there is an NVDA bubble, as the GPUs are being sold at outrageous premiums. It's bubbling left and right with de-facto vendor financing (2000 deja-vue) though. Which is not sustainable since a) competition will inevitably kick in, especially for the enormous profit potential as of today b) only a fraction of GPU functionality is needed for the linear algebra operations for LLMs c) hardware optimized for transformers including KV cache are/will be coming and not only from Nvidia, d) hyperscaler customers are showing reluctance to pay those premiums despite the CUDA bastion, at last going to other sources (AMD as of this week) besides in-house silicon.

I actually bet against NVDA based on the expectation that DGEMM-type matrix multiplication would soon(er) be in hardware optimized for LLMs plus putting part of the transformer into hardware. An expensive undertaking. It's incomprehensible given the market size and overpayment for Nvidia products that things have been slow. I don't get the technical rationale why some of the heavy GPU buyers don't put effort into making PyTorch, Transformers and more of the stack more affine to non-CUDA accelerators. in HPC matrix-matrix, matrix-vector has long long been abstracted away in BLAS libraries. Here, even with KV caching and what not the GPUs are completely overengineered to what LLMs need feature wise.

Emphasize: this is not EU regulatory burden called out here. Actually more EU action would be needed for a liquidity attracting common capital market among other things.

What's to point out is that rather some many countries went overboard with an administrative mindset to everything. It's toxic for innovation, business and regular life. The amount of government vs. private is just out of whack. The amount of GDP being government-related and the sheer number of people in public administration jobs is off. Everything is done with more rules, red-tape and administration.

most EU countries have price control on books, that is every retailer including Amazon has to charge the same price on a given new book. This is braindead as it the prices of books are in a simple demand supply relationship leading to fewer books being bought. The rational arguments given when it was re-legislated last was that it should protect authors and cultural heritage, which is economic nonsense, but works in politics.

"there is a world market for maybe five computers" [Thomas Watson Sr.]
Webbrowser are for consumers, Lotus Notes for Enterprise (in meaning] [Irving Wladawsky-Berger, early IBM Internet Division VP on web browsers] IBM WebExplorer was canceled ...

politics. Not having anywhere near current technology launchers and failure to reach agreement to at least try to catch up, what do we do: PR. Sounds good the effort that's about it.

BS and correction from before it's probably in the O(1 kWh) for a 100k token context SGEMV operations alone

it's now almost 2 years after the release of ChatGPT. There are many vendors pitching x10 and more from ASIC to FPGA to Wafer. Hard to understand that in 2 years matrix-vector optimized for transformer training and inference is not all over the place in lieu of the absurd pricing for CUDA devices.

Submitting a more than 100k context right now - just first step initially processing it is probably more than 0.5 kW.

positively, it provides funding and community to fusion

taking along a bunch of Methane-metabolizing bacteria (methanotrophs) to heat it up comfy for the next visit?

Despite the unfortunate lack of an oxidizer on Titan, anaerobic methanotrophs, might happily reduce sulfates (SO4^2-), nitrate (NO3^-), or metal oxides like ferric iron (Fe^3+) or manganese dioxide (MnO2) in case it is available there. There certainly should be enough Nitrates given the atmosphere is full of it.

if Dune by David Lynch is witty and artistic, this series is just mindless dullness in cinematic desert

That was maybe first time the Indian Government regulated an industry (away) with foresight. Compare this to the EU parliament spawning red sheets and tape over inference which is NOT where the imminent dangerous are.

Model builders, especially those, who infuse structure into the models through recursive/reflective/meta-programming & have 30k + H100 equivalents to work with are subjecting the world to undo risk. It's probably going on in several places at this point stealth/dark R&D, self-modifying LLM coupled with RL learning/optimizing on a reward function being the most obvious to try and entirely doable that is to couple.

Given that its imminent that ais will start to self - modify and spin up to be maybe being emergently completely uncontrollable there needs to be upstream trainers monitoring, not just red tape. Think air-gapped facilities. Trusting red power buttons in case is hilarious against machines that can chess-alike think hundreds of moves ahead against much lower developed creatures (us).

This is outlandish but real as something. The things necessary to stop it now won’t be done in time probably unless something happens on a small scale which makes it obvious to shut the large computers down. Even then then your favorite autocracies of the world will happily continue the path to experiment with AIs - end of story.
https://x.com/elonmusk/status/...

Can someone give a short explanation on how the previous relate to the new ones. Also are the new ones in the category of silicon fix (halve a year cycle time at best) or microcode??