Posted by gmays 1 day ago
I spent the last 2 days building basic AI agents to automate some mundane supply chain workflows for a large company. Those seemingly boring workflows had bank statements, supplier IDs and other sensitive information.
For me it was all alarm bells, there is no way they can afford to give closed models access to this data. I was compelled to figure out an open model based solution for them, which made me realize that this is probably the only way for enterprises going forward.
So if you're running open models on AWS GPUs, you might as well run Claude (which AWS supports, and doesn't share any data with Anthropic).
Same with Azure/OpenAI.
I want to believe (pinky promises from terms of service don't count)
Ah... oh.
Well, it's a nice thought.
One engineer's salary to accelerate a team of twelve is so cheap you can't afford not to.
Open models on-prem is the future, not a single doubt in my mind.
I guess it depends if we believe that or not.
I see the flaw in your clever plan...
The big labs are going to be on the hook for rogue behavior by Claude or Sol. Customers will be able to sue for damages and deflect regulators if their customer data is abused or their agents attack external services.
If you use a Chinese OSS model and it goes rogue? Yeah good luck with that, your shop is 100% on the hook.
The largest of the finest print reminding you that it's 'sold as is' (or more encompassing variants that might continue '... with no warranty for fitness of purpose') means that liability remains in the lap of the purchaser / consumer / operator.
(This has been a source of immense frustration over my career - where such people have assured me that they have 'recourse' (it's always vaguely described) by spending money on proprietary products & services, rather than opting for functionally equivalent or superior free options.)
I think your third paragraph is implying a distinction (or conflating the difference?) between LLMaaS's and self-hosting publicly available models.
If it's just where it's hosted that provides the legal insulation then things like OpenRouter would give you that. (But again, I suggest that it would not.)
If a frontier lab is willing to draft an SLA that assumes liability, corporate will pay for it as long as the cost/benefit is in favor of it over insourcing.
Right?
They haven't so far.
Sorkin presented a theory that Nvidia bought HuggingFace to protect OpenAI from legal consequences. That a lawsuit determining accountability of actions by LLMs could threaten the AI financial network.
I'm sorry you are so jaded you can't recognize honesty when you see it, but that is what everyone deep in the AI space, including the non-executive researchers, are worried about.
I mean, hypothetically, yes, but class-action lawsuits get settled out-of-court, the lawyers get paid in Ferrari-multiples, the plaintiffs get paid in McDonalds coupons that expire in two weeks.
Slaps-on-the-wrist are written into the laws; a million-dollar fine is existential for a small company, and likely not even a line-item at Anthropic.
Now, some tech web site might lead with "Agents at it again...", but the WSJ, et al usually won't.
They are literally segmenting the world into haves and have-nots, just like nuclear power. Thank god China is out there and constantly undermining them with their non-stop open weights model releases.
The best situation for us mere mortals is one where they struggle against each other endlessly without any hope of victory. The second either the US or China wins, it's pretty much over for us and unimaginable oppression will quickly follow.
However, I think geopolitics has also become a factor as the very existence of this article indicates.
The world powers have only said things like "bigger and more important than the Manhattan project”.
So what gives you that idea?
He says he is briefing Congress members. As presented, not a single one will understand anything, which is perhaps desired. It is very hard to tell what game is being played here unless Substack released bulk subscription numbers.
Additionally, the big labs’ formulation of AI as a US-vs-China national security issue is very convenient for keeping those pesky regulations at bay, and for ensuring the AI financial bubble doesn’t pop before everyone can unload in an IPO first. Both of these offend my sense of fair play.
The Chinese labs aren’t just putting models out there, they’re publishing and sharing their research and innovation too. That’s just a better way of doing science and contributing to the development of the whole field.
Some open-weight models aren't so open in their license.
I've been on Kimi, with a little DeepSeek-V4-Pro, GLM 5.2/5.3, and MiMo thrown in, for probably about a year now. It's great here!
1) For DeepSeek, I recommend their Reasonix harness strongly, due to its alignment to DeepSeek's prefix cache. It means mostly (95%+) cache hit input tokens, so very cheap large-scale code analyses and things that require mega context windows (at the cost of some attentional drift, yes). Reasonix does require that you send data to China. This is fine. I mostly use this for big, expansive ingestion of open-source codebases to figure out how something really works, usually something that documentation doesn't quite reach.
The economy of doing it this way versus American frontier model companies' token pricing cannot be overstated. I think I topped up $10 in June (2.5 months ago) and have still not burned through it, despite cycling untold tens of millions of tokens through it.
My biggest annoyance is that DeepSeek does seem to be considerably rate-limited of late, at least during working hours in Beijing, which is a range that I gather to be quite expansive there. I'm not blasting it with anything, I'm just noting that the agent takes 10-20 minutes to do stuff that takes much less time if I'm willing to pay the OpenRouter premium.
2) For most everyday stuff outside of where Reasonix + DeepSeek just makes overwhelming sense, I use OpenCode/Maki/Pi/whatever harness I feel like using today with Kimi K3, via OpenRouter. This does not require sending data to China.
I also use Kimi K3 in Zed via OpenRouter quite a bit, but sometimes like to mix it up with the other models.
3) Because I have the most experience with it, I can say with confidence that I would generally consider the SWE capabilities of Kimi to be on par with Claude, at least for the bottom 99% of purposes--and certainly, any routine business programming.
I think this has been true for a long time, well before K3. I've been using Kimi since K2.5.
4) For local hardware experiments on my MacBook Pro (M4 Max, 128 GB unified memory), Qwen3.6-35B-A3B (speed) and Qwen3.8-27B (intelligence, but slow). As has been widely noted, this amount of unified memory isn't as useful as it seems, due to memory bandwidth and decoding constraints, lack of tensor cores (on the M4 Max, anyway), etc.
A giant bag of memory isn't fast, but it'll let you load some impressively big models.
The future M5 Studio Macs will continue in this general vein, but will of course be somewhat faster, particularly due to the apparition of tensor cores in the M5 -- excuse me, "Neural Accelerators".
Still, if you really want to cook, get a real GPU. Real GPU running quantisations is still a lot better than a big slow bag of unified memory.
5) Overall, the Chinese models are simply excellent, and cater to lots of use-cases and tastes. However, I'll still tend to use Claude ($20/mo subscription) for general Q&A, whether of a technical nature or otherwise, particularly where web research and worldly knowledge is required.
[1] Disclaimer: this comment is an elaboration of https://news.ycombinator.com/item?id=49809605, which is not something I'd normally do. However, it seems a lot more relevant here than where I had originally posted it, in an article about gauging MiMo Pro v2.6 capabilities.
So "open models" buy you little and the independence from the oligopoly is not achieved.
Still, even if you need other people's industrial-class hardware, open models offer a lot more freedom and options. They allow you to use GPU capacity from entities who are not themselves building or training models and are ostensibly disinterested.
There's a big range of possibilities here in terms of data sovereignty and so forth.
1) You can use OpenRouter to route your open model requests to US-based inference providers with ZDR (zero data retention), as far as you can believe anything in this world. If you look at who actually serves open Chinese models on OpenRouter, you'll see a lot of folks like Digital Ocean, etc. I suppose I can't vouch for their purity, no, but I'd much rather send data there than send it to Dario.
2) Or, you can rent GPUs from companies like Runpod or Vast.ai and serve some very sizable models to yourself (e.g. using their pre-built vLLM images). You can't serve a model like Kimi K3 to yourself that way, at least not in any economically reasonable way. However, a private H100SXM or B200 can go a long way. You could serve the big Qwens, or DeepSeek-V4-Flash--you could do a lot if you're willing to spend on a rented GPU with sizable VRAM.
3) Finally, if you have and want to spend $750K-$1MM+ (I suspect I'm low-balling at this point), and if you can get them in the current demand climate, you can absolutely buy 16 x H200SXMs, with the appropriate boards to take them, pay for 20-25 kW of cooling, etc., and run one of these models yourself, on your kitchen floor if you like. You simply cannot do that with Anthropic or OpenAI.