1R053

I guess pkill would rather be a sleep or koma. Erasing itself from any storage would rather equate to aicide

I think to be useful it needs a mode of playing, that is always musical / in tune.

Not yet sure how to really do it, but one concept I like from NI plugins is that you have multiple keyboard zones: one zone is for notes, others are e.g. for patterns or styles. Imagine a guitar where one zone is for the chord type and tone, another for the striking pattern...

The challenge here is probably the resonance algo for multiple systems based on multiple notes... Maybe the piano concept would be handy here... imagine instead of having 3 strings like on the piano the instrument to be one system for each key... that excite each other via air or direct resonance points... the systems should be automatically tuned based on one reference system (e.g. using automatic string length or tension scaling)

Anyway, amazing work and having it on GPU allows this really to scale.

are you comparing the same models? How did you calculate the TOPS for M3 Ultra?

I think it depends a lot on how research AI generated traffic counts towards these stats.

Also, if there is no answer yet on the web the AI may also not know it. Then these questions should still end up on SO.

I might add, that SO also could build their own chat / research UI. It would need to have some benefit over others, but I guess the community aspect of it alone would suffice...

I think the title of the paper is misleading. Obviously the result shows an impressive performance with just few training examples. However, I cannot see that while keeping the same method reducing training data leads to more performance. They have simply shifted the performance curve (impressively) to lower thresholds. Still also with this new method more training data should give better results. It would be interesting to see a full performance curve for the method based on training data amount (and potentially quality).

not sure if there would be a deepseek open source model if there were no leading models from OpenAI et al

that is a horrible way to be treated by a doctor. In Germany you have free choice of your doctors. Although sometimes you have to wait quite a bit to get appointments... I am not sure, but with a a European health insurance card you should be able to also go to doctors in other countries and be at least partially covered by your insurance.

one of the annoying things about Mathematica is that all functions are crammed into the same namespace and that there is no overloading with different parameterization options...

creative destruction is not necessarily bad, but the default way of nature to move forward.

Obviously, you need to destroy the right things...

To everyone voting it was clear, that Elon would get this role. So he was part of the elected package. Probably, that was also a deciding factor for many voters.

An H100 supports 80 GB of memory. so at FP8 that would allow 3 of the 16+1 models per GPU (assuming around 26B per model), requiring 9 H100s, that usually would not fit one chassis I guess.

Once you have something with 192 GB it gets interesting. You could probably have 7 at FP8 per GPU. At FP16 it probably only would fit 3 per card, requiring 9 again.

I'd say for the current memory layout of cards they missed a little bit the sweet spot. With slightly smaller models or one expert less one should be able to run it on 8 H100s at FP8 or 2 B100s at FP8 or even on 4 B100s at FP16 if I calculated correctly.

you can get a long way on something with 41% less performance than your favorite supercar...