TypeSafe AI’s New Model: Jev

LLMs have taken another massive jump forward.

I keep saying I am less bullish on LLMs and more on vector DBs.

Then something like Jev by TypeSafe AI comes along, and I have to eat my words.

The same thing happened when Tool Calls and MCPs came out.

Jev and Reinforcement Learning for Calibrated Decisions (RLCD) take this to the next level.

Models like Jev get fed structured program state instead of sequential messages and spit out Type-safe structured values.

Have you tried to get a legacy model to follow schema rules? Many of them struggle to come close to following your output rules, as training them to do so was an afterthought.

With Jev, that wasn’t just a side feature. It was the sole focus.

Most LLMs generate tokens in sequence. This means that your system prompts and earlier context can get watered down as the conversation goes on. This allows the model to go deep down rabbit holes you had not originally intended it to go.

Jev runs in parallel, generating all outputs in a single query. This means, hypothetically, that it will give appropriate weighting to all input, not just the latest in the sequence. Additionally, according to them, this is more efficient.

Because of the way this model functions, they boast an end-to-end response time of 70ms-500ms for TypeSafe inference.

Considering even the smallest LLM models I have wired into production have added several seconds of average latency, this is incredible.

It also outputs a confidence score with each of its decisions, which gives you a better indicator of when it's hallucinating.

Considering just yesterday I asked a model to give me links to support its arguments when it was clearly gaslighting me, and it just made up a link that went nowhere, this would be a welcome feature.

I can’t wait for this tech to be available on AWS Bedrock so I can play with it at scale.

Personally, I think this approach is clever and could yield really interesting results. The problem they are solving is a real problem… for existing AI models.

What do you think? A significant jump forward or just more AI hype?