Also, the original title is only 77 characters long, so there was no reason to change it. And the title should never be changed to something that misrepresents the article.
I have a question for those who closely follows Cerebras: do they have a future beyond being inference platform based on (an unusual) in-house silicon?
Ever since the recent revelation that Ars has used AI-hallucinated quotes in their articles, I have to wonder whether any of these quotes are AI-hallucinated, or if the piece itself is majority or minority AI generated.
If so, I have to ask: If you aren’t willing to take the time to write your own work, why should I take the time to read your work?
I didn’t have to worry about this even a week ago.
the hardware diversification story here is more interesting than the speed numbers. OpenAI going from a planned $100B Nvidia deal to "actually we're unsatisfied with your inference speed" within a few months is a pretty dramatic shift. AMD deal, Amazon cloud deal, custom TSMC chip, and now Cerebras. that's not hedging, that's a full migration strategy.
1,000 tok/s sounds impressive but Cerebras has already done 3,000 tok/s on smaller models. so either Codex-Spark is significantly larger/heavier than gpt-oss-120B, or there's overhead from whatever coding-specific architecture they're using. the article doesn't say which.
the part I wish they'd covered: does speed actually help code quality, or just help you generate wrong code faster? with coding agents the bottleneck isn't usually token generation, it's the model getting stuck in loops or making bad architectural decisions. faster inference just means you hit those walls sooner.
Large scale Capital is not gonna make any more investments into microelectronics going forward
Capital is incentivized to make large data centers and very high speed private Internet, not public Internet, private Internet like starlink
So the same way in the 1970s it was the main frame era and server side computing, which turned into server side rendering, which then turned into client side rendering which culminated in the era of the private computer in your home and then finally in your pocket
we’re going back to server side model communication and that’s going to encompass effectively the gateway to all other information which will be increasingly compartmentalized into remote data centers and high-speed access
The rich have purchased the means of production (as in, prod, the deployment target), and I hate it. The home deployment target is just so fucked for so long.
11 comments
[ 2.8 ms ] story [ 38.0 ms ] threadIf so, I have to ask: If you aren’t willing to take the time to write your own work, why should I take the time to read your work?
I didn’t have to worry about this even a week ago.
They used amd gpus before - MI300X via azure a year plus ago
1,000 tok/s sounds impressive but Cerebras has already done 3,000 tok/s on smaller models. so either Codex-Spark is significantly larger/heavier than gpt-oss-120B, or there's overhead from whatever coding-specific architecture they're using. the article doesn't say which.
the part I wish they'd covered: does speed actually help code quality, or just help you generate wrong code faster? with coding agents the bottleneck isn't usually token generation, it's the model getting stuck in loops or making bad architectural decisions. faster inference just means you hit those walls sooner.
The era of “Personal computing” is over
Large scale Capital is not gonna make any more investments into microelectronics going forward
Capital is incentivized to make large data centers and very high speed private Internet, not public Internet, private Internet like starlink
So the same way in the 1970s it was the main frame era and server side computing, which turned into server side rendering, which then turned into client side rendering which culminated in the era of the private computer in your home and then finally in your pocket
we’re going back to server side model communication and that’s going to encompass effectively the gateway to all other information which will be increasingly compartmentalized into remote data centers and high-speed access
Your post is terror-resonating with me.