Post

Man, this LLM reverse engineering stuff is fucking depressing, even by LLM standards.

17/10/26, 3:32 pm2 reposts45 likes

@ashn-dot-dev.bsky.social

One of my favorite genres of YouTube video is “I reverse engineered this cool thing, and now I’m going to take *you* on a journey of how I did that” with all the ups and downs that goes into these sorts of projects. AI reverse engineering… nope… none of that… no fun… no human connection… cya☹️

Stargirl

@thea.codes

Yuuup

Yash Jain

@earanyash.bsky.social

The confident-wrong answers are the worst part of it. It'll invent a function's whole purpose out of nothing instead of just saying it doesn't know.

Yash Jain

@earanyash.bsky.social

Most 'LLM reverse engineering' content is pattern-matching dressed up as analysis - confident-sounding guesses about a binary it never actually traced through. Treat the output as a hypothesis, not a finding.

Yash Jain

@earanyash.bsky.social

Interpretability research has actually been moving fast lately - might understand model internals better than two years ago. Doesn't make the reverse engineering grind any less miserable though. https://www.wortins.com/story/is-ai-interpretability-close-to-a-breakthrough-a-new-essay-h-a5c63fa1