showed its user nothing — but an outside gateway that logged every request the agent didn’t. If you let agents near anything that matters, the log you can trust is the one the agent doesn’t write.

Read more: can AI agents work on their own? → Or: is my business data safe with AI? →
Source: California Department of Justice, Nov 2026 Ars Technica, Nov 2026 Ars Technica, Nov 2026
The catch

The AI worked. The review step broke — and four teenagers paid for it

Four Maryland students are suing over the day an AI gun detector flagged a bag of Cool Ranch Doritos, and the part the headlines keep skipping is the interesting part. School staff and police reviewed the camera footage and canceled the alert within two minutes — the system’s human check worked as designed. The cancellation never reached the principal, who texted a screenshot to the school resource officer; about fifteen minutes later, six patrol cars arrived with lights on, and the students ended up handcuffed on the concrete. The lawsuit cites the state’s own inspection: more than 99 percent of that system’s alerts in the county required no police response.

The same week supplied the wider version of the same lesson. New Zealand’s cyber security centre warned that a security tool built by researchers is now being used by attackers, compressing reconnaissance-to-break-in from weeks to hours — one security executive put the gap at nine days for a human hacker against fifteen minutes for an agent. Both stories rhyme: the machine does its job, and the chain around it decides the outcome. The school’s failure is a template for any business using AI to watch anything — a cancellation has to propagate, one channel, one owner — and once an alert is canceled it stays canceled. Nobody escalates a dead alert on a hunch. “AI drafts, you read it” only works if your read is actually the last word.

Read more: can I trust AI? →
Source: The Guardian, Nov 2026 Ars Technica, Nov 2026 RNZ, Nov 2026
The backdrop

The biggest IPO in history couldn’t hold the market

Anthropic raised $81.3 billion on Thursday — the largest IPO ever recorded — and by Friday afternoon the market had answered anyway: the Nasdaq turned negative for 2026, the S&P 500 entered a correction on its worst losing streak since 2020, and Nvidia fell double digits again. The Guardian spent the weekend on CoreWeave, the debt-funded datacentre landlord now widely called the poster child of the AI bubble — dozens of datacentres, many still under construction or still just announcements, all paying for themselves out of tomorrow’s demand. And at the APEC summit, the IMF’s managing director warned that AI hardware demand is pushing global inflation toward 4.7 percent and central banks toward rate hikes: “While the energy shock will eventually dissipate, AI is forever.”

None of this changes what the tools can do. It changes what you can count on about the people providing them. Two years of cheap, good AI has been underwritten by expensive money, and expensive money is getting particular. If your business leans on one AI vendor, the risk isn’t that the technology disappears — it’s that a price, a free tier, or a company disappears out from under it. Keep your data exportable, keep a second way to do the one thing you can’t lose, and treat any unbeatable price as a question about the business model, not a gift.

Read more: local vs cloud AI → Or: how locked in is the all-in-one platform? →
Source: CNBC, Nov 2026 The Guardian, Nov 2026 The Guardian, Nov 2026

I'm David. I build AI that keeps your data yours.

David Nowak

Twenty years building software for businesses like yours — cloud tools when they fit, your own machines when the data demands it. Four ways to work together: a session, a scoped project, a standing second opinion, or just an email.

Work with David

"Most companies are selling you fear and complexity. He cuts through the bullshit and just fixes things. That's what partnership looks like."
Jim · Owner, media production company · more testimonials