Ask HN: Is anyone else bearish on OpenAI?

>It's also very much like crypto where for every one person doing something useful with it, there are 20 trying to exploit the newness and low comprehension the general public have of the tech

This is definitely not correct in terms of numbers, there are many more people using LLMs well than have ever used crypto for any real use case. Also it's worth considering that the only real use cases of crypto are illegal, from noble stuff like busting sanctions to get food to hungry children, through to bribe evasion, bribe payment, tax evasion, drug deals, hiring hitmen, and child sexual exploitation/trafficking. In general, crypto produced close to zero for global society, even when it wasn't being used as an overt and intentional scam.

LLMs are producing significant value for society right now, because OpenAI gave everone API access to a very weird intern who has incredible knowledge breadth and makes dumb mistakes. Interns (or "relatively low intelligence/experience human workers who need handholding for difficult and sometimes easy problems, with an occasional flash of insight") have always been controversial as to whether they actually provide value from the perspective of the person who has to manage the intern, but from the perspective of the company/society it's unquestionable that they do provide significant value. Different people put different value on having a collaborator at all, some people do not want to handhold anyone or work with anyone who's mistakes they ever have to work around. It is nevertheless true that in aggregate for knowledge work, "worker + intern" is more economically productive than just "worker", outside of very, very specialist use cases. This just wasn't possible with GPT-2, and even GPT-3.5 is not quite at a quality where I'd really compare it to a normal intern. No other machine aside from the human brain was even close.

That's the tech now, the worst it will ever be. Whatever comes next with a major leap (GPT-5, Claude 3 or Gemini if they're good, maybe Llama 3 or the next Mistral if they can get improved significantly by the OS community before the next GPT release) is going to be either a reliable version of the same intern, or the same intern with better intelligence and comprehension that still suffers from reliability issues, or a major step up where they're equivalent in productivity to a full-blown knowledge worker in some high percentage of cases. It's already important now, it's only going to get more important.

As for OpenAI specifically, I think they have a very good chance of continuing to lead the pack, particularly with the this cringe-y GPTs/GPT Builder/GPT Store thing. It's pretty transparent that this is them getting data to train an AI on how to spin up agentic AIs to accomplish specific tasks, because they'll have the data on how the GPT Builder is used and the data on how useful and effective the GPTs it builds are, so they can do things like dramatically overweight the most effective and useful GPTs for training their internal "GPT Auto Builder". They'll be running a store for these things as well as effectively controlling the operating system they run in, so purchases, ratings, time using a GPT, sentiment analysis in the GPT text log to detect success, plus explicit in-GPT feedback (the thumbs up and down, feedback submission form) will all be data they can feed into their machine, to make an AI that can build good GPTs for a task and an AI that can evaluate their performance and an AI that can most effectively get good performance out of a GPT. That's going to be huge, particularly the signals that have real economic costs to users (I know they haven't announced it, but I think eventually they're going to make it so you can purchase GPTs) because that starts to pull away rose-tinted glasses and the fog of futuristic sheen and get some more unvarnished data on how much people actually value these specific things. That data means eventually you should be able to just ask ChatGPT to do something for you, and if it can't do it natively it will be trained to be able to spin up a task-specific GPT with access to the correct tools, docs etc, then have the GPT Whisperer AI use it to get the right answer with a bunch of backup data, and return you the answer with the option to see the work. This is also a pretty auditable process, which makes a lot of the legal and AI safety folks happy. I don't see another company that is similarly well-placed in terms of having the tech, talent, compute, product, and roadmap to pull this off.

doesn't it also help with trivial problems?

the goal here is to accelerate productivity, is it not?

Interesting take. Very elitist and not rooted in reality though, I must say. The overwhelming majority of code out there is in languages more verbose than strictly necessary, and less expressive than possible. So in real life, yes AI generated code is a helpful and worthwhile thing. Not everyone can swing their Coq around at work.

That said, I do agree with the general notion. I find the more verbose the language, the better the help. Dense, more expressive languages fare worse. I’m referring to Python and Rust in my case, so one factor is of course massively larger training corpus for Python, and relatively more churn in Rust.

I am.

The underlying tech is amazing. Where LLMs are headed is wild.

I have just lost a lot of confidence that OpenAI will be the ones getting us there.

The chat niche was an instance of low hanging fruit for LLM applications.

But to design the core product offering around that was a mistake.

Chat-instruct fine tuning is fine to offer as an additional option, but to make it the core product was shortsighted and is going to hold back a lot of potential other applications, particularly as others have followed in OpenAI's footsteps.

There's also the issue of centrally grounding "you are a large language model" in the system messaging for the model.

So instead of being able to instruct a model "you are an award winning copywriter" it gets instructed as "you are a large language model whose user has instructed you to act as an award winning copywriter."

Think about the training data for the foundational model - what percent of that was reflecting what a LLM would output? So there's this artificial context constraint that ends up leaving to a significant reduction in variability across multiple prompts between their ChatCompletion and (depreciated) TextCompletion APIs.

They seem like a company that was adequately set up to deliver great strides with advancing the machine learning side of things, but then as soon as they had a product that exceeded their expectations, they really haven't known what to do with it.

So we have a runaway success while there's still a slight moat against other competition and they have a low hanging fruit product.

But I'm extremely skeptical given what I've seen in the past 12 months that they are going to still be leading the pack in 3 years. They may, like many other companies that were early on in advancing upcoming trends, end up victims of their own success by optimizing around today and not properly continuing to build for tomorrow.

If you offered me their stock at a current valuation with the stipulation I wouldn't be able to sell for 5 years, I wouldn't touch it with a 10 meter stick.

This is also my current worry. If you know the concepts (Not the workflow) about the problem you’re solving, I find it easy to get answer, and in the meantime you’ll collect some new knowledge in the process. even when asking someone, they will often point out the knowledge you lack while providing the answers.

Getting straight answers will be detrimental in the long term, I fear. It feels like living in a box, and watching the world on a screen and the person answering my questions is mixing lies and truths.