The last interface.
Technology should adapt to us, not the other way around.
The translation tax
The limits of my language mean the limits of my world.
Ludwig Wittgenstein
We have made a strange bargain with our machines. They get smarter every year, yet using them still means learning what they can and can't do, translating your goal into something they understand, and babysitting them through every step. You start with an intent; what comes back is an approximation of it, filtered through how fluent you happen to be with the tool.
AI promised to remove that filter. A new generation of tools is being built precisely so you never have to learn them: no menus, no shortcuts, no years of practice, you just ask. But asking still happens in words. The result no longer depends on how well you use the tool; it depends on how well you can describe what you want.
The filter didn't go away, it just has a new name: language. You see an idea all at once in your mind, but handing it to a machine means unpacking it into words, one detail at a time, and hoping it gets packed back the way you meant it.
Take the figure above. Two people have the same idea at the same moment: a rocket with a red nose and three red fins, a round window, a stripe across the middle, flames underneath, and a ringed planet hanging off to the left.
The one on the left does what all of us do today. He opens a chat and describes the idea. What comes back is close, but not quite it, so he adds a detail, then another, fixing one thing at a time until the picture on the screen gets near the one in his head. Near, but not the same: what he ends up with is the model's interpretation of his idea, limited to what he managed to put into words.
The problem is not only how long this takes. Some of what we picture has no good word at all: the exact shade of red, the angle of a fin, the way the flames curl. Other details we don't even realize we care about until the result gets them wrong. They were part of the idea from the start; we just never put them into words, because we never had to. Language can only carry what we notice and know how to name.
The one on the right wears The Nutshell. He doesn't describe the rocket, he just pictures it. Everything in that thought, including the parts he would never have thought to mention, comes along with it.
Put numbers on the difference between them and it becomes hard to ignore. Across seventeen languages, speech carries information at roughly 39 bits per second1, and typing is slower. A single photo on your phone holds millions of bits. So Wittgenstein was only half right: language does not limit the world in your head, it limits how much of it you can hand to anyone else. Switch between text and thought prompting below and watch the idea squeeze through.
Generative AI made making almost free. It did not make explaining any cheaper. Drawing the rocket now takes a few seconds; describing it still takes a few minutes and a fair amount of patience.
The bottleneck didn't disappear. It moved into the prompt.
The right input for the job
Language is a wonderful tool, but it is only one of the ways we express ourselves. Thought weaves together intuition, reasoning, emotion, memory, the full range of the senses, and much more that we either cannot put into words or don't even understand yet.
What if we removed the translation and mapped thought to intent directly?
Everything you already do on your phone or computer would still work, from replying to an email to designing a logo, but through an input that fits the task, so it would be faster, more accurate and closer to what you meant.
Furthermore, things you could only imagine doing would become possible.
Think of the things you do on a screen on an ordinary day, and ask a simple question: is language the best way to express each of them? Sometimes, yes. Often, not at all.
- An email. Here words are the work itself, but the right ones don't always come: you know the tone you want, firm but kind, and still stare at the cursor. An interface that reads the intent behind that pause can offer the phrasing you were reaching for.
- Searching. You know exactly what the jacket you want looks like, but not what shops call it, so you try word after word and scroll through results that are almost it. Picture the jacket, and the search starts from the jacket itself.
- A website or app. You see the whole screen at once: where things sit, how much room they have, what happens when you tap. Words have to lay it out one element at a time; a picture arrives whole.
- A presentation. You can see how a slide should look long before you could say whether it is the spacing, the font or the order that is off. Picture the slide as you want it, instead of turning it into a list of fixes.
- A logo. A logo starts as a feeling and a shape that is almost there. "Modern but warm" means something different to everyone; the image in your head means exactly one thing.
- A memory. That afternoon at the beach, just before sunset, your friend squinting into the light. You would know the photo in a second if you saw it, yet you can't name a single thing to search for; remembering it is the search.
- A song. Humming one tune is easy; composing is not. In your head a bass line, a melody and a beat play at once, layered in ways your voice and hands struggle to perform together, let alone describe. Thinking it hands over the whole arrangement, not one line at a time.
- A quick note. Ideas arrive at the worst moments: mid-conversation, on a walk, with your hands full. Saving one should take as long as having it, not as long as unlocking a phone and typing before it slips away.
None of this makes words the enemy. Language is one part of what a thought is made of, alongside images, sounds, shapes, feelings and memories, and in every example above the words were there too, just never on their own. Yet with AI, we have quietly accepted that only that one part gets through. Take the whole thought instead, words included, and what you can express is no longer limited to what you can describe.
One mind, many agents
In the last two years, AI went from answering to acting. It no longer just replies to you; it does the work: agents that plan, write, design and ship on your behalf. That changes your job. You are no longer the operator of every tool. You are the orchestrator of many.
A conductor doesn't play every instrument. They set the direction, keep everyone in time, and can stop the music with a single gesture. Less operation should not mean less agency. It should mean more.
There is a catch, though. An agent that can execute anything still has to understand what you actually mean. Give it a vague instruction and it will confidently do the wrong thing, very fast. Autonomy without a true understanding of intent doesn't remove the friction, it just moves it into the instructions.
The fix is not a smarter agent guessing harder, but a better way for your intent to reach it. When a task can leave your head as fast as you think it, and arrive whole instead of in pieces, you stop being the queue and start being the conductor.
So which input comes closest to a thought? Only the thought itself, which already holds every way we express ourselves at once. That is why we are moving beyond text-prompted models toward thought-prompted ones: a Brain-AI Interface. Not a brain-computer interface that decodes which button you meant to press, but an extension of yourself, the way your arm already is. You don't calculate which muscles to activate when you reach for a cup. You just reach. You form the intention; the interface handles the execution.
How it works
A city of senses
Here is the good news: the brain is not one big blob doing everything at once. It is more like a city, with neighbourhoods that specialise in different things. Some are more involved in touch, others in taste, sound, smell or sight. When you do something, activity shifts across these interconnected neighbourhoods. The busier they get, the more oxygen-rich blood they need. Pick a sense and watch the lights come on.
Now the interesting part. These neighbourhoods don't only work when something comes in through your senses. They can also become active when you replay an experience in your mind: imagining a song engages auditory regions, imagining a movement engages motor regions, and picturing a place engages visual regions. Perception and imagination share much of the same machinery2.
Measuring the brain with light
How do you measure a brain through a skull without surgery? You could use electrical currents, magnetic fields or ultrasound. We choose light. Tiny light sources in The Nutshell shine near-infrared light into the head. At those wavelengths skin and bone are surprisingly transparent (press a torch against your fingers in the dark and you will see the red glow come through). The light travels through the scalp and the first few layers of the brain, scattering along the way. Some of it comes back out where a photodetector measures it3.
Why would light tell us anything about thinking? When an area of the brain gets busy, its neurons start firing, and firing takes energy that is paid for with oxygen. The body answers with a hemodynamic response: over the next few seconds, nearby blood vessels widen and flood the area with oxygen-rich blood, more than the neurons actually used. So the busier an area is, the more oxygenated the blood around it becomes4.
On the way, the light passes through blood. Oxygen-rich and oxygen-poor blood absorb near-infrared light differently, so the returning light lets us measure the brain's oxygen levels, and therefore the brain activity.
Put these pieces together and you can run a simple experiment. Sit someone in front of a screen that stays blank, then shows a picture, then goes blank again, and measure their brain activity. Each time a picture appears, some areas of the brain activate. When the screen goes blank, they settle again. Watch them light up below with every image and dim between them.
Do this between many light-sensor pairs across the back of the head, many times a second, and you get a moving map of what your brain is doing.
From signal to picture
What can we do with a map like that? One thing is to run the experiment in reverse: start from the brain activity and generate the picture behind it. To get there, we train an AI model on pairs: a picture someone looked at, and the brain activity it elicited. Every picture leaves its own trace, and pair after pair, the model learns what each pattern means.
We map the human concepts (the neural latent space) to the model's concepts (its digital latent space). This aligns the model to the user. Every new pair brings what the model understands closer to what you intended; matching it exactly would take infinite data, so it is a limit the model approaches rather than reaches.
Once trained, the model needs only the brain activity. Picture something, and it generates an image from what it measured. Science fiction? Not anymore: research using large brain scanners has already generated images from blood oxygenation5. What remains is taking it out of the lab and putting it onto something you can simply wear. Below is an example of real brain decoding in action: on the left what the person saw, on the right what the model generated from their brain activity alone.
Pictures are only an example. Generating an image from brain activity is one specific application, picked because it is simple to understand: you can see what went in and what came out, side by side. The same approach is not limited to pictures, and other modalities are supported too, such as audio and video6.
Your thoughts stay yours
A device that measures the brain should raise one question before any other: who else gets to see what it measures? Our answer is no one, unless you decide otherwise, and The Nutshell is built around that answer.
- It measures blood flow, not your inner voice. It tracks slow changes in oxygen in the patch of cortex under the band; it does not hear your thoughts as words.
- It's only on when you want it to be. You choose when to wear it and when to start measuring.
- Only your model can decode it. Your brain activity is encrypted by default, and since every brain is different, only your model can decode it.
Your mind is the most private place you have, and it should stay that way.
Conclusion
For as long as we had ideas, we needed to translate them before anyone else could see them. The translation became so normal that we stopped noticing it was there.
That made sense when the only listener was another person. It makes much less sense now. With agents doing more of the work, the way we express intent matters more than ever, and language is not always the most efficient way to do it. The best interface is the one that fits the task. Everything we do as humans comes from a thought, so why not use it directly.
Getting there takes more than a device. Today's foundation models learned from an internet of text, images, audio and video. There is no internet of thoughts. It can't be scraped; it has to be created: deliberately, consensually, and at scale. Building the first reliable way to do that is the work ahead of us.
When intent can reach AI directly, execution stops limiting creation. What you can make will depend less on what you know how to prompt, code or operate, and more on what you can imagine.
That kind of access should belong to everyone.
Be the first to try it.
References
- Coupé, C., Oh, Y. M., Dediu, D. & Pellegrino, F. (2019). Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche. Science Advances, 5(9).
- Kosslyn, S. M., Ganis, G. & Thompson, W. L. (2001). Neural foundations of imagery. Nature Reviews Neuroscience, 2(9), 635–642.
- Jöbsis, F. F. (1977). Noninvasive, infrared monitoring of cerebral and myocardial oxygen sufficiency and circulatory parameters. Science, 198(4323), 1264–1267. The method is known today as functional near-infrared spectroscopy (fNIRS).
- Attwell, D. & Iadecola, C. (2002). The neural basis of functional brain imaging signals. Trends in Neurosciences, 25(12), 621–625.
- Scotti, P. et al. (2023). Reconstructing the Mind's Eye: fMRI-to-Image with Contrastive Learning and Diffusion Priors. NeurIPS 2023.
- Audio: Denk, T. I. et al. (2026). Brain2Music: Reconstructing Music from Human Brain Activity. Nature Communications, 17, 91. Video: Chen, Z., Qing, J. & Zhou, J. H. (2023). Cinematic Mindscapes: High-quality Video Reconstruction from Brain Activity. NeurIPS 2023.