WanderPanda

Tbh I got exactly what I expected when buying the R1, build quality is actually even better than expected.

This! The US hegemony is flawed but:

1. There is no other country (not even close) that could be trusted with that amount of power (especially considering size)

2. Held up the (illusion of) “neutral” international institutions like the UN. They barely worked in the presence of a “benevolent” power, and will probably completely lose relevance to anarchy and the “right of the stronger” (on local levels), shall the US hegemony subside.

Then on the other hand the US has started undermining their own most important principles:

1. 1971: Removing the gold convertability from the $

2. 9/11: Starting to spy on each and everyone, eastern germany/soviet-style

3. Removing personal freedoms during COVID (not as severe as other countries, though)

If it weren’t for silicon valley, the us would already look like a stagnating state where the economy is mainly driven by government spending. The problem is larping EU socialism will only yield even worse results in the US, since the government seems to be even less efficient.

On the other hand the US is also one of the few countries that have turned around non-violently in the past. Attractiveness for international talent is still immense. So with a few adjustments I’m pretty sure it could be turned around

looks like AGI has been achieved externally

I mean to go thinner than 6/6s I can see the 3.5mm causing trouble. Part of me is still sad they bounced they went the other direction when it comes to iPhone thickness

Wait wasn't BERT all about non-causal masking aka predicting words in the middle?!

„I‘m a pretty good local maximum“ that is what any local maximum would tell you if asked how it likes itself

Is this "just" CUTLASS in user friendly?

Him saying this always puts me off. Gives hard old sales-guy vibes. I really wonder who/which demographic is influenced in nvidias favor by this rethoric.

One's "dumb idea" is another marketers "genius stroke". Seems like he is playing the media puppets while he can

Wait but nvidia tensor-cores are exactly the hardware that likes 16x16 tiles, no? I thought that was the whole point? The hardware is already here and I'm sceptical if there is another order of magnitude in performance to be gained from even more specialized designs.

Wasn't a key thing of the xz attack vector that people where encouraged to download the custom source release instead of the autogenerated Github one? I don't know if that is a pattern but it seems like best practices in the (source) supply-chain could prevent a large class of these attacks.

Are we facing a Byzantine generals kind of situation now?

Just that in this case the end-customer is a person that could shop around for what he/she needs. Such a competitive market would instantly vaporate the current, rent-seeking margins

Seems like the best way forward would be to give some smaller amount, lets say $30,000 to the one in need and let them shop around for the best solution and keep the (probably still sizable) difference

Because there are no multinationals in New York State?

Tech giants, while getting worse in terms of lobbying etc. are still much more benign than the old-school industry giants we have in europe. The European Roundtable for Industry [1] is particularly disgusting. The bigcorp lawyers are essentially writing EU law…

[1] https://en.wikipedia.org/wiki/European_Round_Table_for_Indus...

In the end an "idea" is by definition contrarian which is the opposite of the training objective of LLMs. The question is how far fine-tuning and tree-search can go to extrapolate from the data-manifold. And the answer is probably not that far, currently.

TE might have their name really damaged by this. I wonder why they agreed to this colab in the first place

The opposite is the case, with all the advancements, even by doing nothing, Apple (like everyone, including hobbyists) is moving closer to the frontier. Hopefully this trend stays alive!

Is there any small risc-v soft-core with big ass SVE (scalable vector extensions)? I would like to play around with them but the only option seem to be cloud instances like gravitron and small (128bit) like the licheerv nano (c906) which also seems to only support a beta version of the standard.

So be fair, sometimes (e.g. in the case of scatter plots with many dots) pdf renderers become very slow and/or mess up the rendering. In this case the easiest option is rasterizing it (for performance and consistency of the appearance)

I've been working on a C++ deep RL library (RLtools) and also created some WASM examples (https://rl.tools). I didn't pay any attention to the binary size at all but it turns out it is also just around 200-300kb (including everything, deep learning forward/backward, RL algo and dynamics simulation).

Even though it's not prohibitive rn, I'm curious how small it could be and hopefully find some time soon to squeeze it down

Did you at least front run the market and stocked up of 4090ies before this release? Also gamers are probably not too happy about these developments :D

Templates would be nice, though. So much better than macros

This leaderboard is not the best for comparing model architectures, the dataset and finetuning have too much influence. I think perplexity on a particular dataset would be a better way to compare

The opsec of this guy nakamoto is just on another level.

It is a bit worrisome but we manage to deal with subpar human code as well. Often the boilerplate generated by ChatGPT is already better than what an unexperienced coder would string together. I‘m sure it will not be a free lunch but the the benefits will probably outweigh the downsides.

Interesting scalability questions will arise wrt to security when scaling the already unmanagably large code bases by another magnitude (or two), though.

Isn‘t git submodules a much more lean way at accomplishing this than a clunky 500gb repo?

Even this "I hate investing time in proprietary technologies, because I know I can be stopped or locked out" is a hard-gained insight. Hackintosh is one of those things that made me understand this. Nothing like spending weeks to get your hackintosh working smoothly with all the hardware just to find out that the next update breaks everything. I've come to see it as a necessary part of the journey

I’ve been using Github copilot daily for two years and ChatGPT for 1 year now. And I think the tide lifts all the boats. I’ve seen a (perceived) 2-3x productivity increase. I think these tools slightly favor people in front of the learning curve of a particular field. I’ve been dabbling in all sorts of things so if you’re a focused expert (who doesn’t need to explore but just exploit) you probably get less than a 2x boost from using LLMs.

I can see LLMs eating into the expert regime IF they get another 5-10x better. But even in that case human (expert) knowledge will be required to know what is possible and hence what to ask (kind of like reward function design in reinforcement learning)