Isn't kernel development one of their banned topics in their public facing models? The ethics of this are just so inherently dirty. They can enjoy the benefits--the sum value of all that compiled knowledge--but not us. If open models don't succeed, the future is looking pretty bleak. Really. This normally feels like such a tin hat thing to say, but damn if it doesn't feel true right now.
> Isn't kernel development one of their banned topics in their public facing models
I've been doing custom HIP kernels for a while now without any issue whatsoever. I guess it depends what you're trying to achieve? If Claude recognizes that your goal is password cracking or key breaking, pretty sure the session will be flagged, but I've been playing with custom HIP kernels to run open-source/open-weight models without any issues for a while now, even on the 5.5 series models.
That's because that's how it is, and not only that, they can use whatever un-used compute to do it. If they're at 30% capacity they can have 70% of max data center resources do all the things we can't do, with the things we can't and don't have (because of price fixing and circular economies).
Honestly it's probably just a matter of time before we all have a tiny little Rasp-Pi looking thing that powers AI locally, bring back self-hosting IMO after seeing the Office 365 stuff down.
You know damn well they have teams working on some crazy stuff behind the scenes. We're just here to fund them by buying tokens.
Honestly the rise of Chinese open models is the biggest stroke of 'dumb luck' (strategy, whatever you want to call it) that I would say we've ever had in the last 100 years or so. With how locked down the internet has become now, you'd never be able to source a corups of enough training data today.
My example case of not having an open model, is if I had to write half the software I do now, without the aid of agents. It would be beyond impossible to keep up with anyone.
Granted reading code and tests is boring and painful, but it's still much faster than having to write the code, and the tests, then read it all!
That’s why I really don’t care about Huang allegedly helping smuggle chips. Fuck US AI hegemony. I don’t want a handful of shitty American CEOs controlling the world. The future envisioned by Liang Wenfeng is so much more compelling.
> Isn't kernel development one of their banned topics in their public facing models?
I see various fable/opus-vibecoded "kernels" (and even benchmarks) of various flavors just searching for "fable kernel" or "fable kernel gpu".
The barrier to trying this sort of thing yourself is prob more the token consumption * API prices (and expert oversight?) required for a serious project than the security/CYA restrictions.
People constantly hit security guardrails doing kernel and inference engine development. Not too hard to work around them, but its very frustrating. Not to mention the publicly released Fable 5.2 intentionally nerfs its kernel optimization knowledge. Meanwhile, Mythos can write Anthropic engineers whatever kernels they'd like.
AMD doesn't get access to Fable or Mythos. Big problem now that non-Mythos is locked out of doing GPU development work.
Nvidia has access to everything at huge discounts, including unreleased models.
The "moat" is that anyone who isn't a $1+ trillion company or an AI lab is going to be locked out of the cartel going forward because they don't want you competing in their space as they consume more of the economy.
So yes, Mythos can one-shot CUDA, but only Anthropic/OpenAI are allowed to use it.
Just before Anthropic's IPO, just after it looks like nvidia may be struggling. How convenient!
How do we know it's not another lie, or full of conveniently hidden limitations? But I guess that works, investors will be appeased enough either way.
AMD is not AI's bank though, so their cards are the wrong cards to run AI on (and that might be the reason we've not seen a "we can run on AMD too!" announcement earlier)
I doubt whether this will eventually be the undoing of Nvidia -- no, Jensen's affair doesn't even count.
If Claude, which is trained on Nvidia chip, becomes so powerful that it allows developers to get onboard with ecosystems other than CUDA easily, what moat does Nvidia still have?
Remember, CUDA is widely viewed as the greatest -- some even say only -- moat Nvidia has. You can duplicate the hardware, the software, but you can't duplicate the developer ecosystem. Developers are reluctant to change because it really takes time and they don't want to ditch existing, mature toolsets and ecosystem.
But with Claude being so powerful, it becomes vastly easier to learn new skills and to duplicate the toolsets. The switching cost drops significantly, so what is there to prevent China or anyone else to switch to Huawei or AMD or other ecosystem?
Isn't kernel development one of their banned topics in their public facing models? The ethics of this are just so inherently dirty. They can enjoy the benefits--the sum value of all that compiled knowledge--but not us. If open models don't succeed, the future is looking pretty bleak. Really. This normally feels like such a tin hat thing to say, but damn if it doesn't feel true right now.
> Isn't kernel development one of their banned topics in their public facing models
I've been doing custom HIP kernels for a while now without any issue whatsoever. I guess it depends what you're trying to achieve? If Claude recognizes that your goal is password cracking or key breaking, pretty sure the session will be flagged, but I've been playing with custom HIP kernels to run open-source/open-weight models without any issues for a while now, even on the 5.5 series models.
That's because that's how it is, and not only that, they can use whatever un-used compute to do it. If they're at 30% capacity they can have 70% of max data center resources do all the things we can't do, with the things we can't and don't have (because of price fixing and circular economies).
Honestly it's probably just a matter of time before we all have a tiny little Rasp-Pi looking thing that powers AI locally, bring back self-hosting IMO after seeing the Office 365 stuff down.
You know damn well they have teams working on some crazy stuff behind the scenes. We're just here to fund them by buying tokens.
Honestly the rise of Chinese open models is the biggest stroke of 'dumb luck' (strategy, whatever you want to call it) that I would say we've ever had in the last 100 years or so. With how locked down the internet has become now, you'd never be able to source a corups of enough training data today.
My example case of not having an open model, is if I had to write half the software I do now, without the aid of agents. It would be beyond impossible to keep up with anyone.
Granted reading code and tests is boring and painful, but it's still much faster than having to write the code, and the tests, then read it all!
That’s why I really don’t care about Huang allegedly helping smuggle chips. Fuck US AI hegemony. I don’t want a handful of shitty American CEOs controlling the world. The future envisioned by Liang Wenfeng is so much more compelling.
The biggest stroke of dumb luck in the past 100 years was penicillin. Chinese models don't get even close...
> Isn't kernel development one of their banned topics in their public facing models?
I see various fable/opus-vibecoded "kernels" (and even benchmarks) of various flavors just searching for "fable kernel" or "fable kernel gpu".
The barrier to trying this sort of thing yourself is prob more the token consumption * API prices (and expert oversight?) required for a serious project than the security/CYA restrictions.
People constantly hit security guardrails doing kernel and inference engine development. Not too hard to work around them, but its very frustrating. Not to mention the publicly released Fable 5.2 intentionally nerfs its kernel optimization knowledge. Meanwhile, Mythos can write Anthropic engineers whatever kernels they'd like.
Not sure about their policy, but i see LLM-gen PR to opensource RTOS all the time.
AMD doesn't get access to Fable or Mythos. Big problem now that non-Mythos is locked out of doing GPU development work.
Nvidia has access to everything at huge discounts, including unreleased models.
The "moat" is that anyone who isn't a $1+ trillion company or an AI lab is going to be locked out of the cartel going forward because they don't want you competing in their space as they consume more of the economy.
So yes, Mythos can one-shot CUDA, but only Anthropic/OpenAI are allowed to use it.
This will happen to any easily verifiable domain.
So the question becomes - how do we ease verification in hard-to-verify domains?
Just before Anthropic's IPO, just after it looks like nvidia may be struggling. How convenient!
How do we know it's not another lie, or full of conveniently hidden limitations? But I guess that works, investors will be appeased enough either way.
AMD is not AI's bank though, so their cards are the wrong cards to run AI on (and that might be the reason we've not seen a "we can run on AMD too!" announcement earlier)
I doubt whether this will eventually be the undoing of Nvidia -- no, Jensen's affair doesn't even count.
If Claude, which is trained on Nvidia chip, becomes so powerful that it allows developers to get onboard with ecosystems other than CUDA easily, what moat does Nvidia still have?
Remember, CUDA is widely viewed as the greatest -- some even say only -- moat Nvidia has. You can duplicate the hardware, the software, but you can't duplicate the developer ecosystem. Developers are reluctant to change because it really takes time and they don't want to ditch existing, mature toolsets and ecosystem.
But with Claude being so powerful, it becomes vastly easier to learn new skills and to duplicate the toolsets. The switching cost drops significantly, so what is there to prevent China or anyone else to switch to Huawei or AMD or other ecosystem?
I think anthropic mentioned a lot of work on multi vendor chips, so they might have filled the moat already?
Moat isn’t filled when Nvidia is still outselling 10-1 for AI labs
Ironic this is the same company trying to prevent people from modding Skate 3 inside of Minecraft
"Not like that!"
- NVIDIA, probably.
"Anthropic Fills NVIDIA's Metaphorical Moat with Metaphorical Gravel"