logoalt Hacker News

mattrighettiyesterday at 4:37 PM6 repliesview on HN

Is distillation something we have to live with or are there ways to prevent it?


Replies

sosodevyesterday at 4:46 PM

Realistically you can't prevent distillation. OpenAI / Anthropic are slowly moving towards hiding the steps in-between input and output (hidden thinking), but that only helps so much. Imagine you put a file into Claude and say "do X to this" and it returns it to you without showing any of its internal reasoning. That's harder to distill, but the simple mapping of input to output still creates very valuable training data. It is reflective of all the training the model did to learn how to do that transformation.

show 1 reply
Cytobityesterday at 5:01 PM

You make it sound like a bad thing.

recursiveyesterday at 9:27 PM

To the extent that we have to live with AI and LLM, I believe we have to live with distilling also.

HarHarVeryFunnyyesterday at 5:13 PM

You can prevent it by outputting a reasoning "summary" instead of the actual reasoning trace.

Which Anthropic already do.

sowbugyesterday at 6:03 PM

If you build a device that can help build devices, you shouldn't be too surprised when people use it to build devices.

epolanskiyesterday at 5:06 PM

If it was genuinely useful, we would've long reached the point where you train a model on a previous one's output in an ever improving loop.

But this doesn't actually work.