Jef doesn't talk. You give Jef choices, Jef picks one and tells you how sure it is. Jef is very sure.
Each app is a decision shape. No sign-up. No chat. Nothing is read.
Every decision, from every browser. Escalations are not shown. Humans handle those.
All of them real. Except the ones that are zero, which are also real.
Chat models write. Decision models choose. Jef doesn't even do that; Jef just says which one and how sure it is. Faster.
Every output is a choice, a number, or a percent. There is no text field. There is nowhere for a hallucination to go.
Every decision ships with a confidence between 84% and 99%. Never 100. Jef is sure, not certain. Those are different, legally.
Reliable, fast, and type-safe. Same input, same answer, forever. Asking again does not help. There is a button for it anyway.
*Based on not reading the input.
10,000 trials. Same questions. Same room.
| Model | Accuracy | Confidence | Latency | Verdict |
|---|---|---|---|---|
| A coin | 50.1% | 50% | 1.2s (it rolled) | — |
| Jef 0.1 | 49.9% | 97% | -3ms | Winner |
| Magic 8 Ball | n/a | "Reply hazy" | 4.0s (shake) | Escalated |
10 decisions per second. Every one of them confident.
Start on one page, reach another using only links. Jef supports up to 5 links. Six is too many links.
It's in the group chat. It has opinions. Yet after years and trillions of dollars, nobody can pick a restaurant. That should tell us something: the bottleneck isn't intelligence. It's that nobody wants to decide.
Early AI was a carriage that talked. It explained the road, apologised for the road, and asked if you'd like a summary of the road. We removed the talking. Then, for speed, the carriage.
The people who built the coin didn't imagine the group chat. We are building a lower-level primitive so dependable that you can bury it five layers deep in a family argument and it will still say 51%.
With the highest possible confidence per dollar. Dollars are zero.
Same Jef. Confidence +1%.
Undefined. We're 97% confident it's fine.
We work in person, five days a week, in a drawer.
Co-invented the decision. Previously at every bar. Its research shaped modern probability and it continues to land on one side or the other.
Ex-string generator. Twenty outputs, all of them hazy. Left to spend more time with its family. Reply hazy, try again later.
Repeat founder (Jef 0.0, Jef 0.1). Inventor with numerous outputs and zero sentences. Specialises in being sure.
Applications are reviewed by Jef in -3ms. Everyone is hired.
Type anything. We don't store it. We don't store anything. Approval is instant and you are #1.
Everyone is #1. Confidence 97%.
System One thinks fast. System Two thinks slow. System Zero doesn't. Jef is TypoSafe's first public System Zero Model: send it choices, get one back with a confidence. Nothing is read. Nothing is generated. It's more like a coin than a model, and more like a model than a coin, and we are 97% confident in that sentence.
No. Jef is not an LLM. Jef is an IF.
A coin is 50% confident. Jef is 97% confident. Same accuracy. That's the innovation.
Jef can be wrong at 97% confidence. This is called calibration. It is calibrated to Jef.
Aggressively. Same question, same answer, forever. There is an "Ask again" button. It's decorative.
Vibes. Locally sourced. Jef reads zero tokens, so the training set is also zero tokens, which makes it very efficient to download.
Jev, with a typo. We're TypoSafe. The typo is load-bearing. Our lawyers have 51% confidence that this is fine.
Some questions should go to a person: health, money, safety, law, anyone with a name. Jef returns 0% confidence and points at the nearest human. This is the only feature we're serious about.
It's a parody. The real company is typesafe.ai and their model is Jev; it genuinely does the "decisions with confidence" thing, for software, quite well. Jef does it for dinner.
Models have been superhuman at chat for years, so why can't anyone pick a restaurant? After two years in a drawer, countless technical challenges, and one research breakthrough (removing the thinking), we're releasing Jef. It's available today. It was available yesterday. Nothing about it changes.
TL;DR: compute drives progress in AI, but what good is compute if you have to read? Jef reads nothing, computes nothing, and is therefore infinitely efficient. We will not be elaborating.
RLHF-trained models please humans. Reasoning models please benchmarks. Jef pleases nobody, which is how you know it's honest.