Amodei vs open source, like Gates vs Linux. I think it’s a mistake
🔗 Learn more about me, my work and how to stay in touch: maeste.it: personal bio, projects and social links.
A week that opens with a sentence that made my hair stand on end: Dario Amodei declaring that open source models are dangerous and must be limited or blocked. Anyone who has followed me for a while knows how much open source matters to me, and in the deep dive I explain why I remain baffled, disappointed and worried: the move reminds me of Bill Gates, who in the late nineties used to say that Linux was a dangerous system. My thesis, as a European citizen, is that I see no alternative to open weight models, and that our contribution travels along two paths: local inference, the way antirez does it with DS4, and the optimization of harnesses and the surrounding software, which is where I am betting everything, with GLM, Ollama and OpenRouter, and which brings me back to my old fixation with multi-model setups. So much so that on the podcast I cut 90% of my Anthropic subscription. In the links section you’ll find the surrounding themes: Fable 5 returning after the export controls, the new Sonnet 5, Nano Banana 2 Lite and the Thinking Machines interaction models, Anthropic’s drug discovery and DeepSeek’s DSpark. Happy reading.
My agenda
On Saturday “US AI and open source policy makes my hair stand on end” came out: I cut 90% of my Anthropic subscription after the Fable move and Dario Amodei’s words. Listen
Our projects Lince.sh and AntiVocale (Google Play, GitHub), by now you know them well. Also take a look at Agent ready skills, which I talked about on June 24 at AIConf.
We’re thinking of doing short weekly lives on YouTube and X, around lunchtime, to show you practical things about our projects, personal agents and different uses of AI.
On my own:
Finally a quiet period for my public speaking. After all, summer has arrived, but we’re already working on September.
As soon as the videos from the conferences of the past few months come out, I’ll point you to them, because I’d love your feedback.
Open source as a danger? Amodei’s thesis, and why I don’t share it
Anyone who follows the podcast knows I already covered it there too, but I really can’t open this edition with anything other than Dario Amodei’s words, when he declared that open source models are dangerous, cannot be left unrestricted and must be limited or blocked. Anyone who knows me, in person or because they’ve followed me for a while, knows how important open source is to me. And although I can see the problems, and the dangers of handing such a powerful technology to everyone without regulation, I can only remain baffled, disappointed and worried by anyone who wants to put a limit on open source.
I said it on the podcast and I’ll repeat it here: this statement reminds me, in a striking way, of Bill Gates, who in the late nineties used to say that Linux was a danger to cybersecurity and security in general.
I even understand Amodei when he observes that open weight models are only partially an open source version, because they don’t tell us everything about the models. But it is surely better to know something and to be able to control something, rather than dealing with the black box of fully closed source models like Anthropic’s.
If we add to this the American policy, increasingly invasive and dominant in choosing which models can be “exported”, in their words, and therefore usable outside United States citizens, I think it is a duty, as a European citizen, to start thinking about how Europe and the rest of the world can protect themselves. From this point of view, I see no alternative to using open weight models. And I think it is important that European researchers and developers start asking themselves what contribution they can give to the open source world.
It is difficult, today, to contribute to generating new models: it takes capital and long-term planning, and Europe, unfortunately, is far behind on both, with the possible exception of Mistral, which in any case remains decisively further behind than the Chinese models.
The possible paths, then, even while using Chinese-style models, are two. The first is the one antirez, and others with him, are pursuing: optimizing local inference. His DS4 project is very promising and is becoming well known in the community of those who want to run local inference, but it requires hardware that not everyone has.
There is then a second possible contribution: optimizing the harnesses and all the surrounding software around these models, while using them through remote inference. Right now I’m relying on GLM and on Ollama in the cloud, as well as on OpenRouter. GLM is an extremely powerful model, and probably the one among open weight models that, on coding, comes closest to the performance of the main closed source models like Anthropic and OpenAI; as an early adopter I have a discounted invite to try it. Ollama and OpenRouter, on the other hand, let you use and try many different models.
Why is being able to use different models important? Anyone who read last week’s newsletter knows I talked about using multiple models to achieve similar or better results than larger monolithic models. It’s an idea that has fascinated me since the days of my wise-agents PoC, and that today I find again in projects like Sakana’s Fugu or Hermes’ Mixture of Agents (MoA).
In the coming weeks and months I’ll devote much of my experience to figuring out whether, and how much, mixing multiple models can be a viable path at the engineering level. From me, as usual, you’ll know everything: I’ll talk about it here and I’ll release as open source the software I produce.
Links that caught my attention this week
Redeploying Fable 5
It’s the most direct connection to today’s deep dive. Fable had been interdicted precisely by the export control policy I mentioned, and now it’s back, with an updated safety classifier and a framework shared with Amazon, Microsoft and Google to classify the severity of jailbreaks. Good that access has been restored. The lesson, though, remains: if a model can disappear by political decision, the open weight insurance policy is not an option, it’s a necessity.
Claude Sonnet 5
Sonnet 5 closes the gap with Opus 4.8 at much lower prices, and you can feel it. The Sonnet family opened the agentic era, and this version makes a clear leap on planning, tool use and coding. I use it every day, and as a model it’s excellent. I’m still left with the doubt from the deep dive: it remains a closed black box, and that limits how much I can build on top of it in a way that’s truly mine.
Nano Banana 2 Lite
Google is pushing the accelerator on low-cost media generation: Nano Banana 2 Lite generates images in a few seconds for a handful of cents, and Omni Flash brings conversational video editing at accessible prices. The theme for me is another: when components become this cheap and fast, it makes more and more sense to compose multi-model pipelines, the thread I talked about in the deep dive and that I’ll devote the coming weeks to.
Inside Thinking Machines’ Interaction Models
This one I find truly fascinating, and it connects to two themes of mine. The two-model scheme, a fast one for conversation and a slow one for reasoning, is the fast path and slow path pattern I was talking about. And then they move interactivity inside the model instead of pasting it on the outside with a harness of small components. It’s a direction that aligns with where I think the agentic world is heading.
Anthropic launches AI drug discovery program
Anthropic enters drug discovery, and does so focusing on neglected diseases that the market ignores. As a story, it’s a nice one. And it fits into a thread I believe in: scientific research is fertile ground for agents, because a molecule either works or it doesn’t, and it’s where results are verifiable that AI gives its best. The irony, after the deep dive, is that it’s precisely Anthropic, the black box par excellence.
DeepSeek open sources DSpark
I’ll close with a link that speaks straight to the deep dive. DSpark is DeepSeek’s open source speculative decoding, and it speeds up inference by up to 85% without changing the model’s output. It’s exactly the kind of contribution I was talking about: not generating new models, but optimizing the surrounding software, the harnesses, the inference. That it comes from DeepSeek, and in open source, reinforces the point: the open ecosystem is working precisely on the pieces that matter.