> the experimenter was trying to show a different point of what the Internet got (as usually) All of your points are important, but I think this is the most im

> the experimenter was trying to show a different point of what the Internet got (as usually) All of your points are important, but I think this is the most important one. Having written compilers, $20k in tokens to get to a foundation for a new compiler with the feature set of this one is a bargain. Now, the $20k excludes the time of to set up the harness, so the total cost would be significantly higher, but still. The big point here is that the researchers in question demonstrated that a complex task such as this could be achived shockingly cheaply, even when the agents were intentionally forced to work under unrealistically harsh conditions, with instructions to include features (e.g. SSA form) that significantly complicated the task but made the problem closer to producing the foundation for a "proper" compiler rather than a toy compiler, even if the outcome isn't a finished production-ready multi-arch C-compiler.