[Aldo_MX]

This restriction makes no sense.

The US says "18 countries may purchase our AI chips", but what I understand is "90% of the world may purchase Chinese AI chips".

China just needs to infiltrate Taiwan, which is geographically and culturally closer than the US.

[cookiemonsieur]

It doesn't even matter at this point. As evidenced by DeepSeek. A growing portion of the world is distancing themselves from the west, and It's a good thing. People should have the ability to choose who they do business with without a third party strongarming them.

[cocoggu]

Can you please explain to me why AI chips don't matters anymore because of DeepSeek? I thought it was just a better model, but perhaps I didn't get it?

[Gigachad]

Deepseek used older generation chips and developed a model that takes significantly less compute. Making having access to tons of the latest nvidia hardware unnecessary.

[lostmsu]

This does not make sense. If R1 scales similarly to other GPTs, throwing 100x more compute at it will produce an even stronger model.

[Cumpiler69]

Being forced to live with more HW restrictions usually results in more reliance on SW creativity and better optimizations instead of lazy developers bloating SW to fill all available resources.

Just like how it's no surprise that websites developed where everyone has the latest and grates fully loaded M silicon MacBooks also sufferer from horrible lack of optimizations because "it works on my machine" while being a stuttery mess everywhere else.

[iforgot22]

Websites and the like are a different world from ML training, where devs seem to be more performance-conscious. But there's a weird reliance on CUDA because devs (rightfully) don't trust the alternatives.

[Gigachad]

Yeah, seen the same with PC gaming. Minimum specs absolutely exploding for no real reason other than the fact most gamers were buying the latest top tier cards. Then the Steam Deck came out and devs are forced to consider the fact that a 2D pixel art game shouldn't be lagging out on a SoC capable of producing stunning 3D graphics in a properly optimized game.

[talldayo]

Well, the Steam Deck's release coincided with the popularity of reconstruction techniques. The Deck didn't "force" devs to consider optimization so much as it just gave them a low-end reconstruction target to play with. Without FSR and XESS, there's no doubt that the Deck would be a solidly last-gen console.

Strictly speaking, a lot of games really shouldn't be playable on the Steam Deck. Baldur's Gate III and Cyberpunk 2077 are both CPU-bound before reaching 60fps and can barely keep their head above 30fps running at 360p internal resolution. The Deck's saving grace is that it can tap into the same dynamic resolution mode that last-gen consoles depend on for consistent framerates.

[dathinab]

It's not that they matter anymore.

But there has been a long term suspicion in the AI community that the ultra expensive to compute and very expensive to run humongous LLM approach is a dead end, or at lest fully unnecessary (and as such monetary wise a dead end).

I mean think about it, the target crown jewel of AI was never to find ways to train on insane amounts of data, but to be able to get as good as possible results with only as much data as necessary but no more. Because for a lot of use cases there simply isn't that much data.

And from everything we know the structure of language is not so complex that you need this insane amount of data and model size.

It's just we worked around of problems by throwing more compute and data on it instead of solving them proper. Similar we try to reformulate any little-data use case by reformulating it in a way where we hope to take advantage of the mass "causal text" data modern foundational LLMs where trained on and fine tune and instrument the model using the "little data" of the use case.

But conceptually this is ... sub-par and non desirable. And sure that we made it work with this trickery is quite magnificent.

And sure this huge LLMs do more then encode language, they encode miscellaneous knowledge/data, too.

But a messy, hallucination prone, non properly updateable and potentially outright copyright or privacy law violating encoding of data...

So many systems already do use RAG like approaches to get supply the knowledge in a updateable much more well defined fore and "only" use the LLM to find the right search queries and combine things together into human readable responses.

In turn the moment we have small LLMs which still work well for language structure they likely will very reliable win through a lot of reasons (the ones mentioned above and they are also much cheaper) and that even through they are _way_ more complicated to use then "just prompting a LLM". But most advanced assistants are anyway already way more complicated then "just prompting a LLM".

Or in other words the technical breakthrough anyone (including OpenAI) would like the most (OpenAI: financially, as long as it's an internal secret) is one which eliminates the need for having the latest bleeding edge ML chip tech. And DeepSeek is seen by some as a signal that exactly such a change is going to happen. Also I have heard rumors (which I don't believe) that one reason for OpenAI to go non-open was because they realized that, too. And with cheap to run open models they would lose the competitive benefit of competition not being able to do from scratch training even if they want to.

[dathinab]

well I mean the US itself distanced itself from the west

depending on your definition of what "the west" means

or it doubled down on what "the west" means, if you take different and in which case the EU is slowly moving away from "the west"

either way using "the west" to lump together EU+US+Canada+Australia+... seems to be becoming increasingly meaningless

[genewitch]

Per HN guidelines, that "just" is load bearing.

[jjallen]

What does Taiwan have to do with this? Because it’s a US ally China could get GPUs through Taiwan if china invaded Taiwan?

Maybe old ones if this is what you are saying.

[dathinab]

Most bleeding edge chip manufacturing is currently done in Taiwan, Taiwan has both the machinery and know how to produce state of the art AI training chips. And while they are not quite at NVIDIAs knowledge about chip design by far the largest source of performance differences between NVIDIA and China based completion is the manufacturing not the design (which still is better, but not to the margin where it's not something you can somewhat "good enough" bridge by "throwing more money and (electric)power at it", for the manufacturing differences on the other hand "throwing money/(electric)power at it" to bridge the gap isn't viable. The difference is to big.).

Through then if we are realistic both Taiwan and the US did take measurements to make sure the even if China attacks Taiwan and wins they aren't getting their hands on this tech.

Except Taiwan would have been not very clever if they didn't try to find ways to work around this, so that they are a lever to negotiate in case the US abandons them... (it's unclear if they did find ways tho).

[Aldo_MX]

I said "infiltrate" as in "send their best engineers as undercover agents to work at TSMC".

[nisten]

It makes perfect sense if they're laundering Russian assets to circumvent US Sanctions.

[iforgot22]

Isn't this going to erode Nvidia's CUDA advantage? I thought that was way more important than their hardware.

[iforgot22]

Well seems Deepseek still used CUDA at least

[sofixa]

> China just needs to infiltrate Taiwan, which is geographically and culturally closer than the US.

Taiwan is China. Their official name is the Republic of China, and they're the remnants of the losers from the civil war that ran away go to an island and assimilated the locals. They still officially claim to be "the one and only real" China. For Americans, think it the Confederation ran away to Puerto Rico (assuming it used to be American before being occupied by Spain for a few years, before they went there) and was still there.

But today Taiwan is de facto independent. More and more of its people consider themselves Taiwanese, not "the real" China. Ideally, they should be left to self-determinate. Unfortunately China (PRC) considers itself to be the one and only real China too, and wants all of it. And it would consider Taiwan, with which it has a lot of bad blood (its former dictator literally preferred fighting the Communists over defending over the Japanese that were invading and committing mass atrocities; and he started all that with multiple purges of anyone left aligned), becoming "independent"/separate as a big humiliation. But they also know that any war will probably result in TSMC being sabotaged, so it might be all for nothing, economically. The question is will they risk it for "prestige".

[dathinab]

> Taiwan is China.

This is a sever case of word nit-picking and misinterpretation.

When (western) people say China they mean "China (PRC)" _never_ Taiwan and pretending they don't isn't helping anyone.

And just because two countries have the same root in a civil war of a past now gone country which both claim to succeed doesn't mean they are the same country. Nor is today's China the same country as idk. China during the Ming dynasty. Yes they are political successor, yes they use the same name, but no they aren't politically the same country if we really nit-pick. I mean if they where then we also would need to treat Austia and Hungary as the same country. Or say the BRD (i.e. today's Germany) is the same country as the 3rd Reich, Weimar Repulic and the Holy Roman Empire of German Nation. They are all predecessors of Germany, Germany has to take responsibility for some of it's past predecessors, too. But they aren't the same country from a political nit-picking POV. Same for Russia and the UDSSR etc. etc.

And yes it's complicated as both countries still claim sovereignty over the territory of each other. But for Taiwan that is ironically more to appease China, i.e. dropping that claim would signal Taiwan trying to finally fully break away from China (PCR) which China would never allow.

[genewitch]

Are the machines German or am I misremembering? The ones in the video were maybe 14nm and I vaguely recall the machines being manufactured in Germany and flown and assembled for TSMC.

I know, for example, that Motorola and Texas instruments know how to fab, and a dire shortage would be things like 680x0 and 6502 and 386 (see intel quark for the state of 386/486 processors a decade ago...) And ideally there'd be national RISC-V chips or something idk let's keep it fun.

[madspindel]

They are Dutch and insanely advanced and complex machines

https://youtu.be/h_zgURwr6nA?feature=shared

[sneak]

Dutch. ASML is in the Netherlands.

[genewitch]

Oof, my dad would never let me hear the end of it.

[sofixa]

The machines are Dutch, but there's more to it than just the machines (or we'd have cutting edge fabs in Germany, France, US, China, South Korea, and all around).

[tgv]

The bad blood is mutual. Mao's armies stood aside while the NRA tried to hold off the Japanese, with the sole intention of weakening the NRA. They could have prevented several blood baths.

[DiogenesKynikos]

That's a strange reversal of history.

Chiang Kai-Shek's policy was to prioritize the fight against the Communists over the fight against the Japanese. He only agreed to a united front with the Communists against Japan after he was kidnapped by some of his own generals, who forced him to talk with the Communists.[0]

0. https://en.wikipedia.org/wiki/Xi%27an_Incident

[sofixa]

No. As the sibling comment points out, Chiang ignored the Japanese invasion until he was forced by his generals to take action. After that the much smaller Communist forces (diminished after multiple purges and military campaigns by Chiang's Nationalists) fought a guerilla war against Japan, which is the only thing they could do with the forces at their disposal. Meanwhile Chiang had a proper army, and as such fought proper battles against the Japanese.