Oh no the Chinese are doing the useless thing that gives wrong answers cheaper
https://www.cnn.com/2025/01/27/tech/deepseek-stocks-ai-china/index.html
US stocks dropped sharply Monday — and chipmaker Nvidia lost nearly $600 billion in market value — after a surprise advancement from a Chinese artificial intelligence company, DeepSeek, threatened the aura of invincibility surrounding America’s technology industry.
@Lazarou @scalzi "stupid Monkey pictures" - how did they ever catch on? Because #suckers fall for the #grifters line about free money *every time*.
Ever read https://www.goodreads.com/book/show/403452.The_Big_Con ? Certainly worth a look. Otherwise see @molly0xfff
@[email protected] Oh, no, the fake price of their bullshit generator is lower than the fake price of our bullshit generator.
@scalzi who could have foreseen that this would happen? :O
(that said, the way it rains on the parade of an assortment of shitty people is a delight of above-average size)
@pierrotechnique
> a quick look at their benchmarks shows that it's /not/ wronger though
And of course it wouldn't be like Chinese researchers, engaged in a high-profile techne arms race with their US counterparts, to carefully pick the benchmarks to produce the appearance of less wrongness ...
1) researchers all over the world do exactly this all the time, no matter the field
2) i'm not an expert in LLMs but their list of benchmarks looked quite comprehensive
so yeah, not quite sure why that casual racism was necessary
@pierrotechnique
> researchers all over the world do exactly this all the time, no matter the field
That's a strong claim. You're going to need to back it up with specifics and evidence.
> not quite sure why that casual racism was necessary
It's not racist to reference the well-established fact that the CCP value status over accuracy, and researchers under their rule regularly submit fiction to academic journals. It would be racist to say they do it because they're Han. I'm not saying that.
@strypey it's not a strong claim lol it's a "well-established fact" as you say. we all choose to evaluate our work with metrics that support our results.
yeah i've seen shitty work from chinese researchers but i've also seen super interesting stuff. in fact i'd say about the same shit to cool stuff ratio as i've seen out of american or european or any other researchers.
painting all chinese research in such a broad stroke is definitely kind of on the edge of racism imo
@pierrotechnique
> we all choose to evaluate our work with metrics that support our results
Speak for yourself. Actual scientists seek to prove the null hypothesis, specifically to avoid fooling themselves that way.
> painting all chinese research in such a broad stroke is
... stating facts. This is a well-documented problem with the Chinese research system;
"Elite researchers in China say they had ‘no choice’ but to commit misconduct"
#SmritiMallapaty, 2024
@pierrotechnique
> no true scotsman publishes without having proven the null hypothesis!
All I see here is loud noise, made by bashing on empty arguments. If you have a serious point to make, feel free to get to it.
@scalzi Imagine if capitalists used trillions of dollars to do something useful instead.
> Sam Altman CEO of OpenAI, last year said the AI industry would need trillions of dollars in investment to support the development of in-demand chips needed to power the electricity-hungry data centers that run the sector’s complex models.
Or 6 million if you use the optimised chinese code for AIs.
@josephdickson
> AI industry would need trillions of dollars in investment to support the development of in-demand chips needed to power the electricity-hungry data centers that run the sector’s complex models
And here we see the little people inside the Mechanical Turks. The whole "AI" thing is a way of artificially generating huge demand for something the DataFarmers want to build anyway; hyperscale datacentres;
https://techwontsave.us/episode/241_data_vampires_going_hyperscale_episode_1
It's the digital equivalent of corn ethanol.
@scalzi What's hilarious is that it's about 1300 lines of code using pytorch, and the "novel" solution is to train a model on LLM generated data. And the market is like "oh no!"
Scammy turtles all the way down.
@scalzi good, let the U.S. AI companies sink.
And then when the people realize the Chinese one is just as crap, it'll sink too.
Looking forward to them rolling out the next release of R1 Deepseek - R2D2
@scalzi It's interesting IMO because:
- You can run the smallest model on a Ras Pi with no internet connection.
- It prints its "train of thought".
- The models are all open source.
I have no love of China, or of how the tech industry has watered machine learning down to "everything is a shitty chatbot", but I can't deny that this is an impressive breakthrough and I hope it's enough to help drive real innovation in ML.
On the other hand, I also realize this is probably going to buoy the grift even longer.
@faoluin
> You can run the smallest model on a Ras Pi with no internet connection
That is interesting. But I've seen a smaller Llama model run on a MacBook with no net connection. I'm still not convinced that makes it anything beyond a vaguely amusing waste of resources.
> The models are all open source
Including the weights and the data they're generated from?
@faoluin
> [weights and the data being freely licensed] would be preferable
As a number of software freedom activists have pointed out, these are just as necessary to recreate the model as its source code. So it's not just preferable, it's a prerequisite for saying with accuracy that ...
> The models are all open source
@faoluin
> I'm not going to waste time on pedantics on this topic
You specified that the models are "open source" is if it matters. Now you're arguing that distinguishing actual Open Source from openwashing is "pedantics". Not sure how you square that circle.
@scalzi I can already do cheaper, faster and wronger generative AI!
cat /dev/urandom
Attached: 4 images #Deepseek is absolutely the new tankie toy! Congratulations to all #China adoring #tankies! :tankies: Don't get me wrong, the American #AI's are also lying and making mistakes. But I have to admit that #ChatGPT is not so obvious lying as Deepseek. Although that can change. 👀 Imagine a future were AI tools from various power blocks are fighting each-other who is the greatest historical revisionist. All that lying costs of course immense energy, making the climate crisis a climate apocalypse, and destroying all human culture and knowledge. All for who has the best lie. OK, the former paragraph is more of a good science fiction film. But good science fiction films have a grain of truth. Proposed title 'The War of the Lies'. * * This splendid scifi film idea is licenced under the Creative Command licence BY-NC-SA 4.0.
Time will tell, but they do seem to have produced a better mousetrap as well.
Competition is a bitch. Too bad the free market fundamentalist spent the last 40 years stamping out any with that pesky competition.
Because free market fundamentalist actually argue that monopoly is fine.
The marketplace doesn't need actual competitors competing in it in order to be free market.
🧵
"It hasn’t yet proven it can handle some of the massively ambitious AI capabilities for industries that – for now – still require tremendous infrastructure investments. "
neither has ChatGPT lol
Nate Sullivan
Scalzi
Lazarou Monkey Terror 🚀💙🌈
CoolMcCool
Luna chan
Felix Eckhardt
RebelLog
Roger BW 😷
noodlejetski 
raphael
Zeugs
Eva Chanda
Darwin Woodka
Pierre Massé
Strypey
Joseph Dickson
Estarriol, Terrorist Dragon
Mal 🇿🇦🏳️🌈🇮🇪🏳️⚧️🇸🇪
madopal
ManVsXerox: Resistful Dingus
Jonas R. (er/ihm)
Twobiscuits🚴♂️ 
Raj 🇬🇧🇪🇺💻🖥(🌻🇺🇦;🇵🇸)
BunRab Is Old And Tired
Patrick Morris Miller
jiub 
ProScience 🇪🇺
Ben
JavAlps
bureauxx
Softwarewolf
Wendroid 🇺🇦
Elias
Riley Norman
joene 🏴
Daniel M Karlsson
David
Mᴀʀᴋ VᴀɴᴅᴇWᴇᴛᴛᴇʀɪɴɢ
Tom Forsyth
Nicky🦨✌️
GhostOnTheHalfShell
Random_Seed 
💾 🇿🇦
Nick Appleyard
Simon Michalke
Seth Galitzer
Gabs
Darberoom