/ 01 — The problem
What was broken.
Most LLM UIs in 2023 treated model output as if it were a query result — render and done. That worked for demos but failed users the moment the model was wrong, slow, or interrupted.
←Back to work
/ 09jsconf-talk·2023
Conference talk · 2023
JSConf Talk.
Keynote at JSConf.id 2023 on how to design UIs that gracefully handle LLM uncertainty — streaming, interruption, citations, and the 'admit-when-unsure' pattern.
// previewJSConf Talk
views on YouTube1.8k
talk length45 min
patterns covered6
follow-on invites2
/ 02 — The approach
How I tackled it.
Walked through six concrete UI patterns I'd seen work — streaming with interrupt, source-citation popovers, confidence-gated actions, regret-free retries, structured handoffs, and 'show your work' panels. Live coded the simplest one.
/ 03 — The outcome
What shipped.
Talk hit 1.8k views in 3 months. Patterns became the basis of stream-ui (project #3). Invited to give variations at 2 follow-on meetups.
/ Like what you see?