Some more info on that '#AI coded C compiler':
-
Some more info on that '#AI coded C compiler':
> Anthropic's Claude Opus 4.6 spends $20K trying to write a C compiler. AI agents build something that mostly works but worries the project's creator. https://go.theregister.com/feed/www.theregister.com/2026/02/09/claude_opus_46_compiler/
Twenty grand? That's two months of developer time. So maybe a *little* cheaper than hiring a coder. But not hugely cheaper and that's just the AI costs. Humans were still in the loop.
And what if they had to pay for the training data? They *stole* that.
-
Some more info on that '#AI coded C compiler':
> Anthropic's Claude Opus 4.6 spends $20K trying to write a C compiler. AI agents build something that mostly works but worries the project's creator. https://go.theregister.com/feed/www.theregister.com/2026/02/09/claude_opus_46_compiler/
Twenty grand? That's two months of developer time. So maybe a *little* cheaper than hiring a coder. But not hugely cheaper and that's just the AI costs. Humans were still in the loop.
And what if they had to pay for the training data? They *stole* that.
This comment on Anthropic's github repo…
> https://github.com/anthropics/claudes-c-compiler/issues/1#issuecomment-3869799573
-
This comment on Anthropic's github repo…
> https://github.com/anthropics/claudes-c-compiler/issues/1#issuecomment-3869799573
-
This comment on Anthropic's github repo…
> https://github.com/anthropics/claudes-c-compiler/issues/1#issuecomment-3869799573
@jackwilliambell aaaand it's stolen ofcourse: https://github.com/anthropics/claudes-c-compiler/issues/231
-
@jackwilliambell aaaand it's stolen ofcourse: https://github.com/anthropics/claudes-c-compiler/issues/231
@jackwilliambell @SolarDavy Fascinating for two reasons
1. The first reply is LLM output
2. If you'd take a compiler industry person to write this project and they come up with similar abstractions, would it be a license issue, too, and where do you then draw the line
-
@jackwilliambell @SolarDavy Fascinating for two reasons
1. The first reply is LLM output
2. If you'd take a compiler industry person to write this project and they come up with similar abstractions, would it be a license issue, too, and where do you then draw the line
I draw the line with *intent*. An LLM cannot have 'intent', it is simply a stochastic choice machine picking things out of a database and then layering over algorithmic 'fixup' on the results.
Whereas a human being chooses with intent. If the human simply copy and pastes and then makes minor changes, that falls under copyright and the OS license. If the human uses similar abstractions and then writes the code themselves, with intent, then it does not violate the license.
-
This comment on Anthropic's github repo…
> https://github.com/anthropics/claudes-c-compiler/issues/1#issuecomment-3869799573
-
I draw the line with *intent*. An LLM cannot have 'intent', it is simply a stochastic choice machine picking things out of a database and then layering over algorithmic 'fixup' on the results.
Whereas a human being chooses with intent. If the human simply copy and pastes and then makes minor changes, that falls under copyright and the OS license. If the human uses similar abstractions and then writes the code themselves, with intent, then it does not violate the license.
@djh @jackwilliambell I agree that the algorithm is not alive in any way or shape.
But the people that collected the data which it's repeating didn't respect the licenses of that data. So all LLM output is stolen.
-
R relay@relay.infosec.exchange shared this topic
