I mean, even if the ship were sinking, it's not like the plan is stupid per se.
It's a neat plan. Keep prices high, circularly finance so that you still have customers, getting ownership of other firms and effectively having lower prices.
It's a superb plan for NVIDIA, and then of course, people try to get around that plan: we Europeans obviously want to be able to do AI, and if we aren't getting the circular financing will of course want to make our own chips and do our own circular financing, or at least have normal prices. Google, Amazon, etc. obviously don't want to pay with equity, so they make their own chips.
I don't think it has to be nonsense. I think it's a good plan, it's just that anybody who wants to exist or be their own thing is fighting against it as they of course should.
In the end I hope for a competitive DL accelerator market with prices 1/5th or less of what they are now. High prices tend to lead to high prices though.
I am saying the cultivated hysteria is nonsense, not that nvidia’s strategy is.
Nvidia were notorious assholes to work with 20 years ago, but people persisted in doing so because they needed to. It would not be remotely surprising if their strategies involved concepts many fond unpalatable even if highly effective.
What's so magical about selling to the USG that it should create some sort of obligation on the part of a civilian CEO?
Remember, Huang isn't an army officer, or someone in intelligence working at the Pentagon, or a CIA officer.
Surely he gets to do what he likes. You are not at war with China. Treason is not relevant here, nor are graphics cards some sort of ultra-secret domain. What NVIDIA does is easy, it's just that the package is not so easy to reproduce, what with it having many aspects, from software to chips and fast interconnect.
If you sell national security impactful equipment to the USG, you're in a special category, private industry or not.
Do you think the USG doesn't hold the CEOs of Lockheed, Raytheon, General Dynamics, or Northrop Grumman to higher security standards than Delta or Ford?
And I think it's hard to argue that in 2026 Nvidia is not designing and manufacturing national security impactful equipment.
Nvidia is the Cray Research of the last couple decades, and the USG has always given a shit about where frontier-level mainframes / supercomputers can be sold.
Lots of scary stuff has been possible for ages. Much of it is so easy that wouldn't be publishable because of lack of novelty.
I see LLMs as being essentially irrelevant for whether someone can do this. It's lab skills, planning skills, biological experiment debugging skills.
"Why did my plasmid not start getting expressed, how did I screw up?" "How do I make this huge protein" aren't LLM questions, they're university level biology questions, and "why is my protein growth tank not making the thing I want, and what is the grey stuff, do I need to put clean it again" is an experience question.
Putting checks on DNA synthesis puts making the horrible stuff out of the range of what high schoolers can do though.
I remember years ago now, trying to feed a text in twice etc., hoping that the representation in the KV cache of the second would end up better. I don't think it worked all that well.
But maybe that is most of the function of restating things babbling. To look at things multiple times.
You can't seriously be arguing for Isekai's and levelling.
You could fix it. Drop the levelling, keep the portal story-- portal stories can be fun, C.S. Lewis and Astrid Lindgren have done portal or portal-adjacent stories and they can be fun.
I don't think a fixed portal story is necessarily attractive to idiot-readers though. I think popularity of Isekai or levelling stories might have more to do with impatience and inability to tolerate uncertainty.
Unrelated to this slowdown, I could imagine this sort of thing being the right application for things like Taalas, although I'm not quite sure whether it can actually offer what I imagine that it can (i.e. ability to make a cheap, physically small card that can run a large model).
A laptop can't run the latest big models, because memory is too expensive. This thing probably can't run models larger than Mistral Small 4. But it'd be a completely different thing to have a hardcoded Mistral Large 4, than a choice of tiny models.
You probably won't be able to run a sota model on a single Talaas card. You probably need a set. It should theoretically be cheaper, faster, more power efficient than Nvidia at running that model. But then you are stuck with a chip that can run only that model. And I understand that context size will also be a problem.
Yeah, that was why I felt that, while Taalas might not be the right thing in general, it could be really appealing in a laptop, that would otherwise not be able to run any big model at all.
I think I misunderstood what the technology could do a little bit, because I imagined these fixed matrix multiplies encoded into hardware as small-ish, as using routing space etc., not just fast.
I think what I really imagined was useful would be more easily achieved by having a big ROM with model weights connected to an ordinary GPU.
I think a post-training common would have to be somewhat protected, since these things would be created with great effort and would potentially impact people's ability to earn money in their jobs.
I could also imagining them happening as state projects for the sake of competitiveness.
One of the things I'm/we're thinking through is how to do this in a way that’s participatory rather than extractive. It's not necessary that what scientists would choose to automate with agents is the same as what a company trying to automate scientists would choose.
If we can help people automate drudgery while giving them a real say in what gets built and control over how their work is used, I think there’s a possibility of a much better outcome than the current trajectory. I’m not sure we’re in that world, but I think we should try.
I'm not sure I'll do it justice, but one of the general ideas behind open-model initiatives is that the economic impacts of specialized post-training are a fait accompli. OpenAI and Anthropic have spent $many-millions procuring closed posttraining data.
The relevant question here is whether you want open-weight models — projects like Open Athena — to remain or become competitive in the same range of specialties.
> I also wonder when the NSA starts sharing their wiretaps/etc with labs for 'national security'.
What would this benefit the NSA? If they share wiretap records with labs such as Anthropic, then suddenly anyone who can jailbreak a model gets access to this data. You don't need to train a model on data to understand or interpret it, as far as I know - providing others (esp. AI labs) with sensitive data is a breach of security, or I am not understanding.
I should’ve been more specific, I mean more industrial espionage and getting access to internal data that would be useful to frontier labs for training.
See the well-documented case of NSA conducting espionage against EU’s Airbus to help Boeing.
NVIDIA has done great work with the open datasets and they're much appreciated, but it's worth mentioning that a large fraction of their post-training data is not released. cf. https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1... But at leased they disclosed what wasn't released.
Wrong direction. NSA says “give us the weights to Fable and loan us some engineers who can get a TS-SCI clearance so we can train our own model in house, if you don’t well just take them, in fact maybe we already did, it’s for national security”.
No, NSA programs are not SCI, haven't been for a few years. They are SAP.
NSA doesn't have to post train Fable. Maven works well enough and that same infra. can be used for Fable. I imagine that's what's going to happen in 2 years once the current admin. moves out.
reply