this post was submitted on 11 Dec 2024
389 points (99.0% liked)
Technology
59981 readers
4261 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
One thing you gotta remember when dealing with that kind of situation is that Claude and Chat etc. are often misaligned with what your goals are.
They aren't really chat bots, they're just pretending to be. LLMs are fundamentally completion engines. So it's not really a chat with an ai that can help solve your problem, instead, the LLM is given the equivalent of "here is a chat log between a helpful ai assistant and a user. What do you think the assistant would say next?"
That means that context is everything and if you tell the ai that it's wrong, it might correct itself the first couple of times but, after a few mistakes, the most likely response will be another wrong answer that needs another correction. Not because the ai doesn't know the correct answer or how to write good code, but because it's completing a chat log between a user and a foolish ai that makes mistakes.
It's easy to get into a degenerate state where the code gets progressively dumber as the conversation goes on. The best solution is to rewrite the assistant's answers directly but chat doesn't let you do that for safety reasons. It's too easy to jailbreak if you can control the full context.
The next best thing is to kill the context and ask about the same thing again in a fresh one. When the ai gets it right, praise it and tell it that it's an excellent professional programmer that is doing a great job. It'll then be more likely to give correct answers because now it's completing a conversation with a pro.
There's a kind of weird art to prompt engineering because open ai and the like have sunk billions of dollars into trying to make them act as much like a "helpful ai assistant" as they can. So sometimes you have to sorta lean into that to get the best results.
It's really easy to get tricked into treating like a normal conversation with a person when it's actually really... not normal.
This is a really fantastic explanation of the issue!
It's more like improv comedy with an extremely adaptable comic than a conversation with a real person.
One of the things that I've noticed is that the training/finetuning that's done in order to make it give good completions to the "helpful ai conversation scenario" is that it flattens a lot of the capabilities of the underlying language model for really interesting and specific completions. I remember playing around with gpt2 in it's native text completion mode, and even with that much weaker model, it was able to complete a much larger variety of text styles without sliding into the sameness and slickness of the current chat model fine-tuning.
A lot of the research that I read on LLMs is using them in the original token completion context, but pretty much the only way people interact with them is through a thick layer of ai chatbot improv. As an example for code, I imagine that one would have more success using an LLM to edit your code if the context that you give it starts out written like it is a review of a pull request for the code, or some other commentary of a form that matches the way that code is reviewed in the training data. But instead of having access to create that context directly, we have to ask for code review through the fogged window of a chat between an AI assistant and a person discussing code. And that form of chat likely isn't well represented in the training data.
I caught myself thanking GitHub Copilot after getting a response to a question. Felt...weird. For a whole two seconds my brain was operating like I'm talking to another human. You are absolutely correct.