eden-u4

Yes of course for roughly 3 months, after that they'd starve

I don't understand the RMS table, shouldn't it be non commutative? Eg "example 0 vs 1"'s RMS != "example 1 vs 0"'s RMS? Which doesn't seem the case for the checkpoints I checked.

There are not enough samples in that book to generate new "infinite" data.

Guardrailing is usually done with a smaller model (< 1b) to filter out simple "not aligned prompt" and not waste compute.

But what if your rubber duck is actually steering your thought process (since you may not have a consolidated one)? In this way I think the AI as editor is far better than a rubber duck AI. While in the former, it might point out your mistakes and give useful advices (which is similar to what you describe), it might not steer your thought (unless your mistakes are far too severe!), and actually help in your reasoning. But AI as a brainstorming rubber duck (or thinking tool) could be harmful to your thought process.

This is basically a diffusion model: start from a random seed, use a generative process to transform it into something

mine too, but none was such a dick. also, anything related to school (particularly at a young age), is not viewed as something to boast of (at least in my experience in italy, serbia and portugal).

Also, they have all the infra to actually use all that tpus advantage (as well as actual researchers, contrariwise to OpenAI)

what type of experiments did you run in less than a week to be so dismissing? (seriously curious)

Not OP, but I guess based on your comment:

> But the extent and the way in which Zig specifically puts it to use -- which includes, but is not limited to, how it is used to replace other features that can then be avoided (and all without macros) -- is unprecedented.

That MrWhite wanted to knkw an example of Zig's comptime that is not merely a "macro", rather the usage as a replacement of other features (I guess more complex..)

PS just interested in zig, I'd like some pointer to these cool feature :)

wow, thanks for this long explanation.

so they are basically using a similar idea to that of a stirling engine in thermoelectric generator or they use a different mechanism to produce energy?

No open model/weights?

I don't have much experience with ROCm for large trainings, but NVIDIA is still shit with driver+cuda version+other things. The only simplification is due to ubuntu and other distros that already do the heavy lift by installing all required components, without much configuration.

why don't you ask the model about the shrinked system prompt and the original system prompt? in this way you can infer whether the same relevant informations are "stored" in the hidden state of the model.

Or better yet, check directly the hidden state difference between a model feed with the original prompt and one with the shrinked prompt.

This should avoid remove the randomness of the results.

I dunno, these reasoning models seems kinda "dumb" because they try to bootstrap itself via reasoning, even though a simple direct answer might not exist (for example key information are missing for a proper answer).

Ask something like: "Ravioli: x = y: France, what could be x and y?" (it thought for 500s and the answers were "weird")

Or "Order from left to right these items ..." and give partial information on their relative position, eg Laptop is on the left of the cup and the cup is between the phone and the notebook. (Didn't have enough patience nor time to wait the thinking procedure for this)

When they say "in 20-50 years" it is implied that it might not exist ever. See nuclear fusion or AGI.

I think the issue with RL is that, in order for a model to perform well in a task, you have to make it stubborn. In the same way a student that thinks outside the scope of the task might not perform well in a graded exam, but that does not mean he/she is a bad reasoner. With RL and all training procedure you are creating a very focused and very fit to the task thinker, which might not be useful in all applications (consider an open problem, it might need an out of the box kind of thought).

because non-tech friends don't have coding problems, which is the only problem LLM are ok at.

anyone knows which program they used to record the demo?

meta is large enough (and diversified enough), when OpenAI valuation goes to pennies (because LLM won't solve any real problem, nor they'll remain relevant in the long term and open source alternative are more than enough for most chatbot usecases), they can acquire all the relevant technologies from OpenAI, Anthropic and such. The real question is why google didn't release their models.

this project only uses kaggle metadata and abstract from arxiv. Moreover it is "focused" on only 5-6 categories in the arxiv. Therefore, the costs are marginal.

Plus you could use a mixed system: first you index the abstract of the most relevant 50 papers, then embedd the text of those 50 in order to asses which are truly relevant and/or meaningful.

if you try hard enough you can always ~latinize~ ASCIIfy a language

ah, numerical instability in the warmup stage might be the issue then?

that's a wrong simile given that you would get the same end result in both cases.

I'm not using a transformer, just a plain Feedforward, Relu and dropout for a simple classifier.

I don't know, I can be wrong. I hope and some toy experiment shows that even in low case parameters it works fine as well as adam.

Tried the source code on a toy model: adam took 2 epochs to train a 10k parameters model, this didn't achieve anything useful in 20.

Tweaked a bit the hyper parameters and such, but nothing. Probably a bogus implementation?