Maxwell: a new model that speeds up AI models like Whisper and LTX-2 by rewriting their binaries [D]

Came across this today and haven't seen it discussed here yet. It's apparently a new type of model (called Maxwell) that sits under a running program and speeds it up on the fly based on what it sees (hardware and data). The cool thing is that it doesn't change the source code, but instead it modifies the binaries (during runtime!). So kind of like a JIT, except it's a model working out what to change instead of a fixed set of compiler heuristics. What got me is that they seem to be using formal verification to prove the optimized version still does the same thing, which I guess is the only way you could let something like this loose with no one checking it. No idea how that holds up with floating point or anything nondeterministic, to be honest