Denny's has entered the AI safety debate.
Yes. The American diner chain.
It posted a Venn diagram showing that Denny's and NVIDIA both understand "the importance of staying open".
Very good. No notes.
It is also the funniest bit of a much bigger fight. Jensen Huang joined X and used his first post to share NVIDIA's letter backing open-weight AI. Google, Meta, Microsoft, OpenAI, Hugging Face, GitHub and a load of other companies signed it.
Anthropic did not.
What actually happened
On 24 July Jensen Huang shared a four-page letter called Open Weights and American AI Leadership.
Have a read.
The letter argues that America's AI lead cannot sit inside a few closed models. It says businesses, universities and governments need models they can download, adapt and run on their own infrastructure. That means more competition, lower costs and less dependence on one AI provider.
It also takes aim at attempts to restrict open models. The signatories want policymakers to separate illegal extraction from normal distillation and avoid broad rules that could kneecap the whole open-weight market. IMHO: fair.
And then came the list of names. Here it is as of 28th July:
NVIDIA. Google. Meta. Microsoft. OpenAI. AMD. Hugging Face. GitHub. Mistral. Cloudflare. Replit. Basically the entire AI infrastructure world.
Except Anthropic.
That was not a random omission. Anthropic has been making the opposite argument. In February it said DeepSeek, Moonshot and MiniMax used roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude. It says those labs were extracting Claude's best capabilities and that the danger multiplies when the resulting models are released openly.
In their defence - they are probably correct.
So this is the actual disagreement:
NVIDIA's camp says open weights spread access, competition and control. Anthropic says powerful weights can spread stolen or dangerous capabilities beyond anyone's control.
Then the internet got involved. And piled on Anthropic. As we are wont to do!
Anthropic researcher Julian Schrittwieser compared the demand for open weights with asking NVIDIA to open-source CUDA and Microsoft to open-source Windows.
Andrew Ng called that a false equivalence. Correctly, I reckon. Imagine pissing off Andrew Ng. He’s such a sweetie but even he came out to call out the BS.
To be VERY clear: nobody is demanding that Anthropic release Claude's weights. It is perfectly entitled to keep Claude closed. The complaint is that Anthropic appears to want governments to make it harder for other companies to release open models.
Open weights is not open source
Open weights and open source are not the same thing
Quick definition because this gets mangled constantly.
An AI model is basically a giant set of learned numbers called weights. If a company releases those weights, you can usually download the model, run it on infrastructure you control and adapt it under the licence.
That does NOT necessarily give you the training data, training code or the full recipe used to build it. So "open weights" and "open source" are not the same thing. Annoyingly even the Anthropic researcher Julian above got it wrong - Jensen was calling for open-weights. Not open source. Shit, the open letter is titled Open Weights and American AI Leadership.
We’re talking open-weights not open-source.
And also open-weights also does not mean “free”, small or laptop-friendly. The giant Chinese models may need a server rack. Smaller Qwen, Gemma and Mistral models can run locally on normal hardware.
Kimi K3 weights were released yesterday. I did a back of the envelope calculation and you’ll need around $650,000-750,000 worth of computer to run it. That is not “free”.
I’ve talked about this at length here: Local Models 101.
Why should you care?
With a closed model, the provider can change the price, limits, behaviour or access. We have watched this happen repeatedly with Claude Fable over the last few weeks. It’s exhausting.
Hell, Sam Altman is this moment in Washington showing the next set of models (ie. GPT-6) to the White House and Trump administration for a go ahead. We’re at the whims of many forces.
With open weights, a business can keep the model on its own infrastructure, fine-tune it for a specific job, protect sensitive data and move without rebuilding everything around another company's API.
I went through the practical differences in my open-source AI price war guide. The short version is that open models put a ceiling on price and give customers an exit.
Anthropic does have a point
Now whilst Anthropic are pissing people off right now…they still have a point. And they are standing up for their (oft-repeated) principals. They can’t be faulted there.
Once model weights are released, they cannot be recalled.
You cannot patch every copy, turn off a dangerous capability or know who modified it. If a future model becomes genuinely useful for cyberattacks or biological weapons, chucking the weights online would be a fairly permanent decision.
Anthropic also says DeepSeek, Moonshot and MiniMax used around 24,000 fraudulent accounts to generate more than 16 million Claude exchanges for distillation. In plain English: use Claude's answers to help train competing models.
Anthropic's fear is that stolen capabilities then get released as open weights and spread beyond anyone's control. Which…is precisely what is happening. That is not imaginary - that’s legit. It is a proper safety and national-security argument.
BUT…
It still does not follow that open weights in general should be squashed. A small model running private document search on your laptop is not the same risk as tomorrow's frontier model with dangerous capabilities.
Treating them as the same thing is a massive overreaction. There are (as in all things) degrees.
Follow the money
Equally we need to look at the signatories here. They have vested interests in keeping open-weights in play.
Yes yes I’m sure this also aligns with their ethics and philosophy. But those stances are a lot easier to hold when you are also making billions of dollars.
NVIDIA benefits when open models exist because somebody needs to buy the chips that run them. Cloud companies benefit because most businesses do not fancy building a data centre in the garage.
Open-weight models mean more models being deployed. More sales. More compute. More volume. All great for the signatories.
Funny that.
This does not make NVIDIA's case wrong or Anthropic's safety case fake. Both sides have real beliefs AND very convenient business incentives. Calling either side disingenuous or spinning complex conspiracies sounds smart but often isn’t.
So. What do you DO? What should you use? The useful question is what gives you options.
Use Claude or ChatGPT for the difficult work where they are plainly better. That’s not going to change anytime soon. Open-weight (mainly Chinese) models are still behind - no matter what some influencer is screaming at you. RIP ChatGPT.
Test smaller open models for private, repetitive and high-volume jobs. Keep your prompts, source files and business context in one shared AI vault so changing providers is not a total faff.
Also keep watching the Chinese labs. Their cheap open-weight models are why this argument became urgent in the first place. My Chinese AI 101 issue maps the main players.
Use the best model. Keep your stuff portable.
Do not let Dario, Jensen or Denny's decide your whole stack for you.
OK maybe Denny’s.
To the Task,
Kyle