ThoughtsOfChains
← News & Updates

Jev is here!

Jev is here! A new type of AI that is supposed to be cheaper, faster, and better for automation.

This model is inspired by the book "Thinking, Fast and Slow" by Daniel Kahneman, and is supposed to be a "System 1" of thinking.

First impressions

Jev works with some different concepts from your common LLM: It doesn't predict the next token, so its objective is a bit different. It can predict:

  • Noul : A 0-1 probability that the answer is yes
  • Score : A position on an ordered rubric
  • Choice : One option from a predefined set

This brings some advantages, since its outputs are "structured" and defined in advance, making it a LOT cheaper compared to current LLMs:
$0.042 / Million tokens (and free output), while GPT Sol 6.1 is ($2.00 input and $10.00 output per million for short context)
I'm not even looking at smarter models like Astra or Fable, because it's a BIG difference
Also, it's really quick (70ms - 500ms) and parallelizable, because it doesn't need to do autoregressive generation.

The public really likes this model (at least from what I see on X) and most comments online say that this new model is the new 'if' for computer programs.

if (call jev to take a decision with yes or no):
    something
else:
    something

Surprisingly, Jev can generate text. You only need to give it an entire vocabulary with the 'choice' type. It will generate text similarly to an LLM, and it can actually hallucinate doing this process.

It's basically a one-pass LLM trained to be more calibrated in its decisions. The output is similar to a softmax, but it tastes and looks like a 'fancy' and more general classifier.

It's important to note that it brings a new type of reinforcement learning with it.

(RLCD) Reinforcement Learning for Calibrated Decisions

Looks like the algorithm is not public, but it replaces RLHF (Reinforcement Learning from Human Feedback).

It rewards the model for attaching probabilities that are honest. So it trains the model to make 'calibrated' probabilities instead of generating text.

Example: If I have 50% confidence, this result must happen 50% of the time.

There's no public paper I could find, so no way for me to do an in-depth explanation.

Sources:

https://typesafe.ai/blog/introducing-system-one-models-and-jev https://huggingface.co/blog/sora-2/how-to-use-the-jev-ai-model-a-step-by-step-develop https://www.sanity.io/glossary/rlcd-reinforcement-learning-for-calibrated-decisions