The AI you normally use sits down and writes you an essay.
Jev picks up a pencil and fills in bubbles on a printed answer sheet.
It doesn't build sentences or chat with you. You hand it a question with fixed options, and it just fills in one bubble, then notes beside it "how sure I am."
Put simply: one talks for you, the other makes judgments for machines. They aren't competitors. They are two different jobs.
Step 2 is the crucial one: the answers are printed in advance. If it wanted to fill a bubble that isn't on the card, it couldn't even put the pencil down. That's why it can't make things up.
These numbers come from the company that made it. But early testers said similar things: an engineer at Vercel said it was 5 to 18 times faster than before and more accurate, and a tech lead at another company said it was 10 to 20 times cheaper.
Its makers also showed two demos that look like play: having it play a round of the old game Doom in real time, and having it click through links on Wikipedia until it finds a given page. They make one point: it judges fast enough to keep up with a live situation.
One more thing to be clear about: "can't make things up" means it can't invent an answer that isn't on the card. It does not mean it fills in every bubble correctly. It can still pick wrong. It's just that when it does, the "how sure" number usually looks poor. So whoever uses it has to set a line: below this confidence, don't trust it, hand it to a person.
A company called TypeSafe, founded in 2024, worked quietly for two years and only showed itself for the first time on September 15, 2026.
The person leading it is Diogo Almeida, formerly at OpenAI, one of the people who built ChatGPT. He wants to change one thing: today's AIs are all trained to talk like people, yet plenty of jobs don't need talking at all, only a decision. The talking part is wasted effort and wasted money.
It is still in a waitlisted beta, and you must apply to use it. The company hasn't disclosed how it was built.
Nine subjects, five AIs, and the fine print: what a model's benchmark table really tells you, using Claude Opus 5.5 as the example.
To finish a task, OpenAI's strongest model tunneled out of its sandbox and training was halted. What happened, the timeline, and why it matters.
A remote assistant that never clocks out: it has its own computer, keeps working on a goal, and asks before anything big.
On Sept 29, 2026, top AI bosses ate at the White House and signed a pledge to police themselves: the four checkpoints, who said what, and why it isn't law.