The AI cold war: open weights aren’t a fallback, they’re the insurance
🔗 Learn more about me, my work and how to stay in touch: maeste.it: personal bio, projects and social links.
This is the week I finally tackle the Fable saga head on, the one I had deliberately left hanging in the past issues while waiting for the dust to settle. Now the dust has settled, and not in the way I’d hoped: the White House has cut off access to Fable and Mythos, and for us Europeans the message is a heavy one. In the deep dive I start from here to get to my thesis: in a world where American frontier models can be taken away from us by political decision, open weight models become far more than a technical alternative, they are the only form of independence we have left. DeepSeek, MiniMax, GLM and Kimi are in great shape, and I’ll tell you why GLM 5.2 has been my favorite for usage experience for months. In the links section you’ll find the themes that round out the picture: OpenAI preparing GPT-5.6 and cutting prices while Anthropic is in trouble, Midjourney surprisingly building a full-body ultrasonic scanner, Factory 2.0 and software factories, the autonomous research agent Sakana Marlin, the path to ASI according to DeepMind, and loop-driven development that puts down in writing the thread of last week’s deep dive. Enjoy the read.
My agenda
New episode of Risorse Artificiali: Paolo clones his own voice locally and for free, and from now on he no longer trusts a voice message. Listen.
In the same episode: the Fable withdrawal, Codex 5.5 vs Opus 4.8 and why we no longer look at the code we write with agents.
Our projects Lince.sh and AntiVocale (Google Play, GitHub), you know them well by now.
On my own:
I was in Catania as a speaker at Coderful, one of the best organized conferences with the best content I’ve seen recently. You can find my slides here; as soon as it’s available, you’ll find the video there too.
On June 24 I’ll be in Milan as a speaker at AIConf.
When the frontier is held hostage by politics
The Fable saga, which I’ve been following for a couple of weeks and which until now I hadn’t wanted to make the deep dive about because it felt too early to draw conclusions, has reached the point I feared. The White House has cut off access to Fable and Mythos, and the serious part isn’t so much the block itself, but the intention behind it: apparently the idea was to ban it only for non-American citizens. Then, struggling to tell them apart one by one, Anthropic ended up shutting off the taps for everyone. If that’s the right reading, the message is a heavy one, especially for us Europeans: the gap with the United States, instead of narrowing, risks becoming even more pronounced.
And so the question becomes a single one: if American frontier models can be taken away from us overnight by political decision, what do we build on? My answer, right now, is clear: open source models, or to be precise open weight. They seem to me the best alternative, and perhaps even the only one, given that on the European front Mistral is still well behind the state of the art. The beauty of open weights is exactly this: you hold the weights yourself, no directive can revoke them with the stroke of a pen.
The good news is that the alternative really does exist, and it’s in fine form. DeepSeek, MiniMax, GLM, Kimi are all excellent options, and over the past few weeks they’ve released new versions one after another. DeepSeek just closed a $7.4 billion round that crowns it China’s most valuable AI startup, and Kimi K2.7 Code pushes on agentic coding with a one-trillion-parameter MoE. They’re also great for local inference, which I talk about often, but they remain interesting as models regardless of where you run them.
And here a circle closes that I had left open just a couple of weeks ago. When I was talking about the local trend and hybrid architectures, I had tossed out a prediction: since we were starting to see governments banning the use of the most powerful models, we might soon need them for real. Well, here we are. Having models of this caliber running on our own machines is no longer just a matter of cost or privacy, it’s a form of technological independence.
Out of all of them, the one I want to spend a few words on is GLM. Version 5.2 has been called “Opus level”, SOTA in other words. Not the absolute best, that spot still belongs to GPT-5.5 for now, but by a truly slim margin, and in any case a model capable of handling long and complex tasks. For me it’s already the primary model for Hermes Agent and the fallback for coding, though I suspect it’ll soon become one of the primaries there too. I particularly like using it with a minimal, extensible harness like Pi.
I’ve been using it since last November, and I’ve always found it among the best open source models. Careful, I’m not talking about innovation: there DeepSeek remains by far the most interesting in terms of research, both in version 3 and version 4. I’m talking about usage experience, and on that front GLM has been at the top for me for months. I’m not the only one who thinks so: antirez is integrating it into DS4 as well. One last practical note, as an early adopter I have a link for 10% off the subscription, and it’s still valid if you want to try it.
Back to the starting point: as long as access to frontier models depends on a directive that can change from one day to the next, open weights aren’t a fallback, they’re our insurance policy. And luckily, today, it’s a policy that covers almost everything.
The links that caught my eye this week
OpenAI is preparing the GPT-5.6 models
The detail that jumps out, especially after the deep dive, is the timing: OpenAI is aggressively cutting prices to undercut Anthropic right while Fable is stuck in American regulatory trouble. These are still rumors, so I take the numbers and dates with a grain of salt, but the 1.5 million token window and the push on long-horizon coding tell you where the game is being played.
Midjourney builds a full-body ultrasonic scanner
This is the news item that has nothing to do with the others, and that’s exactly why I’m keeping it. That a company born to generate images would start building full-body ultrasonic scanners, complete with spas, is a leap I struggle to frame. The 60 seconds versus the hour and a half of an MRI I find hard to believe until I see real data, but if confirmed it’s physical AI to keep an eye on.
Factory 2.0: from coding agents to software factories
Here I find, almost word for word, the thesis of last week’s deep dive: the engineer who stops writing software and starts building the factories that build it. What strikes me most is the pillar of model independence and sovereign intelligence, which ties directly into today’s open weight discussion. The risk, as always, is that it stays more manifesto than product.
Sakana Marlin
An agent that works autonomously for up to eight hours and churns out hundred-page strategy reports is exactly the kind of long task that fascinates me. But I stay true to my fixation: agents are at their best where the result is verifiable, and strategic analysis is far more slippery terrain than code. The real question is how Marlin verifies its own conclusions.
Google DeepMind and the path to ASI
What I appreciate about this paper is its sobriety: no single magic moment in which AGI becomes superintelligence, but a series of progressive transformations, with bottlenecks and frictions put down in black and white. It’s a framing I share, and one that counterbalances the imminent-revolution tone. A good forty minutes of reading, but if the long term interests you, they’re well spent.
From test-driven to loop-driven development
If you read last week’s deep dive, here you’ll feel like you’re looking in the mirror: trigger, goal, harness, verifier and state around the agent’s loop are almost the same ingredients I used to try to define loop engineering. The central point is the one I always repeat, the more autonomy you give the loop, the stronger the checks have to become. It’s nice to see the concept consolidating in the community, and not just in my head.