It's even possible they converge when trained on different data, if they are learning some underlying representation. There was recent research on face generation where they trained two models by splitting one training set in two without overlap, and got the two models to generate similar faces for similar conditioning, even though each model hadn't seen anything that the other model had.

Got a link for that? Sounds super interesting

That sounds unsurprising? Like if you take any set of numbers, randomly split it in two, then calculate the average of each half... it's not surprising that they'll be almost the same.

If you took two different training sets then it would be more surprising.

Or am I misunderstanding what you mean?

It doesn't really matter whether you do this experiment with two training sets created independently or one training set split in half. As long as both are representative of the underlying population, you would get roughly the same results. In the case of human faces, as long as the faces are drawn from roughly similar population distributions (age, race, sex), you'll get similar results. There's only so much variation in human faces.

If the populations are different, then you'll just get two models that have representations of the two different populations. For example, if you trained a model on a sample of all old people and separately on a sample of all young people, obviously those would not be expected to converge, because they're not drawing from the same population.

But that experiment of splitting one training set in half does tell you something: the model is building some sort of representation of the underlying distribution, not just overfitting and spitting out chunks of copy-pasted faces stitched together.

If not are sampled from the same population then they’re not really independent, even if they’re totally disjoint.

They are sourced mostly from the same population and crawled from everything can be crawled.

That's explanation of central limit theorem in statistics. And any language is mostly statistics and models are good at statistical guessing of the next word or token.

I mean, faces are faces, right? If the training data set is large and representative I don't see why any two (representative) halves of the data would lead to significantly different models.

I think that's the point; language is language.

If there's some fundamental limit of what type of intelligence the current breed of LLMs can extract from language, at some point it doesn't matter how good or expansive the content of the training set is. Maybe we are finally starting to hit an architectural limit at this point.

But information is not information. They may be able to talk in the same style, but not about the same things.

LLMs are a commodity

https://www.investopedia.com/terms/c/commodity.asp

Maybe, but that classification by itself doesn't mean anything. Gold is a commodity, but having it is still very desirable and valuable.

Even if all LLMs were open source and publicly available, the GPUs to run them, technical know how to maintain the entire system, fine tuning, the APIs and app ecosystem around them etc. would still give the top players a massive edge.

Of course realizing that a resource is a commodity means something. It means you can form better predictions of where the market is heading, as it evolves and settles. For example, people are starting to realize that these LLMs are converging on fungible. That can be communicated by the "commodity" classification.

Yea it feels like transformer LLMs are in or getting closer to diminishing returns. Will need some new breakthrough, likely entirely new approach, to get to AGI levels

Yeah, we need radically different architecture in terms of the neural networks, and/or added capabilities such as function calling and RAG to improve the current sota

can't wait for LLMs to dispatch field agent robots who search for answers in the real world thats not online /s

skynet would like a word

The models are commodities, and the API's are even similar enough that there is zero stickiness. I can swap one model for another, and usually not have to change anything about my prompts or rag pipelines.

For startups, the lesson here is don't be in the business of building models. Be in the business of using models. The cost of using AI will probably continue to trend lower for the foreseeable future... but you can build a moat in the business layer.

Embeddings are not interchangeable. However, you can setup your system to have multiple embeddings from different providers for the same content.

There are people who make the case for custom fine tuned embedding models built to match your specific types of data and associations. Whatever you use internally it gets converted to the foundation model of choice's formats by their tools on the edge. Still Embeddings and the chunking strategies feeding into them are both way too underappreciated parts of the whole pipeline.

Embeddings are indeed sticky, I was referring to the LLM model itself.

Or be in the business of building infrastructure for AI inference.

Or be in the business of selling .ai domain names.

Is this not the same argument? There are like 20 startups and cloud providers all focused on AI inference. I'd think application layer receives the most value accretion in the next 10 years vs AI inference. Curious what others think

Excellent comment. Shows good awareness of economic forces at play here.

We are just going to use whatever LLM is best fast/cheap and the giants are in an arms race to deliver just that.

But only two companies in this epic techno-cold war have an economic moat but the other moat is breaking down inside the moat of the other company. The moat inside the moat cannot run without the parent moat.

Intriguing comment that I don't quite follow. Can you please elaborate?

Probably OpenAI running on Azure. But it was still convoluted.

That's not what investors believe. They believe that due to training costs there will be a handful of winners who will reap all the benefits, especially if one of them achieves AGI. You can tell by looking at what they've invested most in: foundation models.

I don't think I agree with that. For my work at least, the only model I can swap with OpenAI and get similar results is Claude. None of the open models come even close to producing good outputs for the same prompt.

The big thing for locally hosted is inference efficiency and speed. Mistral wears that crown by a good margin.

There's at least an argument to be made that this is because all the models are heavily trained on GPT-4 outputs (or whatever the SOTA happens to be during training). All those models are, in a way, a product of inbreeding.

But is it the kind of inbreeding that gets you Downs, or the kwisatz haderach?

Yes

Consider the bulldog: https://youtube.com/watch?v=hUgmkCgMWbg

Maybe true for instruct, but pretraining datasets do not usually contain GPT-4 outputs. So the base model does not rely on GPT-4 in any way.

Even in the most liberal interpretation of prove, it doesn't do that. GPT-4 was trained before OpenAI has any special data or deal with microsoft or the product market fit. Yet, no model has beaten it in a year. And google, microsoft, meta definitely have better data and more compute.

> this proves that all llm models converge to a certain point when trained on the same data

They are also all trained to do well on the same evals, right? So doesn't it just boil down to neural nets being universal function approximators?

The evaluations are not comprehensive either. All of them are improving and you can't expect any of them to hit 100% on the metrics (a la. bayes error rate). It gets increasingly difficult to move the metrics as they get better.

Of course, part of this is that a lot of LLMs are now being trained on data that is itself LLM-generated...