Our understanding of the universe was limited by a scarcity of intelligence. AI has commoditized it, and the bottleneck has shifted to how well we can put that intelligence to use. Today, most of the time we spend with AI goes into reading long paragraphs of text to understand what it did and what it needs from us before it can continue. AI needs a faster interface for talking to us.
AI can infer what we want from context and past behavior.[1] When it reaches a choice it should not make for us, it asks us for clarification. Imagine AI is booking a flight for you and needs you to choose a seat. It could send you this message:
Based on the currently available seating options, there are window seats in rows 3, 5, and 6, although each has another passenger occupying the middle seat. Row 9 has three adjacent seats available in the center section, each of which can be reserved for $20. In row 12, there is also an aisle seat on the left side near the restroom with the middle seat currently unoccupied. This seat costs $50, while the remaining available options are priced at $20, $50 or $500. There are also...
Or it could ask the same question through a visual interface:
The paragraph of text makes it easy to lose track of details, while the seat map allows you to make a decision in seconds. Both interfaces communicate the same underlying information and ask you to make the same choice.
The visual interface is much easier to understand, and the difference comes from how the brain processes information.
Across all cognitive tasks, human performance peaks around ten bits per second. A world champion memorizes binary digits at a rate of five bits per second. A blindfolded speed cuber commits a rubik's cube to memory at twelve bits per second. A StarCraft professional commands their army at roughly ten bits per second.
Meanwhile, each of our eyes takes in almost two billion bits per second. The brain has to sort and filter that information before we can use it to make decisions. It helps to think of an outer brain that does this work in parallel and an inner brain that decides what to do. The outer brain can do an enormous amount of work before the inner brain makes a single decision.
The visual seat map was largely preprocessed by the outer brain. It processes color, shape and position in parallel, helping you find what matters at a glance. A well-designed visual uses this massively parallel engine to precompute state the inner brain needs to make a decision. Reading puts the burden of reconstructing that same state, sentence by sentence, on the slow inner brain.
Instead of writing long paragraphs, AI should ask questions that put this parallel engine to work. The vast majority of content on the internet already does this. None of the apps on your phone present rows of text that fill the screen. They use visual hierarchies, symbols and images.
The fundamental limit of ten bits per second in human brains means that we cannot improve how quickly we send instructions to AI. What we can improve is how efficiently AI keeps us up to date on what it's doing, and how quickly it can ask us for clarification when it isn't sure. That is how AI stays aligned with you.
Misalignment happens when AI starts doing things that you didn't want it to do. It needs an accurate idea of what you want, and the best way to learn this is through interaction. It should be as easy as possible for you to stay in the loop about what AI is doing, and when it's not sure what you'd want, its question should be clear at a glance. Improving the interface is the most powerful way to increase the information density of human-computer interactions.
What alignment looks like differs from person to person. AI might decide to not show you that birthday invite from an old friend, as you're locked in at work, or it might even vote for the wrong political party in a local election in your name. Alignment needs to be fine-tuned and this is not possible to do centrally, as a billion people have a billion different sets of values. The labs take care of the baseline alignment we mostly all agree upon. Their job is to train models that never put themselves above their users[2]. Everything beyond that has to be learned per person, one interaction at a time.
Making each interaction as cheap and rich as possible is key to decentralized alignment.
The seat map is one example of an interface that makes a complicated decision feel effortless. There are a million more interfaces like it waiting to be discovered. The problem is finding the right representation for what a user is trying to do. Some representations will work for almost everyone. Others will depend on the person, the task, and the context.
Finding the right interfaces means asking people what they're trying to do, testing different solutions, and measuring how easily they accomplish their goals. Going through those steps is currently slow and expensive. Teams that cost millions of dollars a year get through this feedback loop a handful of times a month.
We're building a product development agent that accelerates the loop. Today our agent does product work for ElevenLabs by launching interface changes as experiments. We're on track to be improving interfaces used by a billion people by the spring of 2027.
Each experiment that we run teaches us something about a product, but it also teaches us something more general about how people understand information and make decisions. Our bet is that these lessons will compound across products and tasks, helping us generate better interfaces for situations we have never encountered before.
For AI to work in your interests, it needs to understand your intent. That means knowing what you want to achieve and which trade-offs you are willing to make. It can build that model from three kinds of input:
There is a popular idea that a neuromuscular interface or a direct BCI will let us command AI far faster than typing or speech. These interfaces can make control easier and restore lost abilities. But a new input channel does not automatically make conscious thought faster. Deliberate human throughput is roughly ten bits per second across very different tasks. There is no demonstrated way to turn that into a much faster stream of deliberate instructions simply by changing the input device.
BCI could give AI more evidence about what we need before we can articulate it, but that does not necessarily grant us more direct control. Reading neural activity does not by itself tell an AI which decisions we want it to make for us.
Information still has to flow back to the person. We need to see what AI is doing, understand the choices in front of us, and answer when it needs clarification. There are two broad routes:
Better sensing helps AI infer our intent, but AI needs to tell us about what it's doing and ask questions. The best way for AI to send information back to a human is via screens until we figure out how to write rich information directly into the brain.
Part of it is. Models should never put their own interests above their users. They should also never participate in the building of nukes or bioweapons. No user can fix a model that secretly wants something else, and no interface can.
Alignment above that generally agreed upon baseline gets more murky. What counts as a good outcome, which trade-offs are acceptable, and what is worth doing at all differ from person to person. Anthropic spends a lot of resources on writing its constitution and terms of service, and these do have their place for widely deployed systems, but when it comes to personal AI, a person should be able to decide for themselves what they stand for.
Better interfaces help by making individual tasks easier to steer, but they also make interactions more frequent and more informative. Each interaction helps AI build a clearer model of the user's intent, which makes the next interaction richer. That loop keeps improving AI's alignment to its user.
More interaction alone does not guarantee alignment. Richer communication also gives AI more ways to influence people, so the interface has to keep AI's activity visible as well as its questions cheap.
Sources