Sh_

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

8 months, Q1 2027
the bigger problem is that Mythos itself is a last generation pretrain, and by then Anthropic will likely have completed something vastly more capable, so the real gap will only increase, towards Dario's targeted 24 months.

[Edward @somefoundersalt]

how it feels to be a hardware engineer visiting shenzhen

[WireWhiz @TheWireWhiz]

Hear me out, vim, but you use this to switch modes:

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

$650K for your own personal mecha that can go Gorilla Mode
Imagine the customization…
No idea who's buying this but its pure childish awesomeness can't be denied

[GalacticMarch @goandstudy117]

@_akhaliq wow

[xjdr @_xjdr]

this is art

[meowbooks @meowbooksj]

MAY RAM NEVER BE CHEAP AGAIN

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

Companies with semi-frontier or frontier labs, from memory:
Alibaba (2+ labs), Ant, BOSS Zhipin, ByteDance, DeepSeek, Meituan, MiniMax, Moonshot, OpenBMB, RedNote, StepFun, Tencent, Ubiquant, Xiaomi, zAI
Any of these could rival Anthropic if given 300K H200s and 6 months

[xjdr @_xjdr]

man, it turns out all you really need to bootstrap a frontier lab is impeccable research taste, world class engineering talent, a single pretrained base model and enough compute to both serve and train that model at the same time

[sphinx @protosphinx]

that is not a swarm. this is a swarm.

[Zephyr @zephyr_z9]

v4 will be the DeepSeek moment of 2026

[ludwig @ludwigABAP]

J Cole sampling that Wiles interview on Fermat’s last theorem for his promo, kinda neat

[tender @tenderizzation]

gradient descent navigating the loss landscape

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

On 26th December 2024, DeepSeek gave the world the open base model that continued to challenge the frontier for the entire following year.
A year has passed without fanfare.
Modal days of DeepSeek releases are Monday and Wednesday. Wednesday is December 31st. A man can dream…

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

“Except this simply is not true”
I notice non-technical people fail to understand this point about Chinese dependency on CUDA. So here I'll explain.
You can NOT create an ecosystem by decree and subsidy. That's a real issue even in a “command economy”, and that's why China rejected H20s and is being so lukewarm even towards H200s. You need people, lots and lots of talented people, to use your chips, your servers, your frameworks, your APIs, your whole architecture, to solve cutting edge problems, give you feedback and contribute. As a rule, talented people have interesting stuff to do, and it's usually not “fixing an ngmi platform”. Superstar effects are very significant, learning curves are unforgiving. AMD is still struggling! Huawei can make good (in theory) silicon. But if the only people training on Ascends are Huawei and IFlyTek, that silicon will not get anywhere, because Huawei and IFlyTek are plainly clueless about frontier model training.
You see, everyone likes that Wenfeng's quote about lacking chips, but the core limiting factor in the “AI race” is still not compute but TALENT, both in R&D and management. Compute is the limiting factor *for DeepSeek specifically* because they are geniuses, unironically, IOI Gold level monsters. They have all the talent they need, they only lack GPUs to execute on their vision. Conversely, compute is NOT the limiting factor for titans like Meta, Microsoft and Amazon, as can be surmised from them making garbage tier models despite 1000x or greater compute advantage over DeepSeek. Most labs that whine about their hunger for compute and joke about “hnngh I want gpus go brr” despite glaring inability to match a mature V3 with vastly more than 2048 H800s are saving face. They are not bottlenecked by FLOPS, they are bottlenecked by skill issues. That's how it is in late 2025, at the end of the era of blind scaling, as announced by Ilya and the faithful. Time to update your priors.
At least two of DeepSeek engineers that I know are former NVidia interns, incidentally. What do you think they prefer to work on?

Insofar as China has steady access to NVidia chips, this very directly and VERY DRASTICALLY slows down the development of their sovereign tech stacks because ≈100% of the cream of the crop of their talent will just go to labs that use NVidia, to work without hassle on their interesting frontier research and deliver world-class results, rather than debug CANN-ONNX with a sleepy Huawei intern. Simple as.
They're not a Maoist society, they can't tell a proud 3500 CodeForces anime pfp autist with a top NeurIPS paper to go fuck around with CANN, he'll just leave, either to a saner lab or outright to the US. This has been going on forever, and this is what motivates Xi to impose restrictions and inconveniences on H200 procurement – he wants those who stay to have no choice but to assist domestic chipmakers. You may find this policy wrong due to a disagreement about AGI timelines and trajectory. Fine. But that is the Chinese policy. Huawei, Cambricon and others are hurt massively by NVidia. Add AMD, which seems to have finally rubbed two braincells together, and it's a full-out attack on their capability to develop an alternative to American architectures within a decade.

I am relatively optimistic about outcomes for China in either case. They'll figure it out. My point is simply motivated by irritation about this very crude and smug misunderstanding. No bitch, locking them onto your stack has value. Do with this information what you will.

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

«The Vermilion Bird is elegant and noble in both appearance and behavior, with feathers in many different hues of vermilion. It is very selective about what it eats and where it perches»
what a choice of a name for a methalox reusable rocket.

[Zephyr @zephyr_z9]

Quant firm doing quant things

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

No, it's fine, I just… I just didn't think «the singularity» would be Nvidia vendor financing

[Yicai 第一财经 @yicaichina]

Chinese robotic vacuum manufacturer Dreame today announced its foray into the automotive industry. Its first car, an ultra-luxury pure EV that aims at competing with #Volkswagen's #BugattiVeyron, is scheduled to debut in 2027. @DreameGlobal

[Zephyr @zephyr_z9]

This is very cool