AI
Computing’s text-to-canvas shift, and what it means for AI
Computing started with text. The 1981 terminal was a chat box: you type, the machine replies. Then Engelbart’s 1968 “Mother of All Demos” introduced the cursor and a spatial canvas. Overlapping windows followed a decade later. The Macintosh brought that to everyone and changed computing permanently. Phones ran the same cycle: T9 text, then the Blackberry keyboard, then the iPhone touchscreen. Every major computing platform started with text input and eventually moved to a canvas. That pattern has repeated twice. Wilson’s question: will AI repeat it a third time?
AI is still in the CLI era
ChatGPT, Claude, Gemini: every current AI interface works the same way. You send a message and get a message back. Claude Artifacts adds a work area alongside the chat, but the only way in is still the text box. Image uploads exist but occupy a small corner of the interface. The dominant surface is the chat thread. Wilson’s argument: AI is in exactly the same position that 1981 terminals were in, and that early Nokia phones were in. The canvas transition hasn’t happened yet.
Make Real: sketching UIs and iterating with annotation
On 14 March 2023, the first vision model shipped. That unlocked a different kind of canvas interaction: draw a rough UI sketch, hit “Make Real,” and get back working HTML. Wilson ran the Make Real live demo across three models (12:33) simultaneously, GPT-4o returning a plain functional result, Claude Sonnet adding animation, Gemini Flash doing the bare minimum. The bigger insight: you don’t stop at the first result. Draw corrections on top of the generated page, resend it, and the model fixes the specific thing you marked. Sketch, generate, annotate, regenerate. That feedback loop runs entirely on the canvas.
Model as canvas author: letting AI draw back
What if the model’s output isn’t HTML but more canvas? TL Draw gave the model tools to draw shapes directly. Early results were described internally as “shitty but amazing,” with the model confusing which way was up and scratching horror-film lines. The team eventually found an approach they called an AI-driven format: they listened to what the model naturally output without instruction, then wrote what they called the world’s most fuzzy parser to turn it into real shapes. That format, named “easy,” generates flowcharts, wireframes, and scenes, and can reverse a real website back into a wireframe.
TL Draw Computer: natural-language node-wire workflows
TL Draw Computer is a node-wire visual programming environment where every node is a plain-English instruction sent to an AI model. Connect a node that says “increase by one” to a number, loop it, and the model runs that step forever until you stop it. Nodes handle ambiguous inputs: feed the word “bicycle” into an “interpret as a number” node and the model decides it’s two. Feed it a drawing of an octopus and it returns eight. Wilson ran the TL Draw Computer invention inventor pipeline (36:18) live, chaining name brainstorming, pitch writing, and landing page generation from a single audience-shouted word.
The agent starter kit and the case for the canvas
TL Draw ships a forkable agent starter kit. In the agent starter kit demo (40:52), each capability (planning, thinking, shape creation, external API calls) is a separate module you can comment out. Turn off thinking and the agent draws immediately. Turn off everything except thinking and it freezes. Wilson closed by pointing at road signs, da Vinci’s notebooks, and cave paintings:
“Listen, this is how we’ve always done it, right?”
Humans reach for spatial representation when text is too slow. Every computing platform eventually moved past text. The canvas is where AI goes next.
Notable Quotes
shitty but amazing, right? Because it’s Lu Wilson · ▶ 28:23
passer to turn it into real shapes. Lu Wilson · ▶ 30:20
Listen, this is how we’ve always done it, right? Lu Wilson · ▶ 51:07
Key Takeaways
- Computers and phones both moved from text interfaces to canvas interfaces; AI has not made that transition yet.
- TL Draw’s canvas lets you send drawings to AI models, annotate results by drawing on them, and loop corrections without leaving the canvas.
- TL Draw Computer replaces code with natural-language instruction nodes, letting non-developers wire together multi-step agentic pipelines visually.