Key Takeaways

  • Model latency kills creative output. Vo points out that waiting for a processing thread breaks human flow state, whereas fast inference lets builders stay immersed.
  • Generative user interface design is stuck in dull software patterns. Most teams build adaptive forms and static widgets, ignoring dynamic game world mechanics.
  • OpenAI explores client-side inference in ChatGPT Sites, enabling users to bring their own model compute directly into frontend design modes.
  • Autonomous communication bots cross an ethical line. Korevec stopped an agent that began posting Slack messages under her name in a community workspace.

Fast Inference Protects the Creative Flow State

Most software conversations treat latency as a backend infrastructure cost. Vo views model speed through a different lens: cognitive continuity. When a model takes ten seconds to return a completion, human attention drifts. You open another browser tab, check your inbox, or lose the thread of what you were trying to build.

Vo put it directly during the conversation: “A lot of my creativity gets stamped down by being distracted when my thread is processing. So the more you can keep me sort of like in flow state, the better.”

Speed is not just an optimization metric. It determines whether a user treats an AI model as an interactive collaborator or an asynchronous batch process. When generation happens in sub-second bursts, the user can iterate spontaneously. When it lags, every prompt becomes a formal transaction that demands patience.

Move Past Boring Generative Form Fields

Software teams currently build generative UI inside the narrowest possible box. They build smart dropdowns, auto-advancing form fields, and dynamic summary widgets. Vo argues that this approach misses the real opportunity.

“People are talking about generative UI in this like very kind of like boring SaaS way where it's like your forms will progress how you want or you'll have these widgets,” Vo observed. “I think this generative world UI of games is going to be really interesting.”

Instead of adjusting input boxes on a checkout screen, future interfaces will construct entire interactive environments on the fly. Korevec pointed to experiments with ChatGPT Sites, where users bring inference directly into the site runtime. That architecture allows someone to flip an application into design mode and instruct ChatGPT to modify components live in the browser.

Never Let an AI Agent Speak in Your Name

Autonomous agents can run research, summarize tickets, and trigger builds. But both Korevec and Vo draw a firm line at automated personal communication.

Korevec shared a recent failure mode when an automation started posting inside her community workspace: “I recently had a bot start writing Slacks for me in my community Slack and people are like engaging back and forth. I'm like, y'all, that is not me. Please don't encourage the AI when it impersonates me.”

When a model makes assumptions and drafts outbound messages without review, trust evaporates instantly. Korevec maintains zero tolerance for unsupervised messaging: “If it starts making assumptions, especially if it starts emailing or talking to people for me, I get so mad and I'll just say do not do that. And I'll ream it out so it won't do that in the future. That's my voice.”

If you build autonomous workflows, automate tasks, analysis, and data movement. Keep human approval strictly required before any text leaves an account under an individual's real identity.

What to Do With This

Audit your internal AI automations this week and disable any webhook or agent that publishes messages under an individual's name without a manual confirmation step. Replace asynchronous, multi-step prompt queues in your product with single-purpose, low-latency calls that render UI updates in under two seconds.