Then, if you're like me and read this years ago, play around with the Light Mode dropdown which was new to me. :)

EMA helps because it changes more slowly than the learning network which prevents rapid collapse by forcing the predictions to align to what a historical average would predict. This is a harder and more informative task because the model can't trivially output one value and have it match the EMA target so the model learns more useful representations.

EMA has a long history in deep learning (many GANs use it, TD-learning like DQN, many JEPA papers, etc.) so authors often omit defense of it due to over-familiarity or sometimes cargo culting. :)

The online, continuous and lossy version of this problem is more like how our memory works and still largely unsolved.

[0] https://arxiv.org/abs/2112.04035

[1] https://arxiv.org/abs/1803.07770

[2] https://www.nature.com/articles/s41586-018-0102-6

Cyclists have a bad rep in SF because many (not all) ride quite dangerously. It's a common sight to see cyclists running four-way stop signs and lights without even yielding. I live adjacent to a four-way stop and there's an incident where a cyclist fails to yield nearly hourly.

Meanwhile, Waymo has millions of incident-free miles and of all the self-driving car companies generally takes safety seriously, even if they will act to protect their interests here.

Until more evidence comes out I'll be taking Waymo's side here. I want safer vehicles and Waymo is currently the best bet.

Here's the python version of what I think you're looking for. Shouldn't be too difficult to port to rust.

- https://arxiv.org/abs/1909.07528 - https://arxiv.org/abs/2212.10403 - https://arxiv.org/abs/2201.11903 - https://arxiv.org/abs/2210.13382

There are also literally hundreds of articles and tweet threads about it. Moreover, as I said, you can test many of my claims above directly using readily available LLMs.

GP has a much harder defense. They have to prove that despite all of these capabilities that LLMs are not intelligent. That the mechanisms by which humans possess intelligence is fundamentally distinct from a computer’s ability to exhibit the same behaviors so much that it invalidates any claim that LLMs exhibit intelligence.

Intelligence: “the ability to acquire and apply knowledge and skills”. It is difficult to argue that modern LLMs cannot do this. At best we can quibble about the meaning of individual words like “acquire”, “apply”, “knowledge”, and “skills”. That’s a significant goal post shift from even a year ago.

Citation needed. Numerous actual citations have demonstrated hallmarks of intelligence for years. Tool use. Comprehension and generalization of grammars. World modeling with spatial reasoning through language. Many of these are readily testable in GPT. Many people have… and I dare say that LLMs reading comprehension, problem solving and reasoning skills do surpass that of many actual humans.

> They model intelligent behavior

It is not at all clear that modeling intelligent behavior is any different from intelligence. This is an open question. If you have an insight there I would love to read it.

> They don't know or care what language is: they learn whatever patterns are present in text, language or not.

This is identical to how children learn language prior to schooling. They listen and form connections based on the cooccurrence of words. They’re brains are working overtime to predict what sounds follow next. Before anyone says “not from text!” please don’t forget people who can’t see or hear. Before anyone says, “not only from language!” multimodal LLMs are here now too!

I’m not saying they’re perfect or even possess the same type of intelligence. Obviously the mechanisms are different. However far too many people in this debate are either unaware of their capabilities or hold on too strongly to human exceptionalism.

> There is this religious cult surrounding LLMs that bases all of its expectations of what an LLM can become on a personification of the LLM.

Anthropomorphizing LLMs is indeed an issue but is separate from a debate on their intelligence. I would argue there’s a very different religious cult very vocally proclaiming “that’s not really intelligence!” as these models sprint past goal posts.

https://huggingface.co/datasets/glue

https://www.tensorflow.org/datasets/catalog/glue