Not an expert, but I believe it's just the opposite. Even with the same LLM and the same training data, responses diverge. And that can be a problem.

Id think yes.

Intelligence is a model of reality and the future. They'll converge into the same system as a reflection of the laws of physics and human psychology.

And then when they are used as weapons they'll perhaps try to diverge and it will become an arms race to create models of the adversaries models.

_

Another way to look at it is our own history. Intelligent apes all "converged" into our one homo sapien.

I wonder how much of the AI depends on its initial weights? If in coming decades we understand better how neural networks work, it would be funny to look back and realize that Google beat OpenAI because they got lucky with their initial weights or something.

At a basic level it generates a probability distribution of what the next token should be.

There are a zillion questions that can be asked where you can get a prob dist where multiple tokens have the same probability (flat probability distributions). Then it has to randomly pick one and you can get large variation.

I'd guess no. While they have similar training data, there is plenty of novelty and unique data entering each model due to how each user is using it. This is why ideas like model collapse are fun in theory, but don't really play out due to the irregular ways LLMs are used in the real world.

I could be wrong, but I have not heard a convincing argument for what you propose.

In aggregate? Signs point to yes. For the general purpose SFT base models. We see some evidence even with RNNs vs Transformers. You're essentially finding a function that models language. Use the same optimization function, get a similar result.

However, the RL and especially the RLHF does a lot to reshape the responses, and that's potentially a lot more varied. For the training that wasn't just cribbed from ChatGPT, anyway.

Lastly, it's unlikely that you'll get the _exact same_ responses; there's too many variables at inference time alone. And as for training, we can fingerprint models by their vocabulary to a certain extent. So in practical terms there's probably always going to be some differences.

This assumes our current training approaches don't change too drastically, of course.

This is called the The Platonic Representation Hypothesis

https://arxiv.org/abs/2405.07987

We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

There are like maybe <100 people who actually contribute actively to LLMs.

Just treat it like a commodity (like cloud infrastructure) and build cool shit using it.

If the provider can roll that feature into their offerings then you’re not actually adding any value to the world.

I'm fairly certain that wouldn't happen. Unless you were to overfit the models until the error were to drop to zero, which would likely take almost infinite time. If you did get that point, you've managed to achieve lossless compression of the training data into the weights of the model.

Given that AI models are randomly initialized with noise, and the goal of training is to avoid overfit, there will always be variance between the weights of models, even if trained from the same data, due to those initial conditions, and chaos theory.

And all of the above, is for the same model architecture. I expect you could do some principle component analysis and come up with a transform to work between models, again if they were overfit to zero error. (After all, that would be a compression engine instead of an AI at that point)

Upon reflection, it seems to me that free Stanford AI course I took a decade ago actually stuck. 8)