sojuz151

The air in Kraków is fine once you give it a good chew. I don't know why people are complaining.

I don't understand why even put that HBM on top of the core.

From what I understand, in a typical gpu core you put logic and connectors on one side and innert silicon on the other. So unless you drill through silicon you don't get shorter routing.

Why not put GPU one one side and HBM on the other side of the PCB? This would fix the cooling problem?

They lose the competitive advantage. They have nothing more to offer than what Google has in-house.

This means that current tokenisers are bad, and something better is needed if text rendering + image input is a better tokeniser.

Compute has been getting cheaper and models more optimised. So if models can do something it will not be long till they can do this cheap.

Polish is quite good, expected based on the founders' background

And the Hubble tension is around 4%, within a factor of 2 of that 10%. This sounds reasonable.

Also, with renewable energy sources, the problem is not manufacturing energy per se but getting the energy where and when you want.

For example, you could run a gas turbine in Germany in winter at night during a spike of the consumption and run carbon capture on a desert during the day

In a sense you use atmosphere as battery/transportation system.

I am not saying that this must make sense but this is far more complex than napkin math.

Because in a classical theory, where there are no particles, there is still the same short range potential.

What about using multiple open source projects with already existing tests? You check if results of execution are the same between compiler versions? This should also provide you with a better coverage of various features

Remember that convolution is multiplication in the frequency domain, so this also handles different responses at different frequencies, not just delays

This will have same energy density as TNT. I have a feelings that fully chargred battery of this type might explode

>My intuition is that as contexts get longer we start hitting the limits of how much comprehension can be embedded in a single point of vector space, and will need better architectures for selecting the relevant portions of the context.

We are dealing with multi-headed attention, therefore we have multiple points per token. You can always increase the number of heads or the size of the key vector.

This look like something that debuggers would hate, because following pointers would be broken

I would say it would make the most sense to have a replaceable entire ram+cpu+gpu assemble. Just have some standard form factors and connectors for external connectors.

This way, you could keep power consumption low and be able to upgrade cpu to a new generation