Derek Law

Today we're announcing Tap8: interactive video as the interface between humans and AI. The model answers with a small runnable world instead of prose. Every element is exact, and every element is a handle: pull on one and the explanation reshapes around it. Facts render as live code, not hallucinated pixels: tokens for what is true, pixel diffusion for what it looks like. Research bets: visual quality trained by recursive self-improvement, and real-time generation on a Diffusion-LM. Preview coming soon. tap8.ai/research/tap8-… The chatbox had a good run.

I spent two days and a thousand-word essay arguing in text that text is the low-bandwidth bottleneck. A minute of video makes the case. The medium is the argument.

An agent writes visual code, a VLM critiques the rendered output, the agent refines based on feedback, best traces train the next model, and that model powers both agents in the next round. Vision-Guided Iterative Refinement for Frontend Code Generation, ICLR 2026 Workshop on Recursive Self-Improvement. arxiv.org/abs/2604.05839

At the ICLR workshop on recursive self-improvement this year, we showed a complete RSI loop closed in visual coding. Visual code is an unusual RSI domain. There is no verifiable reward. Whether the code runs is a small part of the problem; the reward is subjective aesthetics. What I've learnt in the months since the paper: taste moves. A static judge trains your model into yesterday's taste. The loop only works if the judge evolves too.

If text is too slow, why not pure video? Because pixels lie. A model trained to reconstruct appearance learns what a whiteboard looks like, never whether the math on it is right. Its loss was never a function of the math. More scale buys a more convincing whiteboard, not a more correct one. Correctness and appearance are different problems. Treat them as one and you get beautiful things that are wrong.

Chain of thought made models more capable by letting them think out loud in text. But text is linear, so the reasoning it exposes is linearized. We can't see what the model weighed in parallel, what it discarded, or how its sub-claims depend on each other. A showerthought: train models to render their internal state as inspectable UI. The chain of thought becomes a chain of interfaces. Confident parts stable, uncertain parts annotated, evidence unfolding on tap. Prose becomes the compressed export, not the native format. At that point interpretability and UX stop being different problems.

The question that kept resurfacing at the ICLR RSI workshop: as recursive self-improvement loops accelerate, how do humans stay in the loop without becoming the bottleneck? An RSI loop emits decisions faster than any transcript can be read. Oversight at that speed cannot be a log file. It has to be a high-throughput visual interface, built for steering at the level of policy rather than reviewing at the level of commands. This is the next defining AI safety problem. And it is an interface problem.

I built Generative UI for Amazon Nova in early 2025. Watching the more polished versions arrive since, from Gemini and Claude, says the direction is right: model output as live HTML instead of prose. But a one-shot HTML artifact was never the end state. It's the first rung. On to the next.

Books made information readable. Video made it watchable. Software made it operable. Nearly every AI product still ships answers into fixed containers designed in advance. The information just sits there, displayed. The next medium is all three at once, generated in real time. Information shouldn't display. It should run.

Agents spend most of their output on prose we will never finish reading. Then we decide from a skimmed fraction of what was generated. That is not seen as a failure mode. It is the default of every agent product shipping today. And it compounds. One model outruns us tens of times over. A swarm of ten subagents pushes the gap into the hundreds.

9 years ago at early Alexa, the hard problem was getting a model to hold up its end of a conversation, and perhaps do something useful. We've come so far: models now generate text faster than humans can read, and can rm -rf your disk before you know it. The bottleneck has crossed from the machine's side of the channel to our ability to follow and steer

Compute FLOPs aren't the ceiling to AI. We are. Frontier models can output hundreds of tokens a second. An adult reads about 5 words a second. The constraint on AI usefulness stopped being intelligence. It's the interface to humans. Every model capability gain widens a gap the human on the other end physically cannot close. That's the bottleneck to tackle.