The general population does not care about being 10% more efficient. I have a hot take thesis that this Muse charm is actually less about trying to win the hardware game and it's more about data collection in the real world to fuel Zuckerberg's future metaverse of mapping out the actual world.
We were joking internally like we're days away from an agent messaging someone saying, 'I noticed you weren't that into her, so I went ahead and broke up with her.'
There is massive defensibility around proactivity. I was biking to work. I want to get stuff done and I just start talking to ChatGPT voice and throughout my 30-minute bike ride to work, categorized my emails, submitted them to different labels, sent out calendar invites, got to the desk, inbox zero. But it will be a very fine line because if you cross that line once, you lose all trust with your user.
It does feel like there's going to be infrastructure and products that exist only for agent-to-agent interactions that doesn't exist today. What do you think?
It is inevitable that we're going to see...
All right, welcome to the A16Z show. I'm so psyched to be here with my friend David Pawlan. We hung out a couple of weeks ago. We're both so enthusiastic about everything that's been happening with personal assistants and consumer AI generally. And man, so much has happened in the last couple of weeks. Maybe to tee it up, you know, I sort of think that there's been a few moments where some part of the ecosystem has sort of seen God. And ChatGPT was a big one. And you know, I think for a lot of folks in November '22 and a bit of '23, it was like, what is this thing that can, you know, write emails and poems and, you know, start to have a conversation back and forth and feel like you're interacting with a synthetic person? That was of course the beginning, the big bang. The second moment really was coding agents and everything that developers have been obsessed with around sort of Claude Code and Codex and the ability to trivially create software. And then finally, most recently with Instinct, Muse, and ChatGPT work, we've had this extraordinary enthusiasm around personal agents and what they can do for us as consumers. So, we're here to talk about all things personal agents. We're going to go as deep as we can. We're going to assume that everybody has seen many of the same things that we have seen and we're going to make this the first in a series of conversations. Welcome, David.
Thank you. Thanks for having me. Excited to be here. Let's like zoom out and maybe talk about, you know, the arc of personal agents over the last few months, the last few weeks. What have you been seeing? What's your sort of view on things?
Yeah, so this space has exploded over the past four weeks. We can roll it back. I was an early adopter of Poke. So I want to say Poke, which was kind of one of the first like consumer AI agents to really hit the market. Yes. That launched I believe it was September 8th and then I started using it September 11th, like 3 days later. And like I remember it so vividly because it was just an unbelievable experience. Their whole gimmick was this, you know, negotiation with the agent on how much you're actually going to pay. And it totally blew my mind. Then comes late November and Open Claw gets released and that kind of takes the actual world by storm of like there is something coming. And then a few others have popped up here and there, but most recently we saw Instinct really blow up the tech Twitter bubble. Just absolutely rode a wave, went nuts, and then it was just dominoes. You have Grokbot, you have Muse, you have a bunch of longtail ones like Caddy or Ali or Pi. And it's really been amazing to see, you know, this whole tech bubble is just exploding with interest and like an intellectual curiosity for what can these agents do for us on a day-to-day basis. And there's, I think everyone's kind of in the same boat where nobody really knows, like it's the wild wild west. We're just playing around here.
I totally agree. And we should take a moment to talk about Assistant Bench, which is this new, you know, really important and discussed benchmark that you're the author of. Maybe tee up Assistant Bench a little bit and then I'll ask you a few questions.
Assistant Benchmark is a site that essentially compares all these AI assistants on a use case basis. So we're asking all of these assistants the same prompt: book me a flight to Chicago, find me a vegetarian restaurant within five blocks of my specific location, etc., etc. And we're comparing the actual outcome on a one-shot prompt. Does it actually work? Does it ask follow-up questions? How does it perform across 16 different dimensions? And we compare those all. So, we use the word benchmark. Think of it more as like a consumer-facing benchmark, not necessarily technical behind the scenes, but if I want to book a flight, which one should I actually use? Who is performing the best? How fast are they actually replying? Got into this whole comparison because I was just doing market research myself trying to figure out how are these all comparing against one another? What are the gaps? And published a bit of that and it was almost two weeks ago. It's 16 days ago today that we launched it and it's just absolutely exploded. It's, you know, over 100,000 views on the website. The website has had over 100,000 visitors, have had outreach from basically every single founder. And I think it really just goes to show like everyone's curious. People are looking for answers to where this industry is moving.
And so talk us through, David. That's incredibly exciting. It's cool to actually... It's funny how benchmarks have become such a staple now of the conversation because there's so many launches, it's hard to try everything, which is also an exciting signal. Maybe talk us through everybody who's not Instinct and Muse. We're definitely going to deep dive on those later, but who else do you think is doing interesting work and what are the kind of directions of specialization that you're seeing?
Totally. So, we've got a bunch of longtail solutions out there. I think I would categorize it into two groups. You have like the B2C consumer-facing generalized agents. That's like the Instinct or Muse. Some of the longer tail ones, you have Caddy, you have Ali, you have Season. There's I believe 64 right now just in the general category on the site and then we can break that down even further into a little more specialized within that. So you can have travel-specific ones like Soore or Miso. You can have email-specific ones. The other side of the spectrum is going to be more of like your B2B workflow type assistants. And those are, you know, anyone from the town to Catch to Vellum. And they're all essentially doing the same thing, right? They're your assistant that's actually like more of like an executive assistant, not the consumer-facing approach to it. And that's kind of like the general landscape at the moment. So far to my knowledge there are 122 different ones.
Across all the categories.
Incredible. And what about areas of concentration? Where is the most sort of interesting competitive focus?
So everyone right now by and large is focusing on the generaliz...
Just trying to do everything.
And when you think about something like travel, just to pick on that, that's an infrequent high-value behavior. Do you think that that category ends up being organized more as a connector into the horizontal agent or do you think they have the right to win as a general horizontal agent as well?
I think you're going to maybe see potentially one winner as an independent agent, but by and large I really think it is going to be more of a connector. So I have like seven or eight different group chats now with just individuals obsessed with assistants talking. There's now over 1,200 people across all of the group chats just chatting over and over and over. Travel is the fourth most talked about use case, which is really interesting because when you look at like virality, it tends to be the number one that people are talking about of it checked into my flight for me, I was able to book my flight. But then you actually look at the day-to-day conversations of what people are really using it for and most curious about. To your point, people aren't traveling every day. So, it's...
It's more of a gimmick, right? It's an attention grabber.
What are use cases one to three? So, number one is more just daily admin type stuff, whether it's cleaning out your inbox, whether it's filing forms for you. The things that are really annoying and I think that gets to the broader thesis of where I think the winner of the space will go that we can, you know, touch on. The second is going to be more agent orchestration, which is really interesting. I think these group chats are a little biased as the third one is development focused. Okay, these are all group chats with people in tech Twitter, right? So, it's not necessarily representative of the world at large. Yeah, but people are talking a ton about agent orchestration and how do you find the most efficient agent or how let your agents talk to each other. Development agents, how does this now translate to this coding world on your phone? Whether it be, you know, using Claude Code from a text component or even take it one step further of, okay, if I'm a designer, can I now make edits to my design files by just like texting my agent of like little small iteration changes?
And then the last category being finance. I think that's the most fun one to look into because it is in my opinion the least practical. Like there's so many decisions and right like if you could have an agent make a billion dollars for you all the hedge funds would be doing it. It's not something that like everyone's going to succeed at with their own phone. But it is the most interesting that you give someone an agent, a powerful tool right in their hands and the first thing they think is okay how are you going to make me a million dollars?
Yeah. Yeah. It's really interesting because one of the critiques, you know, Ben Thompson was on TPN a few days ago talking about how most consumers aren't looking for productivity in their lives. They're looking to spend time not save time, which I agree with, but it does feel like finance is an area in which there is a ton of administrative overhead for the average consumer. And if you look at, you know, the number of financial markets or financial sort of profit pools that are defined by consumer apathy or being uninformed, that's a lot of profit dollars that could be delivered back to the consumer, right? So I do think that while I generally agree with Ben that most consumers are looking for entertainment and spending time, I think there are sort of defining aspects of the US consumer's life that are, you know, heavily administrative that these agents can really be a breakthrough for.
Totally. And I think that kind of touches on the general thesis to where I think the winners will sit. And that is that like to Ben's point, I agree. I think the general population does not care about being 10% more efficient.
You hire an assistant to allow it to be invisible, right? Like the best employee I've ever had are the ones that are just doing their work, getting things done, they update you like here, here's what I did, and you're like, 'Wow, amazing.' You don't hire someone because you enjoy managing them.
And the best, most proactive use cases I'm seeing, I think, do fall in the finance category. And I think we're going to see an explosion in a world though where it's not cost makers, it's cost savers. And so it's going to be workflows where like some cool use cases I saw people are using it to go through their past year of receipts and file HSA reimbursements.
Or one that I do is anytime I book a flight, if the flight price drops, I then have my assistant hit up the airline and get me a travel credit.
And I have a friend, this is my favorite use case I think I've seen yet. He hooked up his Grokbot to his sprinkler system at his home and connected it to the weather system.
And so now his sprinklers are dependent on the weather and it's dropped his water bill by 50%.
And so I think these like invisible type agents that are going to be doing these cost-saving type workflows fall within that finance category, but they're really going to lean into these like cost-saving activities that create a world that is a lot more consumer aligned.
Yeah. It'll be interesting too like I think the American consumer nobody wants to hear about saving money. People want to hear about spending money and free money.
And I actually think the hook that's going to get a lot of people here is it's going to feel like free money. Like going back through and, you know, submitting HSA sort of refund requests is going to be free money. You know getting all of your airline tickets where the price has dropped, you know, to refund you the delta is going to be free money. So I think that that's a hugely powerful hook.
It's also really really interesting because this is a sort of infrequent but high-value behavior and that's where my head goes on both travel and finance like maybe these things will be really powerful, you know, billion-dollar businesses that are still vertical connectors rather than general horizontal agents.
Totally. It's going to be a very interesting world in terms of the application phase which I think then, you know, that also something that we talked about last time we went for the walk was okay here's what it can do but now how do you actually interact with that agent. And something that we chatted on, I'm curious your thoughts. Right, so like we both were in the world of iMessage, it's where we live, it's easier for us. You mentioned your wife and her friends prefer a separate application. Yeah. And so like where do you see this world moving? Do you believe that multiple surfaces can exist as a winner or there's going to be one dominant lane?
Yeah, that's a great question, man. I mean, I think this is why there's so much consumer preference that's going to matter here because I've heard the argument that for folks that are Gen Z, they're more interested in just texting and for folks that are millennial or Gen X, they're interested in app interfaces. So, I think part of it may fragment by generation, part of it may actually fragment by gender as well to some extent. I mean, it's a generalization, but the types of things I'm doing versus the types of things my wife is doing, you know, she's big on building a dream board and trying to achieve all of her one-year goals and her lifelong goals. So, she really prefers a visual interface. You know, she does a lot of sort of travel is very visual driven for her, whereas for me, it's much more functional like, hey, how do I just get from X to Y? Ideally, with Starlink on my plane, which works better for me. I feel like iMessage sort of... the iMessage space is more precious. So it feels like a more personal interaction when I'm in that space. And now of course that changes if every second thread is with an agent. But for now I feel like it kind of lives in a privileged place on my phone where, you know, the app on the app grid just doesn't. What do you think?
Yep. I completely agree. I think the personalization that comes with iMessage is just because I already live there, it feels like it's something that's already connected to me. It doesn't feel like an additive workflow.
And I really have only though seen by and large two services appear. It's either the iMessage or a separate application. I've seen a third signal startup actually Sky. They have like an iPhone widget that they're trying to introduce as a different surface which is pretty interesting. But then even yesterday, right, we saw Zuckerberg and the Muse team dropped the Muse Charm. Yes. That's now an entirely different use case for how we interact with an agent. What do you think about hardware?
I think it's fascinating. I mean, a smart person said this, which is that every person who's working on a piece of hardware thinks that their hardware is the kind of be-all end-all. Like the pendant is the ultimate thing, you know, the watch strap is the ultimate thing. So I don't think that that's right. I think again it's going to be very personal preference driven. Just as jewelry is like some people like to wear necklaces and some people like to wear earrings. And I think the same way this kind of form factor is going to manifest is going to be very customer specific. But I think that there's value in having something that is sort of with you and ambient and can capture all the kind of context and requests and, you know, promises and commitments and everything that happens through the day. I mean, I think we're going to have to sort of figure it out together in terms of privacy expectations in society. You know, you go to restaurants in Washington and they take away your phone. Maybe they're going to be taking away your phone and your pendant. But I think that there's actually something there with hardware and I just like that we're being ambitious in that direction. You know, I think after 10 years of startups being discouraged from working on hardware, now we've got a lot of interesting new consumer efforts.
Right? I think it's going to be super interesting. I have a hot take thesis that this Muse charm is actually less about trying to win the hardware game and it's more about data collection in the real world for the Meta team in terms of like how the real world actually functions given it has a camera and multiple microphones and it's primarily... I don't really see it being a mass adoption thing. I think it's going to get adopted by the Twitter bubble and they're going to use it everywhere and it's just going to feed Meta with a ridiculous amount of data to fuel Zuckerberg's future metaverse of mapping out the actual world.
But don't you think that they already get that from the glasses?
I think they do to some extent, but I don't know if I mean, correct me if I'm wrong, but are the glasses always going to be on in the same way that this charm will be because...
I don't think so. Yeah, it's a great point. I think the charm is meant to be more as, you know, as you mentioned, it's an ambient 24/7 companion versus the glasses are more action-oriented.
