Hacker News Reader

"OpenAI will have to continue to raise more money than any startup has ever raised in history, in perpetuity, to survive."

As of 2022, Uber had supposedly raised a total of $25.2B in funding over 32 rounds. So they may be ahead in the money-drain business.

"There is no way to justify the current level of investment, with local models being freely available."

The business problem OpenAI faces is that their systems aren't good enough that users can trust the results. Slightly crappier results are also available at much lower cost. Unless OpenAI can definitively fix the "hallucination" problem, and at least return "Don't know" when appropriate, this isn't going to work at OpenAI's price point.

Chatbots are about as good as low-end outsourced call centers. They can sort of help with programming. The systems that generate pictures can do some impressive things. LLMs produce better blithering than most bloggers. It's really impressive. But, absent a major theoretical breakthrough, that's not enough to fund OpenAI.

One big difference is Uber only has a few peers and became instrumental before the realities of market size caught up to them.

But every hyperscale data center owner is investing for themselves, major companies capable of the R&D are investing in their own science, and the market for LLMs is slow to materialize relative to the hype. That's money that might have gone to OpenAI if their situation was similar to Uber's.

> The business problem OpenAI faces is that their systems aren't good enough that users can trust the results.

It depends what you're doing with it. Answering arbitrary questions is very different from document summarization or from mapping arbitrary questions onto a list of canned questions to then answer.

It also doesn't have to be perfect. Instead the cost of an overall system (AI + error reporting & handling) just has to be cheaper than that same overall system using whatever the AI is replacing.

If it makes errors 3% of the time, and people from Mechanical Turk make errors 2% of the time, its usefulness depends on whether or not those 50% more errors cost more than the cost savings from paying a model provider rather than paying humans.

Humans hallucinate too though. A lot. A good tip for system design is to assume the human operator goes off the rails at some point and does something absurd and nonsensical.

Hallucinations don't need to be fixed in one go as much as improved progressively. There is some magic threshold where they'll stop being an issue for specific tasks, or alternatively they simply become more reliable than humans. The problem will die with a whimper.

I think you have a fundamental misunderstanding of the problem here if you equate human hallucinations with LLM hallucinations.

What's the misunderstanding? The hallucinations I've seen from AI tend to not be that dissimilar to when humans are guessing. They will guess that a function like that exists. The difference being that humans are capable of checking their work and see if this actually is true by referencing some existing source of truth.

An additional difference is that humans know what they dont know. LLMs will happily and confidently make up nonsense.

I'm not entirely cynical on the value of LLMs, but I've yet to see one say "I dont know", or "I'm not sure, but here's my best guess".

> I'm not entirely cynical on the value of LLMs, but I've yet to see one say "I dont know", or "I'm not sure, but here's my best guess".

I've used LLMs to build form autofilling based on unstructured documents. It correctly does not answer fields that it doesn't know, and does not try to guess anything. It has been pretty much error-free.

It's all about your prompting. Without explicitly being given guidance on how not to answer, you're right, they will never say they don't know.

If you've never actually seen that happen, I encourage you to experiment more with LLMs; there's lots that can be achieved with the right prompting.

>A good tip for system design is to assume the human operator goes off the rails at some point and does something absurd and nonsensical.

And your assertion is that when this happens, it's because the human operator is hallucinating?

No. It's that humans will sometimes do the same thing that we've decided to (mis-)label as hallucinating when an AI does it.

I just saw an announcement that Musk has launched the largest GPU compute cluster in the world in Memphis TN. How did he fund this?

I wonder at what point Grok might become a serious competitor to OpenAI and Anthropic?

IIRC A proof was just published for the inevitability of hallucinations in the kinds of statistical models we use.

It seems like OpenAI is trying to build a moat now, each generation appears to be less and less open to the point that you can't even see LLM reasoning steps.

It's been a while since we've had a major tech company commodify/open source a major dependency (and not just layer some rent seeking pay/subscription service over existing assets). Meta seems to have the best long term plan for LLMs.

I like your comment, but it doesn't go far enough! The other issue here is the price of FLOPS is still dropping; so even if training the model weights was a moat it will probably evaporate quickly over the next few years if price trends hold. All this capital spend would have been much cheaper if deferred by a few years.

If they aren't making massive profits right now then there may well be a problem.

History is redolent with first-to-market fails followed by also-ran successes. This would appear to inform Apple's AI strategy, but not just since AI.

Not only that, but these models are at some point superior. Even if AI companies offer embedding services, which some do not or only recently added such capabilities, the best results are often a specifically tuned openly available model like llama.

I do think this moat can still exist though, as providing a product is different than running a model somewhere, even if you have the compute power to allow prompting by a large amount of people. And making a model ready for the business user is quite a bit of work.

Although true, the capabilities of commercial products will probably be eclipsed sooner rather than later. We saw that in image generation. Crowd sourcing the problem did produce better results. Generating videos is probably still locked behind enormous compute power (well, the 400+B model of Llama is too...)

I think the focus of these companies isn't the performance of the model itself, it is providing interfaces to all kinds of systems and maybe solve a specific task.

Sounds like a good plan until the US and EU declare open-weight models "unsafe" and ban them. "It can help somebody create a bioweapon," etc.

I would imagine lobbying efforts and scaremongering is already underway. This is a powerful technology that the general public (and small businesses) might not have access to in the near future.

> Writing code with AI? I could swap to a local model. Costs nothing.

Costs nothing? I'd have to buy a desktop computer with a good GPU to run that model. How much would that cost?

You can get a second hand MacBook Air 16GB with Apple silicon from a couple of years ago for about $1,000 and get a decent performance from Llama 3.1

Right? Or a second hand desktop for $300 and put a $300 3060 in it and get even better performance. So many options.

I think the point is that anyone can enter the AI cloud market to compete with the fat cats if they get too pricy. NVIDIA's de-facto standards are the closest thing to a market moat. If AI ends up highly tied to NVIDIA's standards, then NVIDIA could become the Microsoft of AI. It's hard to know how tool chains will evolve to answer that though, it's new territory. (It could be all moot if the bots take over and eat us humans. I taste like chicken.)

You need to own that or buy someone else's compute to run any LLM. The marginal cost of the open local models is free.

You can get a LOT of compute credits, at the public rates the cloud providers charge (not OpenAI's discount) for the price of an NVIDIA 4xxx with 24GB or more of RAM. Enough to do what most people need AI to do for a long time.

To run a 7b model? Not much. To run the 405b model mentioned, at interactive speeds, at reasonable quantisation? A lot.

> Writing code with AI? I could swap to a local model.

AWS only brings in 80B+ a year, must be from everyone self-hosting. I'm curious why people think this is different?

AWS' moat, if they have one, is not in renting out generic compute. That's been done since before anyone ever heard the term "cloud".

I think the point was that the things LLMs are useful for can increasingly be done with freely available models you can run on your own hardware, hardware you rent from someone else, or via a 3rd party's hosted service. Want to run Meta's models on Google's server? Nothing stopping you.

As it has been mentioned numerous times in the past, we are not the average user. I don't know anyone outside our bubble who would bother running a local LLM, or going into all the trouble building an x8 GPU rig to run one of the biggest models.

Easy to run locally for geeks means easy to package up and sell to non-geeks.

Why would non-geeks want to run a local LLM? Electricity prices alone are quite high. Why not pay a subscription on a cloud service and be done with it.

My point is that you will have many cloud services to choose from if geeks can build them easily, which will keep prices down.

Privacy could be one reason. There are a lot of cases where people do not want to send data to a cloud service.

interesting. I have not found Llama able to do any of the sort of 'reasoning' that the larger foundational models can do. It's too easily tricked up by small things.

The larger models aren't perfect, but worlds apart from Llama.

> If there were no "weights available" models, OpenAI would have an incredible, unbeatable moat

Not really, as Claude is competitive both on quality and price.

(agree with the rest of your comment)