The future of AI products isn’t chat. It’s natural that we started here, but we won’t achieve the high-performance, partnered products we dream of by building everything around messaging experiences.

Where it began for me

In May of 2021, we were putting the finishing touches on the Google I/O keynote. We had skipped I/O 2020 amidst the global shutdown of COVID-19, but with vaccines — and before Delta came ripping through — it was exciting to see colleagues in person again and prepare to talk about a host of new inventions.

My understanding of the potential of LLMs coalesced while prepping to show LaMDA. While the transformer paper and BERT had convinced me that conversational interfaces were inevitable, using LaMDA while writing for I/O was the moment. We knew we had to get this right. We had to tell a story people could understand that flowed from Google’s earlier work on image recognition and ML into what makes conversational computing so different.

It also had a moment that still makes me smile. We were debriefing a rehearsal and Sundar was tweaking his line after showing Google image recognition helping with algebra.

“I wish I had this when I was in school”

The thing about humor — as I’ve explained to my kids — is you have to be right on the line. You have to go for it. The core I/O team was with Sundar, plus a handful of Google SVPs. As you’d expect, there tends to be an aura of respect, calm, and quiet around Sundar, particularly when he’s working on something.

So, I had near total silence for “yeah, then maybe you would have made something of yourself.”

The near silence became total silence, long enough for me to wonder if my estimate of the line was off.

Then Sundar started cracking up, along with everyone else. After the year of lockdown, the pressure to pull off I/O amidst COVID protocols, and the losses so many had felt, it was one of many moments of returning to work that I’ll always cherish.

I/O is one of the most collaborative efforts of any project — or product — at Google. For me, it was a rich source of inspiration and ideas around working with people and technology to solve hard problems.

Conversational and Personal

I/O 2021 was also where Matias launched Material You. Material You and LaMDA, to me, form two foundations of how we should be thinking about products in the age of GenAI, LLMs, and agents. With Material You, we released a user interface that could be personal, quirky, and very you while still achieving high standards of usability and utility.

This matters because while UIs and UXs can be beautiful, memorable, and striking, they aren’t just pieces of art. They exist to support your goals, your ideas, and ensure the product connecting you to them does all the amazing things it promises. You need look no further than the active commentary about Apple’s new Liquid Glass to see how difficult it is to balance these pressures. Like the Material team at Google, Apple has some of the most talented designers and developers in the world, and they’re struggling to get this right.

And getting it right matters. Personal products done right — that feel comfortable and familiar to each of us while preserving high utility — help us get the most out of products and find the right head-space to learn and discover new capabilities and superpowers in them.

LaMDA made it clear that superpowers were coming fast. As Sundar explains in the keynote, conversational products enable you to end up in different places than you started, to collaborate on goals, information, and experiences while exploring solutions. LaMDA was too slow and limited to do this well or at scale, but the potential was clear to everyone working on it.

Beyond Personalized and Chat

This combination led me to think about products that partner with you, rather than simply act on signals to optimize via personalization. A partnered product respects your time, attention, and preferences. Like a real-world partner, it should be opinionated and push the user when needed. It’s on your team, working with you and on your behalf. My current working definition of a partnered product is:

1. A product that actively collaborates with the user and anticipates needs without being asked

2. Clearly and directly embodies the developers’ opinions about the best, most effective approach to the problem

3. Explicitly responds to and reflects explicit and implicit feedback from the user, being clear when that feedback is being balanced against other priorities

What what an actually partnered coding experience be like? Clearly, adding a chat bot into VS Code is reasonably valuable and it is probably a long-term productivity gain as well. But what if they were really partnered? What if rather than being a well read, incredibly overconfident minion, coding assistants were there to help you efficiently deliver the best possible solutions?

What would be different?

First off, they’d always be aware of your (or your organization’s) coding styles and preferences. Second, it would be opinionated, so when those preferences collided with its understanding of best practices, that would lead to a conversation. Third, the interface wouldn’t need to a be a constant series of “hey, AI, I need you do something for me”, instead the partner would be picking up what it was good at all the time — writing tests, looking for refactoring opportunities, spotting inconsistencies, etc.

Now that would be an awesome coding partner!

Partnered products require developers to think in ways resistant to the pressures enshittification (a term made famous by The Other Cory, who has a book coming out about it — I often disagree with Cory on particulars, but rarely about the big picture) tend to apply to products built foundationally on attention reinforcement. How is the product clearly valuable to its audience and customers? How is that value communicated and shared? How easy is it for the customer to tweak and tune that value to their specific needs right now?

What I’m most excited about extends from that early work of 2021, particularly when we start thinking beyond chat conversations. Unsurprisingly, chat has become the de facto interface for AI. LLMs are so charmingly (distractingly? terrifyingly?) human in their responses — thank you, training sets largely built on human communication — why wouldn’t we want to talk to them? Chat interfaces also give natural places to hide latency, partial responses, and reset things when they go off the rails.

But if we think of our highest performance partnerships, the most rewarding and effective, we aren’t using chat as the exclusive — or even primary — communications channel. Bands, teams, negotiating the lineup at The Point — chatting is the fallback, not the primary. What happens when we think about products that way, too?

What are the product equicalents of a glance, a beat, of timing? What could we change when non-verbal, non-textual cues are flowing back and forth between the product and user? In many of these, the AI — the heavy lifting the teams involved will have done — may feel much less obvious than a big chat window.

This is a good thing, so long as we don’t fall back into invisible personalization. Partnered products need both the AI and the user to be expressing and reflecting preferences, experimenting with responses, and working together to create the best possible experiences.

Partnered products aren’t just “chat bot + existing app.” Inventing them is the most exciting product exploration of our time and I’m so excited to be in the midst of it.