For the last three years, the story of AI has been told almost entirely in text. We measured progress in tokens, context windows, and benchmark scores on question-answering datasets. That framing was useful, and it produced genuinely capable systems. But it also quietly encoded an assumption worth questioning: that intelligence is fundamentally a language problem, and that everything else is a downstream application.

I think that assumption is about to break. The most important capability in AI over the next decade will not be language. It will be perception, that the ability of a machine to look at the physical world and understand what it is seeing. Computer vision is not one AI subfield among many. It is the layer that decides whether AI stays trapped inside screens or steps into the world.

Perception comes before reasoning

Human intelligence is embodied. We do not reason in a vacuum; we reason about a world we are continuously perceiving, and a large share of that perception is visual. Exactly how large is genuinely debated, and you will see figures ranging from around a third of the cortex to more than half, depending on whether you count only primary visual areas or every region that touches visual information. The precise number is contested, and anyone quoting a confident single percentage is overselling it. But the direction is not in doubt: vision is the dominant channel through which humans build a model of their surroundings, and the retina itself is, developmentally, an outgrowth of the brain rather than a peripheral sensor bolted on.

The relevant point for AI is ordering. Before a system can reason about a situation, plan a response, or take an action, it has to perceive the situation in the first place. A language-only model is extraordinarily good at manipulating descriptions of the world that humans have already written down. What it cannot do is look at a world nobody has described yet and form its own account of it. That gap is not a rough edge to be smoothed over with more parameters. It is a missing sense.

You cannot build a mind that understands the world if that mind has no way to see the world.

Vision is the foundational layer, not a feature

Look at the domains where AI is expected to produce the most change over the next decade including, autonomy, robotics, medical imaging, industrial inspection, defense, agriculture, logistics, and a pattern shows up immediately. Almost none of them are text problems. They are perception problems with a reasoning layer on top.

This is why I'd argue computer vision plays a role closer to infrastructure than to application. The useful analogy is TCP/IP. TCP/IP was never the thing users cared about; nobody opened their laptop to experience the Transmission Control Protocol. It was the substrate that made every application above it possible. Computer vision occupies a similar position in physical-world AI. The self-driving stack, the surgical robot, the crop-health drone, and the warehouse picker are all different applications sitting on top of the same underlying requirement: the machine has to see, and it has to be right about what it sees.

Get that layer wrong and nothing above it can be trusted, no matter how sophisticated the planning or the language interface looks in a demo.

Vision unlocks the world's largest dataset

Large language models were trained on the accumulated written output of human civilization such as books, articles, code, forums, conversations. That corpus is vast, and we have already scraped most of the easily reachable parts of it. But as a fraction of the information that actually exists in the world, text is tiny. The physical environment generates orders of magnitude more data every second, and almost all of it is visual and completely unlabeled.

Every factory floor, hospital ward, road network, farm field, and shipping lane is a continuous stream of visual signal that no language model has ever touched. Until AI systems can read that signal directly, they are working from a transcription of reality written by humans, not from reality itself.

This is why I think the framing of "we're running out of training data" is only true for text. We are nowhere close to running out of visual data, yet we have barely started. And vision is not just another input modality to bolt on. A system that learns from visual feedback, sees the consequences of its own actions, and generalizes across new physical environments develops a richer world model than any language-only system can. Vision is a training signal for general intelligence, not merely a sensor feed.

The real race is at the edge

For most of modern AI's history, capability lived in the cloud. That worked because the tasks were latency-tolerant, that a chatbot can afford a few hundred milliseconds of round trip. Perception in the physical world usually cannot. A drone avoiding a wire, a surgical robot mid-incision, or a car reading a pedestrian's trajectory cannot ship every frame to a datacenter and wait for a reply. The decision has to happen locally, in the millisecond between seeing and acting.

That constraint is now driving one of the most active areas in applied ML: getting capable vision models to run on-device. The current wave of vision-language-action (VLA) models is the clearest example. Systems like OpenVLA, Physical Intelligence's π0 and π0.5, NVIDIA's Isaac GR00T, and Google DeepMind's Gemini Robotics take visual input, reason about it, and emit motor actions — and a large body of 2025–2026 work is specifically about squeezing them onto resource-constrained hardware. The honest tradeoff is right there in the research: quantization can cut a model's size by 50–75%, but naive compression often degrades exactly the multimodal reasoning you deployed the model for in the first place. Closing that gap through distillation, sparse inference, hardware-aware architecture search, and KV-cache tricks, and that is where a lot of the genuinely hard engineering is happening.

Whoever masters edge vision holds a structural advantage across robotics, autonomy, and smart infrastructure, because they control the layer everyone else has to round-trip to the cloud to reach.

Vision is what gives AI agency

Here is the argument I actually care about most. A system that can only process text is, in a real sense, reading about the world. A system that can see is participating in it. The difference between reading a manual about a room and standing in the room is the difference between passive knowledge and the ability to act.

Agentic AI, the systems that set goals, take multi-step actions, and operate in real environments and needs perception to mean anything outside a browser tab. Without visual grounding, an agent's entire scope of action is digital. With it, the boundary between the model and the physical world starts to dissolve. The robot navigating a hospital corridor, the drone inspecting a wind turbine, the warehouse system fulfilling orders unattended, and every one of them, before it does anything else, sees. The VLA wave is the first time we've had general-purpose systems where vision, reasoning, and action are trained together end to end rather than stapled together after the fact, and that is why it matters more than the benchmark scores suggest.

The counterargument, taken seriously

I don't want to pretend this thesis is unopposed, because the strongest version of the other side is worth stating.

The counterargument runs like this: vision isn't privileged, it's just one modality, and the real engine of recent progress has been scale and the transformer architecture, both of which are modality-agnostic. On this view, "vision-first" is a category error and what matters is a general learning substrate that happens to ingest pixels the same way it ingests tokens. There's also a fair point that a lot of embodied-AI demos are self-reported, run under conditions the authors chose, and not independently reproduced; π0.5's home-robot results, impressive as they are, fall into that category. And you can reasonably argue that world models can be learned from other signals entirely on simulation, proprioception, even language descriptions of physics and without vision being the load-bearing element.

I take those seriously. My honest position is narrower than "vision is magic": it's that vision is the highest-bandwidth channel we have into the unlabeled physical world, that most economically valuable physical-world tasks are perception-bottlenecked rather than reasoning-bottlenecked, and that the systems actually crossing from screen to world right now are overwhelmingly vision-grounded. That could change. If a purely language-and-simulation approach produces robust real-world agency first, I'll have been wrong about the mechanism. I don't think I'll have been wrong about the destination.

Where this leaves us

The question of what AI will ultimately be able to do in the world is, to a remarkable degree, a question of how well AI can see. That's not a philosophical flourish, and it's an engineering roadmap. The bottleneck between today's impressive-but-disembodied models and genuinely capable physical-world agents is perception: fast, reliable, on-device visual understanding.

Language gave AI something to say. Vision is what will give it something to do.

Notes on sources

• Estimates of the share of cortex involved in visual processing vary widely across the literature (roughly 15% for narrow definitions up to ~50% for broad ones); the figure is genuinely contested and is presented here as a range rather than a settled number. The two ends of the range are linked inline: ~a third (Gilbert & Walsh, Current Biology, 2004) and more than half (Perkins School for the Blind, summarizing University of Rochester work). On the retina as brain tissue, see Frontiers in Neuroscience, 2020.
• The "running out of text data" claim is drawn from Epoch AI's analysis of the stock of human-generated public text (Villalobos et al.).
• Current vision-language-action systems referenced — OpenVLA, Physical Intelligence π0/π0.5, NVIDIA Isaac GR00T N1, and Google DeepMind Gemini Robotics — and the edge-deployment tradeoffs (quantization reducing model size ~50–75% at some cost to multimodal reasoning; see also this 2026 survey of embodied-model edge deployment) reflect 2025–2026 published work. Several headline embodied-AI results are self-reported and not yet independently reproduced.