Many users prefer DeepSeek's Wengfeng models over Anthropic's offerings when trained on large B200 chip clusters because they see them as more valuable and better executed.
Based on 3 visible X reactions from 9 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@NunoSempere I don't even say this is kayfabe or a subsidy. EV makers can eventually turn profitable and this is more valuable than another EV. At 2%, large debt could be serviceable. In any case, they repeatedly speak of a 100K cluster as something that's real (or close to being real) https://x.com/teortaxesTex/status/2076745814584500638/photo/1
@teortaxesTex Idk man, I think if you let Wengfeng cook with 100K B200-equivalents I'd prefer his model to the Fable model Anthropic is currently offering
@NunoSempere Link I do like how they write https://app.mokahr.com/social-recruitment/high-flyer/140576#/job/85084089-eaf4-4df8-b3d6-1634a5200218
DS obviously can get the compute for pretraining a "10T" (something-AB) model, as can many more players. It doesn't take a whole lot of compute. The problem is that this number is overrated. What model? How much would it cost for them to discover an *Anthropic tier recipe*? https://twitter.com/NunoSempere/status/2076738479262945318
@NunoSempere I don't even say this is kayfabe or a subsidy. EV makers can eventually turn profitable and this is more valuable than another EV. At 2%, large debt could be serviceable. In any case, they repeatedly speak of a 100K cluster as something that's real (or close to being real) https://x.com/teortaxesTex/status/2076745814584500638/photo/1
@teortaxesTex Idk man, I think if you let Wengfeng cook with 100K B200-equivalents I'd prefer his model to the Fable model Anthropic is currently offering
@NunoSempere Link I do like how they write https://app.mokahr.com/social-recruitment/high-flyer/140576#/job/85084089-eaf4-4df8-b3d6-1634a5200218
DS obviously can get the compute for pretraining a "10T" (something-AB) model, as can many more players. It doesn't take a whole lot of compute. The problem is that this number is overrated. What model? How much would it cost for them to discover an *Anthropic tier recipe*? https://twitter.com/NunoSempere/status/2076738479262945318
that said: > Compare with DeepSeek's USD 7B recent round; this is the right order of magnitude I'm maybe naive but why doesn't Wenfeng just go to the bank? Why is "a round" + whatever profits made the limit of his capital? Chinese banks are known to lend below market to industry.
And that's before post-training As far as I know GPU-rich labs don't even do "a pretraining run", they run several similar jobs and then promote the best one to post-training and prod. The Chinese have to make-no-mistakes every time, it's a whole different regime.
@teortaxesTex DM
Many users prefer DeepSeek's Wengfeng models over Anthropic's offerings when trained on large B200 chip clusters because they see them as more valuable and better executed.
Based on 3 visible X reactions from 9 accounts; directional sample.
Ask a question below.
Published answers will appear here.
And that's before post-training As far as I know GPU-rich labs don't even do "a pretraining run", they run several similar jobs and then promote the best one to post-training and prod. The Chinese have to make-no-mistakes every time, it's a whole different regime.
@teortaxesTex DM