Question 1

Does Talk to Jev contain a language model?

Accepted Answer

No. The only model is TypeSafe's Jev decision model, which can only choose from given options. Every word is selected by classifier questions, with no template, phrase bank or generative LLM in the loop.

Question 2

How do I run Talk to Jev locally?

Accepted Answer

Create a Python virtual environment, install requirements.txt, put an OpenRouter key in .env, then run server.py; the README shows the three commands and notes the demo opens at http://localhost:8787.

Question 3

How is generation broken into individual questions?

Accepted Answer

The same primitive is asked at coarser grains. At reply level: what are they doing, how much should I say, how do I feel about it, should I look something up. At sentence level: what the next sentence does, who it is about, how much it carries, is the reply complete. At word level: one of about 150 function words, one of your words, or a kind of word, and when unsure it imagines one word further and picks the best continuation. Digits are chosen one letter at a time.

Question 4

How does it remember the conversation and use tools?

Accepted Answer

Each answer is written back into the text the next question reads, which is the memory in generation as repeated classification. It holds its mood and what it said about itself across turns, says your own words back including names, does small sums digit by digit, and uses tools the way it does everything else: by being asked a yes-or-no question per tool, such as whether it would help to look up the time, and choosing.

Question 5

What does one reply cost and take?

Accepted Answer

The README is blunt about it: a decision takes about 0.3 s and a few thousand input tokens. A short reply is a dozen decisions and roughly two seconds, while a five-sentence reply can be several hundred decisions and up to a minute. It needs a paid OpenRouter key to reach Jev, served there as typesafe/jev-1.13 through OpenRouter's Decisions API.

Question 6

What happens when the output goes wrong?

Accepted Answer

After a reply, Jev is asked whether it is grammatical, relevant and natural, which sentence is weakest, and how the speaker feels now, and weak sentences get rewritten. When the top pick falls below 0.65 it triggers confidence-driven lookahead. The stated design rule is that the fix is always more information in the state or a better description of an option, never a new rule, and nothing is forbidden.

Question 7

How faithful is this repository to the paper?

Accepted Answer

The README says plainly that this is one implementation of the idea, written quickly and iteratively, not a one-to-one rendering of the paper and not tidy. The paper, Generation is Classification, is on Zenodo at doi 10.5281/zenodo.22940945, and this archived code at doi 10.5281/zenodo.22895303. Credits list Claude as the engineering collaborator with the human deciding at every step.

Talk to Jev

About this project

Highlights

Questions about this entry

Related terms

More Jev resources like this

From the Jev guides

From the Jev community