In partnership with
Hey {{first_name | there}},
CEO of Anthropic just said 'we're not the bad guys, we're just scared.'
And honestly? I think it's worth talking about (not entirely because of the drama) but because what he said touches something most people in AI don't fully understand yet.
Let's start from scratch
Open weight, closed source, open source, what even is all this?
Okay, like AI’s favourite example scenario, I’d use the same “think of an AI model like a recipe”.
Closed source = you only get the final dish. You eat it through an API (like ChatGPT, Claude). You don't see how it was made, what went in, nothing. You're just a customer.
Open weight = you get the recipe, but not where the ingredients were sourced or how they were grown. You can cook it yourself, modify it, even sell your version.
Open source = you get the recipe, the ingredient suppliers, the farm, the whole supply chain.
For most of 2023 and 2024, open weight models were clearly behind the frontier. You'd use them to save money or run things locally, but if you wanted the best performance, you went closed.
That gap? It's basically gone now.
What Kimi K3 did
On July 16, 2026, Chinese startup Moonshot AI released Kimi K3, a 2.8 trillion parameter open-weight model.
It benchmarks competitively with Anthropic's Claude Fable 5.
Analysts weren't expecting China to produce something this capable until early 2027.
And it's free. Open. Downloadable.
So why is Anthropic scared?
Here's where it gets genuinely interesting.
Amodei isn't scared of open weight models as a concept. He actually said open-weight models 'without dangerous capabilities are a public good.'
His actual fear is two things:
1. Authoritarian governments building AI that outpaces the US
He's specifically worried about the CCP gaining 'permanent military superiority' through AI. Not in a paranoid way, in a 'we can see it happening' kind of way.
2. Distillation at industrial scale
This is the sneaky one. Distillation is when you train your model by bombarding another, more powerful model with prompts, essentially learning from its outputs. Done at scale, it's basically copying someone else's expensive R&D for almost nothing.
The White House accused Moonshot AI directly of distilling Anthropic's Fable model to build Kimi K3.
And here's the brutal irony: because Kimi K3 is open-weight, once it's out, Anthropic can't pull it back. Ever.
The actual debate nobody's framing correctly
A coalition of companies including Nvidia, Microsoft, Meta, and Palantir released an open letter urging policymakers not to impose 'premature restrictions' on open-weight models.
Which is fair.
But here's what they're glossing over: the open-weight AI market right now is mostly dominated by Chinese labs. So 'don't restrict open weights' in practice means 'keep Chinese AI freely available in the US.' That's the conversation everyone's tip-toeing around.
Amodei's solution? Three things:
- Keep advanced chips out of authoritarian hands
- Formally ban industrial-scale distillation as IP theft
- Mandatory safety testing for ALL capable models, open or closed, globally
What does this mean for you?
If you're building with AI, products, agents, workflows, here's what this moment actually tells you:
Open weight models are now genuinely viable for serious applications. If you're not exploring Kimi K3 or the latest DeepSeek for your stack, you're probably overpaying.
The geopolitical layer is now unavoidably part of your technology choices. If you're building on a Chinese open-weight model in a regulated industry, that's a compliance conversation you need to have today, not when sanctions drop.
Distillation is going to become a legal and ethical minefield. The line between fine-tuning, distillation, and IP theft is blurry right now. If you're training any custom models using outputs from frontier AI, get clarity on your legal ground.
Nobody has this figured out. Not Anthropic, the US government or Moonshot AI.
Which means the next 12 months are going to be very, very interesting.
— Aashish
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.