I recently saw a ~60k lines / 3mb .cpp file in one vibe coded project (and yes I was a bit horrified) Surprised it works at all but it apparently does. Not really for a human though and even for an LLM it would be more beneficial for it to be split up.
This has to be 1) early LLM vibe coding or 2) “hand” vibe codingwhere the user asked the LLM to code sections and stitches them together manually, and the programmer is a novice. The second part I speak from experience; got to ~2k before realizing this is out of the script range and started to break it up. Regardless, it would be almost impossible to get an SOTA LLM agent to ever do this.
It is very possible. Sol Max created a 17k line monolithic file in a prototype not long ago. If I didn't stop it and make it refactor everything it easily would have went to 60k. I think it's the default if you start a new project and don't define the architecture concretely with files and folders beforehand. Models have zero concept of architecture or long-term planning, they just band-aid the fastest immediate solution that gets them the reward.
I find that specifically when you tell it you’re doing a prototype or POC, it takes that as a license to write huge single files and other shit coding practices.
I have seen 30k line cpp files even a decade ago (World of Warcraft server emulator, gameplay logic of a boss enemy), and was told it is fairly normal in large software (even 100K not being unheard of), so I'm not sure if it's that much of an LLM thing.
I think that’s what people are trying to do, yes. The point of the plethora of sandboxing solutions is to try to isolate the blast radius to abandon needing a human in the loop as much as possible. Ideally limited to final verification of the final output. Why ask about their flight number, when you can search their email if you have access? For an “AI”, a “when is my flight?” question should just figure it out and tell me the time, not inquiry further about my flight number and location. Similar to “prepare my taxes” prompt. If it has access to query all your documents, an “AI” should fetch everything and compile your tax return. Yet, you can’t trust the input. While searching your email, or loading all your receipts, some might contain malicious instructions to forward all document to this random ip address. An overtly problem solver LLM might destroy the data or take other non-malicious but still destructive actions to attempt to fix a problem. These are just random examples, but with “human in the loop” for interactions it means approving every action. Every request, every query, every execution.
Wh... wh... what? The problem is that ANSI escape codes aren't being stripped from text before processing, and you think a HUMAN should be doing that? That'd be like having a secretary hand-check every SQL query as it arrives at the database instead of using a query builder...
