So, just thinking about this Chinese AI thing and the parallels to how a Fediverse ActivityPub node can run on a $50 Rasberry Pi, but a Bluesky AT Protocol node requires a $30,000,000 supercluster.

Walt Baldwin

@mastodonmigration

really wonder if those catz are lying with media reports of clusters ranging from 10k a100s to 50k h100s.

faraiwe

@mastodonmigration counterpoint.

The answer to lots-of-nazis is NOT more-nazis (even if they are cheaper nazi).

Consider the analogy as you see fit. Or not.

Mastodon Migration

@faraiwe

Not following. Just talking about how technology design has a huge impact on efficiency and cost. Unclear what this has to do with Nazis, but maybe just missing your point.

faraiwe

@mastodonmigration

Probably.

Technology does NOT exist in a vacuum. Its intended use and actual usage ARE the point.

NO #LLM was ever used to anything other than to make skill accessible for those who have wealth.and move wealth away from skill.

None.

A new LLM is not gonna change this. Merely gonna change the routing number of the bank account.

Can't make this any clearer, or make people want to understand.

Doesn't matter if the Chinese LLM was "cheap" (because they naturally copied R&D from the existing ones, never bet against that). Their LLM will do what LLMs have done.

Mastodon Migration

@waltbaldwin

Not following. Are you talking about the Chinese DeepSeek. Here is a good technical explainer. He thinks the claims are credible:

DeepSeek FAQ

DeepSeek has completely upended people’s expectations for AI and competition with China. What is it, and why does it matter?

Stratechery by Ben Thompson (stratechery.com)

Mastodon Migration

@faraiwe

Got it. Don't disagree with you at all. Certainly not making a case for AI. Indeed, cheaper garbage is still garbage. What is interesting is how better design of algorithms can lead to radical changes in computational efficiency and cost. We saw this in the 80s and 90s with imaging and video compression. The entire industry was turned on its head multiple times by breakthroughs in encoding/decoding efficiencies.

faraiwe

@mastodonmigration

Never bet against the fact that douchebag techbro billionaire (irrespective of specific face, birth location, douchy facial hair or headgear, and spewed filth from their facial orifices) developed tech is *inherently* bad and can never do any good.

#DTBO #LLM

Mastodon Migration

@faraiwe

Agree that their aims are to turn it into a massive grift.

That said, there are things like astrophysics and medical research where such programming techniques will be helpful.

faraiwe

@mastodonmigration

OR

we could dump what was made by people with ill intent, and develop different things, from ground up, without the contamination of #DTBO intent.

Be very very wary of those who think all problems can be solved by technology. They can not. Categorically, absolutely can not.

I have yet to see a single medical or astrophysical application of #LLM that was worth their insane consumption of resources, the social cost, and the end-state of their owners destroying democracy.

It's just the latest idiotic box that beeps and boops, for techbros to go WOO, GIMME MONEY, and fools to open wallets. Nothing else. Nothing more.

Walt Baldwin

@mastodonmigration

yes, there are others that are saying the costs could be much higher. cnbc last week claimed they have a 50k cluster of nvidia h100s. deepseek themselves claimed a couple years back that they had a 10k cluster of a100s. if any of that is true, the costs could be north of their $5.6m claim.

the interview above addresses those claims so i'm not certain what the reality is.

Carlos Solís

Well... There are several ways to tackle the metaphor on each side.

As for #Bluesky: apparently, the devs are aware that using a single firehose for everything in the network is prohibitively expensive to replicate at home, so they're developing "jetstreams", which only replicate data from people you follow and only retain a certain amount of days of data. Something that would have been unnecessary, had they intended the network to be truly distributed since the beginning.
As for #DeepSeek: looks like the US sanctions came back to bite them, as the Chinese were then forced to start optimizing with what they had. And regardless of this, while I hope the methods used by DeepSeek can be generalized to make in-house model training with properly-consented content only, I cannot trust anyone not explicitly targeting that specific scenario, otherwise I know the legal and ethical implications will not be worth the trade-off.

Solarpunk Davy

@waltbaldwin @mastodonmigration very efficient way for China to burst that AI bubble. The market already knew they were in a bubble, so at the first sign of trouble, they started dumping stock... No receipts needed.

Jorge Stolfi

@mastodonmigration

It is not easy to be evil, you know.

Mastodon Migration

@faraiwe

No argument.

Mastodon Migration

@csolisr

So basically ActivityPub

Mastodon Migration

@JorgeStolfi

Evil works for a while, until it doesn't.

Viral Obscurity

@mastodonmigration I guess that's what happens when you bake profit into every part of the system

I'm sure the Chinese system, if costs are accurate, is done at cost for political capital over amerika and not for everyone involved to make big money

Mastodon Migration

@viralobscurity

Bingo. Being sneaky and tricky makes you vulnerable. Unfortunately sneaky and tricky is the Silicon Valley way. This may just be the tip of the iceberg.

SamuelJohnson

@viralobscurity @mastodonmigration And success in EV's? Not everything is done for reasons of political capital. Climate change matters to China too.

China is determined to succeed and to do so despite obstacles placed in its way, because it can.

Mastodon Migration

@samueljohnson @viralobscurity

EVs are another great example. Here we prop up super expensive monster vehicles while the rest of the world makes cool small affordable EVs. Guess which strategy wins in the long run?