Latest / Agentic Conversations (formally mlops.community) / From Shiny to Strategic: The Maturation of AI Across Industries // David Cox // #303

Transcript

From Shiny to Strategic: The Maturation of AI Across Industries // David Cox // #303 · Agentic Conversations (formally mlops.community)

0:00
So my name is David Cox. I take my coffee black Nespresso
0:04
style though, so I don't actually make it, I just push
0:06
the button. We're back with another
0:09
Envelopes Community podcast. I'm your host, Dimitrios, and
0:11
talking with David, we got in to the ways that you can look at
0:19
machine learning, specifically on the supervised machine
0:22
learning to help you change the way that you interact with this
0:26
world. Behavioral economics was
0:30
something that I learned and I hope you do too.
0:33
Let's get into it with him right now.
0:36
And I will say just as a disclaimer, this was a different
0:41
kind of philosophical conversation.
0:43
We did not get super technical on how you can deploy these
0:46
models and what kind of specs we're looking at, what kind of
0:50
QPSS and all those other fun acronyms that we like to use.
0:54
But I still had a blast talking to him and going on all kinds of
1:00
tangents, and I hope you do too. We should.
1:10
Probably start with you're in Florida and you're wearing
1:15
sweaters. How is that possible?
1:17
Yeah, theory of relativity applies even to temperatures, I
1:22
guess, right? Yeah.
1:24
And I'm in Germany. Where in.
1:26
T-shirts, Yeah, yeah. What is the?
1:29
What is the temp there for you these days I guess.
1:31
I do Celsius. I have been converted so it is
1:36
zero. And the last couple days it was
1:39
like -5 which for the people out there that do Fahrenheit,
1:44
obviously zero is 32, but -5 is like, yeah, in the 20s, I think
1:50
high 20s, mid 20s. That is frigid.
1:52
I don't think I'd leave my house if it were that cold here in
1:54
Florida. Yeah, it's funny how you get
1:57
used to it it. Really.
1:58
Oh, yeah, absolutely. And I mean, I grew up in
2:01
Colorado and a lot of people as soon as they hear that, they go
2:02
like, wait a second. But yeah, I don't know.
2:04
I've been in Florida too long I guess, but.
2:07
You got acclimated to it, man, that's too good.
2:10
Well, the interesting things that I want to talk to you about
2:14
are not at all around LLMS. I think we started off this
2:20
conversation and we really wanted to have this conversation
2:22
because there is so much more happening in AI than just the
2:27
LLM boom and the agent boom and the whatever insert your next
2:33
type word boom in here, right? Yeah.
2:37
What are you working on these days?
2:39
Yeah, so I work and it may help to give it a little context on
2:43
my background. That may help.
2:45
So bro, I got into AI mainly from like behavior science,
2:49
space, technically like clinical work with humans, but you can
2:53
think about what we're doing is more like behavioral ecology but
2:56
applied to humans. That is, you know, you look at
2:58
the environment around people. How might you change that to
3:01
change behavior like workplace settings, right, or clinically,
3:04
things like that? And so then you know how I got
3:07
into AI, stuff I'm working on now.
3:09
As you can imagine, the environment is incredibly rich.
3:12
All sorts of stimuli that we perceive respond to influences
3:16
our behavior. And mid twenty 10s, all of a
3:20
sudden a bunch of research started coming out.
3:22
AI is better at doctors than X&Y&Z.
3:25
And so I was seeing a researcher interested in human decision
3:27
making as like, what is this thing?
3:30
And that kind of pulled me over into the again, kind of where
3:32
I'm working now is, you know, how can we take information from
3:36
the larger environment, sensory modalities, wearable
3:39
technologies, things like that, and from those data, understand
3:44
why people do what they do and then use that to then help them
3:47
make better decisions, healthier decisions with happier,
3:50
healthier lives. So.
3:51
So it's. Not thinking about what the
3:56
diseases are from signals from wearables or from blood tests
4:01
and all that. It's more why are people
4:05
deciding to stay on the couch or eat those?
4:08
Foods that. That pizza that I love?
4:11
Why do I choose that on Saturday instead of some broccoli and
4:15
hummus? Exactly.
4:17
Yeah, that's exactly right. And what's kind of fun about
4:19
that? And the area of kind of
4:22
psychology that I got my training and behavior analysis,
4:24
there's this kind of pocket of literature called say, do
4:27
correspondence. Basically, the idea is that we
4:31
don't, we don't always do what we say or what we, you know, if
4:34
you were to ask me why do you eat broccoli or pizza instead of
4:36
broccoli, I might say something, but that doesn't always match up
4:40
with my behavior. The reasons that I why I might
4:42
actually do something to, you know, if you look at the data,
4:44
you really get into it. And so that's kind of this
4:47
really interesting dichotomy, kind of circling back to LMS,
4:50
right? Language text based,
4:52
Interesting, but may not give us all the information we need to
4:55
understand, you know. Why do you do what you do and
4:57
how can I use that to help you make better choices?
5:00
So what are some things that you look at?
5:03
So it's interestingly, and this is kind of where I also touch
5:06
point, mid twenty 10s kind of brought me into AI reinforcement
5:10
learning. I'm guessing a lot of listeners
5:12
are familiar with there's a whole pocket of biological
5:15
literature, reinforcement learning with biological
5:17
organisms been around, you know, 100 years, 150 years.
5:21
And so we look at a lot of the same stuff.
5:22
You know, you're a rat in an opera chamber, you're a human,
5:26
you know, scrolling through Twitter, whatever.
5:28
There's going to be a bunch of stimuli that are kind of
5:31
presented to you that comes before some behavior you engage
5:34
in. After you engage that behavior,
5:36
something happens, right? You know, I make a tweet and
5:39
someone clicks the like button and that kind of unit, that's
5:42
antecedent behavior, consequence chain is what we kind of talk
5:45
about. You can analyze those over time.
5:48
And from that I can start predicting, you know, based on
5:50
the behavior you engaged in, what's a reward or a reinforcer
5:53
for you, And then I can start using that to predict what
5:55
you'll do next. Twitter meta that they all use
6:00
the same algorithms for human behavior, right?
6:02
We're just talking about doing that in context of like health
6:04
behavior. Yeah, that's why I have so many
6:07
notifications when I sign on the LinkedIn.
6:10
And he and he can't seem to not get distracted by them because
6:13
they're so good at getting your attention.
6:15
Yeah. It is killer.
6:17
And so then how are you trying to help people make the right
6:21
decisions from that? And and basically like, I still
6:25
think break it down. Like the data that you're
6:30
gathering is more on how many steps I'm taking.
6:34
And yeah. It could be.
6:36
Like it's more exercise or deals with the body or it's also how
6:41
much screen time I have. Yeah, could be all related,
6:45
right. So there's you can imagine,
6:47
yeah, I'm not sure how much you sleep 24 hours in a day minus
6:50
hours you sleep you get that that many hours of behavior.
6:54
So you can think within this kind of total set of behavior.
6:59
There's some really that's hold the matching law known
7:02
principles of biological organisms.
7:05
You tend to allocate the amount of time to an activity that
7:09
corresponds with the amount of reward that you get from that
7:12
thing. So, you know, if I'm looking to
7:15
say change your physical number of steps that you take, physical
7:18
activity levels, I would look and see, you know, how much time
7:21
do you walk now? What else do you do in your day?
7:25
What are those kind of things that you find reward Netflix,
7:28
right, or whatever, you know, maybe scrolling Twitter, all
7:30
that kind of fun stuff. And then the question becomes,
7:33
is there a way that you can contact more reward for physical
7:37
activity then maybe you do from say Netflix or something?
7:41
So you can a lot of the, and this you can imagine, has a lot
7:44
of therapeutic implications. You know, if I'm engaging a lot
7:47
of substance abuse, I spent some time doing some research in that
7:49
area. I want to get you to shift away
7:51
from, you know, cocaine to something else that's maybe a
7:54
bit more healthier. So you try to figure out what
7:56
can I create to compete with the reward that you get from that
8:00
maybe unhealthy behavioral pattern to healthy, you know,
8:04
shift it to some kind of healthier behavior pattern.
8:07
But you know, now we're also talking like these are data
8:10
issues, right? Amazing in theory.
8:12
How do I get the data for you that allows me to really
8:15
understand your day? And fortunately for the line of
8:17
work I'm in, the last 10 years has been incredible, right?
8:20
Technology just allows us to collect so much data on so much
8:23
more than we could even 1520 years ago, but still a non
8:27
serving challenge. Apple watches in in the whoop.
8:31
And Oh yeah, yeah, yeah. Me too, but the the part that I
8:38
I'm not clear on still is how are you proactively trying to
8:42
associate more dopamine or serotonin or whatever is
8:46
released in the brain when I exercise more or when I take
8:51
more steps? How do you make that connection?
8:54
Yeah, yeah. So there's kind of a fundamental
8:57
some going back to that idea of the matching law, that behavior
9:00
flows where reinforcement flows is kind of the fun phrase in
9:04
there, right. So if I can look at your
9:05
behavior across the day baseline before I'm doing anything, I can
9:09
roughly see, you know, you spent three hours on Netflix, 10
9:12
minutes walking. There's a lot of value in
9:15
Netflix. What's going on there?
9:16
How can I take something of that and provide it for you can only
9:21
access it, let's say, through physical activity?
9:24
One kind of popular phrase for these kinds of interventions is
9:27
contingency management. Very simple, lot of research
9:30
settings, you know, maybe you'd prefer to sit on your couch, but
9:34
what if I give you $20.00 for hitting your step count today?
9:38
What if I give you $50 a day, $100 a day, right?
9:41
You kind of start increasing the value and kind of putting on
9:44
these extra reward contingencies, they're called,
9:47
right? So they don't occur in the
9:48
natural environment, but you can supplement, augment, add on to
9:52
the reward value for the healthy behavior to get it to shift.
9:57
And then clinically, usually the challenge is all right, amazing.
10:00
You can get changed in physical activity for like less than a
10:02
dollar a day. You get people hitting their
10:04
step counts. The challenge then becomes kind
10:07
of like what you were leaning into a little bit.
10:08
I think is amazing that I'm now walking, but do I now find
10:13
walking intrinsically rewarding? So you can actually fade out
10:16
those extra wards, rewards and things I've added on there.
10:20
Some people that's easier, for others it's more challenging.
10:23
There's kind of research ongoing there.
10:25
That's the basic idea, how you can analyze it and think about
10:28
it. How are you reaching out to
10:34
people? How are you setting up these?
10:36
Do I give you, Do I just say, hey, whatever app it is, you
10:42
have unlimited untethered access to all of my data like the Woo
10:46
worker and my screen time and everything that I'm doing to
10:49
help me become a better person from goals that I've set.
10:52
Yeah. Wouldn't that be nice if you
10:54
would give me that would be great.
10:55
Yeah. No, that would I would.
10:56
Give all of you honestly like I don't even care about any data
10:59
privacy but we could go over how you are keeping my PII clean.
11:02
Yeah, yeah, yeah. No.
11:04
And I should say, like at this point, data challenges a real
11:08
thing. Most of the work that I get into
11:10
is in clinical settings, so like behavioral health settings where
11:13
we have people coming in for therapy.
11:16
I then know the totality what's going on in that session, right?
11:19
I'm presenting maybe different learning trials.
11:22
They might be called things like that.
11:23
You're trying to work on, I don't know, improving your
11:25
speech articulation, right? Someone that seemed like a
11:27
speech therapist or something, bring them in.
11:30
I know the behavior I'm trying to change.
11:32
I have all the data because they're in my context on, you
11:35
know, when I'm presenting you different words to say, how
11:37
exactly are you saying in the back if we're working on like
11:40
different pairings, chains of the sounds that you might admit,
11:44
right? What does that look like?
11:45
Where do you start to make mistakes and things like that?
11:47
So much easier to get the data that you need and the types of
11:51
clinical settings that that I'm kind of working that like pro
11:55
health behavior daily life. I've set up a few systems for
11:58
myself just with my woop time tracking screen time and things
12:03
like that. You know, maybe there's a
12:05
product played down the road. I don't know.
12:06
I'm a researcher, scientist first and foremost.
12:08
But yeah, the data challenge is a real deal.
12:11
But most of the stuff he has a clinical setting.
12:13
It's kind of weird. Because if it does feel like
12:16
what I want means to have such a 360 view of who I am and it
12:24
needs to understand all of my habits, the good, the bad, the
12:27
ugly. And I think about how could I
12:32
have a product in my life that has almost like God Mode?
12:38
Yeah, Yeah, yeah, yeah, that's exactly right.
12:40
Yeah. And you're doing it from your
12:43
side, which is almost like the only way that it can happen,
12:46
right? I know I've had friends who
12:48
would plug into their Aura Ring API and then create different
12:54
scripts that could run off of that, which feels like the only
12:57
way. It's like very hacker esque.
13:00
Right now. Oh, it's incredibly hacker
13:01
esque, yeah. And that's kind of what I've
13:03
done in my own life too, going back this idea like the matching
13:06
mall, right? So I've been for 15 years
13:08
tracking wake up to when I fall asleep.
13:10
How do I spend my time? Whoop data, right?
13:13
Same thing. I have custom scripts pulled in
13:15
Strava and all that fun stuff, Peloton, but it's hard.
13:20
I know there's still gaps in the day that are missing.
13:22
And this is where I get, I get really excited about what kind
13:25
of what's possible five years from now, 10 years from now, and
13:29
how so much of this stuff, right.
13:30
You think physiological data, that's not really an LLM call.
13:35
It's not like text based. It's just raw, you know, physio
13:38
tensors make this beautiful to play with.
13:41
And so it's yeah, but you're exactly right.
13:43
It's God Mode, right? How do you get get data on
13:46
enough stuff? Maybe you don't need data on
13:48
everything, but on enough things from my life that I can impact
13:52
those things that are most meaningful to me when, you know,
13:55
when I think about what are my values like, what does it mean
13:57
to live a good life? Yeah.
13:59
Well, especially if you're like me and you're trying to set
14:03
goals. I'm always trying to get better.
14:06
And I imagine if I knew a lot more about what you know about,
14:11
I would be better at giving myself the rewards to create
14:15
those habits. Like reading Atomic Habits was
14:19
not good enough for me to become the perfect man that I'm trying
14:22
to beat. Yeah, yeah, oh, sadly.
14:25
That did not work. But if I if there were the
14:31
possibility of an app to be able to understand me and have access
14:38
to everything and then help me with the goals that I'm trying
14:43
to get. Like it just a simple thing that
14:45
I'm trying to do these days and I find it very difficult because
14:48
it's breaking habits and putting new ones in is read before bed.
14:53
I read when I wake up really easily, but reading before bed
14:58
for some reason is really hard. I would prefer to scroll TikTok.
15:02
Yeah, that's fair. I mean, who wouldn't prefer
15:04
scroll TikTok? Yeah, I can't follow me for it,
15:08
but still like I want to. Yeah, I feel like I could do it.
15:13
Oh, yeah. And I mean, there is a whole,
15:14
like, body of scientific literature called self
15:17
management comes from the same behavior science literature.
15:20
How can we, rather than having others impose these rewards for
15:24
us to how can we do it for ourselves?
15:26
And they're a bunch of strategies, you know, like off
15:29
the cuff someone might throw out, like, you know, when you go
15:31
to bed rather than have your phone next to you with your
15:34
alarm, it's across the room. So.
15:36
I set my alarm, I set it down and then that even that
15:39
increasing the effort to go get it to then scroll TikTok
15:42
probably going to prevent me, especially if my books right
15:44
here, right. So like setting that
15:46
differential effort is like 1 strategy to make things, you
15:50
know each. Most people choose the less
15:51
effortful option. Oh.
15:53
That's great. There's a whole rich literature.
15:56
What's kind of fun about it for me is going back again, like
15:58
this, say, do correspond, all that kind of fun stuff.
16:01
We've known about all this stuff for decades, implementing it as
16:04
hard. Getting data on your own
16:06
behavior can sometimes be hard. And we are also don't always,
16:09
we're not always aware of the reasons why we might pick that
16:14
habit we're trying to break. You know, I want to do something
16:17
different, but I had to find myself doing the same thing.
16:19
I also have a sweet tooth every night after dinner Regard, you
16:22
know, I just can't break that one.
16:23
But you know, you start collecting data and analyzing
16:25
your behavior over time and you can start to understand, you
16:29
know, it's XYZ. These are the reasons why TikTok
16:32
is so appealing to me. Before I go to bed, Yeah.
16:36
And so how does this play into AI?
16:40
How are you using AI for these type of insights?
16:44
Yeah. So getting going kind of back to
16:47
the clinical space because that's where I've spent most of
16:48
my time. I've done some of this with my
16:50
own kind of quantified self data, but more, more than
16:54
clinical space. So you imagine you have a lot of
16:56
this data. Again, going back to this idea
16:58
of clinical decision making, there's a lot of stuff that's
17:02
come out. I'm not sure if you read the
17:03
book Nudge. Long story.
17:07
A cat, Kat Sunstein and Richard Thaler I believe.
17:11
What was it about? Because I feel.
17:12
Like so the basic idea, yes, this pocket of research,
17:16
behavior science, it looks slightly different area, but
17:18
similar ideas called behavioral economics.
17:21
Basically the idea is that, you know, traditional economic
17:25
theory, all humans are rational agents.
17:28
We optimize, we always choose the best course of action.
17:33
And that's, you know, when we make decisions, we are, we're
17:35
always optimizing, optimizing, maximizing.
17:37
But you know, you look around at humans, I don't know, you know,
17:40
we smoke, we drink, we, we, we make a lot of suboptimal
17:43
decisions is like the fancy phrase.
17:46
And so the question this book nudges, you know, what are all
17:49
the ways that we might nudge ourselves to move, quit making
17:52
these suboptimal decisions to make better decisions for
17:55
ourselves And what not that you can imagine that's also been
17:58
applied to the clinical decision making literature, right?
18:02
There's one study 2016 third leading cause of death in the
18:06
United States was physician errors, right?
18:09
They made the wrong choice, somebody died.
18:13
So you know, going back to this idea, we have a lot of data.
18:16
People are in these clinical to context healthcare
18:18
professionals. They want to do the right thing,
18:20
but they may not always make the best choice and they may not
18:23
always be aware of all the data and the things around them that
18:26
could inform their decision. So where AI comes in and the
18:28
things that that I get excited about and we have one thing that
18:31
we built that's patent pending, you can start, you can imagine,
18:34
I have this data, one thing, first thing I can do
18:36
unsupervised machine learning, right?
18:38
All these different patients, let's just start finding
18:41
different kind of patient groups, patient cohorts, patient
18:44
profiles, whatever you want to call persona and marketing, same
18:46
basic idea, right? Different patterns of clinical
18:49
presentation that might then inform how I intervene, how I
18:54
prevent some kind of clinical intervention.
18:57
So unsupervised machine learning comes in there, allows us to
18:59
identify really interesting groupings.
19:01
Next phase, usually in some kind of healthcare setting, I'm
19:04
trying to maximize patient outcomes.
19:07
Classic supervised learning task.
19:09
All right, I have this kind of a patient profile.
19:12
Given everything I know about their clinical presentation and
19:15
where I want them to be, what can I do clinically as a
19:19
therapist to help move behavior in that direction and not in
19:22
some other direction? Again, clinical decision making,
19:25
helping make optimal decisions verse suboptimal decisions, and
19:29
even those in and of themselves. You can imagine you're starting
19:31
to probably wrap around all sorts of different ML models
19:35
chaining ensembles together. Most people when they enter some
19:38
kind of therapeutic context, it's rarely one goal that
19:41
they're working on. They may have 1020, right?
19:43
So now you like hierarchical things, you have decision trees.
19:46
I mean, it gets incredibly exciting, but this is what, you
19:49
know, coming back to the kind of main theme that we want to start
19:51
of talking about, There's so much complexity there.
19:55
And I also need this thing to be controlled, compliant,
19:58
transparent, all those fun things that you don't
20:01
necessarily need an LLM for. Probabilistic outputs is also
20:04
dangerous in healthcare settings, right?
20:06
I need it to be discriminative. AI needs to be very good at what
20:08
I'm asking it to do. Yeah.
20:11
And that's, yeah, a lot of the work that I spent my time doing
20:13
the last 10 years or so. I imagine there's a lot of room
20:18
for folks to come in and either say that they've changed when
20:24
they haven't, or they have changed, it's just they changed
20:30
for a week and then they fall off the wagon again.
20:33
Yeah, you like weed those out? Yeah, yeah, absolutely.
20:39
So for the first one for those that may say they any and they
20:42
may generally believe it, right. I feel I'm a new man, right.
20:45
I'm a new person that getting data on that behavior.
20:50
That has been me. I will say like, just so we're
20:54
clear. Yeah, yeah, I'm, I'm off my
20:56
sweet tooth kick, right? Yeah, I started ice baths and
20:59
now I'm a new man. I'm for bed.
21:02
It's all good. Yeah.
21:05
And yeah. And so in that context, you
21:06
know, therapeutically, if I was your therapist, love that you're
21:09
telling me that every day, though, I'd be asking you to
21:12
give collect data on your behavior in some capacity.
21:15
Or we'd figure out a way. So you could say that, but then
21:17
we'd look at your data and say, wow, but your behavior like that
21:20
was a spike. Your your behavior hasn't quite
21:22
changed to the level that you're after.
21:24
Whatever. So that's kind of one way.
21:26
The other thing that you're talking about is this idea of a
21:30
popular phrase that people might be familiar with.
21:32
It's like relapse. Right.
21:34
I got better, then I went back to my old patterns of behavior.
21:39
This is a classic behavioral pattern.
21:41
Studying in the behavior science literature.
21:44
You can think of like almost like a time series forecasting
21:47
challenge, right? What are the variables in the
21:49
environment that when they combine, allow me to predict
21:52
that you're going to relapse versus instances that you won't?
21:56
And then you can start again. Same idea.
21:58
You need the data. But I can start saying again,
22:00
amazing that you said you took ice baths.
22:02
You have this pattern of behavior in your history,
22:04
Demetrios, I can tell you're going to relapse next week.
22:07
You know, let's let you stay on the straight and narrow to or
22:10
you know, add in these additional therapeutic resources
22:12
or whatever to get you over the hump this time so you won't
22:15
relapse or whatnot it. It reminds me a little bit, Do
22:19
you ever feel like it's a little Minority Report esque?
22:22
Where Yeah, absolutely. So, yeah.
22:25
So full, full confession at this moment in life, like I'm a kid
22:28
of the Matrix era, you know, kind of came up Matrix Minority
22:32
Report, saw those Ex Machina, all that stuff was going on
22:36
right when I was in my PhD for behavior science, starting to
22:40
work in AI. And I, I mean, I, the idea is
22:44
incredibly intoxicated, right? And I think what's also
22:48
interesting is if you look at a lot of tech companies, arguably
22:52
they're not to that same full extent, but they're using the
22:55
same ideas, right? How can I make my product more
22:58
addicting to keep your eyeballs on it?
23:00
What we're talking about is like, we can do that exact same
23:03
thing, but to help people live healthier, happier, better
23:06
lives. We can use it in kind of both
23:08
ways. But yeah, very.
23:10
I mean, I think that future, that Minority Report future is
23:14
coming at some point. You know we're.
23:17
And I, I'm 100% behind you on I would much rather have what
23:23
you're doing than me being subject to social media or my
23:29
own vices and have it just happen to me passively with or
23:35
without me knowing it, you know? Yeah.
23:37
Oh, absolutely. And I really like that, that
23:40
idea that you can imagine of like, you know, maybe turning
23:42
some of this this system randomly.
23:44
I could turn that into a product and give it back to the user to
23:47
say, hey, connecting what you want, you know, I'm not going
23:50
to. And you, I mean, technology's
23:52
got, we could run stuff on the edge these days.
23:53
Like I don't need to collect your data or save it anywhere,
23:56
but you'll give that back to the user and say connect in whatever
23:59
you want to get the data in there that you need, state your
24:02
values. And then, you know, just build a
24:04
model on Demetrius behavior, right?
24:07
Build a custom model for you and nudge you recommend.
24:10
There's a whole body of literature called just in time,
24:12
adaptive interventions, basic ideas.
24:15
You know, at that moment of choice, right before you make
24:17
some kind of unhealthy decision, you get a nudge or a prompt or
24:20
something to make you make the healthier so.
24:23
So basically when I am right about to grab my phone before
24:29
bed, it's like, oh, remember that?
24:33
Yeah. You know what I always I think
24:36
about a lot as something that I want to start doing is just
24:41
before I read a book to my kids for bed, shut off the phone.
24:47
And just the act of having the phone off and then having to
24:50
restart it is enough of a barrier for me to say, ah fuck
24:56
it, I'll read a book. Yeah, yeah.
24:57
Oh, I love it. Yes.
24:58
You're, you're already thinking about this idea.
25:00
If I increase effort or the delay to the reward, classic
25:04
examples that we know reduce its value.
25:07
So you know what's more immediate lower effort.
25:10
We tend to choose those things. So yeah, put the carrots at the
25:14
front of the fridge with the cake behind it, you know, and
25:17
I'm still going to go for the cake because it's just so good.
25:20
And so I'm constantly trying to figure out those efficiencies.
25:25
And I imagine you start to see areas where you can do more.
25:30
And, and just from me talking to you, I, I really appreciate this
25:33
because I recognize that one thing I want to do and I it's
25:39
that do say dissonance in my own life is I constantly tell myself
25:45
I'm going to go to bed early and before I go to sleep, I'm going
25:48
to read a few pages and then go to sleep and nail all my sleep
25:53
score goals, right? Oh yeah, yeah.
25:57
Be the perfect man. And that only happens when I'm
26:00
traveling on my own. When it's at home, it's a
26:04
disaster usually. And thinking about it in the way
26:09
of how can I set up more friction is something that I
26:15
didn't necessarily actively do. It was more that I would do it
26:23
in a in a way that I just recognized.
26:26
Oh, yeah. It like I don't sleep with my
26:28
phone in the room. I sleep with it outside in my
26:32
living room. And that's one thing that I make
26:35
sure to do because I recognize that I don't look at it in the
26:38
morning and I read in the morning, right?
26:40
And so doing it the reverse way and just shutting it off before
26:44
I read my kids a book is is setting that friction
26:49
consciously and then like regaining a bit more of my
26:55
willpower. Yeah.
26:56
Oh, I love it. The other thing I think that's
26:59
interesting that you said is that it's easier for you to do
27:02
when you're traveling. So the other thing that I'd be
27:04
curious about is like, what happens like as you're leading
27:08
up to the end of the night that makes TikTok that much more
27:12
valuable. Like, are you, I don't know,
27:14
like you already jacked up and you're like, yeah, now I also
27:16
need to get some TikTok. Whereas on the road maybe it's
27:18
you're already calmer. And so you're like, I'm going to
27:20
get into a book. But yeah, those are most
27:23
behavior is what we call multiply controlled, right?
27:26
Dozens or hundreds of things coming to influence it.
27:29
So it's trying to figure out, you know, what are the main
27:32
things or the things that I might be able to tweak or tug on
27:35
to get the behavior I want versus the one I don't.
27:37
But. Yeah, I imagine it's not only
27:41
the behaviors that you can tweak or tug on, it's the ones that if
27:45
you tweak and tug on them, they're going to have a domino
27:50
effect in what you end up doing. Yeah, Yep, exactly right.
27:54
It's funny you say that a a couple.
27:57
So I also, I'm faculty at the Endicott College has some docs
28:01
here. We're working on the paper right
28:02
now. This idea, we call it keystone
28:04
contingencies, if you're familiar with like keystone
28:06
species and like ecology. So the basic idea there is look
28:12
at Yellowstone National Park, right?
28:14
You have some kind of ecological system.
28:16
There's a species within it that if you were to remove them, like
28:19
the wolf, the whole ecosystem would like reorganize and they
28:23
learned this the hard way and they reintroduce whatever.
28:26
So we're playing a bunch of other areas of played around the
28:28
same idea, Keystone social people on social networks,
28:32
Keystone actors and the corporate whatever.
28:34
We're playing out of this idea of keystone contingencies it
28:37
like in your own life, same idea, right?
28:39
Is there one behavior thing that I might change that has this
28:43
ripple effect throughout the rest of my day?
28:45
And for me, like it's running in the morning.
28:48
If I run in the morning, I drink less alcohol the night before.
28:51
My diet tends to be better, more focused at work.
28:54
And it's like that one thing, if I can tweak that more, a better
28:58
day, not optimal. Yeah.
29:00
It's a keystone moment. It's like this pillar.
29:03
I also think there's a lot to be said for how you end up seeing
29:08
yourself and like what you identify as.
29:12
You identify as a runner now and.
29:14
Oh yeah. It is something like, well, if
29:17
I'm a runner, I got to run. Yeah, yeah, yeah.
29:20
Oh, fair. Yeah, yeah.
29:22
And if you're a night time reader, yeah, you're going to
29:24
have to read for it. The night time, yeah.
29:26
That that's exactly it. Like you end up doing these
29:29
things because you identify as that.
29:32
Yeah, yeah. Agree with that.
29:33
Yeah. I can't remember where I heard
29:36
that, but it was in something. It might have been Atomic
29:40
Habits, or it might have been some other habits book on how
29:44
it's much easier to create a habit if you identify as that
29:48
type of. Oh yeah, sure.
29:50
Whatever. Like if you're, if you're a
29:51
smoker and you identify as a smoker, it's a lot harder to
29:54
kick the habit of smoking because it's like, yeah, I'm a
29:57
smoker. Oh yeah, definitely.
29:59
And there's some really decent literature suggesting that even
30:02
kind of like at the same thing, the language that we use can add
30:06
value to activities or whatever, which kind of fits the same
30:09
idea, right? If I call myself a smoker, then
30:12
yeah, it adds more value to the cigarette and in addition to the
30:15
nicotine, as you know, if I'm a non-smoker but I happen to hit
30:18
it, same nicotine, same stuff going in.
30:21
But just that bit of language can change the the reward value,
30:24
which is kind of crazy to think about.
30:26
I mean, humans are nuts. Yeah, and this is also something
30:28
that is fascinating to look at through a machine learning
30:34
model. Oh, absolutely.
30:36
Incredibly fascinating. Are there certain ways or
30:40
certain sentence structures that for you tend to encapsulate you
30:46
in certain behaviors? Thinking about the sentence
30:49
structures and analyzing the way that if somebody comes in and
30:54
they are speaking in certain ways, is that indicative of
30:59
certain actions? Yeah.
31:01
Oh yeah, definitely. And there's a whole whole body
31:03
of research called framing effects, that kind of study.
31:06
This kind of stuff about. Yeah, no.
31:08
And and I think the other thing go back to like how AIML, I'm
31:12
personally just enamored with unsupervised machine learning,
31:15
this idea that I don't know what I don't know.
31:17
And it's going to like tell me stuff like I was just incredibly
31:19
intoxicating to me. So what I think kind of going
31:21
back to this to where AI would be interesting is, you know, why
31:25
you come in just like talk about things that you feel you do
31:28
well, things you don't do well. Can I analyze that behavior?
31:32
Understand like your frame of reference, your perspective?
31:35
And then, you know, there are some cultures, they have
31:38
different colors for white that I've never perceived, right?
31:41
Just this idea that AI can wrap around language as a, a human
31:45
phenomena. And then can I bring that into
31:47
that maybe therapeutic setting and say, hey, here's how you
31:50
talk and perceive the world. Here's an alternative
31:52
perspective that you may not have even known existed that may
31:55
help you. Like maybe we can get you to
31:56
frame things, you know, in this different way, and then we can
31:59
again go back to the data. Does this actually change your
32:01
behavior? Yes.
32:02
No probability of relapse, like all that kind of fun stuff.
32:05
It, it almost feels like, and this is where I'm drifting into
32:10
the AI hype territory because you got to be really damned or
32:14
careful as you are doing this. But I know there's a lot of
32:19
people that will talk to open AIS, real time API or like the
32:26
voice. Component.
32:28
It's very shaky and anybody that's used it knows that.
32:30
It's not amazing, but I see a potential world where you're
32:36
sending a voice memo and you're just blabbering on with a few
32:40
different prompts and then that gets analyzed.
32:46
And potentially you can do it with an LLM at first, but then
32:49
you get more specific and more. Oh yeah.
32:54
Fine tune the models, fine tune for you and what you're going
32:58
through I guess if it's needed. Yeah, absolutely.
33:02
Have you seen the movie Her? Yeah.
33:04
Oh yeah. I feel like that's kind of, you
33:06
know, it's the thing that's with you that hears your language
33:09
parrot sees what he sees and it can get a tap into the
33:12
collective conscious, but then build model specific for you and
33:15
kind of what you're after. And then again, you know, from
33:18
my angle, we can, there's no reason we can't do that and help
33:22
everybody live happier, healthier lives.
33:25
Doesn't always have to be. I mean, there's the downside to
33:27
any technology if you know, if you if this exists, bad actors
33:31
will do what they do. But.
33:32
How are other ways that you've seen and you like what you're
33:37
seeing with unsupervised machine learning?
33:40
Oh, sure. So another kind of similar
33:43
flavor, but now in the education context because I work, you
33:47
know, higher Ed, teach classes, things like that.
33:48
But same idea with students. And I did some work with a
33:52
company called Glimpse K12 when I was doing my data science
33:55
postdoc, but same similar idea where I can take patterns of
33:59
behavior in educational settings, assessment scores,
34:02
whatever from that identify kids that may need more resources to
34:07
be successful in a classroom setting.
34:09
And you know, if you're a teacher or whatever, you have 30
34:11
kids, 20 kids, 30 kids, it's hard to give everybody what they
34:14
need to succeed. And so if you can use any kind
34:17
of unsupervised machine or any kind of method to identify
34:20
learner type, student types, whatever, to then match kids up
34:24
with resources so that they they have a higher probability of
34:26
being successful learning the things they need to learn.
34:30
Personalized systems of instruction are another big
34:32
thing, right? Rather than every kid getting
34:34
the same per, you know, set of tasks on the chalkboard, you
34:38
know Johnny's going to get one set of math problems, you know
34:41
Elizabeth's going to get another set because they're just
34:44
different students, different skill sets, different resources,
34:46
but. Yeah, different ways of
34:48
learning. Yeah, actually back when I used
34:50
to live in Spain, I was teaching and you learn right away that
34:57
each student is very different in how they learn and what they
35:03
need to learn. And some are visual, some are
35:06
auditory, some need to speak and all that.
35:09
And I, I come back to the idea of we see so many or we hear
35:20
about how small groups tend to be the most effective when it
35:26
comes to teaching or just one-on-one.
35:30
And then you can be very personalized.
35:33
And so how can we, as you're saying, create modules that are
35:39
personalized for each individual that are hitting their needs
35:43
instantly? I think about how do you know,
35:46
how do you get the data? How do you have the context to
35:48
know what each individual needs right?
35:52
Because. If I'm just coming into school
35:54
on day one, you don't know that. I like to listen more and I
35:59
learned through auditory methods.
36:02
Yeah, assessment scores is the, I mean, the only thing that a
36:05
lot of places have right now where I've seen and even some of
36:08
my own work where it's been most successful online learning
36:13
platforms, right, Canvas, some of those places where I have a
36:17
bunch of student behavior captured in a browser.
36:20
When are you logging on? What are you engaging in?
36:22
How long are you reading something when you submit an
36:25
assignment, What's the quality of that?
36:27
What grade do you get back? What are the feedback, the
36:29
teachers giving you, those kinds of things.
36:32
But yeah, again, this is where kind of going back to the I know
36:35
some of the things that I've thrown out of maybe in the ML
36:37
OPS community love that people are building some of these crazy
36:40
systems. But a lot of the stuff we talked
36:42
about today are data challenges. And like, how can I start
36:46
getting the data I need for some of these very important real
36:49
world problems to then even I mean, let's we can just start
36:53
classic ML. We don't even have to get crazy.
36:55
Like, let's just do something to help out some of these kids
36:58
because, you know, a lot of times they, you know, fall
37:00
through the cracks and you have a tough start in kindergarten
37:03
that can cascade through elementary school and maybe hard
37:07
to catch up, but. And it is so true that it's the
37:12
data challenge in that if you're in a class, you're a teacher in
37:16
a class with 30 kids, how are you getting the data to know
37:20
what these folks need in these moments?
37:25
And so it makes sense that if they're interacting with a
37:28
website, you can get much richer data.
37:32
Or if you have them do these assessment tests, that's a great
37:36
start to. But I feel like, yeah, you got
37:39
to constantly be capturing that data.
37:42
Oh yeah, and I mean, don't get me wrong, there's some really
37:45
cool research going on. You know, I put cameras in rooms
37:52
right now I'm tracking what is every kid doing throughout the
37:55
day, what are they getting exposed to, what's their
37:57
behavior, all that kind of fun stuff.
37:58
Look like most kids are on iPads these days.
38:00
So now imagine I can start integrating data from multiple
38:03
sources. Again, I've only seen this
38:05
research perspective primarily in maybe a hospital mental
38:09
health setting, large public spaces, right?
38:13
I'm trying to understand where people walk it, whatever.
38:16
So we do have some computer vision technology that I think
38:19
allows us to at least start playing with these data sets.
38:23
The only I've only seen the stuff research wise, never
38:27
haven't at least seen a product, right.
38:28
Nobody offering, hey, you know, K12 school, buy my cameras and
38:31
we'll, you know, do this stuff for you.
38:33
I don't think it's in their budget, even if they exactly
38:36
something like that. But yeah, I think the yeah,
38:39
technology's there. That's got to be the hardest
38:42
part is like, we've got this great technology to help your
38:47
students, but it's super advanced and it probably isn't
38:51
cheap. Yes.
38:53
Oh, exactly. Yeah.
38:55
And that's where, you know, I think some of the people in the
38:57
ML OPS community, they have these skills, work incredibly
39:02
smart, talented, working on stuff.
39:04
And I know half the challenge, like we mentioned is can I get
39:07
the data? Can I make it the ROI worth it
39:10
for a product to be built? And I have to imagine people out
39:13
there in like in the community have the answers.
39:15
It's just, you know, yeah. Yeah, get that data or something
39:19
that yeah, get that context. I really like that idea of like
39:24
how can we get more data to get more context for the kind of
39:29
stuff that is going to make an impact.
39:32
And it doesn't necessarily need to be context for your LLM
39:38
context window. That's.
39:39
Right. That's right.
39:39
Yeah, exactly. Yeah, Yeah.
39:41
But I think I actually love you said that because I think if you
39:44
look at behavior science literature, whether it's a
39:46
human, whether it's a, you know, a frog, an eagle, whatever, we
39:51
understand behavior by understanding that larger
39:54
context within which it behaves. And there's a great analogy with
39:58
LLMS, right? The more you, the LLM
40:00
understands, hey, this is what I need you to do and like, here's
40:02
the information you need better output.
40:05
Same is true with humans, right? It's just, you know, we.
40:09
We talk a lot, we love language. And so often we default to
40:12
talking with people. You know, why do you feel that
40:14
way? Why are you doing what you're
40:15
doing? But there are other ways to
40:16
understand why we do what we do that are often more accurate but
40:20
more data intensive. Requires people like the people
40:24
in MLF community to run those through algorithms to extract
40:27
the insights and put it in front of a teacher who's not going to
40:29
be coded in Python, right? They just need a dashboard or
40:32
something. But.
40:32
Exactly. They need that product, that end
40:35
user product. Exactly.
40:37
Yeah, yeah.