00:002022 was a breakout year for AI
00:04in fact many have even claimed that
00:06Chachi BT is the fastest growing app of
00:09all time so with so much opportunity on
00:12the table AI is the topic of
00:14conversation in every boardroom as CEOs
00:16figure out how to best integrate this
00:18new superpower but they're also asking
00:20really important questions around data
00:22privacy competition cost accuracy and
00:25also doing this all really quickly
00:27because just like your customers really
00:30don't care if your product is built with
00:31angular or react or runs on AWS or
00:34Heroku there will be a whole host of
00:36ways that companies differentiate as
00:38they look to cleverly embed Ai and here
00:41is how hex is thinking about that hex is
00:44a platform for collaborative data
00:46science and analytics our users are
00:48people doing data work as a reminder the
00:50content here is for informational
00:52purposes only should not be taken as
00:54legal business tax or investment advice
00:56or be used to evaluate any investment or
00:58security and is not directed at any
01:00investors or potential investors in any
01:02ACC fund for more details please see
01:05a16cc.com disclosures where is the
01:09competitive Advantage where is the moat
01:11is it in the UI it feels like that can
01:13be replicated very easily is it in the
01:15cleaning of the data the linking of the
01:17data is it something else like how are
01:19you thinking about the fact that
01:21basically you can kind of depend on all
01:23your competitors to replicate what you
01:26do if it's working yeah we've already
01:28seen that yeah uh down to some UI
01:32elements that look very familiar
01:33um uh yeah the models themselves are
01:36Commodities and they're that's going to
01:39accelerate even in the next couple years
01:40we're going to see a ton of different
01:41types of models emerge we're going to
01:43see the costs plunge even further down I
01:45think the ability to plug these in will
01:47become as ubiquitous as like using cloud
01:49services it's just like everyone does it
01:51I I think there's a few places where
01:54potential notes emerge one is we do have
01:57a pretty great data Advantage you know
02:00already having hundreds of customers
02:02thousands of users writing millions of
02:04lines worth of SQL in Python
02:07we can use that sort of Rich information
02:10um we're not using it to train models I
02:12should be very clear like that's one
02:13thing we've been very upfront with our
02:14customers about like we're not training
02:17models where they would expect to ever
02:18have like some code they've written via
02:20completion for someone else which is a
02:21problem in other places but it's more
02:23like using that actually to personalize
02:25for that person and their team like you
02:28know again like which schemas you're
02:29using which what is their code style
02:31been like in the past it's really how do
02:33you put these pieces together and how do
02:36you build a really great user experience
02:37both from you know the pixels on the
02:40screen to thinking about performance
02:43behind the scenes to things like docs
02:45like that all sort of combines to being
02:48a really Superior product experience I
02:50do think that long term generative AI
02:53large language models they will change a
02:54lot of like fundamental assumptions
02:57um and I think that there will be an
02:59advantage to the companies that can be
03:00the first ones to figure out
03:03what those things are data can be highly
03:06creative highly fun but as you are
03:08looking to integrate this technology
03:10what does that look like under the hood
03:12we are doing a ton in terms of
03:14constructing the right prompts and
03:16parsing responses back from the model
03:18apis we're using and so we have
03:20thousands of people already writing SQL
03:23and writing Python and doing data work
03:24and hex every day so we have a ton of
03:26information we've got we're connected to
03:28their database schemas so we see the
03:30structure of their data we see past
03:31queries and pass code difference so we
03:33know which tables and columns are most
03:34frequently referenced right we have
03:36information about the project they're
03:37building so we know like oh this project
03:39is already referencing this part of the
03:41schema and that's probably the relevant
03:43data for this you can even look at
03:44things like this is the typical way this
03:47organization is formatting their charts
03:49and infer how they might want their
03:52visualizations to look and so there's a
03:54ton of context we get because we're
03:55incorporating this in an existing set of
03:57workflows that can help us create the
03:59right context and and create the right
04:01prompt to pass to the model any
04:04learnings in terms of what is increasing
04:06completion rates what is getting people
04:08to actually interface with this new in
04:11some ways superpower within the app one
04:14of the things that we've observed sort
04:15of more on the back end in terms of
04:17increasing completion rates is
04:19when you're building prompts I think
04:22we've been we were tempted early on to
04:23try to shove as much context as we could
04:25in you kind of figure like the more I
04:26can tell this model the better yeah
04:28realize you can pretty easily confuse a
04:31model right in terms of the amount of
04:33context you're passing and so we've been
04:35really thoughtful on building the right
04:37context understanding how you're
04:38iterating on how different models are
04:40going to respond to different types of
04:41prompts different amounts of context how
04:43you're potentially even chaining
04:45together different types of models or
04:46different modalities of models together
04:48in one sequence in ux and then how
04:51you're giving that feedback to the user
04:53yeah I mean I think one interesting
04:55facet of all this is that there's kind
04:57of this realization of the value of data
05:00and and I think we're going to see a lot
05:02more companies retain their data for
05:04longer collect more data from their
05:07products or from their users and so
05:09there's this again this privacy security
05:11posture of like okay we want to have
05:13this data we're probably going to keep
05:14it for longer how do we keep it safe but
05:16then there's also a cost element right
05:19it's not necessarily free to retain data
05:21it's also not free to run these models
05:24um and so how do you it's also not zero
05:26risk a lot of companies intentionally
05:28delete their data like we'll have
05:30retention policies on email right but by
05:35eliminating knowledge and that knowledge
05:38could be useful sort of like in the
05:39first instance just like I want to go
05:41back and search this I think I think a
05:43lot of big companies deal with as they
05:45start to like have record retention
05:46things but there's also like second
05:48order value to that in terms of could
05:50that inform a model or a um
05:53you know potential different
05:54applications for how you can learn from
05:56what your organization has done
05:59I don't know where we're going to wind
06:00up on that I do think you're right that
06:01a lot of companies are realizing the
06:03value of their data and their IP and
06:05thinking about that in a deeper way
06:07I think there's all sorts of really
06:09exciting opportunities on how you could
06:10use data that customers are trusting you
06:13making sure you're continuing to earn
06:14that trust you know as you are applying
06:16these generative techniques with it is
06:18Paramount and something where you know
06:20we expect to be a central theme as long
06:24as we run the company
06:25you like this segment you're gonna love
06:27our next video where we chat with
06:29sourcegraph a company that has spent the
06:31last decade mastering search who thinks
06:34they're uniquely positioned for this AI
06:35wave and if you like this topic we go a
06:39lot deeper on the a16z podcast which you
06:41can find on Apple Spotify or wherever
06:43you get your podcast