Part 1. Holistic AI models
There's been a clear shift from traditional language models (text-only) towards AIs that natively process audio, vision, and language together. The goal is a unified representation space where these signals coexist and reinforce each other.
Text is a versatile medium (you can put an image in 1000 words after all) but most real-world use cases cannot be mapped into text alone. They require correlating multiple types of data. A reliable understanding of a given context is still a challenging task. Even when it works, its behavior is not yet robust and there can be lots of silent errors.
Recent progress in Generative AI has been remarkable. 2025 was the year when models started to generate coherent video and audio from text prompts in a single pass (instead of generating them separately and having to synchronize them afterward).
It's clear that leading labs now have large-scale, mixed-modality datasets.
Yet, generative media on its own is unlikely to be the final business model. It’s eye candy, but people are reluctant when they see it in their feeds and even more so when it comes to paying for it.
So why the massive push in this direction?
Because Gen AI is only one side of the coin. The real push is for its symmetric counterpart: holistic models that can map video and audio back into structured meaning. It's easy to see that if you have the data fro training one, you automaticlaly have the data for training the other. And this general understanding is what enables AI systems to navigate environments, interpret scenes, and eventually act in the world.
Gen AI in itself may not be a sustainable business model but there's huge value in the gathered feedback, interaction data, and edge cases that can be used to refine the datasets used for subsequent models.
My bet for 2026 is that we'll see AI systems that understand physical and digital environments with a level of generality comparable to how language models understood text just a few years ago.
Part 2. We've run out of text. What's next?
In training our Language Models we have exhausted all the texts ever written by humanity. That’s no exaggeration, if we consider quality text data (books, published articles, papers, curated blogs). It’s an unspoken secret, since even the open-source companies rarely go into details about the data they used. And it’s a conclusion that would clearly curb the enthusiasm, since even with all the text, AI has not yet reached the level of intelligence promised to investors. We already saw the shift towards application layers, confirming that the exponential curve of larger models accuracy plateaued logistically.
There’s two directions from here. On one side, we can stick with text and double-down on human feedback and reinforcement learning. This will increase performance on specific tasks, such as coding, but will hit a wall in terms of general understanding. Moreover, we saw that too much reinforcement learning, especially on feedback from chats from general consumers diverges into syncopathy (AI models optimizing for what we want to hear rather for what is factually correct)
So we are left with abandoning text and looking at all the other types of data our world is so rich in. A simple walk in the park with a camera would gather information about how our laws of physics work, how society function, how people move and act and the wide spectrum of cause and effect. After all, this is how humans learn about the world, before understanding or articulating language.
The new data: Next-action prediction
Datasets up to 2025 were structured for textual next-token prediction. Datasets of 2026 and onward will be for next-action prediction. In-context data will become a tradeable commodity. There won’t be the advantage of written data, publicly available or easily scrapable. Contextual data which contains information about how to interact with the world is not something we had our entire history to gather and create.
However, we already see the first attempts at a solution. Humanoid robots, controlled by humans in an initial phase, gathering data for training the AI of the second phase. Human operators showing the AI what to do and how. Smart glasses that see and hear the world from the same perspective as the person wearing them. Just like a decade ago captcha systems were used to create datasets of labeled images (“select the image containing cars”), we are in a process of gathering the raw material for the intelligence of tomorrow.