At this rate we are gonna lose, and it won’t be close
Important to note @Zai_org train on @Huawei Ascend chips, no NVIDA (!)
So you have frontier -3 months on a fully Chinese stack, 90% cheaper.
Would estimate the total cost of this to be $25m, largely post training (80%)
@Zai_org market cap now nearly $100b, $$s in open source!
Many users dismissed reports of Zai Org training a frontier model on Huawei chips without NVIDIA as fake news or unhelpful optimism, warning that US policy failures are letting China pull ahead in AI.
No Digg Deeper questions have been answered for this story yet.
Huge if true. Others have tried (deepseek) and not been able to break from nvidia in the end, so i think this needs confirmation.
If so I think this is probably on-balance good for humanity (cheap compute, cheap intelligence) but is a striking new reality for western tech
Important to note @Zai_org train on @Huawei Ascend chips, no NVIDA (!)
So you have frontier -3 months on a fully Chinese stack, 90% cheaper.
Would estimate the total cost of this to be $25m, largely post training (80%)
@Zai_org market cap now nearly $100b, $$s in open source!
@mattparlmer I think All Chyna needs to win is to not lose in AI *decisively*. Ie not get Super-Stuxnetted within the next 2 years, develop some reasonable dual use cyber AGI then they continue their industrial singularity in peace All American thinking around AI is wunderwaffe coded
Rah rah optimism about America helps nothing, people need to actually comprehend how fucked we are in order to entertain sufficiently drastic policy options
@JWarral90921 @mattparlmer There is no reason to believe that any human cognitive capability is impossible to replicate with AI, except cope. That we haven't yet is mostly an issue of resource prioritization driven by revenues.
I think that one of the facts facilitating the current Trump/Dario conflict is that Trump kind of likes China and doesn't want to Stuxnet them if it can be avoided. My evidence for this is that, if Trump wanted to take offensive action against China at that scale, I think Dario would be willing to make a lot of accommodations to facilitate cooperation there. The fact that they're not all buddy-buddy suggests to me that there's not a great deal of common ground there.
@mattparlmer Yeah, I assume the trump admin won't feel so smug about blowing Anthropic's foot off over a personal beef when China starts pumping out better-than-fable models for domestic state use only with equivalent export restrictions
@mattparlmer I agree that’s a risk. This is zero evidence.
Some of us do, some of us don’t
The other blazingly obvious fact that I think gets lost in the mix here is that Dario's vision of the future is not just lib-coded but an active enshrinement of "gray-tribe" liberalism as the undying order, something that Trump would find completely unacceptable. Dario's likely playing the admin but I don't think that Trump's plausible innate revulsion towards his whole "thing" is inactive here, and I think it's asinine that so many commentators are like "it's bc Anthropic likes publishing whitepapers and Dario is too smart and the Trump admin is full of RETARDS." Do you think Sam didn't slide an executive summary of Dario's public viewpoints to everyone at that side of the table?
@teortaxesTex @mattparlmer Genuine question - do you actually believe AGI is possible? I have to deal with AIs in creative field daily at the new job, and they are notoriously lacking in conscious self-awareness. This leads me to believe that we might crash and burn and find out, there is no I in "AGI".
@LinusMixson @teortaxesTex @rcampeligro @mattparlmer How does Trump know Dario won’t Stuxnet him instead. He hate his just as much as he hates the Chicoms.
Hilary did it to him multiple times during his first administration even without a machine god.
@mattparlmer For too many winning now means turning America into a trailer park CCP.
I truly think the only reason Dario wouldn't do this is that the EV (under their current not-transcended power level) is highly negative, given the catastrophic company-destroying consequences of discovery. Also he may have some sort of cowardly quasi-Yudkowskian thing about "my priors aren't strong enough to violate the norm of not overthrowing the government." I don't know. I can feel myself becoming deranged by hatred so I'm gonna recuse myself from this conversation lol
@mattparlmer it worked for climate change 🤷
Trump is inclined towards moderate geopolitical isolationism + trade surplus, ultimately. His ventures into military enterprise this term have been totally unlike prevailing American war doctrine for the past seventy years. He really just does not like war at all, and if he perceives China as a long-term good-faith partner who we can trade profitably with and who won't undermine our global interests — which he seems to at least more than the modal Washingtonite — he's not gonna be inclined towards the sort of attack I think Dario would be inclined to facilitate. But who knows; maybe he could be convinced that subtle poisoning of research programs and the creation of military backdoors is safe enough to be worth it. It's hard to say.
@teortaxesTex @mattparlmer what if china themselves does a "project glasswing" for the world's governments lol
@mattparlmer do we deserve to win at this point?
@teortaxesTex @mattparlmer 3/3 I think instead of reaching AGI, we'll instead reach a debate of what human consciousness and awareness are, and how true Artificial Intelligence is beyond our current capabilities, at least until we can understand our perception/awareness.
@teortaxesTex @mattparlmer 1/3 I understand that personal experience is a flawed indicator, but here's an example.
I needed AI to change something in a graphic I made earlier in Photoshop. AI messed it up, basically doing an entirely different graphic, whilst technically following the prompt.
@teortaxesTex @mattparlmer 2/3 A human has the ability to re-access a situation based on contextual clues. AIs, as far as I experienced, can't. Unless you input precise prompt (which is limited too), AI doesn't have self-awareness of the situation at which it operates.