JEV / chat

jev as an llm model. it is not good.

how jev talks

  1. 01 / read 
    you: tell me a story
          │
          ▼
conversation + reply so far
  2. 02 / ask

    jev asks one question:

    “which single word comes next?”
  3. 03 / split

    the 1,020 candidates are divided into four parallel groups.

    bucket a  [255 words] ─┐
bucket b  [255 words] ─┤
bucket c  [255 words] ─┼─▶ jev scores each group
bucket d  [255 words] ─┘

    jev can also borrow a few words from your messages.

  4. 04 / shortlist

    the best 48 choices from each group move forward.

    4 groups × 48 choices
          │
          ▼
up to 192 finalists + <eot>
  5. 05 / pick

    the finalists get compared together, then jev makes a weighted random pick.

    the       ██████████  41%
moon      █████       21%
is        ███         13%
<eot>     ██          10%

weighted draw ───▶ moon
                     │
reply: “the moon” ◀──┘
        │
        └────────────── ↺

    recent words get pushed down so jev loops less.

  6. 06 / stop 
    <eot> is picked
      +
probability ≥ 10%
      │
      ▼
reply ends

min 2 words  /  max 24 words
one more number

end-of-chat is a separate check made after the reply. it guesses whether the whole conversation feels finished.

one word at a time. 1,020 candidates. two rounds. still strange.