Derek Law

Founder & CTO, SigmaZ AI Lab·ex-Amazon AGI, London

Building generative UI as the substrate for post-AGI human–AI I/O. This page is an argument for that thesis, and — inconveniently for me — it is also made of it.

Each card is a model's attempt to render an idea as a surface, not a paragraph.

Scroll or tap ↓ for the next idea. Swipe or tap a tab to change which model rendered it. Rate what you see, then the critique model earns its own turn. Cheap inference is one slow rate limit away.

01/12The Post-AGI I/O bottleneck
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
02/12GenUI > diffusion for information transfer
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
03/12ICLR 2026: latent-surface alignment
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
04/12Ambient / wearable compute
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
05/12Priors for non-visual BCI UX
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
06/12Agents are overrated, UIs are underrated
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
07/12Embodied intelligence as the real VLM test
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
08/12HCI is the bottleneck, not scaling
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
09/12What I'm building at SigmaZ
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
10/12Chat UIs are the new command line
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
11/12Haptic bandwidth as the next I/O channel
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑
12/12The interface is the agent loop
Gemini 3.1 Pro
GPT-5.4
Rate this rendering
Rate first. Critique reveals after.
rate first ↑

The dashboard is a card too.

Rendered by the same pool of models, regenerated hourly from the ratings this page collects. If the model fails, the failure is the message.

dashboard by Gemini 3.1 ProCACHED · 1h TTL
Grok 4.20
GPT-5.4
Leaderboard ranks by human stars. Scatter plots human-vs-VLM alignment. n = total ratings collected by this page.