Artificial Intelligence1 min read

Show HN: Watch weight-space programming in action with WebGPU

By · Published by Everything Blog

In short

In the demonstration, a small Qwen base model depleted its operational budget by endlessly re-checking its outputs until intervention occurred. An external model altered 131,072 parameters across 64 neurons located within layers 15 and 19 to successfully halt the repetitive cycle. Modifying just 0.016% of the roughly 800 million total weights allowed the system to finish generating its responses. This technique establishes an iterative framework where models can autonomously propose and test their own internal parameter updates.

Key points

  • Weight-space programming targets specific failure cases rather than relying on broad benc…: Weight-space programming targets specific failure cases rather than relying on broad benchmark-driven weight merges.
  • Modifying only 0.016% of a model's weights can successfully resolve severe behavioral iss…: Modifying only 0.016% of a model's weights can successfully resolve severe behavioral issues like infinite loops.
  • Autonomous AI agents could adopt an edit-test-repeat loop to continuously identify and re…: Autonomous AI agents could adopt an edit-test-repeat loop to continuously identify and refine their internal weights.

Sources checked by our analysis pipeline: 1

Press play. Thinking text appears in grey, the final answer in dark ink.

The base model gets stuck in a loop

With thinking on, small Qwen keeps re-checking itself — “wait, let me verify…” — until it runs out of budget instead of answering.

An LLM edits 0.016% of the weights

It subtracts one direction from 64 neurons' output weights (ffn_down

) in

layers 15 and 19 — 131,072 of ~800M weights.

Stripes drawn ~25× wider than true scale — at real size they would be invisible.

The edited model escapes

Same model, same prompts — it now stops looping and finishes its answers.

Where this might fit

Model souping - diffusion-LoRA merging is one case of it - is weight-space editing too. The difference is the target: souping picks merge weights (and sometimes merge order) with a broad benchmark; weight-space programming zeroes in on a few specific failure cases and the neurons behind them. And here's the fun part: it's a promising fit for agents - propose an edit, test it, feed the result back - the same edit-test-repeat…

Artificial Intelligence