Feb 27, 2025 · 36m · y-combinator
AI Interfaces Of The Future | Design Review · Y Combinator
Aaron Epstein · 17m spoken Raphael Schaad · 14m spoken Vapi AI Assistant · 17s spoken Retail AI Phone Prompt · 1s spoken Retail AI Call Greeting · 0s spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Hosted by Aaron Epstein alongside Notion's Raphael Schaad, this episode of Y Combinator's Design Review explores next-generation AI interfaces spanning voice agents, workflow canvasses, adaptive email clients, and generative video suites. Through hands-on product teardowns, the hosts establish key design principles for building intuitive, trustworthy, and action-oriented AI applications.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 5.9 Guest teaching 5.0 Guest disagreement 0.0 The partners pushing back 0.0
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 22:47 Good news / bad news of open-ended prompting
In a completely collegial review session, Raphael playfully highlights the double-edged sword of open text prompting when Aaron praises open-ended inputs.
Hardest push from the partners ▶ 29:20 Highlighting button stability against adaptive changesAaron points out how static shortcut keys solve the problem Raphael raised regarding users losing muscle memory when adaptive buttons shift.
Biggest teaching moment ▶ 23:02 Designing LLM constraint feedback mechanismsRaphael educates on advanced UX design by proposing squiggly lines or color cues to show which parts of a prompt the LLM successfully indexed vs ignored.
The partners hold their own ▶ 32:43 Explaining asynchronous generation design in deepfakesAaron demonstrates deep product knowledge by explaining why Argil blurs preview video to prioritize immediate audio feedback before launching a 12-minute render.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Design Review Title Sequence | 6 | 5 | 0 | 0 | Aaron introduces the theme of post-chat AI interfaces and tests Vapi with Raphael. Raphael explains the conceptual shift from static nouns to AI verbs and identifies key missing multimodal visual feedback in voice agents. | |
| Interface Review: Retell AI Phone Agent | 6 | 4 | 0 | 0 | Aaron conducts a live debt collection demo call with Retell AI, purposefully trying to trick the bot by changing his name. Both hosts collaboratively analyze the bot's dynamic adaptation and where latency impacts realism. | |
| Interface Review: Gumloop Visual Agent Workflows | 6 | 6 | 0 | 0 | Reviewing Gumloop, Aaron walks through a web scraping template while Raphael analyzes the infinite canvas paradigm, suggesting improvements like color-coded legends and zoom-dependent fidelities. | |
| Interface Review: Answer Grid AI Spreadsheet | 6 | 5 | 0 | 0 | The hosts review Answer Grid's prompt-to-spreadsheet interface. Raphael recommends contextual dynamic prompt buttons and inline footnote citations, while Aaron notes the historical parallel to classic document citations. | |
| Interface Review: Polymet Prompt-to-Design Tool | 6 | 6 | 0 | 0 | Aaron and Raphael test Polymet to generate a dashboard UI. Raphael explores how prompt interfaces could provide modular drag-and-drop design tokens and squiggly-line feedback on which prompt constraints the LLM satisfied. | |
| Interface Review: Zuni Adaptive Email UI | 5 | 6 | 0 | 0 | Reviewing Zuni's adaptive email interface, Raphael discusses levels of abstraction from autocomplete to full autopilot, and raises the interaction challenge of single-letter hotkeys conflicting with active text focus. | |
| Interface Review: Argil.ai AI Video Studio | 7 | 4 | 0 | 0 | Aaron demonstrates his own pre-trained AI deepfake avatar in Argil.ai, explaining the deliberate trade-off of generating audio fast while serving blurry video previews to keep human iteration immediate. | |
| Summary of Key AI Interface Design Lessons | 5 | 4 | 0 | 0 | Aaron and Raphael summarize their takeaways, comparing the current wave of AI interface design to the mobile touchscreen revolution around 2010. |
Statements from this episode (9)
Schaad: AI shifts software design from static nouns to dynamic verbs
“I think a big difference is that software of today or kind of like up until this point was mostly kind of like just clear things you can point out on the screen that are, you know, kind of nouns like text forms, dropdowns, buttons, et cetera. And with AI, what…”
Epstein: In voice AI applications, latency itself is the interface
“That's a great point because the latency is the interface in some ways in that how fast it responds to you, the longer it takes, the less it feels like a natural conversation and the more it feels like you're talking to a robot. The whole point is to make it s…”
Schaad: Voice AI can automate 50% of calls before human handoff
“And one of the interesting things here is that this may kind of be technology used in a first line of sort of like, you know, a defense, and then you can kind of bring a human into the loop when sort of like, you know, maybe 50% of these calls go through robot…”
Epstein: Canvas interfaces will be the standard AI agent UI by 2035
“10 years from now, this may be the standard to interact and control and monitor a lot of these agents. If they're truly everywhere and doing, you know, so much of the work that we as humans don't want to spend our time doing, this is going to be the way that w…”
Schaad: AI UX should mimic flight meta-search progressive loading
“And maybe some good prior art to borrow from there is today's kind of meta search engines for flights they also take a while, and they already show you kind of in a low resolution, some early results, and you can always start to interact with the filters as ki…”
Schaad: Iterative delta edits and sub-prompts are the frontier of AI UX
“Yeah, so, so, so any interface designer or technical team kind of figure out the challenge of how to kind of add sub-prompts or how to only change kind of iteratively that, that, that, that, that's kind of like really the frontier.”
Schaad: Adaptive AI UIs clash with user desire for predictable controls
“And the challenges, of course, are kind of predictability. People love to have like their, you know, billion buttons and like the exact place.”
Schaad: Blurring video previews enables fast human-in-the-loop AI iteration
“They're trading off basically fidelity for
Immediacy and basically putting the human comes back in the loop because if it was just the generate button, right, we would wait for 12 minutes figure out that something is not quite right and then kind of like, you…”
Schaad: The AI UI paradigm shift mirrors the 2010 touch revolution
“It almost feels like back in like 20 10 or so, when touch devices really kind of came on the market and everything had to reinvent the kind of touch first, and we're at one of those moments again, where like all of software, all the components that we kind of …”