Hacker News Reader

Comment by mcpar-land | Hacker News Reader

the current 'normie' understanding of ChatGPT and other current LLMs like Bard is not that they are stochastic parrots that can complete sentences by probabilities, it's that they are 'stupid AGI.' And if we give them more funding and more training, eventually they will become 'smart AGI.'

OpenAI markets their product exactly this way, focusing on how ChatGPT can 'write code' (it cannot) or 'write poetry' (it cannot) or do any number of things that lean in hard on the 'stupid AGI' angle, never clarifying that their actual product is a spicy autocomplete. You'll never hear OpenAI say 'ChatGPT is a stupid AGI' out loud, but they'll never correct any mainstream misunderstandings that it is.

View on HN

The problem with saying an LLM is "just a parrot" is that we don't know how parrots work, so this doesn't mean anything.

You know a parrot isnt going to impart new knowledge to the world.

I don't know that. You can find new conclusions from existing data if nobody else has done it yet.

You can feed new data into it quite easily.

a lot of the dismissive takedown of LLM's intelligence uses their spurious knowledge of the world, but it is quite easy to have them 1. search the internet 2. search research articles 3. read books and have them work to iteratively improve something.

Right now they cannot do much reliably but calling them a parrot is miopic how little we understand how humans are intelligent and to how LLMs today are more intelligent than a non negligible chunk of the population.

I'm not sure I get what you're saying here. ChatGPT and GPT-4 absolutely can write code and poetry. GPT-4 arguably does both better than I do.

Can you expand on your statement that ChatGPT cannot write code? That doesn't match my experience with it.

When you ask ChatGPT to write a piece of code that does a thing, it is not reasoning about what the code should do, modeling the problem in its head, figuring out how to translate that into code, then writing it. The neural weights are spitting up a piece of code it determines would probably come after your prompt, based on similarity to a kajillion StackOverflow answers, GitHub repos, textbooks, etc etc. It is not coming up with a solution, just giving you the average answer to the million most similar questions.

ChatGPT can “write code” in the most literal sense that it will give you code when given a prompt, but (unlike how OpenAI markets-by-omission) it does not reason about code like an AGI could.

gpt is clearly capable of constructing and interacting with a model of the world -- there was a famous example in the microsoft bing paper where they described a complex maze and had it reason its way out. 'predicting what it thinks is most similar to stackoverflow' etc does not explain any of its behavior in a meaningful way, unless you are willing to apply this description to human intellectual labor. gpt4 is clearly not sentient, but what is shocking about it is that it demonstrates that honest-to-god cognition can be decoupled from sentience.

btw, i think i have tended to see these pessimistic takes by people who have only used 3.5, which is impressive but not reliable -- i strongly recommend using 4 in the course of your job to see how useful it is. every day in the course of interacting with it, it produces novel lines of thought which shock me.

I'm not particularly interested in AGIs (I'm kind of hoping they don't happen) - but for my purposes the illusion of LLMs being able to code is now good enough that I'm using them to write good, working code for me on a daily basis.

Here's a recent example: https://simonwillison.net/2023/Aug/6/annotated-presentations...

Reply on news.ycombinator.com