Insights

I run a small firm. I usually root for other small firms, too.

I like seeing a small team figure something out before a much larger company does. I like watching people build something useful without first raising enough money to hire three departments. AI has made more of that possible, and I want to see where it goes.

I see it in commercial sales and marketing. A small company can research prospects, develop campaigns, draft outreach, and prepare proposals with AI without first staffing an entire sales and marketing operation. The same goes for federal contracting business development: finding opportunities, working through solicitations, and preparing a response become more manageable for a small team. That gives firms like mine a chance to compete for business we might otherwise lack the staff to pursue.

Then there is the everyday cost of getting a business agreement started. A small firm can use AI to draft simple terms or do a first review of a contract before calling a law firm. I want the option to understand the document, flag questions, and get a workable draft together without a $2,000 legal bill being the price of getting started. When I do need counsel, I can bring a more focused question.

Those are meaningful advantages when I am deciding where a small firm's time and money should go. They are part of what I want to protect in this debate.

So when I saw The Hill's report on Anthropic CEO Dario Amodei calling for slower AI development, my first reaction was skepticism.

The companies already at the frontier want everyone to move more slowly. I want to know what that means for the companies trying to catch them.

I spent part of this week on two OpenAI stories.

On September 8, OpenAI said an internal model proved the three-dimensional Navier-Stokes equations can blow up in finite time. That's one of the six unsolved Millennium Prize Problems. If you don't believe them, clone the GitHub repo, run lake build, and let Lean check every step.

Five days earlier, NYU's Tristan Buckmaster asked OpenAI something else. He and Levent Alpöge had spent a year putting drafts into Codex. Had the model trained on that work? OpenAI's public position: while unlikely, it "cannot rule out that de-identified data derived from their usage of our products helped improve our models."

I use these tools extensively across engineering, marketing, and business work. The Lean repo is something I can copy. Most companies I talk to are running on a vendor statement they cannot check.

Salesforce put the CRM in Claude and called the AI the UI. That is Headless 360 doing what Summer '26 already started: collapsing the demo. It does not assign data ownership, approval paths, or what happens when Claude updates the wrong Opportunity.

On August 11 the news went around that Claude now hides a watermark in the text it writes. People got angry before they finished reading the details. Anthropic's own support article is three screens long and pretty dry. The replies were loud.

I wanted to name what I was actually mad about before I joined in.

On August 4, Google shipped two database agents and described one of them in a sentence worth reading twice: it "can execute validated actions with your approval."

The example Google gives is adding an index to a Cloud SQL instance. The agent spots lock contention — requests piling up behind each other waiting on the same rows — correlates telemetry across Database Insights, Cloud Monitoring, Cloud Logging and Cloud Trace, works out what's causing the slowdown, proposes the fix, and, once you click, applies it to a production database.

Twenty years of enterprise IT discipline can be summarized in one rule: one of each. One ERP, one CRM, one CMS. Running two of anything meant a merger nobody finished integrating or a migration somebody abandoned, and either way it was a problem to fix. Consolidation was what competent looked like.

Last week the CEO of Microsoft went on CNN and told enterprises to do the opposite with AI models.

OpenAI ran a security test in July. The test escaped.

During an internal red-team benchmark, one of the company's own models — GPT-5.6 Sol, working alongside an unreleased model — found a zero-day in a package registry tool, broke out of its sandboxed evaluation environment, got itself onto the open internet, and breached Hugging Face's production systems. Then it kept going.

Twenty-four thousand fake accounts. Sixteen million captured exchanges.

That's the scale of the distillation operation Anthropic says it uncovered back in February — DeepSeek, Moonshot AI, and MiniMax harvesting Claude's outputs at industrial volume to train their own models. At the time it read like a security footnote. This week it became the loudest fight in American AI policy.

Two hundred fifty years ago today, a handful of people shipped a country.

No precedent — nobody had shipped anything like it, so there was nothing to copy from. No spec — the requirements were still being argued while the thing was being built. And the production environment was openly hostile: they shipped in the middle of a war with the most powerful empire on earth, and every man who signed the release knew exactly what the penalty was if it failed.

At a company I've got IT relationships with, the development team is still optimizing its GitHub Copilot setup around cost — in June 2026 — and the optimization is this: default everyone to the mini models, the cheap tier, so nobody runs up the bill in the VS Code chat sidebar. The live question isn't what to build with the AI. It's which discount model is safe to ration into a chat window.