Sunsetting decomp.dev
52 points by FedericoSchonborn
52 points by FedericoSchonborn
"What if I could extract the value earned from collaborating with other people without actually having to collaborate with people?" seems to be the running anti-social thread through most applications of LLMs. Speed is the only thing that matters; output is the only value generated.
As someone deeply invested in this scene, I'm saddened, but the writing has been on the wall for a couple months now. LLMs are (predictably) very good at generating matching C code from assembly once you give them the right toolchain. The Super Smash Bros. Melee decomp is evidence of this: they pushed the last few percentage points with LLM-generated output (I am not sure how many, exactly, since I'm on the outside of that project looking in). The distinction, however, is that the maintainers meticulously hand-reviewed the output for quality.
Many AI-driven decomps / recomps / ports aren't even doing that much. They point the harness at the repository and let the agents get to work. It's been very damaging to the collaborative aspect of it. I worry about how the hobby might recover from this, if it even can.
Moreso, though, I worry about how corporations will respond to this new world where any closed binary can be ported to available source code for what amounts to pennies. The Reverse Engineering scene has skirted on tenuous legal precedent for a long time, purely on the basis of the amount of effort required to do it. I fear that that era may be over.
decomp.dev's repository has never had an open source license
Lame.
Is it? From reading the article, it looks like the entire community has been ripped to shreds by AI trained on the very open source projects it once nurtured. Totally understandable to wish that the source of the platform itself remain closed in the wake of that (very likely gut-wrenching) outcome.
Just because AI has hurt the community doesn’t justify hurting the community more by threatening to sue anyone who tries to fork the website.
And AI training doesn’t (and shouldn’t) violate copyright, so it’s not like an “all rights reserved” label protects from AI training.
And here I am barely able to do basic stuff in Ghidra for CTFs. News like this makes me wonder what's the point of it all...