Field Note: The open-source argument I'd been missing

In June I wrote about Banco Santander open-sourcing its AI tooling — the bank published the code that tests whether its own models discriminate against people. I gave the usual reasons that’s a good thing: outside scrutiny, verifiable claims, a governance document you can read for yourself.

Jofish Kaye wrote back with a better one. He’s a friend, a subscriber, general mensch, and he’s VP of Research at Inflection AI, so he’s seen this from inside more than one large organisation:

Most every argument about open source runs either on ethics or licence cost. His is about staffing — and it’s arguably the one that can survive contact with a finance director. A proprietary system means everyone who can work on it has to be taught it on your time. An open one means the skill already exists out in the world, and you’re hiring from a pool instead of a puddle.

It applies below enterprise scale too. When something in my setup breaks, the answer usually already exists somewhere, because I’m running the same tools as everyone else. Had I built it all bespoke, every problem would be a brand new problem.

His site is jofish.com, which you should definitely check out — even though I’m biased!

What’s in the news

​From custom code to conversational prompts — Grok now lets subscribers build a whole app by describing it. That’s the third or fourth handoff in a decade of who’s allowed to build software, and the arc tells you more than the tool does. I don’t generally talk about Grok or any Musk-driven projects, but the gap to creating software on a personal level continues to narrow, so it’s interesting to watch.

​OpenAI published its homework on exactly the question I keep asking — not “can an agent be tricked into leaking a password”, which is well understood, but what an agent can reach once it’s already authenticated and trusted. Their Codex Security release is an answer, and it’s worth taking seriously because it reads as an admission rather than a victory lap.

​The share button is a publish button — around 600 Claude conversations turned up in Google and Bing. Nobody was breached. People clicked Share, and the page had no noindex on it. API keys and wallet details among the contents.

​Donkey work doesn’t need a genius — a nine-billion-parameter open model, about $500 of GPU time, beat every frontier model they tested, on one narrow job. But on the other hand…

​Kimi K3 is open. I still can’t run it. — open weights existing and open weights being usable by an ordinary person are two different claims, and most of the coverage only established the first.

What I’m reading

While I'm on the subject: Jofish’s own team has published research worth your time. It’s a three-part study of who is actually using AI chatbots in 2026, and, unusually, of who isn’t.

• ​Who’s actually using chatbots in 2026? — the survey
• ​Five kinds of chatbot users — and the roles people want next
• ​Taking “no” seriously: the frustration is real — the non-users
• ​The full report​

If you’re only going to read one (shame on you!) then I suggest the third one first. Asking people who avoid AI why, properly rather than to score a point, is rare, and the answers are open and honest.

— Jim