Latest / The Road to Accountable AI / DJ Patil: AI's Steering Wheel Challenge

Transcript

DJ Patil: AI's Steering Wheel Challenge · The Road to Accountable AI

• 0:01
  DJ, glad to have you on the podcast.
• 0:03
  Thanks, good to be here.
• 0:05
  You were the first chief data officer of the US, now something like 15 years ago, back in the Obama administration, when I was also in government.
• 0:14
  Is that a role that governments should still have today?
• 0:19
  So the short answer is yes, but one clarification is the chief data scientist and it's actually it's a You'll appreciate this because you're in public service like these subtle
• 0:28
  distinctions actually make a Big important deltas in the role and part of the reasons to chief data scientist
• 0:38
  is that the mission is to responsibly help the president and the president's staff think about how to responsibly unleash the power of data to benefit all Americans.
• 0:50
  And then in that, you're providing the advice versus sort of saying, what is the control of all the data in the US?
• 0:57
  And now a lot of that has been codified into the chief data officers at the federal level, and the federal CIO has it.
• 1:05
  But if I think about oh
• 1:07
  Why do you want this role?
• 1:10
  What's important about it?
• 1:12
  You know, all the conversations around AI, like what's the foundation of this?
• 1:17
  What's the substrata of this?
• 1:19
  It's data.
• 1:20
  It's making sure you have good data.
• 1:21
  It's making sure that you're using that data responsibly.
• 1:24
  How do you get the value out of it?
• 1:26
  What are the application layers?
• 1:28
  And then what's the data that the AI produces?
• 1:30
  And how do you think of it and utilize all of that?
• 1:33
  And so if anything, I think the role is substantially more important than it had been previously.
• 1:43
  Today though, we're in an era where you don't need a whole lot of training to be a data scientist, or at least to do data science when you can just put in a prompt and it'll
• 1:52
  write code for you and so forth.
• 1:54
  Does that change the nature of thinking about data science and its role in policy or in business?
• 2:01
  I think it is changing.
• 2:03
  to your point, one of the really fine, it's like, finally, we have these tools that are getting rid of the yuck work.
• 2:11
  It's still too often, still cleaning a lot of data, bringing data sets together, a lot of that just like the very painful thing.
• 2:20
  And now we're starting to get into a space where big data won.
• 2:25
  And what I mean by that is like, we got tons of data in the data warehouse.
• 2:30
  We have tons of dashboards and now we're getting paralyzed.
• 2:34
  We don't get to decisions.
• 2:37
  We don't actually have the ability.
• 2:40
  We have the ability to parse tons of data, but what's the decision we're trying to make?
• 2:43
  Are we asking the right question?
• 2:46
  And this is the place where the data scientist is now moved to this higher order function.
• 2:52
  And it's like, great.
• 2:53
  How do we cut through that stuff?
• 2:55
  and actually get to an answer that's there.
• 2:59
  actually, in fact, we just announced one of the companies I've been working on for the past year or so called Sundial, which is a lot of the former Facebook team is really
• 3:09
  building this analytics layer that says, hey, what is the AI native version of dashboarding and business intelligence and all those other things?
• 3:18
  Because now it's how it's time to get to decision, not time to
• 3:25
  to a chart.
• 3:28
  What exactly is the difference though?
• 3:30
  Yeah, so right now, what you see very often is people have the ability, like, what is a day in the life of a data scientist?
• 3:38
  Like, right now.
• 3:40
  Right now, it's still like, hey, I'm trying to get a graph that shows sales figures.
• 3:46
  I'm trying to get a graph that shows pick your favorite thing.
• 3:49
  And then you go, man, I got to go find that data in the data warehouse.
• 3:54
  If a data warehouse happens, I might have to combine external data and I have to put all this stuff together.
• 4:00
  And then you get a chart and then somebody says, oh, not that chart.
• 4:03
  And you go back and you kind of redo it again and again and again.
• 4:06
  And so you ended up with all
• 4:11
  And many times people want to automate that.
• 4:12
  So you end up with dashboard after dashboard after dashboard.
• 4:15
  But then what happens is people really get to say, and the data scientists have this when they have bandwidth, is to say,
• 4:22
  What chart, what is the right question to ask?
• 4:26
  Should we be asking that type of question or another one?
• 4:30
  I'll give you a more concrete version.
• 4:32
  Is you saw this very quickly in the early days of something like the COVID crisis, where people were saying, here's a model.
• 4:39
  And then you say, well, here's a strategy.
• 4:42
  And then you say, what does the model say?
• 4:44
  Oh, no, it's going to take us like two months to go build that into the model.
• 4:47
  And you go back and so you can't keep up.
• 4:51
  with the questions that are being asked.
• 4:54
  Or what happens is your feedback cycle isn't sufficient that you're able to say, hey, I think you mean to ask this.
• 5:03
  This is the more fundamental, interesting question to ask.
• 5:08
  So to get back to the first question I was asking you about government, presumably it's the same challenge.
• 5:13
  If you're a federal agency or a business, you need to ask the right questions.
• 5:19
  People, excuse me, at this call.
• 5:24
  People often suggest that government is just inherently too slow for everything.
• 5:29
  In this kind of era, where presumably there's going to be even more pressure for who can get to the right question and answer it more quickly, what should government be doing?
• 5:39
  Well, one, think, and especially what we're seeing right now, unfortunately, is we're dismantling a lot of the key infrastructure that gives us good data.
• 5:49
  That could be weather data, and we've seen how important that is just very recently with the tragedy in Texas and other type of weather events that have been happening.
• 5:59
  You know, that's high.
• 6:00
  We have a very good system to create high quality data in those worlds.
• 6:05
  Our economic, much of our economic forecasting is built off a pipeline of aggregating data from lots of sources, making sure it's very clean.
• 6:15
  understanding traffic fatalities, is our food system clean and acceptable?
• 6:21
  Is our delivery system of pharmaceuticals good?
• 6:27
  All of this stuff has gone through decades of hardening in the sense of making sure that's very, very good.
• 6:35
  Could it be improved?
• 6:36
  Absolutely.
• 6:37
  But what happens is as soon as you start cracking that infrastructure,
• 6:44
  we learn about how much is actually built on top of it.
• 6:48
  Think about how much actually depends on census.
• 6:52
  Not only the way we distribute Congress is based on census.
• 6:58
  And if we do that, if you don't do a good job of that, you're not allocating representation correctly, but so much more financial and the way we just allocate
• 7:09
  resources is done with that.
• 7:11
  So that kind of gets to one thing.
• 7:14
  Now to get to your other part is government.
• 7:18
  So there is a slowness that you want on government on certain things.
• 7:22
  There's a speed on other aspects and how do you unleash that data more quickly?
• 7:27
  And with that, for example, we should be able to go a lot faster with medical data.
• 7:34
  It's still crazy that in this day and age that how much effort is to go get your own medical data.
• 7:41
  You know, that was a lot of work that that President Obama drove and and it was subsequently driven by both the Trump administration and the Biden administration, but
• 7:51
  it's still too hard.
• 7:52
  And so if you want to use an AI system to look through all of your medical information between multiple doctors and hospitals,
• 8:04
  systems or something where you may have complex care, that's non-trivial to do.
• 8:10
  You have to download stuff, it's just tough.
• 8:13
  So there's a lot more to be done.
• 8:15
  And then the other thing that's there is there's a tremendous amount of data that sits inside the federal government.
• 8:21
  that has yet to be unlocked.
• 8:23
  We haven't even used it yet.
• 8:26
  And the opportunity to use that data to get massive benefits for the American public is very real.
• 8:33
  To be very specific, there's a project in the VA called the Veteran Affairs called the Million Veterans Project.
• 8:41
  Unbelievable data set of biological samples getting contributed by veterans to improve health care.
• 8:47
  That is one of the great, you know,
• 8:51
  collections and what is sitting in that, that data could fundamentally change the trajectory of the speed at which of figuring out other diseases.
• 9:01
  And so the fact that we haven't even begun to tap into that is, in my opinion, just a loss like that that needs to take place now.
• 9:11
  What stands in the way of unlocking that government data?
• 9:14
  One part is funding.
• 9:16
  A big part of it is funding, because these things are just not resourced appropriately.
• 9:20
  uh Too often, we found, uh I think over when I was in that role, as well as subsequently, what we found is these are unfunded mandates.
• 9:32
  We say, we should get that data out there.
• 9:34
  And the people who are responsible for that data, they're stuck behind five other tasks.
• 9:40
  That would be nice if it's only five.
• 9:43
  But we haven't really embraced it uh in that sense.
• 9:47
  Two is you have a process that does need to change.
• 9:52
  You have CIOs and other general councils whose job fundamentally is to say no.
• 9:58
  Because no is a safe thing to do rather than asking, how could we do this?
• 10:04
  What could we do to, what are the right safeguards to do this that people could benefit from?
• 10:09
  And then the third is Congress many times.
• 10:13
  What happens is the government opens up some data and then somebody else gets really angry that you're doing that because they have figured out how to get a contract or something
• 10:26
  for that data and now they feel threatened
• 10:28
  that everyone's, and it's the public's data.
• 10:32
  And so there is this version that you see that people go to Congress to complain and Congress has hearings about why are you opening up this data?
• 10:40
  And you're like, wait, weren't you just yelling at us about not moving fast on the data?
• 10:46
  Now we can.
• 10:47
  So uh it makes it a very, we should really be aligned, I think fundamentally.
• 10:54
  on what does it mean to responsibly unleash this data to benefit all Americans.
• 10:59
  And if you do that as your true North, I think we get to a set of data that are very obvious that that investment is going to pay off massively in the next five, 10 years.
• 11:12
  So it's interesting to compare what you're saying in the public sector with government data with
• 11:21
  analogous kinds of issues in the private sector.
• 11:23
  uh Maybe not as much with unlocking data, but there's still, I think, these incentive problems in companies.
• 11:31
  Sorry, let me take this again.
• 11:34
  Apologies, normally I'm not this bad.
• 11:37
  Let me just take that question from the top, sorry.
• 11:41
  It's interesting what you were saying about uh the challenge of unlocking that data in government, because it makes me think that there are some similar issues in the private
• 11:49
  sector.
• 11:50
  It's still a question of alignment of incentives.
• 11:52
  have general councils and risk and compliance people who want to say no.
• 11:56
  uh You have different groups who have silos and so forth.
• 12:01
  But at least in government, you can ultimately point to, have a mission.
• 12:05
  This is to serve the public.
• 12:07
  In a company to turn to the private sector, what are the leverage points to get organizations to unwind on taking advantage of this AI opportunity?
• 12:20
  So I think you raised a number of really interesting points when we moved to the private sector.
• 12:25
  So first is companies may not want to release that information for various versions.
• 12:31
  One is there have been plenty of cases where people have tried to release data and then somebody finds something in that data that shouldn't have been released and then that just
• 12:40
  causes...
• 12:41
  a problem, it's a headache for the company, and then they say, hey, why are we doing this?
• 12:46
  And most often that benefits researchers.
• 12:50
  There's another side that people don't want to, they find that if they open that data up, they're gonna have a degree of transparency that people are uncomfortable with.
• 13:00
  One of the recent cases about this is Facebook and transparency around...
• 13:05
  the what type of information is how much misinformation is happening, other things.
• 13:11
  And they've continued to pull back from that data.
• 13:14
  And it's not just Facebook, it's the problems on YouTube.
• 13:18
  is pick any of platforms that this is an issue around this topic.
• 13:23
  What type of harms are happening to kids?
• 13:26
  We know that there have been studies that have been able to do this, but there's very hard to get the data to access it.
• 13:32
  So you see that form of it.
• 13:33
  You see another form, which is protectionist, in that you're seeing a version of this happening with AI right now around a Gentic system.
• 13:42
  So it's been reported on is there are a number of the big platforms like Salesforce, Atlassian, others are saying, sorry, we don't want you to have access to everything in
• 13:53
  Slack or some other platform because we, their argument is that even though it's your data, we're protecting it.
• 14:02
  And it sounds very much like that.
• 14:03
  healthcare conversations, I think most companies are saying, wait a second, that doesn't feel like the contract that we agreed to that says our data, we're using it, we want to it
• 14:13
  used to her.
• 14:14
  And so they hold it close.
• 14:16
  And we saw this with the email providers of not allowing export, could you also take your network, your social network and graph and move it?
• 14:24
  So this is like a well-known version of things that have been going on.
• 14:29
  And then you see another side, is very much exactly what you see in the government is one team says, Hey, we need data from over there.
• 14:40
  And that would make this great product over here or we need to do it.
• 14:44
  And those parts of the organization are literally in a fight.
• 14:48
  And it's, it's, it's exact.
• 14:49
  It's, it looks exactly the same.
• 14:52
  And the bigger the company, the bigger the set of issues that, that, that are there.
• 14:56
  And usually it takes like some special initiative or somebody to
• 15:00
  to get there.
• 15:01
  And then the other one which I would say is very often the infrastructure of some of the bigger companies are not ready to leverage this data in the way that they technically need
• 15:11
  to be able to have in place.
• 15:14
  All right, so let's take a step back.
• 15:16
  What excites you most about AI today?
• 15:21
  Couple of things that I see that are the most exciting in AI right now.
• 15:24
  It's one, how much, if I, actually, sorry.
• 15:29
  What excites me the most about AI right now, if there's one thing above all else, it's actually how many people are working on it.
• 15:36
  If you look at...
• 15:38
  you talk to any college student kid, they're playing with something.
• 15:42
  They're not just playing with one thing, they're playing with all these things.
• 15:45
  You say, hey, how do you use AI?
• 15:47
  You learn all these incredible things about what they're doing.
• 15:50
  And you're just like, I never thought that.
• 15:52
  And it feels like the same way when mobile.
• 15:55
  was happening, you're like, I'd never thought about using the mobile device that way.
• 15:59
  They were building apps and doing all this.
• 16:00
  So we see just the raw number of human calories that are going into using this kind of foundation, foundational technology.
• 16:11
  to apply it to different areas.
• 16:12
  So that's one.
• 16:14
  Two, I am constantly surprised, both in my personal life and talking to other people and I hear from, in the various ways that they're using AI, whether it's something mundane like
• 16:26
  planning, you know, what to do with the kids on a rainy afternoon to taking care of somebody that, you know, they're trying to interpret.
• 16:36
  lab results or they're trying to understand what the doctor is saying because they're talking so fast.
• 16:42
  And what I found, you when I was working on precision medicine initiative, Cancer Moonshot, is we often dismiss the patient or the caregiver and say, oh, they need to have
• 16:55
  the doctor to understand this.
• 16:57
  But they get so much value from turning to Google.
• 17:01
  or pick your favorite search engine, that they now have jumped to that with the AI systems to manage the complexity or to say, what questions should I ask my physician?
• 17:13
  What should I ask my primary care physician?
• 17:15
  What should I ask the oncologist?
• 17:18
  And I see those things.
• 17:19
  I see people who are preparing flashcards and other things.
• 17:23
  And so the amount of innovation here is extraordinary.
• 17:28
  There is that flip side also that I see those people who are being left out because of the cost structure of these tools.
• 17:36
  see people whose careers are getting about, if not already getting disrupted, you see it coming very soon.
• 17:44
  And so that combination is front and center right now.
• 17:52
  Although do we deal with the challenges if people are using it for all of these situations, but we know these systems are imperfect.
• 18:01
  They may hallucinate bias, all the problems that data scientists are very familiar with.
• 18:08
  But now it's everyone using it for everything.
• 18:12
  So I think a few things that are there and there.
• 18:14
  One is, and this is getting really hard, the more we adopt these systems.
• 18:19
  There is an EI laziness, a complacency that happens, no different than when we trust, know, map software.
• 18:28
  Like you trust Google Maps, Apple Maps, pick your favorite one.
• 18:33
  You know, people just blindly drove off cliffs or people drove into lakes just because they were like ignoring reality.
• 18:39
  I used to do that in my class.
• 18:41
  There was an Uber driver who in San Francisco wound up stuck on a stairway because it showed that there was a road there, it actually literally a staircase and the car couldn't
• 18:49
  move.
• 18:49
  Yeah.
• 18:50
  they ignored all reality to do this.
• 18:55
  And so what I think there is a very real concern about what does this look like?
• 19:02
  when those type of issues happen.
• 19:05
  we saw this with search engines, like with people said, I want to harm myself or do something where self harm would happen.
• 19:13
  You'd get a great recipe to do so.
• 19:15
  mean, it was terrible until these issues were highlighted.
• 19:19
  And so what is that level of responsibility in the platforms for what you might call base level sort of duh kind of protections, like the obvious things.
• 19:32
  a set of things as you move up and I think it is very clear that these models are improving, getting better, substantially better, but the thing that has to happen in
• 19:43
  there, like if you're thinking about making a medical decision with this, you should be always thinking of this as a co-pilot, not the pilot, and I think that's easy for us to
• 19:54
  say.
• 19:55
  but as more and more people adopt it, and like the example you said of people, know, the Uber driver, the example they use in your class, that...
• 20:03
  exemplifies the problem we're going to have in society here.
• 20:08
  And there isn't an easy answer overall, but I do think of it as a steady, very aggressive testing approach, trying things out, red teaming.
• 20:19
  We actually wrote a book on this, a small book called Data and Ethics, to really kind of give practical ideas that you can do to get a sense of
• 20:29
  how dangerous the technology might be as you're building it, and what do you want to do and put guardrails in place for it?
• 20:37
  Yeah.
• 20:38
  What's the argument though to convince organizations to do that given that there's so much pressure to move as fast as possible?
• 20:44
  Yeah, so I think what we're in a version which I might call breaks versus gas.
• 20:50
  And there is a version of what I would say, know, previously, you know, over the last several years, we've been in a largely at least from policymakers as a breaks modality is
• 21:03
  like, how do we put breaks on this?
• 21:04
  How do we slow down?
• 21:05
  And many technologists were saying, we should stop development on AI.
• 21:08
  So hard breaks.
• 21:10
  We're right now in a moment where I think I would describe it as almost all gas, right?
• 21:14
  Everyone's like, it's a race for China.
• 21:16
  It's a race against each other.
• 21:18
  There's new opportunities.
• 21:19
  So it's go, go, go, go.
• 21:21
  We don't have the steering wheel.
• 21:24
  That's the problem.
• 21:25
  Policy is a steering wheel.
• 21:27
  And what does that happen at
• 21:30
  the state level, what has a federal level.
• 21:32
  So at the state level, there's some, what you might call duh legislations also, like these AI systems should not be allowed to use to generate porn or harms of kids or impersonate
• 21:47
  people without their uh permission.
• 21:50
  It should not be allowed to use to die healthcare or things like that.
• 21:54
  There's like a bunch that are obvious.
• 21:57
  despite all of those being put in, the number of bills that have been turned into law across the country, there has been next to nothing on the enforcement mechanism.
• 22:10
  And so how do you control this?
• 22:14
  There has to be that.
• 22:16
  And then the other portion of this, and this is where I think it's really important in academia, is that we have to teach everybody
• 22:26
  ethics, the notions of morality, civics, all of these things have to be integrated as we talk about what does it mean to not only build these technologies but use them in
• 22:37
  different examples and cases.
• 22:42
  No question.
• 22:42
  that's good for me to hear, because that's what we do in business school, at least at my institution.
• 22:47
  um
• 22:52
  the thing, you're exemplifying one of the biggest challenges that we have, is there is so much being done in pockets.
• 23:03
  We have yet to figure out where the right convening, the right molding place, how do we get the right people, the right organizational structures to take place inside the
• 23:15
  building of these AI systems to allow this to happen?
• 23:18
  Just in the same way that we went through in Web 2.0, we really realized that, hey, you can't have the general counsel way often on one side of the building.
• 23:29
  You have to actually embed the lawyers and policy people into the team.
• 23:35
  These became known as product counsel.
• 23:36
  A new model had to emerge to allow building these things correctly.
• 23:41
  That same version of iteration is now needed for these kind of technologies.
• 23:47
  Yeah, it's interesting.
• 23:48
  It's a good point.
• 23:49
  You're talking about organizational innovation.
• 23:50
  It's not the technical innovation.
• 23:52
  It's how we wrestle with this new complexity.
• 23:55
  I completely agree.
• 23:56
  That's right.
• 23:58
  oh What's your view about the speed of change that we're going to see going forward with AI?
• 24:04
  Will the next five years be as fast as the last five years or what do you think will determine that?
• 24:10
  Yeah, so here's my framework of how I've gotten my head around it is, you know, I think of these things in rate of change.
• 24:18
  So how fast is society changing?
• 24:20
  So one way to think of it is a basic company, a big company has like one major release a year.
• 24:27
  Car companies like, hey, this is our 2025 model or 2026 model, those type things.
• 24:32
  Apple, Google, they're like, here's our phone for the year, or here's our big operating system update.
• 24:37
  That's a 1x.
• 24:39
  Startups have an advantage because they typically release three times a year.
• 24:43
  So they are operating at a 3x versus a 1x.
• 24:47
  The crazy thing is AI companies are releasing at 7 to 10x.
• 24:53
  So startups are already on an exponential compared to the sort of flat, if you will, of traditional companies.
• 25:02
  AI companies, and they're using AI themselves to kind of keep on that 7 to 10x rate of progress.
• 25:11
  So the AI companies are rapidly outpacing everybody else.
• 25:15
  Meanwhile,
• 25:17
  policy changes roughly every 20, 30 years.
• 25:23
  So we're at a oh 0.02X versus a 10X.
• 25:29
  So that delta, it means an incredible imbalance.
• 25:32
  Culture inside companies changes decadally.
• 25:36
  So that's a 0.1X.
• 25:37
  So we go from 1X uh on the base level company, culture changing at 0.1.
• 25:44
  oh 0.02x for policy and 7 to 10x.
• 25:48
  So that imbalance means that we're set up for one of three things.
• 25:53
  Are we going to have a hard landing of society as the technology changes?
• 25:58
  Are we going to have a soft landing?
• 26:00
  And what opportunities are there?
• 26:02
  What accelerants are there?
• 26:03
  And the accelerants go to the companies that are leveraging the technology.
• 26:09
  But now we have to think about what does it mean for policy and other things to happen for mitigating a soft landing.
• 26:17
  And so the conference that I've been arguing for that I think we need to be having in every state level is to say, and the federal level to say, here are our 10 biggest
• 26:29
  economic drivers.
• 26:31
  What is AI going to do for them?
• 26:34
  and what is AI going to change and possibly hurt on those economic ones.
• 26:40
  Then we should have our 10 societal issues and what could AI do for it and what could it hurt.
• 26:47
  Here in California at least, agriculture, Hollywood, classic technology, a la Silicon Valley.
• 26:56
  What can AI do and not do there?
• 26:58
  Societal issues, homelessness, uh climate change.
• 27:03
  uh education, healthcare access, where can AI help?
• 27:08
  And that should define where and what populations are going to get impacted and what we can do both to create opportunities and mitigate that hard landing.
• 27:19
  It makes a lot of sense, but then I think even when you get to that point, so let's take education, we can see lots of incredible things that AI can do for learning.
• 27:30
  We can also see all kinds of problems, starting with students not having to think and just using the, pushing the button with the AI and doing a prompt as a substitute.
• 27:40
  And so just identifying that doesn't necessarily get us any closer to solving the problem.
• 27:47
  Yeah, so you're absolutely right.
• 27:49
  So from that, we need kind of concrete action plans and attempts to do things.
• 27:54
  The thing that I think is the most interesting right now, especially in the AI space, given the pace at which it's working, is every school, like just to stick on education,
• 28:03
  every school is an experiment.
• 28:06
  And different teachers are different experiments.
• 28:09
  And so some are finding ways that it works well, others are finding where it's a dud.
• 28:15
  And what we need to figure out is what are the templates from that?
• 28:20
  What can scale out of this?
• 28:22
  In the same way, I'm sure you've seen, I have seen in the higher academia, like in the university levels, other places, I'm always like, wow, that's a really cool use.
• 28:35
  I didn't think of incorporating AI that way into the class.
• 28:39
  And then once I see that as a pattern, I go, oh, I should do that.
• 28:45
  It's still hard to leverage that learning.
• 28:49
  And part of it is because I think we've, a lot of those other places, especially like in education or places where there's been the unions have largely said, I'll get all breaks.
• 28:59
  And there's also parts where it's good reason to say breaks, to be clear.
• 29:03
  But they've been so hard on breaks, they haven't given space for people to say, how are you using it?
• 29:08
  What could be done?
• 29:10
  And what might work?
• 29:13
  We saw this with Wikipedia where people were like, Wikipedia bad, you should never use it.
• 29:18
  You should go to the library and use Encyclopedia Britannica.
• 29:21
  And then finally we were able to say, oh wait, this actually works pretty well.
• 29:27
  Yeah, it's interesting.
• 29:29
  Similar to what we were talking about at the beginning about government, there is this tension though.
• 29:34
  On the one hand, experimentation is great.
• 29:38
  On the other hand, you can't just have everyone running in different directions.
• 29:42
  Ultimately, the organization needs to get everyone on board.
• 29:46
  You need to say, if you're getting a college degree, this is some aspect that you're getting out of your education.
• 29:52
  It's not just we happen to be in my class, you learn well, some other class you don't.
• 29:57
  So how do we take advantage of that potential of diversity, but at the same time, get the learnings out
• 30:06
  Yeah, so the evaluation frameworks for how we think about what is qualified, so to speak.
• 30:14
  That is a rapidly changing surface.
• 30:21
  And I would argue it's been changing before this.
• 30:24
  A lot of people were saying, how do you evaluate if somebody is good in one of these newer roles?
• 30:29
  What does it mean to be good as a software developer?
• 30:31
  Hmm
• 30:31
  Especially when you're like, you don't have to, you know, we used to have these things where he said, here's a laptop, write this code from scratch.
• 30:40
  And then we realized that's stupid because no one actually writes code from scratch.
• 30:45
  You go pull code from this GitHub repository in this place, you cut and paste templates together to go fast.
• 30:52
  It's a whole reason open source got created.
• 30:55
  You would never do it again.
• 30:57
  Like we don't write assembly code.
• 30:59
  We have apps that on top of it.
• 31:01
  em That same way we've had to move to the different frameworks of evaluation to the same way I think we're going to say what does it mean to be good at something.
• 31:12
  The thing that I think that is more so in at least education for me is portfolio of work.
• 31:19
  What have you done?
• 31:21
  Like, can you show things that are project-based?
• 31:23
  What have you built?
• 31:25
  Can you show your skills in practice?
• 31:28
  And, you know, I think that version that is very true in society, but we don't talk about it, is no one is really looking at your college grades anymore.
• 31:38
  People are looking at what you did in the internship.
• 31:42
  What did you perform there?
• 31:44
  Especially in the business school, I imagine, at least in my case, everyone calls a person's manager where they interned and said, where do you rank this person?
• 31:54
  What is that?
• 31:55
  I'm not even evaluating what classes they took per se.
• 31:59
  I'm putting them through a very different process to understand what can they do.
• 32:03
  And I'm really looking at what have they done to inform the...
• 32:10
  How do we set them up for success in an organization and cultural fit to all those other things?
• 32:15
  Sure.
• 32:16
  So this gets to a broad question about what the role is of humans.
• 32:22
  What you're talking about in this, particular example is you still get value out of having the manager give you that color.
• 32:29
  One of the things that's happening going forward with AI is, you know, we have this notion of value of humans in the loop, but as we give them more autonomy with agentic systems,
• 32:39
  there's less of that, at least in the process of getting to the result.
• 32:44
  So,
• 32:45
  How do you think about that step change for AI and what it means?
• 32:49
  Yeah, so there's a couple things that first, let me go back to some of the about ethics is people come into companies and think about the roles.
• 33:00
  One of the things that when any student is interviewing or anybody is interviewing in general, if you're ever interviewing anybody else, there's interviewing with me, they're
• 33:08
  going to get an ethics question.
• 33:11
  They're going to get an ethics question in there.
• 33:13
  And the only answer that's bad is, I don't know what to do.
• 33:21
  These are designed to be non-trivial, hard problems oh that you have to talk to, think through.
• 33:29
  And so the same way any person who's interviewing should ask the company, how do they handle uh ethical issues?
• 33:40
  What's their process?
• 33:41
  What do they do?
• 33:42
  And when people do that, start to, the company goes, wow, this is important and starts to really focus on it.
• 33:47
  There's a similar version of that with AI systems, agentic systems.
• 33:52
  And we definitely see that starting to merge where people are coming to companies are saying, which AI systems am I allowed to use?
• 34:00
  What am I not allowed to use?
• 34:02
  It's kind of like that version of like, do I get to use Microsoft Word back in the day, or do I have to still use a typewriter?
• 34:08
  You know, like,
• 34:10
  uh Maybe we just dated ourselves here.
• 34:14
  you got measured on how many words you could type for a while.
• 34:17
  And then that became a stupid measure.
• 34:19
  They were like, no, we don't even need to look at that anymore.
• 34:22
  You didn't put it on your resume back then.
• 34:25
  So people are saying what tools there are.
• 34:27
  And I think there's a flip version where also companies are like, which ones do you use?
• 34:31
  And what are you familiar with?
• 34:33
  That same version we're going to start seeing with agentic, which is like, you should be asking, well, how is the company thinking about this role with relation to agentic systems
• 34:45
  or other type things?
• 34:46
  And that tells you a little bit about the longevity of those kind of things.
• 34:51
  So in a hospital situation, they might say, what are you talking about?
• 34:57
  We're not going to be using agentic here.
• 34:59
  But other places, they may say, yes, this is really a clearly
• 35:03
  a temporary, temporary maybe two years kind of thing.
• 35:09
  Now,
• 35:10
  That's like one side of what the handshake is as you try to understand what the company is thinking about on agentic.
• 35:17
  Overall on agentic systems, think the thing and you know, what I think we're still in the very early beginnings to understand what agentic is.
• 35:26
  And the version I would tell people is when we had the opportunity to write the first national strategy on AI and autonomous systems, all these type things, what did we get
• 35:38
  wrong?
• 35:38
  We kind of thought, blue collar
• 35:40
  workers are the ones that are going to get disrupted first.
• 35:43
  And we realized, oh, dexterity is really hard.
• 35:47
  Creating robots is hard.
• 35:49
  Self-driving cars, way harder than anybody thought.
• 35:53
  And so there's a version in this which is, Egentic looks like, it's like, okay, and here it is, it's going to wipe out all this stuff.
• 36:00
  It's not that that's not yet the case.
• 36:02
  And I'd say at least for three to five years, we're going to be in this process of figuring it out.
• 36:10
  And there may be new companies that are fully agentic and take advantage of it.
• 36:14
  But there's a lot of things elsewhere.
• 36:16
  It's not going to be there.
• 36:18
  But no matter what I think of it as agentic.
• 36:22
  Well, a lot of these things is going to help you put stitch together your various apps or workflows.
• 36:28
  And in ideal, it's going to, it should make your workflow a lot easier because like one of the things that I see in customer service centers over and over again is somebody's like,
• 36:38
  or, know, when you call, they're like, sorry, I got to transfer you to this over to this person because they handle this.
• 36:44
  And then you're like, but I still need to get transferred back for this other thing.
• 36:47
  And you're kind of going back and forth.
• 36:49
  In theory, the agentic stuff should be able to pull all of that together so that it makes that customer service person's ability to be effective more so.
• 37:00
  And I think that customer service person is no longer a customer service person.
• 37:05
  They should be more of a concierge person that is there to give you the taste and the feel for all these things that make it work.
• 37:15
  It's fascinating how you've now come back to what you started when I asked you the question about 15 years ago in government, this challenge of putting together data as
• 37:27
  breaking through silos.
• 37:28
  We've come a tremendously long way, but we're still struggling with that same issue.
• 37:33
  That's absolutely right.
• 37:35
  if you think, as I mentioned earlier, the healthcare one is just front and center.
• 37:39
  It's like, what prevents us from moving faster on cancer?
• 37:45
  And I tell people all the time, it's like, it's not just a database.
• 37:49
  It's like 10,000s of databases scattered across the country where all this data sits.
• 37:55
  And the belief is, is if you pulled all that together,
• 38:00
  there are natural, like there are already solutions to different types of cancer out there.
• 38:07
  We just don't know what question to ask.
• 38:10
  And we've seen this over and over again.
• 38:12
  And then people ask like, well, why should we do it now?
• 38:15
  And you go, well, the next big, like not only is there, like we're all gonna experience this, whether we...
• 38:23
  suffer directly from cancer, we're going to have a loved one who goes through it.
• 38:25
  So there's a human direct aspect to it because this is a species disease, this is a genetic disease, a genomic disease.
• 38:33
  But then there's this other portion that is there is, well, who gets to these cures first?
• 38:40
  And you do have these big systems.
• 38:43
  know, this is why places like in China, they're trying to bring all this data together to kind of bring this next nouveau type of precision medicine in place.
• 38:53
  When we launched precision medicine, initially with like a couple hundred million dollars in funding, China responded with like eight or nine billion.
• 39:02
  You know, we finally moved our marker up to one billion, but they're still going hard at this because the next set of transformations in
• 39:10
  healthcare they want to own and they need it for their aging population.
• 39:15
  We need it.
• 39:15
  We get productivity.
• 39:16
  The healthier our population is, the cost bankrupt people.
• 39:20
  So if we can use AI the right way, is this.
• 39:24
  Can I go on a high horse for a second?
• 39:27
  So this is my angry moment.
• 39:29
  So if we think about healthcare, one of my biggest...
• 39:36
  problems about the space is we've done an amazing job of using healthcare data with the purpose of denying people care, charging more, and getting economic efficiencies.
• 39:49
  We're making physician burnout a problem.
• 39:52
  We're making the lines longer.
• 39:54
  All the things are wrong because of using data.
• 39:58
  We have yet to create the incentive structure to use this data to improve quality of care, provide better treatments, make efficiencies, better long-term outcomes.
• 40:11
  We have shown we can do it economically.
• 40:15
  by squeezing people for more, doing more, but we have not used data in any material way.
• 40:21
  We've got lots of great examples.
• 40:23
  We can identify sepsis earlier.
• 40:25
  We can do chronic kidney disease identification faster.
• 40:27
  We can do tons of stuff.
• 40:30
  We won't do it because the incentive structure isn't there to use data to help people.
• 40:37
  No, it's a very, very deep, hard, human problem.
• 40:40
  uh we still have a long way to go, but as you say, the opportunities are there.
• 40:46
  Unfortunately, we've got to wrap up, but DJ, it's really been a pleasure talking with you.
• 40:50
  As always, thank you.