UX and not LLM - Chris Pedregal, Granola's founder, on the real moat for AI products

Why sweating the small stuff matters

Chris Pedregal, founder of AI note-taker Granola, is one of the sharpest AI product thinkers out there. No surprise, then, that he's turning Granola into the de rigueur AI note-taker across companies. So when he talks about AI product design, it’s worth listening.

In a recent episode of Invest Like the Best (EP.412), hosted by Patrick O’Shaughnessy, Chris laid out a key idea that’s often underrated: AI hasn't changed the fundamental differentiator in a consumer app—it’s still the UX and not the models.

All the app layer products run on top of the same foundational models. What separates a product that you love from one that's OK is how well they sweat the details. Chris gives a simple example of catering to edge cases:

"A lot of what makes Granola great is sweating the details of all these technical edge cases, stuff you'd never think of. It's like you're in the middle of a meeting and you take off your AirPods and it's on a Zoom call that has multiple channels. And all of a sudden, Granola needs to do something very specific to make that feel seamless that you never would have thought of until you built it and you realize it felt crappy if you didn't do that."

He makes another point about the evolution of the user interface. As ChatGPT is the dominant consumer AI app, all other AI apps just copied its chatbox UI. While chatboxes are okay for conversational tasks, which was the use case for the early versions of ChatGPT, they are not the best UX for making an image/video. In fact, some of the current ChatGPT features, like reminders, don't fit at all into the chatbox UI construct. So, the app layer's interface has to evolve, which is the other opportunity area for an app-layer product. Chris talks about this through an analogy.

Today’s AI interfaces are like the first generation of cars, which didn’t have steering wheels—just a stick you’d turn. That worked fine when driving slow, but at higher speeds, it was impossible to control. The steering wheel changed everything. I think we still have to invent what the steering wheel is for when you're working with AI and collaborating with AI. Right now we have some very coarse controls and it's turn-taking right now. It's like I write something, then the AI does something, then I react back to it. And I think it's going to be a lot more fluid and a lot more collaborative once we figure that out.

This aligns with my experiments with AI videomakers.

I’ve tested the top ones—Heygen, Veed.io, Deepbrain, InVideo. The differences between them aren’t massive—some models are ahead today in terms of expressions and movement, but the rest will quickly catch up. What’s more interesting is that after generating an AI avatar video, I still have to export it into Canva to create something useful.

The real moat for these AI avatar platforms won’t come from having the best AI-generated talking head. It will come from how seamlessly they can evolve into a full-stack AI-led video editor—which, at its core, is all about UX.

And this is where another key factor in making AI useful comes in: context.

The best AI models in the world still need the right context to perform well. ChatGPT now has a memory feature that allows it to retain information about your past interactions. Without context, even the most advanced LLMs are just guessing.

Chris argues that providing context is actually part of UX. If you compare it to Google, in a search engine world, we simply type what we need into the box. But in a ChatGPT world, we often need to give it more details first.

He illustrates this with a simple story:

"I was trying to barbecue some shrimp for the team. We bought some shrimp. This was in Spain. I've never barbecued shrimp before. I'm typing into ChatGPT, like, okay, how do you barbecue shrimp? And Voss was like, no, give it the right context. So he's like, take a photo of the barbecue and take a photo of the shrimp. And he was totally right. So I was like, yeah, yeah, yeah, give it the context. So I did this. Turns out the shrimp was already cooked. We didn't realize it because it was in Spanish. So we didn’t have to cook it at all, just heat it up—which we never would have figured out if I had just typed in a question."

The key point: AI tools require a new kind of user intuition, which a new type of User Experience layer has to match—the same way earlier generations didn’t instinctively use Google, AI natives will grow up understanding how to give AI the right context to get better results.

But this leads to another challenge: Even if an AI tool has all the context about a person (from emails, meetings, wearables, etc.), how does it decide which context to use for each task?

Chris points out:

"Gathering the data itself isn't difficult; it's a matter of time before various data sources (emails, notes, documents, tweets) can be integrated into LLMs like Anthropic or ChatGPT. The real challenge, however, lies in determining which context is relevant for a specific task."

This is where the final battle for AI usability lies. AI products cannot win by just bolting on the model ranked highest on the benchmarks table—they’ll compete on having the best way to collect, organize, and apply the right context at the right time.

As AI product users shift from early adopters to late adopters/early majority, they won't care if the product uses the fastest LLM or the most advanced reasoning model. All they'll care about is how seamlessly it integrates into their workflows and solves their problems.

The winning products will be the ones that solve for friction—whether that’s through a better-designed interface, a more intuitive way of handling context, or by sweating the small details that turn an AI from “impressive” to indispensable.

Link to the episode: https://joincolossus.com/episode/building-granola/

Thanks for reading.

Rohit