News for hackers. Stuff that matters.
Re:500B Tokens Later: Letting AI Agents Decompile a First-Pe... by burcsahinoglu on Sunday October 11, 2026 @01:38PM (#50045712)
In my project I've been practicing a similar loop as you describe: an agent writes code, SMT solver checks against a spec. A few things I observed:

1. Solver is great at the bugs tests miss (overflow, off-by-one, empty inputs...), its output can be used by the agent to fix it.

2. Solver struggles where you'd assume: loops, non-linear arithmetic and pointer aliasing blow up solve time, floating point is painful.

So your step 2 would cover some of it well, but since the game code is a lot of floating point, SMT would probably have a hard time (get stuck/timeout).

Re:500B Tokens Later: Letting AI Agents Decompile a First-Pe... by hedgehog on Monday October 12, 2026 @12:47AM (#50051491)
It depends, the way I set the problem up I define equivalence as Z3 finds a solution within an rlimit of 8M, anything that fails gets machinery added to subdivide or reduce the problem (e.g. assume an index is in the bounds of the buffer it's associated with and add a bounds check). Floating point will in many cases require software emulation but I think that's unavoidable. That might be less of an issue if source and destination are both fairly conformant to IEEE 754 and use modern sizes.