Latest / Expert Insights Podcast / #52 – What AI Really Is, And Isn't (Craig Martell, Cohesity)
Transcript
- 0:06SPEAKER_00: Hello and welcome to the Expert Insights Podcast.
- 0:11SPEAKER_00: I'm your host, Joel Witts.
- 0:13On this episode of the show,
- 0:14I interview Dr.
- 0:15Craig Martell,
- 0:16the Chief AI Officer at Cohesity.
- 0:19Craig recently served as the Chief Digital and Artificial
- 0:22Intelligence Officer for
- 0:23SPEAKER_00: the US Department of Defense,
- 0:25where he was responsible for accelerating AI adoption across
- 0:28the department.
- 0:30SPEAKER_00: Before joining the DoD,
- 0:31SPEAKER_00: he's had several senior roles at leading tech companies,
- 0:33SPEAKER_00: like Lyft,
- 0:34SPEAKER_00: where he was the head of machine learning,
- 0:36SPEAKER_00: and Dropbox,
- 0:37SPEAKER_00: where he was the head of machine intelligence.
- 0:39SPEAKER_00: In this episode,
- 0:40we discussed Craig's journey into AI and the challenges of
- 0:43building trust in AI systems.
- 0:45I have to say,
- 0:46this is one of my favorite podcasts that we've recorded so far.
- 0:49SPEAKER_00: We cover some fascinating topics.
- 0:51SPEAKER_00: I hope you enjoy.
- 0:52SPEAKER_01: My name is Craig Martell.
- 0:53SPEAKER_01: I'm the Chief AI Officer for Cohesity.
- 0:55SPEAKER_01: Cohesity is the world's largest data protection company,
- 0:58SPEAKER_01: measured by the number of terabytes under management.
- 1:01SPEAKER_01: We have hundreds of exabytes under management.
- 1:03SPEAKER_01: In my immediately prior job,
- 1:04SPEAKER_01: I was the Chief Digital and AI Officer for the US Department of Defense.
- 1:08SPEAKER_01: As you can imagine, that was a fascinating job as well.
- 1:10And I've held a number of positions in the AI field,
- 1:12starting back with LinkedIn,
- 1:14Dropbox,
- 1:15Lyft.
- 1:16And I was a professor at the Naval Postgraduate School for
- 1:18about 15 years.
- 1:19Your question about what drew me into AI is a really
- 1:21fascinating question,
- 1:22SPEAKER_01: which might lead us far afield here.
- 1:24SPEAKER_01: But I started off studying philosophy,
- 1:26and I thought that artificial intelligence might be a useful
- 1:29way to do what I
- 1:30SPEAKER_01: called testable philosophy.
- 1:32SPEAKER_01: Like, what does it mean for humans to behave in certain ways?
- 1:35And maybe we could build these artificial systems where we
- 1:37could test human behavior.
- 1:39SPEAKER_01: But what I learned very,
- 1:40SPEAKER_01: very quickly is that humans,
- 1:42SPEAKER_01: number one,
- 1:42are not as special as we think we are because simple
- 1:45statistical patterns is what
- 1:47SPEAKER_01: solves most of AI's problems.
- 1:49SPEAKER_01: And even today, like modern AI is just statistics at scale.
- 1:52SPEAKER_01: You just gather massive amounts of data from the past.
- 1:55SPEAKER_01: You use it to build a model and use that model to predict the future.
- 1:58And so even though there's lots of stories floating around that
- 2:01modern AI is
- 2:02thinking or it's reasoning and we say things like mixtures of
- 2:05experts and we say
- 2:06SPEAKER_01: reasoning model,
- 2:07SPEAKER_01: None of that's true.
- 2:08SPEAKER_01: That's all just naming conventions.
- 2:10SPEAKER_01: We're just picking names that sound good.
- 2:12SPEAKER_01: Just like artificial neural net.
- 2:13That's a name that sounds good,
- 2:15but it's got nothing to do with the brain.
- 2:17SPEAKER_01: It's got zero to do.
- 2:18SPEAKER_01: It's so far from the brain.
- 2:21And so we just sort of pick these terms that make them sound
- 2:23sexy and cool and like
- 2:25SPEAKER_01: they're thinking things,
- 2:26SPEAKER_01: but they're not.
- 2:27SPEAKER_01: They're just massive statistical pattern recognition systems.
- 2:30SPEAKER_01: But the fascinating thing from a philosophical point of view is,
- 2:33isn't it amazing how much pattern recognition and pattern
- 2:37generation seems like intelligence?
- 2:39So this view of ourselves as these special things that are
- 2:42generating brand new
- 2:43SPEAKER_01: thoughts that no one's ever had before and brand new strings of words.
- 2:47SPEAKER_01: It turns out that if you just gather all the data on the internet,
- 2:49SPEAKER_01: count it all up and then regenerate it,
- 2:51SPEAKER_01: it sounds an awful lot like you anyway,
- 2:53SPEAKER_01: Joel.
- 2:53SPEAKER_00: Yeah,
- 2:54I think this is probably a record for how deep we've got as
- 2:57early as possible in
- 2:57SPEAKER_00: the podcast.
- 2:58SPEAKER_00: Well, you asked me why I got into AI.
- 3:00SPEAKER_00: This is why I got into AI.
- 3:01SPEAKER_00: I love that.
- 3:02SPEAKER_00: That's fantastic.
- 3:03SPEAKER_00: That is fantastic.
- 3:04Yeah,
- 3:04because I mean,
- 3:05there was a story right before Gemini,
- 3:07Google Gemini released.
- 3:08SPEAKER_00: I think it was Bard at the time or whatever.
- 3:10that a couple of the engineers or one of the engineers thought
- 3:12it was real.
- 3:13SPEAKER_00: Sentient.
- 3:13SPEAKER_00: Yeah, sentient.
- 3:15SPEAKER_00: And,
- 3:15SPEAKER_00: you know,
- 3:17SPEAKER_00: I wonder,
- 3:18do you think that we'll ever get to that point or do you think
- 3:20it's always just
- 3:20SPEAKER_00: going to be statistics?
- 3:22SPEAKER_01: Joel, so it's not clear that those two things are separate.
- 3:25SPEAKER_01: Sentience might be just statistics.
- 3:27SPEAKER_01: We don't know the answer to that.
- 3:28SPEAKER_01: I don't think it is, but it might just be.
- 3:30SPEAKER_01: Will we get to a thinking, a conscious sentient machine?
- 3:35SPEAKER_01: That's a fascinating question.
- 3:36SPEAKER_01: I don't think so if we, just on the current path.
- 3:39The current path of statistical modeling is,
- 3:42I can't imagine that getting us to sentience.
- 3:44SPEAKER_01: There's a really great philosopher, John Searle from Berkeley.
- 3:46SPEAKER_01: I think he's emeritus now.
- 3:48SPEAKER_01: And he wrote this great book, The Rediscovery of the Mind.
- 3:50SPEAKER_01: And essentially what he argues, and I'm just, I'm aping that argument.
- 3:54Essentially what he argues is,
- 3:55if we really want to build artificial conscious beings,
- 3:59SPEAKER_01: it's likely that we have to build actual artificial brains.
- 4:03SPEAKER_01: And what we're building now are simulated brains.
- 4:05SPEAKER_01: Simulated brains are not the same thing as artificial brains.
- 4:08So maybe we should start looking at the brain science and
- 4:10thinking about how brains
- 4:11SPEAKER_01: actually work if we want to get to conscious,
- 4:13and how brains give rise to consciousness if we want to get to
- 4:16conscious machines.
- 4:18SPEAKER_01: I doubt very seriously just counting up stuff from the past,
- 4:21SPEAKER_01: building a massive model that causes,
- 4:23SPEAKER_01: that uses all the electricity on the planet,
- 4:25SPEAKER_01: and then generating strings is going to get us to consciousness.
- 4:29SPEAKER_01: By the way,
- 4:29SPEAKER_01: Joel,
- 4:30SPEAKER_01: since we're in a philosophy class now,
- 4:31SPEAKER_01: which is why I dislike the term hallucination.
- 4:35Because the term hallucination makes it sound like a cognitive
- 4:39being gone awry.
- 4:40SPEAKER_01: Like it's a cognitive being that just made a mistake, right?
- 4:43SPEAKER_01: It misinterpreted its own internal signal.
- 4:45SPEAKER_01: I don't think anything could be further than the truth.
- 4:46SPEAKER_01: It's not a cognitive being.
- 4:47SPEAKER_01: So I just prefer the term error, right?
- 4:49SPEAKER_01: The thing got it wrong in that context.
- 4:51SPEAKER_00: Yeah, yeah.
- 4:52SPEAKER_00: A machine has made a mistake.
- 4:54SPEAKER_00: It's a bug or something like that.
- 4:55SPEAKER_00: Yeah, yeah.
- 4:55SPEAKER_01: It's a statistical system and it's just wrong.
- 4:57SPEAKER_00: Yeah.
- 4:58SPEAKER_00: We met a couple of weeks ago just for listeners at the HumanX conference.
- 5:02And this was something you mentioned a couple of times in our
- 5:04conversation and also
- 5:05in the panel you did around this issue of hallucinations and
- 5:09how humans aren't really...
- 5:12I guess evolved to when something sounds like a human that you
- 5:17sort of want to trust it.
- 5:18SPEAKER_00: I wonder if you can kind of dig in and kind of explain that perspective.
- 5:20SPEAKER_01: Yeah, yeah.
- 5:22SPEAKER_01: So this really is,
- 5:24SPEAKER_01: this will lead back to what I think we have to do in modern AI,
- 5:27SPEAKER_01: which is really think harder about how do we do error evaluation
- 5:31SPEAKER_01: And how do we let the end user know that the system might be wrong?
- 5:35SPEAKER_01: That's extremely important.
- 5:37SPEAKER_01: And I don't think we do that well enough as a field.
- 5:39SPEAKER_01: But look,
- 5:40SPEAKER_01: before,
- 5:40SPEAKER_01: say,
- 5:41SPEAKER_01: three or four years ago,
- 5:42the only things on the planet that spoke fluently were
- 5:45cognitive reasoning beings
- 5:47SPEAKER_01: called humans.
- 5:48SPEAKER_01: And you can argue the degree to which they're cognitive and reasoning,
- 5:50SPEAKER_01: but let's just assume human beings are cognitive and reasoning.
- 5:53SPEAKER_00: Yeah.
- 5:54SPEAKER_01: There were cognitive reasoning beings called humans.
- 5:56And these humans spoke fluently and in a conversational and
- 6:00sometimes authoritative manner.
- 6:02SPEAKER_01: And we reacted accordingly.
- 6:03SPEAKER_01: We, as other humans, reacted accordingly when someone was speaking.
- 6:07SPEAKER_01: Then suddenly out of nowhere comes this other thing that seems to speak.
- 6:11SPEAKER_01: I'm putting speak in quotes because it writes.
- 6:13SPEAKER_01: But they speak now, too.
- 6:15SPEAKER_01: Seems to speak.
- 6:16SPEAKER_01: fluently,
- 6:17SPEAKER_01: cogently,
- 6:17SPEAKER_01: consistently,
- 6:18SPEAKER_01: like in a way that makes a consistent argument and speaks with authority.
- 6:22So I think our own cognitive science by default is predisposed
- 6:26to interpret that as
- 6:27SPEAKER_01: a cognitive being.
- 6:28Because the only other beings that we met who speak fluently,
- 6:32humans,
- 6:32are cognitive.
- 6:33SPEAKER_01: So we sort of have a mapping of fluent speech to cognitive being.
- 6:37SPEAKER_01: So I think it's really important that we escape that.
- 6:40And the way we need to escape that is actually to realize that
- 6:43these systems are
- 6:44SPEAKER_01: statistical systems.
- 6:46SPEAKER_01: These statistical systems make errors.
- 6:48SPEAKER_01: By the way,
- 6:48they're guaranteed to be sometimes wrong because they're
- 6:51probabilistic and it's not
- 6:52SPEAKER_01: a probability of one.
- 6:53SPEAKER_01: So they're guaranteed to be sometimes wrong.
- 6:56SPEAKER_01: And that sometimes could be pretty big.
- 6:57Like some research suggests for some fields,
- 7:00they're wrong 17% of the time.
- 7:01SPEAKER_01: For other fields, they're wrong 30% of the time.
- 7:04SPEAKER_01: By wrong, I mean they generate sentences which are false.
- 7:06SPEAKER_01: Which, by the way, for some use cases might be great.
- 7:09SPEAKER_01: For marketing, that might be a great way for you to get started.
- 7:12SPEAKER_01: For your first draft, that might be a great way for you to get started.
- 7:15SPEAKER_01: But if it's question answering, well, maybe that's a problem, right?
- 7:18SPEAKER_01: And the thing that I think we're missing,
- 7:19SPEAKER_01: and we're spending a lot of time thinking about this at Cohesity,
- 7:22SPEAKER_01: I think the thing that we're missing is,
- 7:24how do you know when you're using these systems if it's getting
- 7:27it right or wrong?
- 7:28So if you're using an autonomous car,
- 7:30you're driving a car that drives itself,
- 7:32SPEAKER_01: Well, how do you know when it's wrong?
- 7:34SPEAKER_01: Well, it does something that you don't like, right?
- 7:36SPEAKER_01: So you know when it's wrong,
- 7:37it's too close to the car in front of you or it's too close to
- 7:39the car beside you
- 7:40SPEAKER_01: or it tries to just steer into the wrong lane.
- 7:43And so you can easily detect your cognitive load if you're
- 7:45paying attention and you should be.
- 7:47Your cognitive load to detect when it's wrong is pretty minimal
- 7:51because you already
- 7:51SPEAKER_01: know how to drive.
- 7:52SPEAKER_01: And your input system to change it is very easy.
- 7:55SPEAKER_01: There's a steering wheel,
- 7:55there's a brake and there's a gas pedal and you already know
- 7:58how to use those.
- 7:59SPEAKER_01: So you can quickly correct this system if it gets it wrong.
- 8:03SPEAKER_01: We haven't quite figured that out for Gen.ai.
- 8:05Like for recommendations for online shopping,
- 8:08well,
- 8:08if it gets wrong,
- 8:09you don't care.
- 8:09SPEAKER_01: You just don't buy the shirt, right?
- 8:10SPEAKER_01: You don't have to correct it.
- 8:11SPEAKER_01: You just don't buy that shirt.
- 8:13SPEAKER_01: And for older fashion blue link search results,
- 8:16SPEAKER_01: you just look at them and we've gotten really good at being able to say,
- 8:19SPEAKER_01: yeah,
- 8:19SPEAKER_01: those aren't very good results.
- 8:20I'm going to search again and I try new words or a different
- 8:23way of asking the question.
- 8:24SPEAKER_01: But when it generates a paragraph,
- 8:27the cognitive load necessary to be able to determine whether
- 8:30that paragraph's right
- 8:31SPEAKER_01: or wrong is really difficult.
- 8:32Now,
- 8:33one thing I do like is in some of these new systems,
- 8:35you can click show thinking.
- 8:37Essentially,
- 8:38it's giving you a sort of output of its state as it's going along.
- 8:42And that's really helpful because you can think,
- 8:44okay,
- 8:44that's a path I don't want to go down.
- 8:46SPEAKER_01: So that's the beginning of being able to see,
- 8:48SPEAKER_01: is this thing generating output down a path that's effective for me?
- 8:52SPEAKER_01: I was going to say thinking.
- 8:53SPEAKER_01: I don't want to say that, right?
- 8:54SPEAKER_00: Yeah, no, I've seen that as well.
- 8:56SPEAKER_00: And I do have to be honest, I wonder how many people would click on that.
- 9:00SPEAKER_01: I do every time.
- 9:01SPEAKER_00: Yeah, I think maybe it's one of those where 10%, 20%, I wonder.
- 9:07So it's like how you build in those safeguards,
- 9:09as you say,
- 9:09is such an interesting question.
- 9:11In order for these things to become as widely useful as they
- 9:14are going to become,
- 9:15I think we need to solve that because we need to be able to
- 9:17make it clear to folks
- 9:19SPEAKER_01: it could be wrong here.
- 9:19SPEAKER_00: Yeah.
- 9:20SPEAKER_00: And I suppose, I think there's two things as well.
- 9:23One thing that I've been thinking about a lot since our
- 9:25conversation is that trust,
- 9:26SPEAKER_00: I suppose,
- 9:27SPEAKER_00: goes both ways in that,
- 9:28SPEAKER_00: you know,
- 9:28SPEAKER_00: I think a user trusts the output kind of as that's the baseline,
- 9:33SPEAKER_00: but also trusts it to input data,
- 9:35SPEAKER_00: right?
- 9:35SPEAKER_00: To think I'm in this one-to-one chat with this chat bot, you know,
- 9:40SPEAKER_00: The default position,
- 9:41SPEAKER_00: I suppose,
- 9:42SPEAKER_00: isn't that the data I share is going to be accessible to others.
- 9:45SPEAKER_00: This conversation is not isolated.
- 9:47I can trust it to upload this sensitive data so it can write an
- 9:51email quicker for me.
- 9:52SPEAKER_00: I wonder,
- 9:52SPEAKER_00: do you agree with that,
- 9:53that maybe some psychological trust there and that it kind of
- 9:57feels like you're
- 9:57SPEAKER_00: talking to a person?
- 9:59SPEAKER_01: Yeah, I had never thought about it.
- 10:00SPEAKER_01: I had never thought about that, but I think that's right.
- 10:02I think by default,
- 10:04we are treating it like,
- 10:05you know,
- 10:06just someone else we're talking to.
- 10:07SPEAKER_01: But it's probably the case if,
- 10:09SPEAKER_01: I mean,
- 10:09SPEAKER_01: you still have to read the terms of service,
- 10:11SPEAKER_01: I guess,
- 10:11SPEAKER_01: is the answer,
- 10:12because it's probably the case that that data going up is being
- 10:15used to train the
- 10:16SPEAKER_01: system unless it expressly says it's not,
- 10:18SPEAKER_01: right?
- 10:18SPEAKER_01: I mean,
- 10:19SPEAKER_01: we've already exhausted all of the data on the internet,
- 10:21SPEAKER_01: so anywhere else you can get new data,
- 10:23SPEAKER_01: right?
- 10:24SPEAKER_01: By the way,
- 10:25SPEAKER_01: that's one of the reasons I moved to Cohesity,
- 10:27SPEAKER_01: because Cohesity has hundreds of exabytes of data.
- 10:30SPEAKER_01: Now, all that data belongs to individual companies.
- 10:33SPEAKER_01: That's not accessible to anybody else but those companies.
- 10:36SPEAKER_01: And that's the trust that those companies put in us.
- 10:39SPEAKER_01: But for that company,
- 10:41the general internet data is probably not the best data to
- 10:45build their model.
- 10:46The historical data of their company is probably more likely to
- 10:50be much better data
- 10:51SPEAKER_01: for them to build models.
- 10:52And you normally keep in expensive primary storage just about a
- 10:56year's worth.
- 10:57But if you're going to,
- 10:59let's say the world is moving towards a recession.
- 11:01SPEAKER_01: Well,
- 11:01SPEAKER_01: you can't look at the last five years and say,
- 11:03SPEAKER_01: let me model a recession because the economy has been pretty good.
- 11:06But if you go all the way back to 2007,
- 11:072008,
- 11:07then maybe you have a good business model
- 11:11SPEAKER_01: A good model of how your business did during that recession.
- 11:15SPEAKER_01: So having this longitudinal data that's specific to your company,
- 11:19SPEAKER_01: that actually,
- 11:20SPEAKER_01: I think,
- 11:20SPEAKER_01: can be a real windfall for companies.
- 11:22So we're working really hard at Cohesity to make that backup
- 11:26data also accessible
- 11:27SPEAKER_01: to the model building for your company.
- 11:29SPEAKER_00: Wow, that's really, really interesting.
- 11:30SPEAKER_00: Yeah, I mean, there's huge value in that for companies.
- 11:34And I wonder if now would be a good time for you to just kind
- 11:36of give a bit of a
- 11:37SPEAKER_00: background of Cohesity,
- 11:38SPEAKER_00: kind of what Cohesity does,
- 11:40SPEAKER_00: the services you offer,
- 11:42SPEAKER_00: and kind of what your role looks like within Cohesity.
- 11:44SPEAKER_00: Sure, sure.
- 11:45SPEAKER_00: So Cohesity is a backup company.
- 11:46SPEAKER_01: And if you say that just generically, it sounds boring.
- 11:49SPEAKER_01: Backup sounds boring.
- 11:50SPEAKER_01: So what does backup actually do?
- 11:52SPEAKER_01: Backup,
- 11:53SPEAKER_01: before you can even backup your data,
- 11:54SPEAKER_01: a lot of really cool things from a nerd perspective have happened.
- 11:57SPEAKER_01: One, you've gotten funding for data resilience.
- 12:00So you have funding in your pocket to be able to do the right
- 12:03thing with your data.
- 12:04SPEAKER_01: You've identified the important data for your company.
- 12:07SPEAKER_01: So you already know what matters to your company.
- 12:09And you've actually identified how that data is relatively
- 12:13important because you
- 12:14SPEAKER_01: say this data only has to be backed up once a month,
- 12:17SPEAKER_01: but this data has to be backed up every 15 minutes.
- 12:19So you even have this notion of how the data changes over time
- 12:23with respect to its
- 12:24SPEAKER_01: importance for your business.
- 12:26SPEAKER_01: Okay,
- 12:26SPEAKER_01: so you've done all that,
- 12:26and then you've done the hard work of connecting whatever
- 12:30backup system you're using.
- 12:31You've connected the connectors,
- 12:33the system,
- 12:34to all of the apps that generate your data.
- 12:36SPEAKER_01: Okay, so that's backup.
- 12:38SPEAKER_01: So you do all of that, and then you put in a corner.
- 12:41SPEAKER_01: Okay.
- 12:41SPEAKER_01: And then Cohesity does another, that's called data protection.
- 12:44SPEAKER_01: And, you know, we think that's more than just the corner.
- 12:46SPEAKER_01: You should also have a data vaulted in an air gap place.
- 12:49SPEAKER_01: So, you know, even that backup can't be attacked.
- 12:53SPEAKER_01: It should be immutable.
- 12:54SPEAKER_01: There's a whole set of really great,
- 12:56SPEAKER_01: you know,
- 12:57SPEAKER_01: practices about how to make sure that data is protected.
- 12:59SPEAKER_01: We do all of that and we're excited about it.
- 13:00SPEAKER_01: But the next thing you do is data security.
- 13:02SPEAKER_01: So as that data is sitting there,
- 13:04SPEAKER_01: you ask really interesting questions like,
- 13:07SPEAKER_01: is there any malware?
- 13:08SPEAKER_01: Is there any PII?
- 13:09SPEAKER_01: Are there any credit card information?
- 13:11SPEAKER_01: And as that data is sitting there,
- 13:13SPEAKER_01: our data security tools scan through the data,
- 13:16SPEAKER_01: identify which data...
- 13:17SPEAKER_01: Which data has PII?
- 13:19SPEAKER_01: Which data has credit card?
- 13:20SPEAKER_01: We call it sensitive information.
- 13:21SPEAKER_01: Which data is sensitive?
- 13:22SPEAKER_01: We mark all of that data as sensitive.
- 13:24SPEAKER_01: So now you know, and you might not have known that, right?
- 13:27SPEAKER_01: So you might not have known that in some unstructured document file,
- 13:31SPEAKER_01: somebody put a bunch of credit card information,
- 13:33SPEAKER_01: right?
- 13:34And that,
- 13:34or worse,
- 13:35from like a GDPR or a privacy perspective,
- 13:38somebody put a bunch of
- 13:40SPEAKER_01: social security numbers, right?
- 13:41Knowing that would really matter to you because you have
- 13:44different governance rules
- 13:46SPEAKER_01: for different kinds of information.
- 13:48SPEAKER_01: Okay.
- 13:48SPEAKER_01: So we scan all of that.
- 13:49SPEAKER_01: We find all that out for you.
- 13:51And then we do things like scan for malware and we scan for
- 13:55malware via signatures,
- 13:56SPEAKER_01: but we also scan for anomalous behavior.
- 13:58SPEAKER_01: So we look, are any of these files behaving anomalously?
- 14:02And then we do this really interesting thing,
- 14:04which is a ransomware detection.
- 14:06And we do that by looking for an increase in entropy because
- 14:11the higher the entropy,
- 14:12SPEAKER_01: the more likely something is to be encrypted.
- 14:14So if we look at the bits and there's an increase in entropy
- 14:17and it's a anomalous
- 14:19SPEAKER_01: increase in entropy,
- 14:20SPEAKER_01: meaning it's more backed up than normal and look out,
- 14:23it's the stuff that's credit card information or PII or
- 14:26whatever is being encrypted
- 14:29SPEAKER_01: more quickly than we expect.
- 14:31SPEAKER_01: Then we can warn you and be able to say things like,
- 14:33SPEAKER_01: hey,
- 14:33SPEAKER_01: look out,
- 14:33SPEAKER_01: there might be a ransomware attack.
- 14:35SPEAKER_01: Now,
- 14:35SPEAKER_01: if there is a ransomware attack,
- 14:37one of the nice things about Cohesity is your data is in a file
- 14:40system that's
- 14:40SPEAKER_01: directly accessible.
- 14:41SPEAKER_01: So we can do something that's called instantaneous mass restore.
- 14:45SPEAKER_01: So we can go back to the last known good of your data,
- 14:48SPEAKER_01: and you can point your computers directly to our storage layer,
- 14:52and you can run your computers off of our storage layer while
- 14:55we're restoring your
- 14:56SPEAKER_01: hard drives to where they need to be.
- 14:57SPEAKER_01: So that's data protection,
- 14:59SPEAKER_01: data security,
- 15:00SPEAKER_01: data resilience,
- 15:02SPEAKER_01: being able to get the data back where it needs to be.
- 15:04But think about from an AI perspective,
- 15:06that's just how a backup system works.
- 15:08SPEAKER_01: But from an AI perspective, think about what we have.
- 15:11SPEAKER_01: We have essentially an accidental data lake.
- 15:13We have all of the data that matters to your company in one
- 15:17location with an
- 15:18SPEAKER_01: accessibility layer.
- 15:19SPEAKER_01: That means you could, and it's immutable.
- 15:22SPEAKER_01: So you can make copies of it for model building.
- 15:24You could,
- 15:25for each experiment you do,
- 15:27you might want slightly different sets of data.
- 15:30SPEAKER_01: So per folder,
- 15:31SPEAKER_01: you could add different sets of data,
- 15:33SPEAKER_01: each folder being a different model building experiment.
- 15:36SPEAKER_01: And I say model builder and not just AI,
- 15:38SPEAKER_01: because this idea is valuable for AI,
- 15:41SPEAKER_01: ML,
- 15:42SPEAKER_01: biz analytics,
- 15:43SPEAKER_01: anyone that's using past data to make predictions about the future.
- 15:46SPEAKER_01: So we already have that accessibility layer, what we're working on now.
- 15:51And we already have the ability for you to select some chunks
- 15:55of that data to have
- 15:56SPEAKER_01: a chat GPT-like interface with it.
- 15:59SPEAKER_01: That's called Gaia.
- 15:59SPEAKER_01: We'll talk about that in a second.
- 16:01SPEAKER_01: But we're also working on a discovery layer,
- 16:03SPEAKER_01: which means,
- 16:04SPEAKER_01: and our goal would be something like,
- 16:06SPEAKER_01: show me all of the data
- 16:08about this topic between these two dates and include
- 16:13conversations between Craig
- 16:14SPEAKER_01: Martell and Joel Witz,
- 16:16SPEAKER_01: right?
- 16:16SPEAKER_01: And so that we could actually get that data back.
- 16:18SPEAKER_01: And maybe we're doing an e-discovery,
- 16:21SPEAKER_01: trying to find out,
- 16:22SPEAKER_01: like,
- 16:22SPEAKER_01: did Joel leak some information that he wasn't supposed to leak?
- 16:25SPEAKER_01: Or it's compliance.
- 16:26SPEAKER_01: Did we make sure that we all did the right behaviors?
- 16:29SPEAKER_01: So there's lots of reasons you might want to
- 16:32SPEAKER_01: query your data.
- 16:32Now,
- 16:33we have this Gaia product,
- 16:34which we're very excited about,
- 16:36where you can already select.
- 16:37What you do now is you select the files you want to ask those
- 16:40questions of.
- 16:41So you say,
- 16:41I want to look at the following 10 email boxes over the last
- 16:45two years.
- 16:46SPEAKER_01: And so you select that two-year period.
- 16:48SPEAKER_01: And then you can have that conversation like,
- 16:51Show me any emails where there was an attachment and where
- 16:54someone from within
- 16:55Cohesity sent it out to someone in the press or someone at a
- 16:59particular press outlet,
- 17:01SPEAKER_01: for example.
- 17:01SPEAKER_01: Guy is a year old now.
- 17:03SPEAKER_01: We've had some really fascinating uptake.
- 17:06And one of the things we've learned is we want to make the
- 17:09discoverability process
- 17:10SPEAKER_01: a little bit better.
- 17:10Right now,
- 17:11you have to select the files that you want to have a
- 17:13conversation with.
- 17:14SPEAKER_01: We would like to make it so you could just ask, show me all files.
- 17:19There's a trade-off there because every one of these files that
- 17:22you want to have a
- 17:22SPEAKER_01: conversation with,
- 17:23SPEAKER_01: you have to do embeddings,
- 17:25SPEAKER_01: you have to index differently,
- 17:26SPEAKER_01: and that increased costs.
- 17:27So we're figuring out with our customers now the best way to
- 17:30maximize value.
- 17:31SPEAKER_00: Yeah.
- 17:32SPEAKER_00: Yeah, absolutely.
- 17:33SPEAKER_00: And I've seen...
- 17:34like legacy archiving products where that e-discovery feature
- 17:37is like such a headache.
- 17:38SPEAKER_00: It takes hours of time,
- 17:40even with a great search tool to find specifically what you're
- 17:42looking for.
- 17:43So you're saying you can actually have a conversation and say,
- 17:45hey,
- 17:45I'm looking for this.
- 17:45SPEAKER_00: No, that's not quite right.
- 17:46SPEAKER_00: I'm more looking for this sort of thing and get really granular with it.
- 17:49SPEAKER_00: That's awesome.
- 17:50SPEAKER_01: It is.
- 17:51SPEAKER_01: But let's be clear for things like e-discovery and compliance,
- 17:53you still need a human in the loop to validate the things that
- 17:56the system returned.
- 17:57SPEAKER_01: Yes.
- 17:58SPEAKER_01: are the right things.
- 17:59SPEAKER_01: But our goal is for the first stage of e-discovery,
- 18:01SPEAKER_01: which is the information gathering stage,
- 18:03SPEAKER_01: we want to make that as easy as possible.
- 18:05SPEAKER_01: Got you.
- 18:06SPEAKER_01: Okay.
- 18:06SPEAKER_00: Fantastic.
- 18:07SPEAKER_00: And,
- 18:07SPEAKER_00: uh,
- 18:08SPEAKER_00: you know,
- 18:08SPEAKER_00: in your,
- 18:08SPEAKER_00: when you mentioned earlier about making that more widely accessible,
- 18:12SPEAKER_00: is that something on,
- 18:13SPEAKER_00: on the roadmap so that,
- 18:14SPEAKER_00: yeah,
- 18:14SPEAKER_00: I guess for other business processes,
- 18:15SPEAKER_00: you can leverage that data for,
- 18:18SPEAKER_00: you know,
- 18:18whatever other use case that you might want to have access to
- 18:20all of that information.
- 18:21SPEAKER_01: Well, at this point, I'm just telling you what we think is important.
- 18:24SPEAKER_01: And then I will tell you to stay tuned for any particular announcements.
- 18:30SPEAKER_01: Right now we have Gaia.
- 18:31SPEAKER_01: It's a great product.
- 18:31SPEAKER_01: We're very excited about it.
- 18:32SPEAKER_01: And people are loving it.
- 18:33There are enhancements that you should probably stay tuned to
- 18:37see what and when.
- 18:39SPEAKER_01: Very, very exciting.
- 18:40SPEAKER_00: Okay, fantastic.
- 18:41So I wanted to kind of circle back to your time at the
- 18:44Department of Defense and to
- 18:46talk about some of the challenges you faced when you were there
- 18:48with AI and AI governance.
- 18:51SPEAKER_00: You know, what were your big takeaways from that role?
- 18:53I can imagine it being just an incredible time of change in the
- 18:56industry and what
- 18:58SPEAKER_00: an amazing position to be in.
- 18:59SPEAKER_00: So that,
- 19:00SPEAKER_01: you know,
- 19:01SPEAKER_01: ironically,
- 19:01I've taken the Cohesity job in part because of the challenges
- 19:04at that role.
- 19:05Because the story I just told you about Cohesity and what we
- 19:07want to offer is
- 19:09SPEAKER_01: something that would have been amazing to have had in that role.
- 19:12SPEAKER_01: So I was the chief digital and AI officer,
- 19:16SPEAKER_01: and I believe what they wanted is magical AI solutions yesterday.
- 19:20SPEAKER_01: But that's just not how it works.
- 19:22SPEAKER_01: Like when you walk into a new job,
- 19:24SPEAKER_01: you have to say,
- 19:24SPEAKER_01: okay,
- 19:24SPEAKER_01: what's the fundamental blockers to getting to where you want to be?
- 19:28SPEAKER_01: And at the DOD, and I've said all this publicly, so none of this is new.
- 19:31SPEAKER_01: You can find this on YouTube or in the press.
- 19:33SPEAKER_01: But the fundamental blocker at the DOD was getting the data right,
- 19:37SPEAKER_01: making the data discoverable,
- 19:39SPEAKER_01: clean,
- 19:40and accessible in a way that could be accessed from anyone
- 19:44anywhere who had
- 19:46SPEAKER_01: permission to do so.
- 19:47SPEAKER_01: So there were serious governance issues as well.
- 19:48You can imagine data in DOD has even stricter governance than
- 19:52inside of your company.
- 19:54SPEAKER_01: And so we spent most of our energy working on getting the data right.
- 19:57SPEAKER_01: So my argument was the following.
- 19:59SPEAKER_01: For the organization, the IP of the American people was the data.
- 20:04SPEAKER_01: And the real intellectual work for the DoD was the metrics.
- 20:09SPEAKER_01: What would qualify as success?
- 20:10SPEAKER_01: You have a set of problems.
- 20:11SPEAKER_01: What are those problems?
- 20:13SPEAKER_01: And how do you evaluate whether those problems are being solved or not?
- 20:16SPEAKER_01: Forget technology.
- 20:17SPEAKER_01: Forget which technology you use.
- 20:19If someone could write a Python script that solved your problem,
- 20:21you'd be happy.
- 20:22SPEAKER_01: It doesn't have to be a large language model.
- 20:24SPEAKER_01: What's the problem?
- 20:25SPEAKER_01: What data do you believe will solve that problem?
- 20:27SPEAKER_01: And how do you evaluate whether it got it right or wrong?
- 20:30So in this hierarchy of needs,
- 20:31we had data and then metrics and analytics.
- 20:34SPEAKER_01: How do we actually decide we're doing a good job?
- 20:37SPEAKER_01: And finally, at the top, we had AI.
- 20:39That doesn't mean to say that AI is the least important or even
- 20:42the most important.
- 20:43SPEAKER_01: The actual AI tools,
- 20:45there are thousands of vendors who are going to knock on the
- 20:48door and sell you an
- 20:49SPEAKER_01: AI solution.
- 20:50SPEAKER_01: And let me say it differently.
- 20:53SPEAKER_01: They're going to sell you a technological solution.
- 20:55SPEAKER_01: The current hot way of describing it is AI.
- 20:58right but eventually this will just become the norm like large
- 21:01language models will
- 21:01become the norm and it won't be called ai anymore like ai will
- 21:04be the magical next
- 21:05thing and so they're going to send you sell you a technological
- 21:08solution to your
- 21:09problem but how do you know whether that technical
- 21:11technological solution is going
- 21:12to be effective well you need to know how to measure its
- 21:14effectiveness you need the
- 21:16SPEAKER_01: metrics that's your job dod and then
- 21:19SPEAKER_01: How,
- 21:20SPEAKER_01: and if you just take that vendor and throw them at your pile of data,
- 21:24SPEAKER_01: well,
- 21:24what they tended to do was take that data and scoop it up into
- 21:27a stovepipe and they
- 21:29SPEAKER_01: would massage it and they would make changes to it.
- 21:30SPEAKER_01: And they should get paid money to make those changes to the data.
- 21:33SPEAKER_01: That's their job.
- 21:34SPEAKER_01: They're in this to make money.
- 21:36SPEAKER_01: But then it was locked in a stovepipe for a single problem,
- 21:38SPEAKER_01: a single solution,
- 21:40SPEAKER_01: a single problem set.
- 21:41SPEAKER_01: And then if you sold it again,
- 21:42SPEAKER_01: like you built it for the army and then you sell it to the air force,
- 21:45That same data had to be rebuilt again,
- 21:47and the American people were being charged again.
- 21:50I argued we need to have a separation of the data layer from
- 21:53the app layer.
- 21:54SPEAKER_01: We get the data layer right,
- 21:56SPEAKER_01: and then the app layer can use any data anywhere in that data layer,
- 21:59and not just stuff that's in a stovepipe that you're going to
- 22:02charge us for over
- 22:02SPEAKER_01: and over and over again.
- 22:04SPEAKER_01: And that was a pretty big mental shift.
- 22:08SPEAKER_01: So the argument when I entered was commanders,
- 22:11SPEAKER_01: four-star generals,
- 22:12SPEAKER_01: were looking for a single pane of glass.
- 22:14They wanted to have a view of their environment as quickly and
- 22:18as readily as possible.
- 22:19SPEAKER_01: We switched from a single pane of glass to a single plane of data.
- 22:23SPEAKER_01: And if you think about data as a plane of data,
- 22:26SPEAKER_01: then you could have any apps running on top of it.
- 22:29SPEAKER_01: And my argument then,
- 22:31SPEAKER_01: I still hold it,
- 22:32is that that also opens up for the ingenuity of the American
- 22:35people because our
- 22:37SPEAKER_01: industrial and academic complex is just completely innovative.
- 22:41SPEAKER_01: And two folks in a garage might solve this problem.
- 22:44SPEAKER_01: But if they're not big enough to actually tackle all that data,
- 22:47SPEAKER_01: then they're not going to be able to compete with the big guys.
- 22:50SPEAKER_01: So if you have this separation between data and apps,
- 22:53then even two folks in a garage could build an app that can
- 22:57consume that data and
- 22:58SPEAKER_01: deliver some real value.
- 22:59SPEAKER_01: And that change has continued.
- 23:00Like the people who followed on from me have continued to push
- 23:03that change.
- 23:04SPEAKER_01: And we'll see what happens under the new administration.
- 23:06SPEAKER_01: Evidence seems to indicate that they're not changing that direction.
- 23:10SPEAKER_00: Right.
- 23:10SPEAKER_01: I see.
- 23:11SPEAKER_01: Wow.
- 23:12SPEAKER_00: Yeah, I think.
- 23:12SPEAKER_00: Sorry if that was too much, Joel.
- 23:14SPEAKER_00: No, no, no.
- 23:14SPEAKER_00: I mean, there is complexity there.
- 23:17SPEAKER_00: But yeah, I can imagine the debates internally are just off the scale.
- 23:24SPEAKER_01: My argument was go slower to be more effective later.
- 23:27And nobody in a warfighting environment wants to go slower to
- 23:29be more effective later.
- 23:30SPEAKER_01: They have to go fast now.
- 23:31SPEAKER_01: And so that was a tight line to walk.
- 23:34SPEAKER_01: But let me say, why did I join Cohesity?
- 23:37SPEAKER_01: That story that I just told you where...
- 23:39SPEAKER_01: where Cohesity is an accidental data lake,
- 23:42SPEAKER_01: the act of backing up gets your data ready to do AI,
- 23:45SPEAKER_01: right?
- 23:46SPEAKER_01: That's what we want to build.
- 23:47We want to build the fact that you have to back up and you can
- 23:50get funding to back up.
- 23:51There's going to be a board mandate or a congressional mandate
- 23:53or some government mandate,
- 23:55SPEAKER_01: depending on the industry that you're in,
- 23:56SPEAKER_01: to back up.
- 23:57SPEAKER_01: The act of backing up,
- 23:59SPEAKER_01: if we do this right,
- 24:00will create that data layer that allows for easy accessibility
- 24:04for all of these apps.
- 24:05SPEAKER_00: That's the goal.
- 24:06SPEAKER_00: Absolutely.
- 24:06SPEAKER_00: Absolutely.
- 24:07and um yeah i think definitely we're seeing more and more
- 24:09regulation for backing up
- 24:11there's been some talk around federal regulations like gdpr but
- 24:16that seems very
- 24:17unlikely in the us from my perspective anyway or do you think
- 24:20on the topic of
- 24:20regulation and i suppose more back on the ai side do you think
- 24:24we need more
- 24:24regulations around how ai is used within companies used
- 24:29internally do you think
- 24:29SPEAKER_00: that's likely to happen i mean from your perspective
- 24:31SPEAKER_00: from your previous roles, do you think there's any more needed there?
- 24:34Or really,
- 24:35are we at the stage of innovation and pushing the innovation first?
- 24:38SPEAKER_00: You know, there's a lot of debate around what Europe's done on AI.
- 24:40SPEAKER_00: That's right.
- 24:41SPEAKER_01: I actually go back and forth here.
- 24:43SPEAKER_01: So on the plus regulation side, we have to keep people's data safe.
- 24:49SPEAKER_01: That has to be the case.
- 24:51SPEAKER_01: And it also,
- 24:51SPEAKER_01: I mean,
- 24:52SPEAKER_01: as an author myself,
- 24:52SPEAKER_01: it also has to be the case that we should figure out the copyright.
- 24:55SPEAKER_01: Like,
- 24:56is it the case that my copyrighted data is consumable by these
- 24:59large language,
- 25:00SPEAKER_01: my copyrighted stuff is consumable by these large language models?
- 25:02SPEAKER_01: You know, maybe it is.
- 25:03There's arguments on one side that says copyright was designed
- 25:07to help the American people,
- 25:08SPEAKER_01: not to help the individual authors,
- 25:10SPEAKER_01: right?
- 25:10So we'll have to see where that goes,
- 25:11but we have to answer that question.
- 25:13And we have to figure out a way where individuals are not
- 25:16endangered in any way by
- 25:17SPEAKER_01: this data being gathered.
- 25:20SPEAKER_01: On the other hand,
- 25:21SPEAKER_01: you know,
- 25:21Europe may overregulate and their default assumption may be to overregulate.
- 25:25SPEAKER_01: And we need to allow for innovation.
- 25:28SPEAKER_01: So what's that middle path?
- 25:29SPEAKER_01: I'm not 100% sure.
- 25:31SPEAKER_01: I go back and forth.
- 25:32SPEAKER_01: It's probably not as robust or as onerous.
- 25:36uh as as europe but it's probably not as wild shouldn't be as
- 25:39wild west as where i
- 25:40think we're going now like i think it has to be somewhere in
- 25:42between i know that's
- 25:43sort of a weak answer but i don't actually know where i where i
- 25:45fall yeah yeah and
- 25:47i think to be fair that that does seem to be i mean it's just
- 25:50such a fast moving
- 25:51area isn't it how can you actually regulate in an effective way
- 25:54where you're
- 25:55actually gonna do you want to regulate away the innovation
- 25:57that's yeah that's yeah
- 25:59right if you say you can't move unless you're safe well then
- 26:02we're not going to
- 26:02SPEAKER_01: build anything new ever
- 26:03SPEAKER_00: Yeah, absolutely.
- 26:04SPEAKER_00: Yeah, absolutely.
- 26:05SPEAKER_00: And you'll end up with the innovation happening elsewhere.
- 26:08SPEAKER_00: Yeah.
- 26:10SPEAKER_00: Okay, I'll just ask you one final question, Craig.
- 26:12You know,
- 26:12looking five years into the future,
- 26:14what's one advancement in AI?
- 26:16SPEAKER_00: I suppose my original question was going to be AI safety or governance,
- 26:20SPEAKER_00: but I suppose more broadly,
- 26:22SPEAKER_00: what's one advancement in AI that you're excited to see?
- 26:25SPEAKER_00: And what's one thing you're maybe concerned about?
- 26:27SPEAKER_01: Well, it's really, how about two sides of the same coin?
- 26:29What I really want to see is better theory about how these
- 26:33models work so we can
- 26:34SPEAKER_01: figure out ways to make them smaller.
- 26:36SPEAKER_01: I think a federation of small,
- 26:39SPEAKER_01: and there's lots of work pushing in this direction,
- 26:41SPEAKER_01: and I'm very excited about that.
- 26:42SPEAKER_01: But a federation of small models allows for a number of things.
- 26:46SPEAKER_01: One, more competitors.
- 26:47SPEAKER_01: Right now,
- 26:48SPEAKER_01: to build these models,
- 26:49SPEAKER_01: you have to be one of the big players,
- 26:51and you have to have hundreds of millions of dollars that you
- 26:53can use to try and experiment.
- 26:55SPEAKER_01: But smaller models,
- 26:57SPEAKER_01: federations of smaller models,
- 26:59SPEAKER_01: universities,
- 27:00SPEAKER_01: smaller companies,
- 27:01SPEAKER_01: folks in a garage maybe can start building some models.
- 27:04SPEAKER_01: You can rent compute, right?
- 27:05SPEAKER_01: Compute's not the problem anymore.
- 27:07It's the,
- 27:07I mean,
- 27:08access to it is it's the amount of compute that you need.
- 27:11SPEAKER_01: That's the problem.
- 27:12And so I would like to see smaller models federated that are
- 27:15targeted at particular use cases.
- 27:17SPEAKER_01: I'm not a fan of even the goal of AGI.
- 27:20SPEAKER_01: That's fine if we get there.
- 27:21SPEAKER_01: That's great.
- 27:22SPEAKER_01: But I don't really care about the general.
- 27:24We have a whole bunch of problems that we want to solve with
- 27:26this technology.
- 27:27SPEAKER_01: We should focus on solving the problems.
- 27:29SPEAKER_01: I don't actually care what technology solves the problems, right?
- 27:32SPEAKER_01: I mean,
- 27:32the scientist in me finds it interesting from a technological perspective,
- 27:36SPEAKER_01: but the business person just wants to solve these problems.
- 27:38SPEAKER_01: So let's be clear on our problems.
- 27:40SPEAKER_01: And let's try to find,
- 27:41SPEAKER_01: let's focus on smaller,
- 27:43SPEAKER_01: more targeted ways and not have to try to solve,
- 27:46SPEAKER_01: like we're boiling the ocean.
- 27:47SPEAKER_01: We're trying to solve,
- 27:48SPEAKER_01: we're trying to release the next,
- 27:50SPEAKER_01: you know,
- 27:50SPEAKER_01: N plus one model that now can do everything all at once.
- 27:55SPEAKER_01: Why?
- 27:56SPEAKER_01: Why are we doing that?
- 27:57SPEAKER_01: Why do we care that can do everything all at once?
- 27:59SPEAKER_01: I mean,
- 28:00SPEAKER_01: it's a good,
- 28:01SPEAKER_01: it's good marketing,
- 28:02SPEAKER_01: but I'd rather have you show up and say,
- 28:04SPEAKER_01: here's 10 things we do great.
- 28:05SPEAKER_01: And we're going to solve these business problems for you.
- 28:07SPEAKER_01: We're going to solve these,
- 28:08SPEAKER_01: you know,
- 28:08SPEAKER_01: personal problems for you,
- 28:10as opposed to having one thing that's going to solve all of
- 28:13your problems.
- 28:13SPEAKER_01: Like, I don't like this monolithic view of everything.
- 28:17SPEAKER_01: And that's the flip side.
- 28:18SPEAKER_01: The dangers of that, besides, I just don't think it's very effective.
- 28:21SPEAKER_01: I mean, it's not becoming as effective as quickly as we hoped.
- 28:24SPEAKER_01: The danger is the costs.
- 28:26SPEAKER_01: And because of the costs,
- 28:28SPEAKER_01: nevermind the potential environmental impacts of all the electricity,
- 28:32SPEAKER_01: right?
- 28:32SPEAKER_01: We've kind of ignored,
- 28:33we seem to have given up on the environment because we enjoy
- 28:36talking to these bots.
- 28:37So we've seen this,
- 28:38oh,
- 28:38well,
- 28:39it's worth the carbon if we can talk to these bots.
- 28:41SPEAKER_01: It's really weird the way we've switched on that.
- 28:44SPEAKER_01: But only a few companies.
- 28:46SPEAKER_01: It's monopolistic.
- 28:46SPEAKER_01: It's an oligarchy of a few companies that can actually do this work.
- 28:51SPEAKER_01: And I'd rather see it broader and have competition.
- 28:53SPEAKER_00: Yeah, I suppose it's a bit like the general internet, right?
- 28:56Where you started off with everyone can build a website,
- 28:58everyone can own their own business.
- 29:00SPEAKER_00: And then suddenly it's like,
- 29:00SPEAKER_00: well,
- 29:01SPEAKER_00: there's only three,
- 29:01SPEAKER_00: four,
- 29:02SPEAKER_00: five big internet companies that have dominated.
- 29:04SPEAKER_00: You've got Amazon and Microsoft.
- 29:05SPEAKER_00: And, you know, I guess the danger is the same thing may happen in AI.
- 29:08SPEAKER_00: Yeah.
- 29:09Except,
- 29:09I think that's 100% right,
- 29:10Joe,
- 29:10but let's go back to our governance question.
- 29:13The governance of the internet is such that they're not allowed
- 29:16to block other
- 29:17SPEAKER_01: folks from using it and being innovative.
- 29:20SPEAKER_01: So governance like that,
- 29:21SPEAKER_01: like net neutrality,
- 29:22SPEAKER_01: things that allow,
- 29:23SPEAKER_01: like the backbone of the internet,
- 29:25SPEAKER_01: when it moved from the NSF net to the internet,
- 29:28the folks who got to own pieces of the backbone were required
- 29:32to make sure they
- 29:33SPEAKER_01: transported all traffic and couldn't block any of it.
- 29:36So that kind of governance,
- 29:38which allows for innovation to flourish,
- 29:40I'm on board.
- 29:41SPEAKER_01: Absolutely.
- 29:43SPEAKER_00: Craig, thank you so much for joining us today.
- 29:44SPEAKER_00: It's been an absolutely fascinating conversation.
- 29:46SPEAKER_00: Yeah, thanks.