By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
EVERY SPOKEN WORD
20 min read · 3,726 words- 0:00 – 2:36
A true second brain
- KMKevin Moody
There's actually this opportunity for information to not need to be organized
- SSSteph Smith
It's not as simple as just plugging into GPT-4 and calling it a day.
- DXDennis Xu
The same search for multiple different people should actually yield different results.
- SSSteph Smith
I want an AI that calls me out on my [beep] .
- KMKevin Moody
I think we're gonna have to add that as a potential knob.
- DXDennis Xu
Does it actually adapt to who you are? Does it learn about who you are? That's where I think the magic really starts to happen.
- SSSteph Smith
Some people spend more time organizing their second brain than actually using it.
- DXDennis Xu
There's a, there's a crazy amount of innovation that's happening in the space right now that, that's, that's just been really exciting to see.
- SSSteph Smith
Since the rise of the computer, humans have been gripped by the idea of having a second brain. But has reality held up to that promise? How many people spend more time organizing their second brains instead of leveraging the information within them? How many people constantly look for a better to-do app instead of actually checking the to-dos that they're organizing? And how many people have an endless stream of data that they'd love to one day process, but they just don't have the right tools or time to do so? Due to the structured, directive nature of computers to date, computers have been a shell of maybe the second brain that we've long hoped for. Still incredibly powerful, but also requiring discrete instruction from the director. But within the last year, consumer AI has shown up to the party, now capable of processing simple language prompts and interfacing with unstructured data, possibly fundamentally changing this game. So today, we chat with the founders of Mem, Kevin Moody and Dennis Xu, as we explore what's gotten in the way of a true second brain, and how AI might actually change the game. And Mem itself is one such company trying to reinvent the space, so they can speak to the why now, but also the very real cost of making this vision a reality. But first, here is a wild stat about just how much is lost in productivity within organizations. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a16z fund. For more details, please see a16z.com/disclosures. I'd love to start out with
- 2:36 – 4:49
Time spent searching for information
- SSSteph Smith
a sta- a statistic, which is crazy in my mind. I couldn't believe this. But it's basically knowledge workers apparently spend 2.5 hours a day, which is around 30% of their time in a workday, searching for information. And this stat is from 2016, but it really wouldn't surprise me if this is still, you know, the reality of working in the information age.
- KMKevin Moody
I almost think that that's probably an, an understatement, the 2.5 hours number. I think that it, it's, it's definitely, uh, still relevant, if not increased since 2016. Um, I think in part because it's so much easier to produce unstructured information, share information, the barrier for creating content and sharing it within your team, um, and being productive as a part of a team has, has gone down. And so you just get more and more information that then your coworkers and, and teams, um, are trying to parse through to get through their day-to-day jobs. But I think something interesting there, you mentioned 2.5 hours per day just searching, is this probably hidden time sink too, which is you don't even know that you are looking for some sort of shared piece of information. You just begin doing your work for the day or for the week, and at the end of that week, you realize you've just reinvented the wheel, that somebody else on your team or some other team in your company did that exact same work, and it already exists in some spreadsheet somewhere or some document somewhere.
- DXDennis Xu
And then I think the question I have in my head is how much time are, like how much time are people actually spending then taking those things and then performing really mundane transformations or, or, you know, tasks on top of what they found? Um, and I, I would imagine that that's kind of, you know, at least another few hours, right? So, um, I think what's really interesting is we're now entering this, this world where a lot of that work, um, and, and not just search, right? Search is kind of just the, the, the beginning of, of really where, um, like where people are gonna get value out of, out of these new technologies. But, um, I, I, I think it's gonna be really interesting to see how the entire workflow of a knowledge worker changes over the next few years.
- 4:49 – 8:46
LLMs and technology unlocks
- SSSteph Smith
Exactly. And I feel like if we kind of go back and look at the arc of knowledge management, like let's truly go all the way back. You get the ability to write, and people were doing this on stone at first. At some point, the printing press arrives. You know, then you have individualized typewriters. The computer shows up at some point. The internet gives you access to almost everyone around the world these days. And so you have this promise of technology which has rung true. It has, you know, given people the ability to, uh, codify and share information and knowledge in, you could say, exponentially better ways. Yet, at the same time, you know, if we bring ourselves back to that stat, it's kind of surprising that there still are just incredible inefficiencies in collaboration. What do you think is preventing us from having this, you know, really streamlined back and forth between, you know, folks within an organization or even as an individual? I feel like a lot of people have this image of a second brain, but ultimately, you know, it, it doesn't quite live up to the hype.
- DXDennis Xu
You know, what's really interesting is all of the things that you mentioned, um, all of these devices in the past, they've actually just primarily been broadcasting devices. If you really think about what is core to collaboration, um Sharing and broadcasting is only one aspect o-of that. It's, it's only one component. The other component is how do you actually, um, digest and how do you actually consume that, that piece of information, right, that, that someone else is actually broadcasting? We've never truly had really hyper-personalized consumption devices. Search is very much kind of a, a personal problem, right? It, it's very much about, "Hey, I'm looking for this piece of information." And the same search query, right, the same search for multiple different people should actually yield different results because you think of, you think about the world differently. Um, and the tools that we've used up to this point really just don't model that reality.
- KMKevin Moody
If we, if we look at it from the perspective of one piece of information and its life cycle, when information in an organization especially is captured, it's oftentimes really hard to know all of the people and all of the contexts in which that one piece of information could be useful. And so if you think about a book being written, it's useful when somebody chooses to read it. And how people come to discover that book is usually it's recommended or it's marketed. Um, but when you think about an organization, so much information is getting created on a daily basis, and oftentimes, the times where that knowledge or institutional knowledge would be best used is weeks, months, even years down the road. And so the best effort attempt historically to organize that information is to use a traditional knowledge base. And the person that's capturing that information is putting in a lot of time and effort into trying to organize it into a certain folder or a certain file structure. And, you know, despite their best efforts, at the end of the day, that structure makes a lot of sense to them and varying levels of, uh, comprehensibility to everybody else on their team. Um, a-and usually by the time, you know, they move on to another team or another company, that organization makes zero sense. And so I think kind of one fundamental unlock and, and trend we're seeing and, and that we're pursuing is that historically, um, after information was created, the onus was on the creator of that information to not only distribute but also organize that information, um, within a company, for example. And now we think that there's actually this opportunity for information to not need to be organized, and instead for some automated system or some AI to be able to come in and actually match every piece of information to the context and the situation in which it would be best used. That way, you can kind of get past all of those inefficiencies that happen.
- 8:46 – 12:46
Organizational structure of knowledge management
- SSSteph Smith
Totally. And I'm just thinking through, like, it seems like there's maybe two aspects here. One of them is this idea that humans want to document, but in order to document, you really just-- like you said, you have to spend so much time. Some people spend more time organizing their second brain than actually using it, right? So that's one part. But then there's also another part, which is just there's already this vast amount of unstructured data that other people have created. And so I guess given the technology that's evolved, like, what is the opening here? Are there, are there new approaches that exist now that we have access to these LLMs? Or how are you guys thinking about kind of turning the page?
- DXDennis Xu
The, the organizational structure that we rely on is still very much so the folder. Uh, and, and the folder is, is, is kind of [chuckles] this thing that, um, it was this like skeuomorphic representation of this filing cabinet, right, that, that we used to have, you know. And then when computers were first developed really, like, in, in the 1950s and '60s, we just said, "Hey, well, you know, we... People understand how filing cabinets work, so why don't we just put filing cabinets, you know, on this computer so that it's a little easier for people to actually adapt to this new technology?" Um, and then, and then that kind of begs the question, well, you know, it's been 60 years since, since that time, right, where, where computers have been around. Um, and what, what makes this moment really special where we can actually say, "Hey, you know, can we actually eliminate the folder?" And, and we, we think the answer is yes. A-and part of that is I think the missing ingredient has always just been a lack of contextual understanding by machines of human language and of the, the, the knowledge that humans produce, right? Computers, uh, I think we, we've known for a long time are far better than humans at, uh, very specific things like, you know, computation. It's, uh, it can store vastly more memory and all of these things, but it's never really been able to actually reason. It's never really-- They've, they've never really been able to actually think. Um, and this new generation, right, of, of LLMs, uh, what they actually demonstrate is reasoning capabilities. And that's where I think the magic really starts to happen, when you can marry the reasoning capabilities of an LLM with kind of like the, you know, the, the vast storage capabilities of just, you know, computers from the past.
- KMKevin Moody
One way to think about this in terms of, you know, people wanted faster horses and instead the answer was cars, is instead of having manually organized folders, you have automatically generated folders. And at one point, I think that's the direction we thought made most sense. But I think that what has become apparent and, and what large language models have allowed is to flip that model on its head and, and recognize the fact that you actually don't want to organize a piece of information into one or even five folders. You wanna organize it into almost an arbitrary number of folders, which are the different ways in which you're gonna need to recall or retrieve that information And the thing that large language models now allow us to do is, in a very generalized fashion, take any piece of text from any domain, it could be, you know, in the medical realm, um, it could have to do with real estate, you know, it could have to do with technology, and be able to comprehend that in such a way that then when it comes time for somebody to need to make use of that knowledge, they can ask very simple natural language questions, communicate in their own voice, and find the right answer without any intermediate categorization step even required. And that's something that before kind of the level of depth of comprehension that these large language models affords, it would not have been possible.
- SSSteph Smith
Something
- 12:46 – 15:29
Proactivity and interactions
- SSSteph Smith
that's really drawing my attention is this idea of proactivity, uh, where not only can it reorganize your information in different ways, but it can actually step up and say, "Hey, this is gonna be useful at this time. I know you're working on this kind of project. Have you considered this?" Which again, someone else in the organization maybe has already done. And so how are you guys thinking about that dynamic, which also feels fundamentally new?
- DXDennis Xu
Yeah, I, I, I think the running joke here is, you know, Clippy is, was just ahead of its time, uh, right?
- SSSteph Smith
[chuckles]
- DXDennis Xu
But, uh, you know, what's, what's really interesting is, is the-- fundamentally the idea of, of Clippy and, and, and this, you know, proactive assistant, obviously, I think people have been, have been attracted to this for, for a long time. It's, it's so e-embedded in, in, you know, the sci-fi fantasies that, that society has, uh, for the world. But in order for proactiveness to, to not be annoying, uh, and to kind of cross that boundary into, into truly useful and, and then delightful, you actually have to have a, a deep, um, understanding and, and, and the ability to reason about someone's life and, and, you know, what they're thinking about at that moment and, and all of those things, right? And I think, um, the-- if you think historically, uh, and, and honestly even today, right, with, with, with tools like, uh, with ChatGPT, um, the quality of our interactions with computers are still very much limited by the instructions that, that we explicitly provide them, right? And we used to have to provide, you know, instructions, you know, in code, and obviously this is hyperbole, but you know, now a lot of the time we can just provide instructions in English, but we're still providing instructions. A-and this is where things I think will get really interesting. If you actually have an understanding, you know, of truly of who this person is, of what they care about, of what they know, um, then what you can actually start doing is, uh, kind of flip that model on its head and say, instead of, you know, me constantly having to instruct this, this machine, um, what if the, the machine will just do things for me and then confirm afterwards like, "Hey, you know, is this, is this what you want done?" Or, um, uh, "What else, what else do you need?" Right.
- SSSteph Smith
Yeah. You're evolving with it, right? You can like prompt it initially, but then it can also run on its own. You can give it feedback, and it can iterate instead of the really like strict rules, "Hey, deliver me this specific email reminder every morning at 9:00 A.M." Like that you can't really iterate on because there's, there's not any gray area, there's not any nuance or context like you guys have mentioned. Um, but something that's coming to mind is just the cost of
- 15:29 – 19:34
What is the cost and value?
- SSSteph Smith
all this, right? We're talking about so many different applications that you're tying into constantly, uh, working with data, maybe organizing it to some degree, parsing it, saving it somewhere it has to be retained. I g- I guess in a way it's so much more powerful that maybe a new business model can be built around this idea. Like can you imagine how much people p-pay EAs and this is, you know, kind of paralleling that. But yeah, just how do you guys think about this evolution and how much it costs to do this, but then also maybe how you monetize it on the other side?
- KMKevin Moody
Cost is a really interesting question, and we, we think a lot about this because, uh, at the most fundamental level, we really want to democratize access to this sort of intelligence and, and give every person the power essentially of their own personal AI that can help them think better and do more. And so there's, there's really two ways that we're currently focused on monetizing, and the first way is via the individual plan, and I think you can actually similarly equate it to some of the other, you know, products on the market in terms of a, you know, free to use plan as well as a pr- a paid, you know, premium all access plan. And, and that's one that we want to continue to figure out how to drive, you know, the price down. Um, but one of the really promising trends is that the cost of compute and the cost of actually running these models is coming down by an order of magnitude, um, every several months. And so given that that trend is likely to continue, we think we're going to be able to actually serve, um, all of these things at just an incremental cost above storage.
- DXDennis Xu
On, on separately, one of the things you really touched on that, that I completely agree with is this idea of, you know, what if we could actually-- like essentially what if we could actually give each person kind of their personal, like a personal EA, right? And then that really changes because these-- i-if you think about the value traditionally of, you know, whether it's note-taking apps or knowledge bases, it has been almost exclusively focused on information retrieval, right? It's, it's been, hey, you know, it kind of, the value ends at search. After you receive a list of documents, that's where, you know, you're done. Um, but we can actually go far beyond just retrieving a list of documents now, we can actually, you know, do things with them, right? We can, uh, help you think through problems, right? Um, and, and, and that is a much more valuable problem to solve than just pure information retrieval. And now you can kind of start framing, you know, uh, the costs away from like, oh, this is, uh, you know, this is, uh, relative to kind of like a- A typical software subscription that I would make to rather comparing it to wha- instead of hiring an employee or hiring someone to do this thing, can I have, you know, the help of a truly personalized, um, AI assistant, right? And, and, and I think that is, is where the really exciting, um, revenue opportunity also comes, uh, in the future.
- SSSteph Smith
I tweeted this a while ago, but I said, "I want an AI that calls me out on my bullshit." That is like, you know, it's hard to hear that from a human, right? It's like, "Steph, you know, you're moving in the wrong direction. You're not focused on the right things." You said, and even in my personal life, like, "You said you wanted to exercise this amount. You said that you wanted to, like, you know, go to sleep at these times. You're not doing it." And I think, you know, I, w- there's an opportunity here also to, you know, personalization is so inherent in some of the developments in AI, but to personalize the type of support that you're getting. Like
- KMKevin Moody
I love, I love this example you're giving. And we're, we've been thinking a lot about personalization with the chat with Mem and conversational experience and, and Mem as the assistant. And, uh, I think we're gonna have to add that as a potential knob is, you know, call me out on my bullshit factor.
- SSSteph Smith
[laughs] Exactly, yeah. This is a spectrum. Oh, man.
- DXDennis Xu
Yeah, and, and I think, I think one of the, one of the perspectives that we've developed is, um, personalization is actually not just about, you know, does this AI, uh, you
- 19:34 – 20:58
Personalization
- DXDennis Xu
know, can this AI operate across the data that I provided? Even if this piece of technology or, or this assistant actually can, can, you know, do things for you based on the knowledge that you provided, it still doesn't feel personalized because it doesn't feel like it's adapting to, to you. It doesn't feel like it's learning from the interactions that you're having with it, right? It just feels like it is able to kind of search, right, across a bunch of information and then, and then, and then reply. Um, and, and so I think a key fundamental unlock in how we really think about personalization is not just does it have access to my personal data and my personal knowledge, but rather does it actually adapt, right? So internally, we call this adaptive, adaptive AI. Does it actually adapt to who you are? Does it learn about who you are, and does it then apply that in future interactions with you like a human would?
- SSSteph Smith
Thank you so much for listening to the a16z Podcast. If you've made it this far, don't forget to subscribe so that you are the first to get our exclusive video content. Or you can get started with this video that we have hand-selected for you. [outro music]
Episode duration: 21:03
In the uListen app, the transcript lights up word by word as you listen. Get uListen
Transcript of episode jzV_7kPPYPE
