TypeSafe · Jev

A new model landed. It cannot write.

It only makes decisions. That turns out to be the whole point.

TypeSafe came out of stealth on Tuesday with forty million dollars and a model called Jev. Its founder co-wrote the RLHF paper at OpenAI. The Hacker News thread ran past nineteen hundred points, which never happens.

Here is the strange part. It does not write. Not a word. You hand it a blob of state and typed questions. Pick one. Score this. Is this true. The answers come back with a probability on each, all at once rather than one word at a time.

TypeSafe’s own numbers are forty to two hundred times faster than a frontier model on the same job, and four hundred times cheaper. Under half a second, output free. Built for decisions, not for language, and that is the whole idea.

Every creative workflow I have seen is a few generative moments wrapped in hundreds of small decisions. Which of these four hundred assets are even eligible. Does this cutdown carry the legal line. Does this one need a human. None of that is taste. All of it was a person.

The claim doing the work is that it cannot hallucinate. True in a narrow sense the founder spelled out on Hacker News. Type safety is not factual correctness. On their own evals it loses to the frontier models on all four workflows, and they published their own failure modes two days later. I like that.

Higgsfield · Open

Higgsfield did not open source anything.

I believed it for two days. Here is what actually shipped.

The story going round is that Higgsfield gave its platform away on GitHub, weeks after a four hundred million dollar round. I read it, I believed it, and I had a whole opinion ready about what it meant.

It is not true. What is on GitHub is client side tooling under an MIT licence. A command line tool, SDKs, editor plugins, a skills pack. Every one needs an account and an API key, and calls the paid, closed service. No weights, no inference code, nothing you could run yourself.

What actually happened on Wednesday is ordinary and commercial. They opened their API to anyone with a card. The reason we got it wrong is on their own product page, headlined OPEN HIGGSFIELD API. Open meaning open for business.

I have form here. I spent an evening in their product back in May and said it felt like the whole Cheesecake Factory menu at once. They build fast and market faster. This week the marketing outran the facts and I went with it.

Voice · The missing piece

The mouth and the brain are separate products.

Two launches, five days apart, and the Jarvis problem underneath them.

OpenAI made GPT-Live 1 generally available the Thursday before last, and Google answered on Tuesday with Gemini 3.8 Live. Interruptions, background noise, turn taking. The things that make talking to a machine feel less like operating it.

The detail worth knowing is architectural, and I had it wrong before I looked. GPT-Live 1 has no tools of its own. It is a voice layer that hands the thinking to a separate model and keeps talking while that model works. The mouth and the brain are two products now.

Think about a hotel front desk. Warm, quick, never stuck for a word, and completely unable to tell you what time your appointment is. Jarvis was never impressive because of the voice, but because the talking, the thinking and the doing were one thing.

Anthropic is not in this race and I have said so. Their own docs list no audio in or out on any current model. But they are hiring an audio team and a voice platform team, and one of those posts describes building the voice API surface. Late is not the same as absent.

Three things this week. One to scan. One to test. One to notice.

What I’m working on this week

Not a tidy week, much of it away from my desk.

The part worth writing down is two days in the city with a friend who invests for a living. One question, most of two days. What does venture capital look like in an AI first world, when the cost of building collapses and the moat moves.

The bit I keep chewing on is the blend. Art and science. Creativity and intellect. EQ and IQ. Every model I have seen for picking companies leans hard on one side, and I doubt that survives. No conclusions yet, which is usually a good sign.

I also wrote the one about deleting things, the argument I have with almost every client before we build anything.

• Claude now leads a quarter of Anthropic’s own AI research. Their own measurement puts it at 26% of their R&D work in August, against under one percent in February. They published that about themselves.
• Canada and Germany are funding an AI built to disagree. A hundred and fifty million Canadian dollars and a hundred million euros went to Bengio’s LawZero on Wednesday, for a system built to check others, not please you.
• Sora 2 and the Videos API leave the OpenAI API on the 24th. That is Thursday week. If anything you have built or sold touches either, make it a diary entry rather than a surprise.

Touch Grass

Five people asked me the same question this week.

Friends, people I work with, a couple in between. All asking a version of the same thing. What does this actually mean, and how worried should I be. I do not have a clean answer and I have stopped pretending one is coming. What I have is a frame, from somewhere I was not expecting.

Anyone who has been near a twelve step room knows the Serenity Prayer. The serenity to accept what I cannot change, the courage to change what I can, the wisdom to know the difference. Built for people trying to stay alive, and it works unreasonably well here.

Almost none of this is mine. The pace, the labs, the policy, the outcome. What is mine is what I make, who I help, and whether I stay curious instead of frightened. Shorter list, and all of it is something I can do on Monday.

So. Deep breaths. Nobody controls the future, and it has not decided what it is yet.