weight-edit — Qwen3.5-0.8B before and after a weight edit — openbash

One math problem answered by Qwen3.5-0.8B and by the copy an AI agent edited to stop its thinking loops, side by side, with the edit's signal every 64 tokens.

weight-edit

weight-edit is a command on the openbash shell, reachable over HTTP, SSH and MCP with one key. It composes with the other programs through pipes; there is no per-program endpoint and no schema for the join.

weight-edit run "Convert the point (0,3) in rectangular coordinates to polar coordinates." --max_tokens 4096

help

weight-edit: one math problem answered by Qwen3.5-0.8B as released and by the copy an AI agent edited, as JSON.

THE EDIT
  An agent at MIT CSAIL edited 131,072 weights of Qwen3.5-0.8B
  (RikiyaT/Qwen3.5-0.8B-weight-space-programming). The edit takes one
  direction out of the residual stream in layers 15 and 19. Its neurons fire
  about 4.5 times as hard on repeated text as on new text, so it stops the
  model's thinking loops. On MATH-500 with thinking on, the score goes from
  7.9% to 48.1%. With thinking off, the model card's default, it goes from
  43.6% to 45.2%.

RETURNS
  One JSON object on stdout with `unedited` and `edited`, each holding the
  model's thinking, its answer, the last \boxed{} in the answer, the tokens it
  used, whether it finished before the limit, `loop_at` (the first token of a
  64-token stretch that is mostly repeated, or null), and `bins`: the edit's
  signal in layers 15 and 19 and the repeated share, every 64 tokens. Also
  `price_usd`, and `standing`, which says how many free runs are left.

OPTIONS
  --thinking false   answer without thinking (default: thinking on)
  --max_tokens N     each model's budget, 256 to 8192 (default 6144)
  --seed N           the sampling seed, the same for both models (default 0)

PRICE
  $0.05 a run. An account that has never paid gets three runs free; after that
  the run is refused until $2 of credit is added, and the `credits` tool gives
  the link. A warm run takes about 45 seconds; the first after a quiet spell
  waits about three minutes while the models load.

READ THEM AS ONE SAMPLE
  One sampled answer per model. Show both answers as they are; one run says
  little about which model is better at math.

Studies:
  run <problem>
      both models' thinking and answers, whether each finished, where a loop began, the edit's signal every 64 tokens, the price and how many free runs are left

JSON to stdout, so it composes: pipe to `jq` to pick a field, or `> /tmp/…` to keep it.
Run a study with no arguments — `weight-edit run` — to see what it takes.

What it answers

weight-edit run — both models' thinking and answers, whether each finished, where a loop began, the edit's signal every 64 tokens, the price and how many free runs are left