I agree that incentives are misaligned but there’s several competing model providers. If one gets funny with their costs people will jump ship, especially if the gap between the top 2 labs and everyone else keeps shrinking.
These safeguards already exist when they get a whiff that you might be using Claude to fix security issues. Doesn’t seem farfetched given the incentives I outlined that they would apply to this kind of abuse.
How loose those controls are becomes a market force.
But just so we're both entirely clear on what an LLM is... it's a token prediction system.
it genuinely can't do things except recall things that have already existed.
People are having great success composing things together in new ways, but just like the english language has a finite number of sentences, and music has a finite number of chords: LLMs too are just combining things that have existed.
I don't want to sound condescending, it is remarkable how useful this technology is, but please don't evangelise them on capabilities that they genuinely can never have.
Laptop computers have incredible processing capabilities but nobody expects them to be able to walk your dog, no matter how useful they actually are at doing other things.