Bee AI：可穿戴式环境智能体

Hey everyone, welcome to the Latent Space podcast. This is Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host, swyx, founder of Smol AI.

Hey, and today we are very honored to have in the studio Maria and Ethan from Bee. Welcome.

Hi. Thank you for having us.

You are, I think, the first hardware founders we've had on the podcast. We've been looking to have an AI wearable hardware founder on for a while. I think we're going to have 2 or 3 of them this year, and you're the ones whose product I wear every day.

Thank you for making Bee.

Thank you for all the feedback and usage.

I've been a big fan. You were the speaker gift for the Engine No. 1 Sphere. Let's start from the beginning. What is Bee Computer?

Bee Computer is a personal AI system. You can think of it as AI living alongside you in the first person. It can capture your real-life context, and with that understanding, it can help you in significant ways.

The obvious one is memory, but that's really just the base use case: recalling things and reflecting. I know, swyx, that you like the idea of journaling, but you don't want to do it, while still having some kind of reflective summary of what you experienced in real life.

It's also about having the whole context of a human being and giving the machine the ability to understand what's going on in your life—your attitudes, your desires, and specifics about your preferences.

That means it can not only help you with recall, but also with anything else you need it to do, because it already knows what you would want. Think about somebody you've worked with or lived with for a long time. They just know, without having to ask you, what you would want.

It's clear that this is the future. Personal AI is just going to be much more valuable with personal context.

One of the things we're really passionate about is understanding this personal context, because it will make AI more useful. Think about a best friend who knows you so well. That's one of the things we're seeing from users: they use Bee from a companionship standpoint and for professional use cases. There are many ways to use Bee, but companionship and professional use are the ones we're seeing the most right now.

It feels so dry to talk about use cases.

Yeah, well.

It's really an investor question. But at its best, don't you want your AI to know everything you've said and everywhere you've been? Wouldn't you want that? You shouldn't have to repeat every time what you like. It already knows that, and it does things for you based on that. I think that's really cool.

Great. Do you want to jump into a demo? Do you have any other final questions?

Before we do that, maybe we should cover the origin story. How did you two meet? Was this the first idea you started on, or was there something else before?

I can start. Ethan and I have known each other for 6 years. He had a company called Squad, and before that it was called All About, which was a personal AI company.

Yeah, maybe you should start with that.

That's how I know Ethan. He was pivoting from personal AI to Squad, which was a co-watching-with-friends product. I had experience working with TikTok and video content, so I helped with the pivot, and we launched Squad.

That was really successful. In the end, the founders decided to sell it to Twitter, now X. Both of us joined X. We launched Twitter Spaces and many other products, and we continued working together until we started Bee.

The interesting thing is that this isn't our first attempt at personal AI. In 2016, when I started my first company, it started out as a personal AI company. This was before transformers—before BERT even, just RNNs. You couldn't really do any convincing dialogue at all.

I met Esther, who was my previous co-founder, and we were both really interested in the idea of having a machine model and understand a dynamic human. We wanted to make personal AI. This was more geared toward younger people, because we had much more limited tools at the time.

I don't know if you remember, but in 2016 there was a brief chatbot boom. It was way premature, but it was when Zuckerberg went up at F8 and talked about the Messenger platform. People thought, “Oh, bots are going to replace apps.”

That lasted for about 6 months, and then everybody realized that these things were terrible and weren't replacing apps. But that's when we got excited and tried to make something where you could teach the AI about yourself. It was just an app that you chatted with. It would ask you questions and then give you some feedback.

What we learned was that some people really loved chatting and answering questions, but it was a lot of work just to manually teach an AI all these things about you. We had sentence similarity and other things we could try, but the technology was premature at the time, so we pivoted.

How did the first version get launched?

We started in the same office as Hugging Face because Betaworks was our investor. They had a program called Botcamp. Betaworks is a really cool VC because they invest in things that are out there and way ahead of everybody else.

At the time, Botcamp had 6 companies. It was us and Hugging Face, and I think the other 4 are dead. Hugging Face was the one that really took off. A 30% success rate is pretty good.

It was just the 2 founders. It was a chat app for teenagers. A lot of people don't know that Hugging Face was originally like, “Hey friend, how was school? Let's trade selfies.” They built the Transformers library, I believe, to help make their chat app better. Then they open-sourced it, and it blew up. They realized that this was the opportunity, and now they're Hugging Face.

There were some people who were super passionate about it. Teenagers, for example, really like to talk about themselves, so they would reply to a lot of questions and talk about themselves. But most people don't really want to spend that much time talking.

We went through Y Combinator. Long story short, we pivoted to consumer video, and that went really viral and got a lot of usage quickly. We ended up selling it to Twitter, worked there, and left before Elon—not related to Elon, but we left before that.

I should mention that this was the famous time when Elon had just come in. Esther was the famous one.

Yes, she was my former co-founder. She was the one who was sleeping in the office. She stayed. We had left by that time, but she stayed for a while. Later, she left, or I think she got laid off. I think the whole product team got laid off. She was a director of product.

Then we thought, “Oh my God, things are different now.” We really started working on Bee again right before ChatGPT came out. We had an app version and were trying different things around it, but ultimately it was clear that there were some limitations.

A good question to ask any wearable company is: why isn't this an app?

Because we tried the app at the beginning.

The idea behind Bee comes from “ambient.” If it were just around you all the time, rather than requiring you to open the app and make the effort to enter data, that led us down the path of hardware.

The sensors on this are microphones, so it's capturing and understanding audio. We started with hardware that also had a vision component, and we can talk about why we're not doing that right now.

If you wanted continuous audio understanding with your phone, it would monopolize your microphone. It would get interrupted by calls, and you'd have to remember to turn it on. That little bit of friction is actually a substantial barrier to the experience of having it with you all the time and living alongside you.

We do have Apple Watch support, so anybody with an Apple Watch can use Bee right away without buying any hardware. We worked really hard to make a version for the watch that can run in the background without draining your battery too much.

Even with the watch, there's still friction because you have to remember to turn it on, and it still gets interrupted if somebody calls you. You have to remember to go back and turn it on. We send a notification, but you still have to go back and turn it on because that's just the way watchOS works.

One of the things we're seeing from our Apple Watch users is that they like the Apple Watch integration. Many people start using Bee from the Apple Watch, and after a couple of days they buy the Bee because they just like wearing it.

We're learning from that, and it's really cool.

Fundamentally, we like to think that personal AI is the mission. It's about understanding, connecting the dots, and making use of the data to provide some value.

The hardware is the ears of the AI. It's about integrating the incoming sensor data, and that's really what we focus on. If we can do it well and have a great experience on the Apple Watch, that's just great.

I mean, but there are some platform restrictions that existing hardware makes it hard to overcome.

What do people do in 2 or 3 days that then convinces them to buy the product? This was a product where, after you use it for a while, you have enough data to start to get a lot of insights.

For Apple Watch users, I believe that's because every time they receive a call, they need to go back to Bee and open it again. Or, for example, every day they need to charge the Apple Watch, and it reminds them to open the app every day. They feel like, “Okay, maybe this is too much work. I just want to wear the Bee, keep it open, and that's it. I don't need to think about it.”

I think they see the potential just from the watch, because even if you wear it for a day, we send a summary notification at the end of the day about the key things that happened to you during your day. I didn't even think—I’m not a journaling-type person. I was like, “Oh, I just lived the day. Why do I need to think about it?” But it's actually pretty interesting. Sometimes I'm surprised by how interesting it is to me just to be like, “Oh, yeah, that,” and see how it fits together. I think that's something people get immediately with the watch, but they're like, “Oh, I'd like an easier way to do this.”

It's surprising because I only know about the hardware. But I use the watch as a backup when I don't have the hardware. I feel like, because now you're beamforming and all that, this is significantly better.

Yeah, that's the other thing. We have way more control over the hardware. With the Apple Watch, you're limited: you can't set the gain, you can't change the sample rate, and there's very limited framework support for doing anything with audio. Whereas if you control the hardware, you can optimize it for your use case.

The Apple Watch isn't meant to record this, and we can talk, when we get to the part about audio, about why it's so hard. This is audio at the hardest level because you don't know what environment it has to work in. This environment is great—we're in a studio—but afterwards, at dinner in a restaurant, it's a totally different audio environment. There are a lot of challenges with that. Having really good source audio helps, but there's still a lot that has to be done with machine learning to account for it. You can tune something for one environment or another, but it will make one good and the other bad. Making something flexible enough is really challenging.

Do we want to do a demo just to set the stage, and then kind of talk about it?

Yeah, I think we can walk through the product.

For listeners, we'll be switching to video that will superimpose onto this video. If you want to see, go to our YouTube. Like and subscribe as always. And buy the B. Yes, and buy the B. While you wait for the video to come out, buy the B. Maybe we should have a discount code just for the listeners. Sure. If you want to offer it, I'll take it. Yeah. All right. Discount code SWX.

An important thing to mention is that the hardware is meant to work with the phone. If you look at Rabbit or Humane, they're trying to create a new hardware platform. We think that the phone is just so dominant, and it will be until we have the next generation, which is not going to be for 5 years. Maybe some Orion-type glasses that are cheap enough and light enough will come along, but that's going to take a long time. So we work with the phone rather than trying to replace it.

In the app, we have a summary of your days, but at the top is what's going on now, and that's updating continuously. Right now, it's saying that I'm discussing the development of personal AI, and that's just the ongoing conversation. Then we give you a readable form with little segments of the important parts of the conversations.

We do speaker identification, which is really important because you don't want your personal AI thinking you said something and attributing it to you when it was just somebody else in the conversation. You can also teach it other people's voices, so if there's somebody close to you, it can start to understand your relationships a little better.

We do conversation endpointing, which is a task that didn't even exist before because nobody needed to do this. If you had somebody's whole day, how do you break it into logical pieces? We use not just voice activity, but other signals to try to split it up, because conversations are a little fuzzy. They can lead into one another, and one can start before the next. We also use the semantic content of the conversation.

When a conversation ends, we run it through larger models to try to get a better sense of what was actually said, then summarize it, provide key points, describe the general atmosphere and tone of the conversation, and identify potential action items that might have come from it. At the end of the day, we give you a summary of your whole day, where you were, and a step-by-step walkthrough of what happened and what the key points were.

That's the base capture layer. If you just want to get a glimpse, recall something, or reflect, that's there. But really, the key is that all of this now feeds into generating personal context about you. We generate key facts known to be true about you, and there's a human-in-the-loop aspect: you have visibility into that. I have a lot of facts about technology because that's basically what I talk about all the time, but I also have some hobbies that show up.

I measure my day now by asking, “What is my token output of the day?” As a human, how much information do I produce? It's measured in tokens, and it turns out to be around 200,000 tokens a day.

In the recall case, we have a chat interface. The key here is recall: I probably have 50 million tokens of personal context, so how do you make sense of that and make it useful? I can ask simple recall questions, like details about a recent trip I took to Taiwan, where we were with our manufacturer. In real time, it has various capabilities, such as searching through my memories, searching the web, or looking at my calendar. We have integrations with Gmail and Google Calendar, so it can connect the dots between in-real-life and digital life. I just asked it about my Taiwan trip, and it gives me a breakdown of the details, what happened, and the issues we had around certain manufacturing problems. It also goes back and references the conversation, so I can return to the source.

Yeah, and not just the conversation, but the integrations as well. We also have Gmail and Google Calendar. If there's something there that's useful for more context, we can see that.

I never use the word “agentic” because it's strange, but it can search through your conversations, search through email, and look at the calendar. If I'm brainstorming about something that spans across all of those, it can search through my conversations, search through email, look at the calendar, and then, depending on what's needed, synthesize something with all that context.

I love that you did Spotify Wrapped. That was pretty cool.

Yeah, one thing I did was make a Spotify Wrapped for my life in 2024. It's surprisingly good. It gave me metrics like, “You visited 3 countries and shipped X many beta devices.” It gives a lot of personal insights and reflection points.

That's fascinating. So that's the demo. We can show something that's in beta if we want to.

The vision is not just about AI being with you, passively understanding you through your experiences, but also proactively suggesting things to you at the appropriate time. It's not just a tool; it can step in and suggest things to you.

You’re asking for a recommendation for an Italian restaurant. Would you like me to look up some highly rated Italian restaurants nearby and send her a suggestion?

What I did was just send Ethan a message through WhatsApp on his personal phone.

Basically, Bee is watching all my incoming notifications, and if a notification meets 2 criteria—is it important enough for me to raise a suggestion to the user, and is there something I could potentially help with?—this is where the actions come into play.

Because Maria is my co-founder and because it was a restaurant recommendation, something Bee could probably help with, it proposed that to me. I can respond either through the chat or with another push-to-talk, walkie-talkie-style button. It's actually a multipurpose button to toggle Bee on or off, but if you push and hold it, you can talk. So I can say, “Yes, find one and send it to her on WhatsApp.”

Bee is an Android cloud phone, so it has access to all my accounts.

We're going to abstract this away, and the execution environment isn't really important. We can go into technically why Android is actually a pretty good one right now. But it's searching for Italian restaurants, and we don't have to watch this. I could have my AirPods in my ears and my phone in my pocket. It's going to go to WhatsApp, find Maria's thread, send her the response, and then let us know.

Oh God. What's a good suggestion? I mean, it's not an Italian restaurant.

Yeah. What did you say?

It's easy to say.

Exactly. It's easy to say. Successfully found and shared. Let's see what the AI says.

What does the AI say?

Bottega. I think it's called Bottega. It said it twice. I've been to one called La Cocina, I think. That was good.

Beretta's on Valencia Street. It's fine. The pizza isn't good.

It's not good?

Some of the pastas are good. I'm sorry—Beretta's, sorry. But there's this place, Delfina. Everybody here says La Pizzeria Delfina is amazing. I'm like, this is not—I don't know. It's great. North Beach Cafe.

For the record, since you're all Italians, what's the best Italian restaurant in SF?

Oh my God. I feel like I don't have one. No. I'm not sure.

The place you took us with Michele last night.

Vega.

The guy at Vega just happened to be Italian. It's in Bernal Heights.

He's nice.

What's the name of the place?

Vega.

Vega.

Okay. Cool, cool. We got the name.

But Vega—is that in Italian? What does it mean in Italian?

Vega means... It doesn't mean anything to me.

The last thing I'll mention is also the wake-word detection. The phone can be off, and you can just say, “Hey, Alfred,” and then it just...

Yeah. Being able to have wake words enables some form of voice-agent features that even ChatGPT can never have, because they don't have the recording layer.

Yeah, I think we have some other ideas, even beyond wake words, but I think it's interesting to see how people use the voice side of it. We're going to see a lot of innovation around hardware and stuff, but the real core is being able to do something useful with the personal context.

You always had the ability to capture everything. We've always had recorders, camcorders, body cameras, stuff like that. But what's different now is that we can actually make sense of and find the important parts in all of that context.

And then one last thing—I'm just doing this for you—is that you also have an API, which I think I'm the first developer advocate, because I had to build my own app.

We need to hire a developer advocate.

Or just—yeah, hire AI engineers.

The point is that you should be able to program your own assistant.

I tried Omi, the former Friend, the knockoff Friend. Real Friend doesn't have an API, and Limitless also doesn't have an API. I think it's very important to own your data and be able to reprocess your audio, although by default you don't store audio.

Yeah.

And then also just to do any corrections. There's no way that my needs can be fully met by you.

Yeah.

The API is very important.

Yeah. I've always been a consumer of APIs in all my products.

We are API enjoyers in this house.

Yeah, I am. It's very frustrating when you have to go build a scraper.

This whole combination—you have my location, my calendar—it's really, for me, the sort of personal assistant: just to write into it or to have it take action on external systems?

We're expanding it. It's right now read-only. In the future, very soon, when the actions are more generally available, it'll be fully supported in the API.

Mhm. Nice. I'll buy one after the episode. The API thing, to me, is the most interesting.

We do have real-time APIs, so you can even connect a socket and connect it to whatever you want it to take actions with.

Yeah. When I look at these apps—and there are so many of these products being launched—it's great that I can go on this app and do things, but most of my work and personal life is managed somewhere else. So being able to plug into it is nice.

I have a bunch of more human questions that I think people might have. One is: Is it good to have instant replay for any argument that you have? I can imagine arguing with my wife about something. There are these commercials now where it's basically 2 people arguing, and they can throw a flag, like in football, and have an instant replay of their conversation.

I think this is similar, where people can almost not argue anymore or lie to each other, because in a world in which everybody adopts this. I don't know if you've thought about it. And also, all the lies that all of us tell—there are sometimes things that contradict each other, because I might say something publicly and think something privately that I tell someone else. How do you handle that when you think about building a product like this?

I would say that I like the fact that the AI is an objective point of view. I don't care too much about the lies, but I care more about the fact that it can help me understand what happened and the emotions in a really objective way—a really critical and objective way.

If you think about humans, they have so many emotions. Sometimes something that happened to me—I don't know—I will feel really upset about it, or really angry, or really emotional. But the AI doesn't have those emotions. I can read the conversation, understand what happened, and be objective. I think that level of support is the one that I like more, instead of, “Did this guy tell me a lie?” I think that's not exactly what I find curious for me in terms of opportunity.

Are you going to interject in real time? Say I'm arguing with somebody. Will the AI say, “Hey, no, you're wrong. That person actually said...”?

Proactivity is something we're very interested in. Maybe not specifically for settling arguments, but more generally. A lot of the challenge here is that you need really good reasoning to pull that off, because you don't want it constantly interjecting—that would be super annoying. You also don't want it to miss things that it should be interjecting.

It would be a hard task even for a human, to come in at the right times when it's appropriate. With the personal context, it's going to be a lot better, because if somebody knows about you. But even still, it requires really good reasoning to not be too much or too little, and just right.

The part about some things is that you say something to somebody else, but afterward I change my mind and send something. Every time I have a different type of conversation and data about me.

One of the things that we're learning is that humans evolve over time. For us, one of the challenges is actually understanding whether this is a real fact. So far, we have a human in the loop who can say, “Yes, this is true. This is not,” or they can edit their own fact. In the future, we want to have all of that automated inside the product.

Your question also hits on privacy, which I know we'll talk about. If you have some memory and you want to confirm it with somebody else, that's one thing. But it's for sure going to be true that in the future—not even that far into the future—it's just going to be normalized.

We're in a transitional period now, and I think one of the key things for us is to navigate that and make sure we're thinking of all the consequences and how to make the right choices in the way that everything's designed. It's more beneficial than it could be harmful, but it's just too valuable for your AI to understand you.

If it's Meta Ray-Bans or Google Astra, I think people are going to be more used to it. People's behaviors and expectations will change. Whether that's something that is going to happen now or in 5 years, it's probably in that range.

We adapt to new technologies all the time. When the Ring cameras came out, that was quite controversial. But now people understand that a lot of people have cameras on their doors. We're in that transitional period for sure.

I'll press on the privacy issue, because that's the number 1 thing that everyone talks about. Obviously, I think in Silicon Valley people are a little more tech-forward and experimental, whatever. But you want to go mainstream. You want to sell to consumers, and we have to worry about this stuff.

The baseline question—the hardest version of this—is the law. There are 1-party-consent states where this is perfectly legal. Then there are 2-party-consent states where it's not.

Yeah, the EU is a totally different regulatory environment. But in the US, it's basically on a state-by-state level. In Nevada, it's single-party; in California, it's 2-party.

But it's kind of untested. There are different laws depending on whether it's a phone call or whether it's in person. In a state like California, anytime you're in public, consent doesn't come into play, because the expectation of privacy is that you're in public.

But we process the audio, and nothing is persisted. Then it’s summarized with speaker identification focused on the user. Now, it’s kind of untested legally—and I’m not a lawyer—but does that constitute the same as a recording? So it’s kind of a gray area and untested in law right now.

I think the bigger question is, because if you had your Ray-Bans on and were recording, then you have a video of something that happened. That’s different from having an AI give you a summary focused on you that’s not really capturing anybody’s voice. I think the bigger question, regardless of the legal status, is: What’s the ethical situation with that? Even in Nevada—or many other U.S. states where you can record everything and don’t have to have consent—is it still the right thing to do?

The way we think about it is that we take a lot of precautions not to capture personal information from people around you, both through speaker identification, through the pipeline, and then through the prompts and the way we store the information, to be really focused on the user. We know that’s not going to satisfy a lot of people, but if you do try it and wear it, it’s very hard for me to see anything that somebody wearing a Bee around me could capture that I would ever object to as a third party.

We’re in this transitional period where the expectation will become more normalized that it’s an AI. It’s not capturing a full audio recording of what you said; everything is fully geared toward helping the person understand their state and providing valuable information to them, not logging details about people they encounter.

You know, I’ve had the same question with Zoom meeting transcribers. I think there’s kind of a personal impact. There’s a Fireflies.ai recorder, and I just know that it’s being recorded. It’s not like I know whether I’m going to say anything different, but intrinsically, you kind of feel different because it’s not pervasive.

I’m curious, especially in your investor meetings, whether people feel differently. Have you had people ask you to turn it off in a business meeting and not record? I’m curious whether you’ve run into any of these behaviors.

What’s funny is, on my end, I wear it all the time. I take my coffee at Blue Bottle with it, or I work with it. Obviously, I’m working on it, so I wear it all the time. So far, I don’t think anybody has asked me to turn it off.

I’m not sure whether it’s because they’re really friendly with me or because they know I’m working on it, but nobody really cared.

This is because you live in San Francisco.

Actually, I’ve been in Italy as well, and Italy is super privacy-concerned. Europe is super privacy-concerned. Again, nothing. I don’t know. That, for me, was interesting. Nobody’s ever asked me to turn it off, even after giving them full demos and disclosing.

I think some people have said, “Well, in a personal relationship, my partner was initially kind of uncomfortable about it.” We heard that from a few users, and that was more in a personal relationship situation. The other big one is people say, “I do like it, but I cannot wear this at work,” because they think they’ll get in trouble based on policies.

If you’re wearing it inside a research lab or somewhere where you’re working on things that are sensitive, we’re adding certain features like geofencing, so you can set it to never be active at a particular location. We’re also looking at context fencing, so you can say, “If these topics come up, don’t capture anything.”

I’ve often explained it the other way: Maybe you only want it at work. So you never take it from work, and it’s just a work device, like your Zoom meeting recorder is your work device.

Professionals have been big early adopters. You say San Francisco, but our daily shipments of over 100 are going to addresses in Texas, which I think is our biggest state, and Florida—just the biggest states. A lot of professionals talk for a living, and we didn’t set out to build it for that use case, but I think there’s a lot of demand from white-collar people who talk for a living. We’re just starting to talk with them. I think they want to be able to improve their performance, understand what they were doing, and figure out how they can do better.

How do you think about Gong.io and some of these sales-training tools where you put on a sales call and then it coaches you through it, which are more verticalized, versus having a more horizontal platform?

I’m not super familiar with the space. Like I said, we weren’t building for that use case, so it’s kind of a surprise to us. But I think those are interesting. I’ve seen there are a bunch of them now, right? It kind of makes sense.

I’m terrible at sales, so I could probably use one, but it’s not my job fundamentally. Maybe in a lot of situations, it could be useful. We’ve also heard from people with restaurants that, if they’re able to understand whether they’re doing well, that can be valuable.

In general, I think a lot of people like to have that double-check: Did it do this well, or can you suggest how I can do better? We had a user who said he used Bee for job interviews and then asked Bee, “Actually, how do you think my interview went? What should I do better?” I like that. I’m like, “Oh, that’s actually a personal coach in a way.”

But I guess the question is: Do you want to build all of those use cases, or do you see Bee more as a platform where somebody’s going to build the sales coach that connects to Bee, so you’re kind of the data feed into it?

I don’t think this is just a data feed. It’s more of an understanding engine. Definitely in the future, having third parties use the API and build out all the different use cases is something that we want to do, but the initial case we’re trying to build is that layer for all of it to work.

We’re not trying to build all those verticals, because no startup could do that well. But it’s been fascinating to see. I’ve done consumer for a long time, and consumer is very hard to predict. It’s hard to know what’s going to be the killer feature. We really believe that’s the future, but we don’t know exactly what process it will take to gain mass adoption.

The killer consumer feature is whatever Nikita Bier does. Social apps for teens. Yeah, well, I like Nikita, but he’s good at building bootstrap companies and getting them very viral, then selling them, and then they shut down.

Okay, so you just came back from CES?

Yeah. Crazy. It was my first time in Vegas and my first time at CES. Both were overwhelming.

First of all, did you feel like you had to do it because you’re in consumer hardware?

We decided to be there and have a lot of partner and media meetings, but we didn’t have our own booth, so we decided to skip that. But we decided to be there and have a presence, even just us, and speak with people.

Hard to stand out.

Yeah, I think it depends on what type of booth you have. I think if you can prepare a really cool booth, it can be pretty cool.

Have you been to CES? It can be pretty cool. It’s massive. It’s like 80,000 or 90,000 people across the Venetian and the convention center. To me, I always wanted to go, even just as a fan of—

Yeah, you wanted to go. Growing up, I think CES kind of peaked for a while, and it was like, “Oh, I want to go there. That’s where all the cool gadgets, everything, is.”

There are a lot of cool vacuums and pet tech, and dog stuff.

There’s a lot of robot stuff, new TVs, and new cars that never ship.

I think this time last year was when Rabbit and Humane launched at CES, and Rabbit kind of won CES. Now, this year, there were no wearables except for you guys.

It’s funny because it’s obviously AI everything. Every single product—like, a toothbrush with AI. We saw a hair blower, literally a hair dryer with AI. That was cool.

But another difference around us is that we didn’t want to do a big, overhyped, promised kind of Rabbit launch. I mean, hats off to them on the presentation and everything, obviously, but we wanted to let the product speak for itself and get it out there. We were really happy with the interest we got from the media and some of the partners there, so it was definitely worth going.

I would say that if you’re in hardware, it’s just about how you make use of it. To do a big Rabbit-style launch or have a huge show there, you need to plan that 6 months in advance, and it’s very expensive. But if you go there, everybody’s there. All the media are there, and there are a lot of pre-show events where it’s great to talk to people in the industry.

All the manufacturers and suppliers are there, too, so we learned about some really cool stuff. We met with them; they have thermal energy capture.

And it's like, oh, could you maybe not need to charge it because they have a thermal that can capture your body heat.

What?

Yeah, they're here. They're actually here in Palo Alto. They have a Fitbit thing that you don't have to charge.

How much power can you get from that? What's the power draw for this thing?

It's more than you could get from the body heat, it turns out, but it's quite small. I don't disclose the technical details, but I think solar is still more realistic. They also have one where the face of it is just a solar cell, and that is more realistic. Or kinetic. Kinetic, apparently—they seem to think it wouldn't be enough. Kinetic capture is quite small, I guess.

Well, I mean, watchmakers have been powering things with kinetic energy for a long time. We don't have to talk about that. I just wanted to get a sense of CES. Would you do it again?

I definitely would. Okay, you're just a fan of CES. From a business point of view, it doesn't make sense.

I happen to be in the conference business, right? So I'm kind of just curious.

Yeah, so I would say that, without the booth and with really straightforward conversations that were already planned, 3 days was okay. I think it was okay. But if you need to invest in a booth, that's not cheap.

Which is how much?

A 10-by-10 is $5,000. But on top of that, you have the financial costs. A 10-by-10 is like one of the super-fancy things, and some companies have—I think we'd probably be more in the 6-figure range to get a booth.

I mean, I think that, yeah, it's very noisy. We heard that it's very, very noisy. Obviously, everything is being launched there, from cars to cell phones, so it's hard to stand out. But I think going in with a plan of who you want to talk to was worth it. We had a lot of really positive media coverage from it, and we got the word out, so I think we accomplished what we wanted to do.

I mean, there's some world in which my conference is kind of the CES of whatever AI becomes.

Yeah, I think that—don't do it in Vegas. Don't do it in Vegas. That's the only thing I didn't really like there.

Okay, that's great. Amazing. Those are my favorite ones. You cannot fit 90,000 people in San Francisco.

That's really the problem. You need to do multiple locations. You can do Moscone and then have one—

That's what the Salesforce conference is. GDC is how many?

That might be 50,000, right?

Okay, form factor, right? My way to introduce this idea was that I was at the launch in Solaris. What's the old name of it? Newton? Of Tab when Avi first launched it. He was like, “I thought through every form factor. Pendant is the thing.” And then we got the pendant for the original one—the first one, which was a pendant—and I took it off and forgot to put it back on.

So you went through pendant, pin, and now bracelet, and maybe there are AirPods—a sort of earphone—in the future. What was your iteration to that?

Yeah, so we had, I believe, 3 or 4 iterations, and one of the things that we learned is that people don't like the pendant. In particular, women don't want to have anything here on the chest because maybe they have another necklace or other stuff.

You just ship a premium one that's gold. We're talking about some fashion—some big fashion. There's something there. This is where it helps to have an Italian.

Exactly. Some big Italian luxury. I can't say anything, sorry.

The bracelet actually came from the community because they were like, “I don't want to wear anything as a necklace or a pendant.” Also, the one that we had—I don't know if you remember—was a circle and was really bulky. People didn't like it.

And I actually don't dislike it. We were running fast when we did that. Our thing was that we wanted to ship them as soon as possible, so we weren't overthinking the form factor or the material. We just wanted to be out.

But after the community organically told us, basically, all of them were like, “Why don't you just do the bracelet? It's way better. I will just wear it, and that's it.” So that's how we ended up with the bracelet, but it's still modular. I still want to play around with the fact that it's modular, and you can take it off and wear it as a clip. In the future, maybe we will bring back the pendant, but I like the fact that there is some personalization.

Right now, we have 2 colors, yellow and black. Soon we will have other ones, so we can play a lot around that.

I think the goal for the form factor is for it to be not super invasive, right? Something that's easy. In the future, smaller and thinner—not like Apple's obsession with thinness, but it does matter, the size and weight. We would love to have more context because that will help, but to make it work, I think it really needs to have good power consumption and good battery life.

With the Humane, swapping the batteries—I have one. I think the Humane is pretty incredible, some of the engineering they did, but it wasn't geared toward solving the problem. It was just too heavy. The swappable batteries are too much to manage—the heat, the thermals. It's too much for a light-interface thing.

Yeah, like that. That was cool.

It's cool. It's cool, but if you have your hand out here and you want to use your phone, it's not really solving a problem because you know how to use your phone. It's got a brilliant display, but you have to learn how to gesture with this low-resolution laser. The laser is cool—the fact that they got it working in that thing, even though it did overheat—but it's too heavy, too cumbersome, and too complicated with the multiple batteries. So something that's power-efficient and thin, both in the physical sense and in the edge-compute kind of way, so that it can be as unobtrusive as possible.

Users really like it. I like when they say, “Yes, I like to wear it and forget about it,” because I don't need to charge it every single day. On the other version, I believe we had 35 hours or something, which was okay, but people just prefer the 7-day battery life.

Oh, this is 7 days?

Yeah.

Oh, I've been charging it every 3 days.

Yeah, you can keep it full for 7 days.

The other thing that occurs to me is maybe there's an Apple Watch strap.

Yeah.

So that I don't have to double-watch. I have—

Yeah, that's the other one.

Yeah, I thought about it. I also saw the ones that you can put back on the phone. There are a lot of possibilities.

There's a competitor called PLAUD. It's not really a competitor; they only transcribe, right?

Transcribe, but they're very good at it.

Yeah, no, they're great. Their hardware is really good, too, and he just launched the pin, too.

Yeah, I think the MagSafe kind of form factor has a lot of advantages, but also some disadvantages. You can definitely put a very large battery on that, and so the power consumption isn't as much of a concern. The downside is that the phone is in your pocket. I think form factors will continue to evolve, with more sensors, less obtrusiveness, and easier use. We have a new version.

Okay, looking forward to that. Whenever we launch this, we'll try to show whatever, but I'm sure you're going to keep iterating.

Last thing on hardware, and then we'll go on to the software side, because I think that's where you guys are also really strong. Vision—you wanted to talk about why no vision?

Yeah, I think it comes down to the fact that when you're a startup, especially in hardware, you work within the constraints. Vision is super useful and super interesting, and it's what we actually started with. There are 2 issues with vision that make it not the place we decided to start.

One is power consumption. You have to trade off your power budget. Capturing, even at a low frame rate, and transmitting over the radio actually takes up the majority of the power. So you would really have to have an unacceptably large and heavy battery to do it continuously all day. We have some novel alternative ways that might allow us to do that, and we have some prototypes.

The other issue is form factor. Even with a wide field of view, if you're wearing something on your chest, it's obviously not going to capture the field of view of what's interesting to you. The wrist isn't really that much of an option, and if you're wearing it on your chest, it's often not going to capture what's interesting to you. That leaves you with your head and face.

Anything that goes on the face has to look cool. I don't know if you remember Snap Spectacles. That was kind of like the first—

Yeah, but they weren't very successful, and I think one of the reasons is that they were so weird-looking. Their camera was so big on the side.

If you look at the Meta Ray-Bans, where they're way more successful, they look almost indistinguishable from a Ray-Ban. Meta invested a lot into that, and they have a partnership with Qualcomm to develop custom silicon. They also have a stake in Luxottica now, so they're coming from all angles to make glasses.

I don't know if you know Brilliant Labs. They're a cool company. They make Frame, which is a cool, hackable pair of glasses, and they're really good on hardware.

But even if you look at the frames, which I would say are from the most advanced kind of startup, there was one that launched at CES, but it isn’t shipping yet. The one that you can buy now still isn’t something you’d wear every day, and the battery life is super short. So I think the challenge of doing vision right off the bat would require quite a bit more resources. Audio is such a good entry point, and there’s also the privacy around audio. If you had images, that’s another huge challenge to overcome. Ideally, personally, I would have all the senses, and we’ll get there.

Okay, one last hardware thing, because I have to ask this before we move to software. Were either of you electrical engineers?

No, I’m CS, and I’ve taken some EE courses, but prior to working on the hardware here, I had done a little bit of embedded systems—very little firmware—but we luckily have somebody on the team with deep experience.

Yeah. I’m just like, you have to become hardware people. I learned how to worry about supply chain, power, and radio. I would say this about hardware—and I know it’s been said before—but building a prototype, learning how the electronics work, learning about firmware, and developing this is fun for a lot of engineers, and it’s all totally achievable, especially now with the tools we have. Stuff you might have been intimidated by, like, “How do I write this firmware?”—now, with Claude Sonnet, you can get going and actually see results quickly. But I think going from a prototype to actually making something manufactured is an enormous jump, and it’s not all about technology: supply chain, procurement, regulations, costs, and tooling.

The thing about software that I’m used to is that it’s funny: you can make changes all along the way and ship them. But when you have to buy tooling for an enclosure, that’s expensive. Own tooling?

You have to.

Don’t you just subcontract out to someone in China?

We make the tooling? No, no. You have to have CNC and a bunch of machines. Nobody makes their own tooling, but you have to design it and submit it, and then they go 4–6 weeks later. If there’s a problem with it—

Right.

Well, then you’re not making any of your enclosures.

What resources or websites are most helpful in your manufacturing journey?

I think it’s different depending on the product, because hardware is so specialized in different ways. I would say that, for example, you should choose a manufacturing company and speak with other founders. They’ll give you the language and some tips about who is good and who is not, who specializes in something versus somebody else. Some people are good in plastics, and some people are good in PCBs. For us, it really helped at the beginning to speak with others and understand who is around. I worked in Shenzhen and lived almost 2 years in China, so I have an idea about different hardware manufacturers and all of that. Soon I’ll go back to check things out. I think it’s good also to go in person and check and see the right people.

We did some stuff domestically, and if you have that ability—the reason I say ability is that it’s very expensive—you can build out some proofs of concept and do field testing before taking it to a manufacturer. Despite what people say, there’s really good domestic manufacturing for small quantities, at extremely high prices. We got our first PCB and assembly done in LA, because the defense industry supports a lot of good manufacturing that can do quick turns. It’s like, we need this board, we need to find out if it’s working, we have this deadline, and we want to start; if you want to have it done and fabricated in a week, they can do it for a price. But I think everybody’s trending, even for prototyping, toward moving offshore now, because in China you can do prototyping and get it within almost the same timeline. The thing is, with manufacturing, it really helps to go there and establish the relationship.

Yeah, my first company was a hardware company, and we did our PCBs in China. It took a long time. Now things are better, but this was, I don’t know, 10 years ago or something like that. I’ve heard this too: if it’s something where you don’t have the relationships, they don’t see you, they don’t know you, you might get subcontracted out, or they’re not paying attention. But if you have the relationship and a priority, it’s usually good.

We ended up doing the fabrication and assembly in Taiwan for various reasons, but it really helped that you went there at some point.

Yeah, we’re really happy with the process. But the whole process of choosing the right people—choose the right people—but also just sourcing the bill of materials and all of that stuff: I guess if you have time, it’s not that bad, but if you’re trying to really push the speed, it’s incredibly stressful.

Okay, we’ve got to move to software. The hardware may be hard for people to understand, but what software people can understand is that running transcription and summarization, all these things, in real time every day for 24 hours a day, isn’t easy. So you mentioned 200,000 tokens for a day. How do you make it basically free to run all of this for the consumer?

Well, I think that the pipeline and the inference—people think about all these tokens, but as you know, the price of tokens is dramatically dropping. You guys probably have some charts somewhere that you’ve posted, and if you see that trend, 250,000 input tokens isn’t really that much, right? The output layers?

You do live?

Yeah. Speech-to-text is the most challenging part, actually, because it requires real-time processing and then later processing with a larger model. One thing that’s fairly obvious is that you don’t need to transcribe things that don’t have any voice in them, right? Good voice activity detection is key, because the majority of most people’s day isn’t spent with voice activity. That’s the first step to cutting down the amount of compute you have to do, and voice activity detection is a fairly cheap thing to do—very, very cheap.

The models that need to summarize don’t need a Claude Sonnet-level model. You do need a Claude Sonnet-level model to execute things like the agent, and we’ll be having a subscription for features like that because, although now with DeepSeek R1, we’ll see—we haven’t evaluated it. There are already models that can perform at that level. I was going to say in 6 months, but—

DeepSeek R1?

Yeah, I mean, not that one in particular, but they’re already there and can perform at that level. Self-hosted models help with the things where you can.

So you’re self-hosting models? You’re fine-tuning your own ASR?

Yes. I will say that I see everything trending down in the future, although I think there might be an intermediary step where things become expensive, which we’re really interested in because the pipeline is very tedious and requires a lot of tuning. That’s brutal because it’s just a lot of trial and error. Wouldn’t it be nice if an end-to-end model could just do all of this and learn it? If we could do transcription with an LLM, there are so many advantages to that, but it’s going to be a larger model and hence more compute. We’re optimistic that maybe we could distill something down, and we kind of focus more on reducing the cost of the existing pipeline or trying the next generation, because it’s very clear that all ASR, all speech-to-text, is going to be pretty obsolete pretty soon. Investing in that is probably a dead end because it’s just going to be obsolete.

It’s interesting. I think when I initially invested in Tab—this shows you how wrong I was—I thought, “Oh, this is a sort of razor-and-blades model where you sell cheap hardware and make up a subscription, a monthly subscription.” Now I just checked: Friend is a 1-time sale, $99; Limitless is a 1-time sale, $99; these guys are a 1-time sale, $49. And Friend is free? What? When you probably invested, how much was 1,000,000 input tokens at that time, and what is it now? It’s a fascinating business, and there’s a lot to dig into there, but just getting that perspective out there is not something that people think about a lot. You obviously have thought a lot about it.

What about memory? I think this is something we go back and forth on: you’re just memorizing facts and then understanding what is a preference, and adjusting the facts that you think about a person? Any learnings from that? I know there are a lot of open-source frameworks now that do it. Did you build all of your own infrastructure internally?

Yeah, we did. I evaluated and used a lot of them in other projects. I think there are a few different tasks or things that revolve around memory. One is retrieval, obviously, and when you need to find something—even if you have a large corpus—how do you find it? I think existing RAG pipelines will probably also be obsoleted. The frameworks—I have not found one. There’s no general way to do RAG that works; it’s really highly dependent on the data.

So, if you're going to be customizing something that much, you get more bang for the buck from designing it all yourself. A lot of these frameworks are great for getting going quickly, but I think it's really interesting when you're trying to do memory for a person, because memories decay, right? I'm going to London, then I come back; I'm not going to London anymore.

What we've learned is that doing traditional embedding and RAG is suboptimal. We built our own using small models to do massively parallel retrieval, which I think is going to be more common in the future. To represent a person, we still require some human loop. This is an ongoing project, and we're learning every day. How do you correct the model when it gets something wrong about you?

Right now, we have things that are super confirmed—ground truth about you because a human accepted it—but ideally that step wouldn't be necessary. Then we have things that are fuzzier. The more stuff that we know is true, the more accurate we are when we're trying to decide whether this fuzzy stuff is true, because if you have the context, it's probably not true.

So, I think one of the core challenges is how to handle both retrieval and modeling, especially when you're dealing with noisy source data. Even in an ideal world, if you had perfect transcription and were going off that, that's still not enough information, right? Even if you had visuals, it's still not enough. There's still going to be some misunderstandings, so how do you not let that damage the value of it and make it unrecoverable and uncorrectable?

Yeah, one way I think about it is that I usually like to order the same thing from the same restaurant if I like it, but I'm not saying that out loud. Are these the types of behaviors that you can capture? When you ask about a favorite restaurant, I would want it to give me restaurants I've already been to and liked. Or if I'm like, “Hey, just order something from this place,” it should reorder the same thing because it knows that I like to get the same thing again. But I feel like today, most agent memory things that I see people publish are just, you know, write down the data thing.

Yeah, I think that's why the reasoning—in our case, giving it time to consider all of the sources it has—is really important. Look at the emails, see the receipts, and then look at the conversations to see what I've mentioned. Then be able to take enough time to search through all the context and connect the dots is really important. I don't know; some of the agent memory stuff is like key-value with RAG on top, and the results are just not complete enough when you have a growing corpus, managing decay, and hallucinations that might be in the source material.

So, this is where people usually bring in knowledge graphs.

Yes.

And do you use them? Is it for speed, or what are the issues?

We don't extensively use knowledge graphs. It's something—we didn't talk also about the potential future social aspects—but the problem with knowledge graphs that we found is that they're great for representing the data, but using them at inference time is challenging. It's just the LLM understanding the graph input. It's not in the training data, for sure. I think the graph is the right kind of way to store the data, but then you need to have the right retrieval and format it in a way that doesn't overwhelm or confuse what you're trying to do.

Should we ask about social?

Yeah, I know. I thought you were going to go into it.

Yeah, what's that?

Not directly related to graph retrieval or graph knowledge bases, but the idea is that you have your personal context, and then other people can query it. It can divulge some things that you would have full control over. Then Maria and I are trying to negotiate where we're going to have dinner. There can be an exchange between the agents.

Yeah, there can be an exchange between the agents.

So, how can my agent speak with Ethan's agent? Both of them know our location, what we like, and where we went in the past. Even if we have our calendars integrated, they know when we're free. They can interact with each other, have a conversation, and decide on a place to go for us.

Wow.

And we did that. It was really cool for me because they suggested a nice French restaurant that we went to in the end.

That you'd never been to?

That we'd never been to. They saw that we both liked French food, and we were both in Pacific Heights. This was really trivial.

Yeah, it's a trivial toy use case, but I guess, in terms of having used it for a while, if I wanted to buy you a gift—

Oh my God, you bought me a bunch of candles, now that I think about it. This is another use case. I was like, “Yeah.”

When we were testing the agent, a bunch of candles from Amazon showed up at her door.

Yeah, because I really love candles, but I didn't expect 20.

Yeah, it's a lot—extreme, but how do you manage that? What's okay for your Bee to divulge, and to whom? Shouldn't you get an authorization request every time for personal context?

Yeah, yeah, yeah. A human would have to sign off on it, but then I wouldn't have to guess.

There's this culture that's very alien to everyone else outside of San Francisco, and outside of the Gen Z bubble in San Francisco, which is sharing location. I can tell exactly where my close friends are right now in the city. It's normal here, and it freaked out everyone who's not here.

Yeah.

So maybe we can share preferences—who we like, or even small updates about your day. My parents would love that because I don't do that.

Yeah, so now there's no friction. It can just be more or less automatic.

Dating? I was always trying to avoid dating as a startup founder.

Everyone hates it?

We thought about it. Sometimes people ask us, “Oh, you know so much about me. Can you measure compatibility with somebody else or something like that?”

Yeah, probably there is a future, and maybe somebody should build that.

I think on our end, we were like, “No, this is—I’ll build on your API.”

My sister's actually a personality psychology professor, and she studies personality. We were at Thanksgiving with my parents, and I was like, “Give me my Big Five,” which is the personality type. “Does it know my Big Five?” You just ask it to consider everything and give you your Big Five. My sister said it was pretty accurate. I didn't agree with it because it said I was disagreeable, but she seems to think I'm agreeable.

You disagree that you're disagreeable?

Yeah. What other proof do you need, then? I think I'm very agreeable.

But I think we did get some users who were like, “Oh, if we're a couple—” We had couples actually buy the product together. Both members of a couple bought our hardware, so there is something there.

Another test is the Myers-Briggs. I know you don't like that one.

No, no, no. OCEAN is cooler than Myers-Briggs. Everyone stop using my MBTI. Use my OCEAN.

For me, it was on point every time.

Awesome.

Anything else that we didn't cover? Any cool underrated things?

Go to B. dot computer $49.99 and buy the device. That's the call to action. You're hiring? We are hiring—for sure, AI engineers. Nice. What is an AI engineer? Somebody who's scrappy and willing to work with us. I think you coined the term, right? So you can tell us. People have different perspectives, and what is useful for you is different from what is useful for me.

So, anyway, it's useful. I think always-on AI is really going to explode. There's going to be a lot from startups, but incumbents, too, and there are going to be all kinds of new things that we're going to learn about how it's going to change all of our lives. I think that's the thing I'm most certain about. And be in the AI. Well, thanks very much. Yeah, this was a pleasure. Yeah, we'll see you launch whenever that launch is happening. Yeah, thanks. Thank you.