leumassuehtam

The result is still surprising even if the result is Astra can edit and write weights of a publicly known model. It means that models can, at least in principle, be trained on a task that it wasn't considered possible before.

My implementation: https://github.com/keyehzy/thinking_with_looped_flows

It looks like the B programming language.

> All memory must be statically allocated at startup. No memory may be dynamically allocated (or freed and reallocated) after initialization. This avoids unpredictable behavior that can significantly affect performance, and avoids use-after-free. As a second-order effect, it is our experience that this also makes for more efficient, simpler designs that are more performant and easier to maintain and reason about, compared to designs that do not consider all possible memory usage patterns upfront as part of the design. > TigerStyle

It's baffling that a technique known for 30+ years in the industry have been repackage into "tiger style" or whatever this guru-esque thing this is.

> The artificial intelligence chip leader's chief in October said that the U.S. can win the AI battle if the world, including China's massive developer base, runs on Nvidia systems. He, however, lamented that the Chinese government has shut it out of its market.

Laughable

People believe that more parameters would lead to overfit instead generalization. The various regularization methods we use today to avoid overfit hadn't been discovered yet. Your statement is mostly likely about this.

text-davinci-001 still feels the more human model

I hope this time they can come up with a model at least or par with open source offerings.

Thanks for the answer!

Genuine question: is this a wrapper around Google's gopacket?

The author has a nice series on compiling a Lisp [0], but unfortunately his search engine fails to find it by querying it with "lisp" or "Lisp".

[0] https://bernsteinbear.com/blog/compiling-a-lisp-0/

Since you've mentioned, which hash functions are better while being faster than XXH3?

One way of doing that is by each time step delta perform two integrations, one which give you x_1, and another x_2, with different accuracies. The error is estimated by the difference |x_1 - x_2| and you make the error match your tolerance by adjusting the time step.

Naturally, this difference becomes big when two objects are close together, since the acceleration will induce a large change of velocity, and a lower time step is needed to keep the error under control.

You're right that he is just rederiving electromagnetism through local U(1) gauge symmetry. He define his metric as g_{\mu\nu}=A_\mu A\nu, which is a gauge dependent metric that gives you Maxwell's equation in the covariant formulation when you identify the gauge field A_\mu with the vector potential. Sprinkling geometric algebra in gives a feel of novelty but these results is at least one hundred years old.

*typo

I like to think of an algebra the same as a language. You want to define and constrain your language such that you can describe elements of your universe is clear and consice.

For example, if I am trying to describe a art piece in the museum, I need concepts and words that connect what I am seeing to reality. In the same way in electromagnetism, proper time is part of the reality and it is an experimental fact, so it must be included in the algebra. Another example is a functional programming language where you identify high-level functions as the building concepts of many computer science problems.

Defining an algebra tailored for the problem you're working with actually highly constrain the space, and this is what makes many problems trivial.

I'm convinced that "completion" models are much more useful (and smart) than "chat" models, being able to provide more nuanced and original outputs. When gpt4 come out, text-davinci-003 would still provide better completions with the correct prompt. Of course this model was later replaced by gpt-3.5-turbo-instruct which is explored in this post.

I believe the reason why such models were later deprecated was "alignment".

What do you mean peer reviewed? This is the standard textbook way of numerically studying Ising models.

That's a fallacy. Do you really believe that paying taxes (essentially what you mean about buying things for everyone else) would stop one to want to become a billionaire? It's about incentives, not taxation.

It's easy to notice, even on this website, when a comment starts with "It's worth noting that...". It sounds very LLM-esque.

The latest incarnation of Fourier series is in AI, in the so called RoPe technique.

Yes, it was tested in 8:08 of https://youtu.be/D5u7trVY5Ho?si=n1N_sm3iy0dTx4b5

The barrier of publishing a paper in arxiv is much lower than in academic journals. That said, you still have to convince them that your idea is relevant. They may require some sort of credentials, but if the idea is solid and the paper is well-written you have good chances of being accepted anyway.

edit: typos

You will not find anything as good as the paid version of ChatGPT. You might try Phind [0].

[0] https://www.phind.com/

Did they use this to train Gemini? Which raises the question, where is Gemini?

I had great results with Phind. Their newest fine tune model (V7) has been a pleasant experience and better than most open source models out there.

Nit: your link has a trailing "s" which makes it 404 :)

Quite the opposite, the 4 straight lines represent all that math. Notation is very powerful and get you quite far.

Do you mind if I reach you out to ask some questions, I'm moving to Shanghai soon.

Shameless plug of a similar idea to embed lisp into LaTeX files.

https://github.com/keyehzy/lispylatex

Last commit to the repo seems to be from 7 years ago. Previous post from 2016: https://news.ycombinator.com/item?id=13230382

Interesting application for the RWKV family of models.