00:00there's actually this opportunity for
00:01information to not need to be organized
00:04it's not as simple as just plugging into
00:06gbt4 and calling it a day same search
00:09for multiple different people should
00:11actually yield different results I want
00:13an AI that calls me out on my
00:16I think we're gonna have to add that as
00:17a potential knob does it actually adapt
00:20to who you are to learn about who you
00:22are that's where I think the magic
00:24really starts to happen some people
00:25spend more time organizing their second
00:27brain than actually using it there's a
00:29there's a crazy amount of innovation
00:31that's happening in the space right now
00:32that's that's just been really exciting
00:38since the rise of the computer humans
00:41have been gripped by the idea of having
00:43a second brain but has reality held up
00:47to that promise how many people spend
00:49more time organizing their second brains
00:52instead of leveraging the information
00:53within them how many people constantly
00:56look for a better to-do App instead of
00:58actually checking the to-do's that
01:00they're organizing and how many people
01:02have an endless stream of data that
01:04they'd love to one day process but they
01:07just don't have the right tools or time
01:09to do so due to the structured directive
01:12nature of computers to date computers
01:14have been a shell of maybe the second
01:16brain that we've long hoped for still
01:18incredibly powerful but also requiring
01:21discrete instruction from the director
01:23but within the last year consumer AI has
01:26shown up to party now capable of
01:28processing simple language prompts and
01:30interfacing with unstructured data
01:32possibly fundamentally changing this
01:35game so today we chat with the founders
01:38of men Kevin Moody and Dennis Zoo as we
01:41explore what's gotten in the way of a
01:44true second brain and how AI might
01:47actually change the game and mam itself
01:49is one such company trying to reinvent
01:51the space so they can speak to the why
01:53now but also the very real cost of
01:56making this Vision a reality but first
01:59here is a wild stat about just how much
02:02is lost in productivity within
02:07as a reminder the content here is for
02:09informational purposes only should not
02:11be taken as legal business tax or
02:13investment advice or be used to evaluate
02:15any investment or security and is not
02:17directed at any investors or potential
02:19investors in any accz fund for more
02:22details please see a16c.com disclosures
02:35I'd love to start out with a statistic
02:37which is crazy in my mind I couldn't
02:40believe this but it's basically
02:42knowledge workers apparently spend 2.5
02:45hours a day which is around 30 of their
02:49time in a workday searching for
02:51information and this stat is from 2016
02:53but it really wouldn't surprise me if
02:55this is still you know the reality of
02:57working in the information age I almost
03:00think that that's probably an
03:02understatement the 2.5 hours number I
03:06it's it's definitely uh still relevant
03:09if not increased since 2016.
03:12um I think in part because it's so much
03:14easier to produce unstructured
03:16information and share information the
03:18barrier for creating content and sharing
03:22um and being productive as a part of a
03:24team has has gone down and so you just
03:26get more and more information that then
03:28your co-workers and teams are trying to
03:31parse through to get through their
03:32day-to-day jobs but I think something
03:34interesting there you mentioned 2.5
03:36hours per day just searching is this
03:39probably hidden time sync 2 which is
03:43you don't even know that you are looking
03:44for some sort of shared piece of
03:46information you just begin doing your
03:48work for the day or for the week and at
03:50the end of that week you realize you've
03:52just reinvented the wheel that somebody
03:54else on your team or some other team in
03:56your company did that exact same work
03:58and it already exists in some
04:00spreadsheet somewhere or some document
04:01somewhere and then I think
04:03the question I have in my head is how
04:05much time like how much time are people
04:07actually spending than taking those
04:10things and then performing really
04:12mundane Transformations or you know
04:14tasks on top of what they found
04:17um and I I would imagine that that's
04:19kind of you know at least another few
04:21hours right so I think what's really
04:24interesting is we're now entering this
04:26this world where a lot of that work
04:29and not just search right search is kind
04:32of just the the beginning of really
04:34where yeah like where people are going
04:37to get value out of out of these new
04:41um I I think it's going to be really
04:43interesting to see how the entire
04:45workflow of a knowledge worker changes
04:47over the next few years
04:49exactly and I feel like if we kind of go
04:51back and look at the Arc of Knowledge
04:54Management like let's truly go all the
04:56way back you get the ability to write
04:58and people are doing this on Stone at
05:00first at some point the printing press
05:02arrives you know then you have
05:04individualized typewriters the computer
05:06shows up at some point the internet
05:08gives you access to almost everyone
05:10around the world these days and so you
05:12have this promise of technology which
05:14has rung true it has you know given
05:17people the ability to codify and share
05:21information and knowledge in you could
05:24say exponentially better ways yet at the
05:27same time you know if we bring ourselves
05:28back to that stat it's kind of
05:30surprising that there still are just
05:32incredible inefficiencies in
05:34collaboration what do you think is
05:35preventing us from having this you know
05:38really streamlined back and forth
05:40between you know folks within an
05:42organization or even as an individual I
05:45feel like a lot of people have this
05:46image of a second brain But ultimately
05:48you know know it doesn't quite live up
05:50to the hype you know what's really
05:52interesting is all of the things that
05:54you mentioned all of these devices in
05:56the past they've actually just primarily
05:58been broadcasting devices if you really
06:00think about what is core to
06:06sharing and broadcasting is only one
06:08aspect of that it's only one component
06:11the other component is how do you
06:17digest and how do you actually consume
06:19that that piece of information right
06:21then that someone else is actually
06:22broadcasting we've never truly had
06:25really hyper personalized consumption
06:28devices search is very much kind of a
06:32problem right it's very much of that hey
06:34I'm looking for this piece of
06:36information and the same search query
06:38right the same search for multiple
06:40different people should actually yield
06:42different results because you think of
06:43you think about the world differently
06:46and the tools that we've used up to this
06:48point really just don't model that
06:51if we if we look at it from the
06:52perspective of one piece of information
06:54and its life cycle when information and
06:58organization especially is captured it's
07:01oftentimes really hard to know all of
07:03the people and all of the contexts in
07:05which that one piece of information
07:07could be useful and so if you think
07:10about a book being written it's useful
07:12when somebody chooses to read it and how
07:14people come to discover that book is
07:16usually it's recommended or it's
07:18marketed but when you think about an
07:20organization so much information is
07:23getting created on a daily basis and
07:25oftentimes the times where that
07:27knowledge or institutional knowledge
07:29would be best used is weeks months even
07:33years down the road and so the best
07:36effort attempt historically to organize
07:39that information is to use a traditional
07:41knowledge base and the person that's
07:43capturing that information is putting in
07:45a lot of time and effort into trying to
07:47organize it into a certain folder or a
07:49certain file structure and you know just
07:51despite their best efforts at the end of
07:53the day that structure makes a lot of
07:55sense to them and varying levels of
07:58comprehensibility to everybody else on
08:00their team and usually by the time you
08:03know they move on to another team or
08:04another company that organization makes
08:06zero sense and so I think kind of one
08:09fundamental unlock and and Trend we're
08:12seeing and that we're pursuing is that
08:14historically after information was
08:17created the onus was on the creator of
08:19that information to not only distribute
08:21but also organize that information
08:23within a company for example and now we
08:26think that there's actually this
08:27opportunity for information to not need
08:30to be organized and instead for some
08:32automated system or some AI to be able
08:35to come in and actually match every
08:37piece of information to the context and
08:39the situation in which it would be best
08:41used that way you can kind of get past
08:44all of those inefficiencies that happen
08:47totally and I'm just thinking through
08:48like it seems like there's maybe two
08:50aspects here one of them is this idea
08:53that humans want to document but in
08:56order to document you really just like
08:59you said you have to spend so much time
09:00some people spend more time organizing
09:03their second brain than actually using
09:05it right so that's one part but then
09:07there's also another part which is just
09:09there's already this vast amount of
09:11unstructured data that other people have
09:12created and so I guess given the
09:16technology that's evolved like what is
09:18the opening here are there are there new
09:21approaches that exist now that we have
09:23access to these llms or how are you guys
09:24thinking about kind of turning the page
09:27the the organizational structure that we
09:30rely on is still very much so the folder
09:33uh and and the folder is is kind of this
09:38um it was this like skeuomorphic
09:40representation of this filing cabinet
09:42right that we used to have you know and
09:44then when computers were first developed
09:46really like in the 1950s and 60s we just
09:49said hey well you know we people
09:51understand how filing cabinets work so
09:53why don't we just put filing cabinets
09:55you know on this computer so that it's a
09:57little easier for people to actually
09:58adapt to this new technology
10:00um and then that kind of begs the
10:02question well you know it's been 60
10:03years since since that time right where
10:06computers have been around
10:08um and what what makes this moment
10:12really special where we can actually say
10:15Hey you know can we actually eliminate
10:20we we think the answer is yes and part
10:24I think the missing ingredient has
10:26always just been a lack of contextual
10:29understanding by Machines of human
10:33language and of the the the knowledge
10:35that humans produce right computers I
10:38think we we've known for a long time are
10:41far better than humans at uh very
10:44specific things like you know
10:45computation it's uh it can store vastly
10:49more memory and all of these things but
10:51it's never really been able to actually
10:52reason it's never really they've never
10:54really been able to actually think
10:57um and this new generation right of llms
11:01what they actually demonstrate is
11:04reasoning capabilities
11:06and that's where I think the magic
11:08really starts to happen when you can
11:10marry the reasoning capabilities of an
11:13llm with kind of like the you know the
11:16vast storage capabilities of just you
11:20know computers from the past one way to
11:23think about this in terms of you know
11:27people wanted Faster Horses and instead
11:29the answer was Cars is instead of having
11:33manually organized folders you have
11:36automatically generated folders and at
11:40one point I think that's the direction
11:41we thought made most sense but I think
11:44that what has become apparent and what
11:46large language models have allowed is to
11:50flip that model on its head and
11:51recognize the fact that you actually
11:54don't want to organize a piece of
11:55information into one or even five
11:57folders you want to organize it into
11:59almost an arbitrary number of folders
12:02which are the different ways in which
12:03you're going to need to recall or
12:04retrieve that information and the thing
12:07that large language models now allow us
12:08to do is in a very generalized fashion
12:11take any piece of text from any domain
12:14it could be you know in the medical
12:16realm it could have to do with real
12:18estate you know it could have to do with
12:20technology and be able to comprehend
12:22that in such a way that then when it
12:25comes time for somebody to need to make
12:27use of that knowledge they can ask very
12:29simple natural language questions
12:31communicate in their own voice and find
12:33find the right answer without any
12:35intermediate categorization stuff even
12:37required and that's something that
12:39before kind of the level of depth of
12:41comprehension that these large language
12:43models affords it would not have been
12:45possible something that's really drawing
12:47my attention is this idea of proactivity
12:49where not only can It reorganize your
12:53information in different ways but it can
12:55actually step up and say hey this is
12:58going to be useful at this time I know
13:00you're working on this kind of project
13:02have you considered this which again
13:04someone else in the organization maybe
13:05has already done and so how are you guys
13:07thinking about that Dynamic which also
13:09feels fundamentally new
13:12yeah I I think the running joke here is
13:14you know clippy is was just ahead of its
13:16time uh right but you know what's what's
13:19really interesting is is the
13:22fundamentally the idea of clippy and and
13:24this you know proactive assistant
13:26obviously I think people have been have
13:28been attracted to this for for a long
13:32embedded in in you know the Sci-Fi
13:35fantasies that that Society has uh for
13:38the world but in order for a
13:39proactiveness to to not be annoying uh
13:42and to kind of cross that boundary into
13:44into truly useful and and then
13:46delightful you actually have to have a
13:50um understanding and and the ability to
13:52reason about someone's life and and you
13:55know what they're thinking about at that
13:57moment and all those things right and I
14:00um the if you think historically uh in
14:04honestly even today right with with with
14:06tools like uh with chat gbt
14:09um the quality of our inner actions with
14:13computers are still very much limited by
14:15the instructions that that we explicitly
14:17provide them right and we used to have
14:19to provide you know instructions you
14:22know in code and obviously decide
14:23verbally but you know now a lot of the
14:25time we can just provide instructions in
14:26English but we're still providing
14:28instructions and this is where things I
14:31think will get really interesting like
14:33you actually have an understanding you
14:36know of truly of who this person is of
14:38what they care about of what they know
14:41then what you can actually start doing
14:43is uh kind of flip that model on its
14:45head and say instead of you know me
14:48constantly having to instruct this this
14:52um what if the the machine will just do
14:54things for me and then confirm
14:56afterwards like hey you know is this is
14:58this what you want done or
15:00what else what else do you need right
15:03yeah you're evolving with it right you
15:05can like prompt it initially but then it
15:07can also run on its own you can give it
15:09feedback and it can iterate instead of
15:11the really like strict rules hey deliver
15:13me this specific email reminder every
15:16morning at 9am like that you can't
15:18really iterate on because there's
15:20there's not any gray area there's not
15:23any Nuance or context like you guys have
15:25mentioned but something that's coming to
15:28mind is just the cost of all this right
15:29we're talking about so many different
15:31applications that you're tying into
15:33constantly working with data Maybe
15:35organizing it to some degree parsing it
15:39saving it somewhere it has to be
15:40retained I I guess in a way it's so much
15:43more powerful that maybe a new business
15:45model can be built around this idea like
15:48you imagine how much people pay EAS and
15:50this is you know kind of paralleling
15:52that but yeah just how do you guys think
15:54about this Evolution and how much it
15:56costs to do this but then also maybe how
15:58you monetize it on the other side
16:01cost is a really interesting question
16:03and we we think a lot about this because
16:06at the most fundamental level we really
16:09want to democratize access to this sort
16:12of intelligence and and give every
16:16the power essentially of their own
16:18personal AI that can help them think
16:19better and do more and so there's
16:22there's really two ways that we're
16:24currently focused on monetizing and the
16:26first way is via the individual plan and
16:28I think you can actually similarly
16:29equate it to some of the other you know
16:31products on the market in terms of a you
16:34know free to use plan as well as a pre a
16:36paid you know premium all access plan
16:38and and that's one that we want to
16:41continue to figure out how to drive you
16:43know the price down but one of the
16:45really promising Trends is that the cost
16:47of compute and the cost of actually
16:49running these models is coming down by
16:52an order of magnitude every several
16:55months and so given that that trend is
16:57likely to continue we think we're going
16:59to be able to actually serve all of
17:01these things at just an incremental cost
17:03above storage on on separately one of
17:07the things you really touched on that I
17:09completely agree with is this idea of
17:11you know what if we could actually like
17:13essentially what if we could actually
17:15give each person kind of their personal
17:17like a personal EA right and that really
17:22these if you think about the value
17:24traditionally of you know whether it's
17:27note-taking apps or knowledge bases it
17:29has been almost exclusively focused on
17:31information retrieval right it's it's
17:33been hey you know it kind of the value
17:35ends at search after you receive a list
17:37of documents that's where you know
17:41um but we can actually go far beyond
17:44just retrieving a list of documents now
17:46we can actually you know do things with
17:48them right we can help you think through
17:53um and and that is a much more valuable
17:56problem to solve than just pure
17:58information retrieval and now you can
18:00kind of start framing you know the costs
18:04away from like oh this is uh you know
18:07this is uh relative to kind of like a
18:10typical software subscription that I
18:12would make to rather comparing it to
18:14instead of hiring an employee or hiring
18:18someone to do this thing can I have you
18:20know the help of a truly per
18:22personalized AI assistant right and and
18:25I think that is is where the really
18:29Revenue opportunity also comes
18:33in the future I tweeted this a while ago
18:35but I said I want an AI that calls me
18:37out on my that is like you know
18:40it's hard to hear that from a human
18:41right it's like stuff you know you're
18:43moving in the wrong direction you're not
18:44focused on the right things you've said
18:46and even in my personal life like you
18:47said you wanted to exercise this amount
18:49you said that you wanted to like you
18:51know go to sleep at these times you're
18:53not doing it and I think you know
18:54there's an opportunity here also to
18:58you know personalization is so inherent
19:00in some of the developments in AI but to
19:02personalize the type of support that
19:05you're getting I love I love this
19:07example you're giving we've been
19:09thinking a lot about personalization
19:10with the chat with mem and
19:12conversational experience and and ma'am
19:14as the assistant and I think we're gonna
19:16have to add that as a potential knob is
19:18call me out on my back exactly
19:21yeah this is a spectrum yeah and I think
19:24I think one of the perspectives that
19:27um personalization is actually not just
19:30about you know does this AI uh you know
19:34can this AI operate across the data that
19:36I provided even if this piece of
19:38technology or this assistant actually
19:40can can you know do things for you based
19:43on the knowledge that you provided it
19:46still doesn't feel personalized because
19:47it doesn't feel like it's adapting to to
19:50you it doesn't feel like it's learning
19:52from the interactions that you're having
19:53with it right it just feels like it is
19:55able to kind of search right across a
19:59bunch of information and then and then
20:04I think a key fundamental unlock and how
20:07we really think about personalization is
20:09not just does it have access to my
20:12personal data and my personal knowledge
20:14but rather does it actually adapt right
20:17it's an internal would be called this
20:18adaptive adaptive AI does it actually
20:20adapt to who you are does it learn about
20:22who you are and does it then apply that
20:24in future interactions with you like a
20:26human would thank you so much for
20:29listening to the a16z podcast
20:31if you've made it this far don't forget
20:33to subscribe so that you are the first
20:35to get our exclusive video content or
20:38you can get started with this video that
20:41we have hand selected for you