I Read Claude Code's Leaked Source – Here's What's Inside

"380K lines, 55+ directories".

" - wait, this is A LOT. For context, gpt-1 was <1000 lines in python (these days on github...), hundreds of times more for more limited domain, it's even suspicious to a point you can argue this is symbolic/statistical engine and not "AI learning". I mean in theory you can add heuristics ad neauseum and generate output, it's not smart per se but likely it would work and it would be O(n) or whatever, it's very surprising especially in high-level language like typescript...if it was assembly at least where most of the job would be working with memory.