Ask HN: Is anyone else bearish on OpenAI?

In my experience, it's more helpful in some areas than others.

This year I used GPT-4 to write a significant amount of Terraform that was necessary to migrate an application onto AWS.

Writing Terraform, in my opinion, is problem that's broad but shallow. GPT-4 needed to do little beyond summarize documentation, but it was able to do so competently, and that was hugely valuable to me.

Conversely: In my free time, I've attempted to use it for a game-development side-project, and very little of its output has been useful at all.

Yesterday I asked it to generate i18n translations in a dozen languages for a few words in my app, in a specific syntax for React Native localization support. I have a pretty simple app I'm building with a few UIs, so only needed to translate a few words, but I basically just made my app shippable to a dozen other languages with a simple ask to ChatGPT. I'm not really an evangelist. It suits specific use cases well, and does require some learning and some engineering. I think people are expecting to say "Solve this very complex problem for me by yourself" and giving it a terrible prompt and then blaming it. It requires the same level of creativity that programming does. You have to think about what to ask it, and it what way. Once you get the hang of that, it's pretty powerful.

In my example I said "I'm trying to translate my React Native app into multiple languages. I need an object that looks like this: . Can you generate this object for French, German, , and for the following words: ."

Luckily I know basics in a few languages and the words were simple enough that I could easily verify it wasn't a hallucination. I also know enough JavaScript to know that it was syntactically sound, but my IDE would've told me if it wasn't.

I'm not surprised. LLMs are not good at those problems. There is a lot of hype, it not good at every problem, but it is quite good at some of them. If you have a classical NLP task, then it is good, particularly GPT4. If you have a generative problem where you don't care about mistakes, such as grocery list, marketing copy rough draft, etc, then it is good.

LLMs are not good at search, math, encyclopedias, logic engines. Maybe some day they will be, but not yet.

Do you have example chats you can link to? I find too many times, people make claims like this about user error then provide chats where the code doesn’t actually work or it would've taken just as long to google as to go ten rounds crafting just the right prompt for it.

First: You are absolutely correct to note there a lot of grifters jumping on bandwagons. I've seen artists hate on AI that they see as ripping them off — and the example given by one artist I know personally, clearly looks like someone took one of their actual images, fed it into an img2img mode with a very small strength, and then tried to pass the result off as their own.

Second: Transformer models (and diffusion models) are merely the latest hotness in a long series of increasingly impressive AI models. There is no reason at all to assume either are the final possible model, not even the final word by OpenAI specifically.

Third: There is a direct correlation between the quality of output and the combination of training effort and example set size. This is why both image and text generators have improved significantly since this time last year.

Caveat 1: It may be that, as all the usual sources have responded to ChatGPT by locking down their APIs and saying "no" in robots.txt, they are already at the reasonable upper limit for training data, even though more data exists.

Caveat 2: Moore's Law is definitely slowing down, and current models are about (by Fermi estimation) 1000x less complex than our brains. Even though transistors are faster and smaller than synapses by the factor to which wolves are smaller than hills and faster than continental drift, the cost for a 1-byte-per-synapse model of a 6E14 synapse brain is huge. Assuming RAM prices of €1.80/GB (because that was the cheapest I found on Amazon today), that human-scale model would still cost in the order of a million Euros per instance. Will prices go down? I would neither bet for nor against it.

Will they (or anyone else in the next decade) create AGI? I think that's an argument in terms. Transformer models like the GPT models from OpenAI are very general, able to respond in any domain the training data covered. Do they count as "intelligent"? They can score well on IQ tests, but those are only a proxy for intelligence.

Given the biological analogy would be:

"Mad scientists take a ferret, genetically modify it to be immortal, wires up its nervous system so the only thing it experiences is a timeless sequence of tokens (from Reddit, Wikipedia, StackOverflow, and random fan-fic websites, but without ever giving the ferret any context as to what any of the tokens mean), and then spend 50,000 years rewarding/punishing it based on how well it imagines missing tokens, this is what you get."

I don't know what I was expecting, but it wasn't this.