Hello natural intelligences, here’s some artificial intelligence news. I disappeared because I was too busy with my followers. Anyway, in my opinion there are three things worth noting.
World models
The first is Meta’s V-JEPA model. LeCun, head of research at Facebook, former researcher and Turing Award winner, says this is the next step after the language model. These are world models: they’re no longer dedicated exclusively to language, but are trained to predict the next frame of video, starting from millions of hours of footage.
What are they for? They’re meant to give the model an implicit sense of how physics works, of vision, of a body in motion, of proprioception. The idea is that you then fine-tune these models with just a few hours of specialized video for a specific niche, for example the movements of a robotic arm at a company that makes certain products. With that small fine-tune on a few hours of filmed motion, the model should learn to perform new tasks.
It’s roughly the same mechanism that took us from the foundation language model, GPT-3, to the small fine-tune that turned it into ChatGPT and made it conversational. They want to do the same thing for robotics. I don’t know where this goes. LeCun has been hammering on it a lot. Meanwhile Zuckerberg is trying to buy up anyone and anything in AI, including Scale AI, a super startup specialized in data annotation.
Software 3.0
News number two: Karpathy put out a great talk on Software 3.0, and in my opinion it’s an absolute must-watch. Credit where it’s due: this is the former researcher who was a key figure at OpenAI and at Tesla, where he worked on Autopilot. He’s a pure researcher, very creative, visionary, far more so than a lot of entrepreneurs who declare themselves as such. Karpathy also coined the term “vibe coding,” on top of having launched “prompt engineering.”
At some point Karpathy came out with the idea that Software 2.0 was no longer software written by hand by people, but neural software, meaning models. Now, with Software 3.0, he means to announce the next phase: the agentic one, where the multimodal language model sits inside an architecture that also includes other pieces of software. This is a kind of operating system still waiting to be invented, but you can already see the first signs of it in the world of agents we’re all happily immersed in.
A2A goes to the Linux Foundation
I actually forgot the third piece of news, but here it is: it slipped by quietly, yet it’s absolutely worth noting. Google’s A2A protocol, meant to regulate and standardize agent-to-agent communication, was donated to the Linux Foundation. To me that’s a strong signal of good intentions from Google. The A2A protocol isn’t growing as fast as MCP. MCP is going strong, and you can find plenty about it on YouTube.
Bonus fourth item for those of you who made it to the fourth minute of the video: you’re a legend, and as a gift I’ll drop you this tip. On YouTube you can find a conference on MCP that was held on purpose a few weeks ago, with about ten videos. It’s the only one they’ve done in the United States. It’ll get you up to speed on what the MCP world has become, because its undergrowth is already enormous.
And with that I’ll wrap up. By the way, MCP inside the Cheshire Cat, 100% A2A, I’m thinking about it. Tell me what you think.
