November 20, 2024 • 35 mins
This time my guest is Dr Brandon Rohrer. He is a data scientist who originally started in researching robots to assist with stroke recovery, and who now works on solving data challenges. Brandon, whom I tracked down via his online discovery work on https://e2eml.school/blog.html where I was captivated by his journey in understanding AI. It is a great resource, and I did show it to my team who also loved it.
We had a fascinating and wide-ranging chat that went from the demise of Twitter to the challenges of getting robots and AI to be more human.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to another episode of the Dead Revolution podcast.
(00:06):
My guest today is Brandon Roura, who I found online and cyber-stalked.
Welcome, Brandon.
Hi, Kate.
Thanks for having me.
I really wanted to have a chat with you because, you know, I actually found you by, I was googling
ML back in the day a couple of years ago, and I found your e2eml.school website, and
(00:29):
it talked about your learning journey about machine learning, and I was really fascinated
by it and showed it to my team.
We found it an excellent resource.
So I found you there, and I started cyber-stalking you.
And I think I started following you on Twitter when it was still a theme, and now we're context
on Mestodon.
So that's how I found you, and I'm really pleased that we get to have this chat.
(00:50):
I love that story.
To me, that highlights the very highest promise of social media because I've never been to
Australia.
I live in Boston, and the chances that we would have bumped into each other in a coffee
shop are probably not high.
So yeah, and that's one of the things, and it's one of the things I think is really
(01:12):
so tragic about the demise of Twitter is how everything's fragmented now.
Yes.
I'm so sad.
I enjoyed Twitter a lot.
I had connections with the data, data science, machine learning communities that honestly
are gone now.
There's a handful on Mestodon, and there's LinkedIn, but the tone is different.
(01:40):
It's like all you got to make sure it's suitable for work and pretty on the up and up.
And then just informal channels outside of that, but Twitter was, I'm still mourning
that as well, still looking for the next incarnation of what that could be.
And the real problem for me is people are fragmented.
So some people have gone to Mestodon, like the entire infoset community, Hollis-Bollis
(02:04):
moved to Mestodon in November 2022 because I had a Mestodon account from 2017, but there
was no one there.
Suddenly in November 2022, my entire cyber community was there, and I was like, oh, I've
got a reason to move now.
And then a bunch went to Blue Sky, and then LinkedIn's the surprising winner, which is
(02:25):
really weird with people kind of using it like Twitter.
But anyway, it's just a strange world now.
So why did you come to be working with data?
Because I loved your website, and it was like, it took through your journey about learning
ML, which was really fascinating.
It was not my original destination.
(02:46):
My first technical love is robots and robotics.
And so that's what I did my graduate work in.
My degrees are in mechanical engineering.
My dissertation is in using a robot to rehabilitate people who have had a stroke.
But in the process of doing that, I was, this was back in like 2000, I had a whole bunch
(03:12):
of human movement data, someone moving a joystick around for an hour, three times a week, times
six weeks, times 50 people.
And back then, that was gigabytes of data.
Oh, too much.
That was a terrible amount.
And I had to find patterns in this.
(03:34):
And I had a rough idea what I was looking for, but I wasn't certain.
So I had to hack together database.
I had to hack together some routines.
I was working in MATLAB.
And it was like, anyway, so it was like, I was doing algorithmic stuff and data stuff
without any training that was formal to it.
(03:57):
And but this ended up tapping into my love because it, through the course of the research,
revealed some telling things about how humans moved, which led to some interesting hybrid
hypotheses about how we could help robots move in a way that was more, more robust, more
(04:19):
able to work in an unpredictable environment, which led me to say, great, well, maybe I
can start making some robot learning algorithms and control algorithms that are based on this.
And then pretty soon, I'm doing things that look a lot like existing machine learning
techniques and someone stands up in a talk and says, Hey, isn't this just, you know,
(04:40):
a certain algorithm?
And you suddenly go, huh?
And then I was like, Oh, shoot, I should learn about all that stuff.
So I don't look dumb.
It's like kind of a bunch of machine learning textbooks and just started kind of reading
through them.
I mean, this is over years.
It was not, there's no Rocky montage here.
But and I realized also that a lot of this stuff, I'd had this experience over and over
(05:02):
again where I'd read through a handful of papers and posts and book chapters.
And when the realization finally clicked for how it worked, I was like, Oh, well, that
was less complicated than I thought.
And I realized like, well, what if I wrote the thing that I wish I had when I first started
(05:23):
learning this?
And so I started doing that for one of my early ones was for convolutional neural networks.
When I was back at Microsoft, I'm turned it into a video and it got a bunch of views and
it's still one of my most popular ones where someone's like, the comments a lot of times
will be like, I got it.
(05:43):
Like I've seen a half dozen other tutorials, but I got it.
And it's full of errors and oversimplifications.
But I am proud of it because it took some important, some hard to grasp concepts and just cartoonized
them and made them easy to ingest.
And I, it's not really out of any deep altruism that I've done this, but I find it saddest,
(06:07):
very satisfying for me.
And you know, if you read between the lines, whenever I'm published something new, you
can always tell that that's what I'm confused about in a particular month.
I'm trying to figure it out and so I write it up.
And then it's just kind of like over years kind of congealed into this big collection
of posts and videos and whatnot.
(06:30):
Oh, yeah, it's a really great resource.
I shared it to my team because we were grappling with, we did our first machine learning curriculum
concept back in 2019.
And we were trying to predict which students would fail a course during the course.
Oh, wow.
Yeah.
Because it's highly correlated, students fell in courses means they failed degrees
(06:52):
and it's highly correlated to them self harming.
So we actually wanted to find out who might be on a trajectory towards self harm.
Got it.
So the course failure was like the leading indicator for the thing that you were really
concerned about that was more rare.
Yeah.
(07:13):
And the thing was, you know, we had data on who was in the counseling service and was
getting treatment, but all the people who weren't were the ones who would go on to self
harm.
So, so it was kind of a real, really big problem that we wanted to solve.
So and we was efforts for a intuit.
And we're on the Microsoft platform and they were really helpful to us and we did a great
(07:37):
proof of concept and took six weeks to build it, build this model.
It's still running in production now.
Oh, wow.
And it just went on, which is great.
That is, that is such a good story because, you know, it's rare that someone actually
uses machine learning to make the world a better place.
(07:59):
But it sounds like actually the case here.
I am curious.
Did you have to stretch to like figure out what was important to measure or to include
in the model?
It was hilarious.
So, so the teachers had five factors that they thought were important and we tested
(08:21):
all sorts of factors because we had all the data and it turned out to be the five factors
that the teachers thought.
So they were right.
So they were very chuffed that they were right.
But there's really basic things.
Has the student logged on to the learning management system?
You know, have they done their first assignment?
(08:42):
All of those sorts of things that you would think are really predictive are and it's been
running at 95 plus percent accuracy since since then.
So it's pretty good.
That was that was our first for it and I had my development team were like, what?
And I found you found your website and showed it to them and they were like, oh my God,
this is so helpful.
(09:03):
I'm so glad.
It's funny because a lot of times like, I mean, I did a deep dive into transformers because
I wanted to understand about transformers and LLMs and all of that.
And at the very end, in my own mind, the punch line after doing these deep dives usually
(09:24):
comes down to, you know what?
It really comes down to how good the data is going in.
And you can have these huge models and dump in all of these unrelated features and the
results are met.
But if you have even a very simple model and you have domain experts like the teachers
saying like, you know what, my gut says it's this handful of things.
(09:47):
Let's look at it.
That wins every time.
Well, the interesting thing is, you know, every organization needs to get better at
data governance, which includes data quality.
Because you know, if you're working with poor quality data, you're going to have some standard
outcomes, especially if you can be driving actual decision making by these technologies.
Yeah.
(10:08):
I could not agree more.
It's easy for algorithms to get a lot of airtime.
But issues that really get my attention are, get my positive attention when someone's collecting
more data about something or higher quality data or more careful data and gets my negative
attention sometimes when someone's making data harder to get, less visible or somehow
(10:33):
corrupting it or calling it into question.
Like that, at that point, like there's not really anything that you can do.
I'm thinking about like government, public entities, you know, the environment.
In Australia, we had a particularly bad thing called RoboDat, which was the government using,
(10:58):
and it wasn't machine learning algorithm or anything.
It was just Excel basically, but it was still using a formula to determine if people owed
it money.
They were doing really egregious things.
Some people committed suicide over being harassed by the government saying that the government
owed them money.
It was really an evil program.
(11:22):
The fascinating thing about that was the public servants were just saying, oh, we're just
impersonally executing government policy, but it was a really bad policy, but it was
driven by this underlying algorithm and you couldn't prove that you didn't have a debt
for the most part.
If you had a lawyer, if you could lawyer up and just go, my lawyer will be in touch,
(11:43):
you got off, and if you were a poor person, you would end up having your pay garnished
to that government debt.
Wow.
That sounds like exhibit A for like how not to use automation and data to make the world
a better place.
Yeah, it is.
It is.
(12:04):
I spoke out about it when it was running and people were like, oh my God, you're saying
it in public and I'm like, yeah, I'm saying it in public.
It's really bad.
It's not good for people.
It's so, so bad.
Yeah.
The ideal use case for machine learning is when there's a lot of decisions that are
(12:25):
small and there's not a very high penalty for getting them wrong, but there are just
so many that a human either gets bored or just can't move fast enough to do them, or
if the information is just like so broadly dispersed to be able to pull it in.
That is the absolute sweet spot.
(12:47):
But the whole, the book, Weapons of Mouth Destruction, Cathy O'Neill, some great case
studies in there about how similar to this, it's like when you have an algorithm, there's
going to be mistakes and when there are people, it's making important decisions about people
on the other end.
(13:08):
Those mistakes are unacceptable.
If those decisions have to be made, you have a person make them.
Always like you just get out of it.
You can have the query it.
So, I mean, the whole thing with the advent of generative AI, and you know, it's been
a kind of a thing for a while, but it's sort of popped into the public perception back
in 2021-22.
(13:30):
And I think the interesting thing about that is ordinary people could understand what AI
could do for the first time.
Like people like my husband, who's a high school maths teacher who is not technology
driven at all.
He uses MATLAB and stuff, but he doesn't know any programming.
(13:50):
But he could see it and see what it could do for the first time.
And I think that sort of captured people's imagination.
But you know, the problem of hallucinations is an interesting one.
I think imagination is the key word there.
It's like in our household, we name our automobile, we name our devices.
(14:15):
It's just a cute thing and it's fun.
But it's easy to personify an inanimate object no matter how dumb it is.
And to start to ascribe motivations to it and ascribe internal processes to it that
simply aren't there.
Yeah, if you like the classic ELISA computer program from I think it was the 60s.
(14:42):
Where it just followed some kind of rope.
You say something, it asks you a follow up question based on that and it sounded plausibly
human.
That was the original blow people's minds.
It's like, oh my goodness, computers are thinking and it was in this first round of
AI sci-fi, it's going to take over the world.
(15:05):
But now with so much more data at its disposal, like LLMs are so much more plausible, even
people who are familiar with how they work on the inside have to be really, we have to
remind ourselves, there is no ghost in the machine.
This is really, really good autocomplete.
(15:28):
And yeah, if you give it enough data, it will give you very good answers and maybe even
the right answers in some fraction of the cases, which it does.
It's great for a lot of things, but it doesn't have this rich.
It shouldn't be like those lawyers who just took the generative AI, created case summary
(15:48):
or whatever it was and sent it to the judge and had it completely invented, completely
hallucinated citations.
Yeah.
Yeah, I mean, so my wife is in the education world.
She is a PhD candidate.
So she gets to see students and professors working, dealing with generative models.
(16:14):
And they're really good for cases where you have to produce words and you don't care what
they are.
And there are, sadly, like that is a common use case in even in education.
And it's great for that.
It can be quick.
It's also really good.
I've seen really good use cases where you wanted to hallucinate.
(16:38):
You're like, hey, I'm thinking about this thing.
I just, can you give me some suggestions, some ideas, and then I will filter through them.
Yeah.
It's great for that.
I think the process is an important concept I think people need to get.
Yeah.
I was just going to ask you, though, do you think that the RAG models, so retrieval or
(16:58):
augmented generated models are going to be able to help reduce the amount of the hallucinations?
I think so.
And my kind of like one sentence summary of RAG models is they're really good at taking
what you say and then doing like a really quick local search and pulling in actual real
(17:21):
texts that are related to it and using that.
So it's kind of like some kind of a hybrid between an internet search and generative.
But I think they're effective because they are a step closer to sticking with the raw
data.
And so they, I am still waiting to see some mechanism by which they could say like, well,
(17:49):
even with what I've been able to retrieve, I'm not confident that this fact is the case
or that this legal precedent has occurred.
I'm just going to sit back.
I'm just going to say nothing.
LLNs are really good at nothing.
I genuinely think that they're actually throwing into disarray what we've conceptualized as
(18:12):
intelligence because a lot of what we've thought of intelligence is just goodness.
Yeah.
So 15 years ago, I was because of a line of work I was in, I was attending these AGI
(18:34):
conferences and this is back when it was very much a fringe activity.
It would like co-locate with these AI conferences but be like in a little venue, like not even
co-located with me.
Yeah, like two days before the real conference and you kind of said you were going there
(18:55):
and like winged and be like, you do it ironically.
But yeah, even then it was hard to pin down what we meant when we were talking about intelligence.
And the thing that really drove this home for me is there was one big Q&A session where
(19:17):
we were getting into it and I just stopped.
There's room of like 100 people.
I was like, so I feel like I'm hearing two things.
Intelligence is it operates with the same mechanisms as the human brain or intelligence
is it does approximately what a human can do, like functional.
And if you could have just one or the other, which would it be?
(19:41):
What's core to you?
And the room was split 50-50.
So like this gathering, like we were all looking for something different.
We all meant something different when we were talking about intelligence.
And if there is a place where we had a hope of defining it, it was there.
And we weren't even close.
Yeah, but I'm just fascinated because you know, generally AI is capable of string together
(20:08):
really good words that sound really good.
And we I think a lot of us have mistaken that for intelligence for a very long time.
And we probably need to dig a bit deeper because we now know that generally AI is often quite
wrong.
Yeah.
(20:29):
Yeah.
I mean, there are the there are the jokes about, you know, there are there are humans
who are very good at stringing together words that sound great.
And then when you pull them apart mean nothing or contradict themselves or have no grounding
in reality, in which case, LLMs are extremely human.
Yeah.
Yeah, yeah, they are very intelligent.
(20:52):
Whether they're intelligent is a separate question.
But I think that that's a big question for our time.
What do we really mean by intelligence?
And when we're probably going to be, like you said, 50 50 split on what it is.
Yeah.
And any room that we're walking to.
This is something that has has been in my mind for a while because even going back.
(21:17):
So set aside reasoning, production of words, you know, playing chess, things like that.
Even if you go down to something very mechanical, like moving an arm.
So back when I was doing human movement, stroke rehabilitation, there were some very large,
(21:37):
ambiguous, unresolved questions about how the human brain even moves our limbs.
And it's just so, I mean, it's just so hard to measure what's going on.
We don't even know what to look for.
There was even a debate, like, are we looking to measure like one neuron that controls something
or somehow it's in the distributed set of neurons that are doing something, or maybe
(22:01):
it's in how those patterns of neural activity change over time that's doing something.
And there's so much going on and it's hard to measure it accurately.
Like no one could agree even on what we were even looking for.
And this is something where we can measure movement very, very accurately.
We can all see it and agree on it.
(22:21):
There's no disagreement there, no confusion there.
And we still don't know what what motor control is at its core.
And so abstract that out a little bit to like reasoning and empathy and model building and
like, you know, building metaphors and analogy and communicating, building theory of mind.
(22:49):
And it's like, I love it because we are so utterly clueless.
Like it is in my mind, the frontier.
We all do it.
We all see other people do it.
And we don't know how.
Yeah.
Yeah.
I only recently became aware of how difficult it is to build bipedal robots.
(23:15):
Oh, yeah.
Yeah.
So I had friends in the MIT Lake Lab, which has done some great work with bipedal robots
going back into the 80s.
But back then it was fun because I was fascinated with robots and they're doing this great research.
(23:38):
But even then it was like their their failure, their B-roll failure reel was hilarious because
these robots would just just like fall over or jump and land on their head or do a flip,
but but not quite clear it or just stumble over a chord.
And it's easy to forget that now because that line of research has since evolved into
(24:00):
Boston Dynamics.
And we have, you know, the big dog and we have Atlas and some robots that do some just
amazing things.
But that's decades of focused research from some very dedicated people, which I have immense
respect for.
(24:20):
And even then my understanding of the current state of the art is you change that hardware
a little bit and the policy that they have is not very helpful anymore.
Like it's very specific.
So things that humans do without even thinking about it are really, really hard to program.
(24:43):
Like we're really complicated.
Yeah.
Yeah.
It's kind of funny too, because we stack the deck in our favor when, you know, chess,
for instance, for a human, that's hard.
Like you have to study and practice for a long time.
That's a great problem for a computer.
Like just small finite set of moves and rules and finite possibilities.
(25:08):
The three heavily rules based world.
Yeah.
Yeah.
So when we initially set up checkers and chess as measures of machine intelligence,
like we were really stacking the deck in their favor.
But things that a five year old does, you know, like stumbles around on a shag carpet
(25:28):
with sticky sold shoes.
Like, oh my gosh, that's a nightmare for a robot being controlled by a computer.
I mean, one thing I did want to talk to you about is this idea of where you think Generative
AI is going.
Because you've heard people talk about a genetic AI, which is the idea of the next generation
(25:54):
of Generative AI with agents that can daisy chain their own prompts.
And so I've been talking at conferences about how I think things like AI, Internet of Things,
robotic process automation are all joining together to make autonomous systems, which
now people seem to be calling a genetic AI.
(26:17):
And I sort of see that as the next interesting thing that people are going to do things with.
I like that line of research.
I think that has a lot of possibility.
Let's see a couple of thoughts.
So LLMs are not perfect, but they are a tool that we have not had before.
(26:40):
So like any tool, there's some things that they are very good at and some use cases that
they just really nail.
But like any tool, if you use it places where you shouldn't, it's going to look silly.
So we're kind of exploring where that boundary is right now.
But to your comment before about, you know, a lot of us, a lot of us, a lot of the times
(27:01):
are just stringing words together in a way that makes sense given the context.
And so for a lot of non-critical situations, like if there's a computer program that can
do that, like that can meet certain needs.
Like if you're going to an information desk and asking for help, like how great to be
able to do that in natural language and get a reasonable answer.
(27:24):
So I think we'll get better at helping LLMs, like not say dumb stuff.
I really think that being able to train an LLM on the data that's specific to what we
want it to do, as opposed to having like giant.
General purpose LLMs.
(27:45):
So we think that we're going to get away from general purpose to specific purpose ones.
Yeah.
And it may not even have so much to do as a difference in architecture.
It's just the training, training data.
But yeah, like domain specific.
And if we can get a better at being confident in what it produces and not producing things
(28:07):
in time.
There was a smaller training data set with a more constrained set of questions that it
can answer confidently.
Yeah.
Yeah.
Then we could treat it like any other human where it's like, you know, I meet someone new
and I interact with them and I'm building my own assessment of whether to trust what
they say, whether they know what they're talking about.
(28:29):
And yeah, I think there's some really great opportunities there.
Yeah.
I see some really good use cases there.
It's also really fun.
Like on the data side, like right now, the easiest thing is to train on words, characters
(28:52):
from the internet.
That looks, you know, different depending on like, if you limit yourself to one language
versus considering lots of different languages versus considering like very localized slang
and including emojis and kind of like making it broader.
But it's still just like a sequence of symbols to a sequence of symbols.
(29:16):
There now they're including like steps to expand that to audio, some great demos on
the most recent GPT I have to give a nod to like a great job doing audio to audio and
text to speech and speech to text.
(29:38):
And then there is the promise of doing this with video too.
The trick being that with printed text is already symbolic.
It's already reduced to ASCII or Unicode.
And with speech and video, there's this extra step of how do you take this high dimensional
analog continuously variable input and reduce it to symbols and then get it in there.
(30:03):
So I expect that.
Yeah, I was talking with a startup that's trying to do some stuff in that space at the
moment and they're just trying to try to automatically translate it to from video spoken from language
to language.
And that's already a pretty hard problem at the moment.
(30:23):
Yeah, yeah, it's super hard, which coming back to robots.
It's fun.
Like a lot of early robot work was done in simulation where you could chunk the world,
you can make the world symbolic and make it so therefore straightforward, more like a
chess game.
(30:45):
When you're dealing with physical robots, you're dealing with actual continuous.
There is audio, there is video, there is these weird continually varying sensors.
And you have to figure out how to get them into this symbolic state, how to make them
symbols so that then you can get them into the computers native language like the computers
(31:05):
playing field.
But that's to me like those data issues, data quality, data transformation.
Those are hard problems.
That's not something that a company is going to plow through to get the next feature out.
Those are things that will take years to sort through and trial and error before we really
(31:27):
make much headway there.
Which kind of goes back to something that I used to keep saying at work is that data
engineers aren't going anywhere in the near future.
We're still going to need all the data engineers to do all the data munging to make the AI
work.
Yeah.
The form of that that I hear a lot is like, well, we have co-pilot, we have coding assistants,
(31:53):
coders are going to be out of a job soon.
And another person that I follow on Mastodon, Jason Gorman, he's like, typing is the easiest
part.
Like actually deciding what, yeah, writing the code.
That's the easy part.
Understanding the underlying data.
If you've got a data set that you don't understand, you can't do anything with it.
(32:17):
You need to actually understand that data, understand what drives what behavior.
You need a human being in the loop for that too.
So it's a real challenge for organizations.
Like I told my developers that if they were thinking that they would be building Power BI
dashboards in the future, they need to get that out of their head that people would be
(32:38):
able to build those using co-pilot.
That'll be.
If you're seeing your value in that, you need to get that out of your head.
The real stuff is the value add of understanding the data, getting the data over into our world,
transforming it and translating it.
So it's ready to use FAI.
That's the real value add.
And that's where we want to play going forward.
(32:59):
Yeah.
The domain expert, the person who has a gut feel that measuring this and discretizing
it a certain way and combining it with this other thing, like the teachers who knew the
five things that they were already kind of watching for.
Like that value is we don't have any kind of a computer that can do that.
(33:21):
Humans are where that comes from.
And that's what takes any problem and changes it from intractable to all of a sudden just
a challenge.
Yeah.
And that was the best thing.
So we had some expertise from outside helping us with that proof of concept.
We had my team learning on the job.
And then we had the people in the education space working with us.
(33:45):
And then we had the teachers input.
So it was really a collaborative thing.
So one of the things that I'm really seeing will drive a lot of this in the future is
people working together collaboratively, bringing different bits to the table.
Yeah.
The projects that I've worked on that I've enjoyed the most have had these in the same
(34:06):
room.
You have the data engineers and you have the algorithm person and you have the person who
has spent a lot of time with users and has a good sense of kind of like what they value.
And speaking different languages, valuing different things coming from backgrounds where
there's totally different incentive systems, but taking the time to build the bridges to
(34:30):
understand like, okay, here's where the real sticky points are and what we're trying to
do.
Here's the hard parts.
Here's what, and here's what it'll take to make it good.
Let's kind of do this.
And it's confusing and it's messy and it doesn't look like a homework problem that you get
a thrill out of like just nailing the right answer.
(34:53):
But when you're done, you feel like you have made something, like you've created something
real.
And that feeling is like, that's one of my favorite feelings.
Yeah.
I love that too.
That's one of the things that is really exciting me about being able to build things.
I'm still a techno utopian at heart, even though I've been crushed many times.
(35:16):
So I still hold out hope that AI can make the world a better place.
I am fully confident that AI can make the world a better place.
It all depends on the hands that's holding it.
And that seems like a good place to draw this lovely conversation to a close.
(35:36):
Thanks, Brandon.
Really enjoyed our chat.
Thank you so much for having me on, Kate.
It has been an absolute delight.
Welcome to another episode of the Dead Revolution podcast.
(00:06):
My guest today is Brandon Roura, who I found online and cyber-stalked.
Welcome, Brandon.
Hi, Kate.
Thanks for having me.
I really wanted to have a chat with you because, you know, I actually found you by, I was googling
ML back in the day a couple of years ago, and I found your e2eml.school website, and
(00:29):
it talked about your learning journey about machine learning, and I was really fascinated
by it and showed it to my team.
We found it an excellent resource.
So I found you there, and I started cyber-stalking you.
And I think I started following you on Twitter when it was still a theme, and now we're context
on Mestodon.
So that's how I found you, and I'm really pleased that we get to have this chat.
(00:50):
I love that story.
To me, that highlights the very highest promise of social media because I've never been to
Australia.
I live in Boston, and the chances that we would have bumped into each other in a coffee
shop are probably not high.
So yeah, and that's one of the things, and it's one of the things I think is really
(01:12):
so tragic about the demise of Twitter is how everything's fragmented now.
Yes.
I'm so sad.
I enjoyed Twitter a lot.
I had connections with the data, data science, machine learning communities that honestly
are gone now.
There's a handful on Mestodon, and there's LinkedIn, but the tone is different.
(01:40):
It's like all you got to make sure it's suitable for work and pretty on the up and up.
And then just informal channels outside of that, but Twitter was, I'm still mourning
that as well, still looking for the next incarnation of what that could be.
And the real problem for me is people are fragmented.
So some people have gone to Mestodon, like the entire infoset community, Hollis-Bollis
(02:04):
moved to Mestodon in November 2022 because I had a Mestodon account from 2017, but there
was no one there.
Suddenly in November 2022, my entire cyber community was there, and I was like, oh, I've
got a reason to move now.
And then a bunch went to Blue Sky, and then LinkedIn's the surprising winner, which is
(02:25):
really weird with people kind of using it like Twitter.
But anyway, it's just a strange world now.
So why did you come to be working with data?
Because I loved your website, and it was like, it took through your journey about learning
ML, which was really fascinating.
It was not my original destination.
(02:46):
My first technical love is robots and robotics.
And so that's what I did my graduate work in.
My degrees are in mechanical engineering.
My dissertation is in using a robot to rehabilitate people who have had a stroke.
But in the process of doing that, I was, this was back in like 2000, I had a whole bunch
(03:12):
of human movement data, someone moving a joystick around for an hour, three times a week, times
six weeks, times 50 people.
And back then, that was gigabytes of data.
Oh, too much.
That was a terrible amount.
And I had to find patterns in this.
(03:34):
And I had a rough idea what I was looking for, but I wasn't certain.
So I had to hack together database.
I had to hack together some routines.
I was working in MATLAB.
And it was like, anyway, so it was like, I was doing algorithmic stuff and data stuff
without any training that was formal to it.
(03:57):
And but this ended up tapping into my love because it, through the course of the research,
revealed some telling things about how humans moved, which led to some interesting hybrid
hypotheses about how we could help robots move in a way that was more, more robust, more
(04:19):
able to work in an unpredictable environment, which led me to say, great, well, maybe I
can start making some robot learning algorithms and control algorithms that are based on this.
And then pretty soon, I'm doing things that look a lot like existing machine learning
techniques and someone stands up in a talk and says, Hey, isn't this just, you know,
(04:40):
a certain algorithm?
And you suddenly go, huh?
And then I was like, Oh, shoot, I should learn about all that stuff.
So I don't look dumb.
It's like kind of a bunch of machine learning textbooks and just started kind of reading
through them.
I mean, this is over years.
It was not, there's no Rocky montage here.
But and I realized also that a lot of this stuff, I'd had this experience over and over
(05:02):
again where I'd read through a handful of papers and posts and book chapters.
And when the realization finally clicked for how it worked, I was like, Oh, well, that
was less complicated than I thought.
And I realized like, well, what if I wrote the thing that I wish I had when I first started
(05:23):
learning this?
And so I started doing that for one of my early ones was for convolutional neural networks.
When I was back at Microsoft, I'm turned it into a video and it got a bunch of views and
it's still one of my most popular ones where someone's like, the comments a lot of times
will be like, I got it.
(05:43):
Like I've seen a half dozen other tutorials, but I got it.
And it's full of errors and oversimplifications.
But I am proud of it because it took some important, some hard to grasp concepts and just cartoonized
them and made them easy to ingest.
And I, it's not really out of any deep altruism that I've done this, but I find it saddest,
(06:07):
very satisfying for me.
And you know, if you read between the lines, whenever I'm published something new, you
can always tell that that's what I'm confused about in a particular month.
I'm trying to figure it out and so I write it up.
And then it's just kind of like over years kind of congealed into this big collection
of posts and videos and whatnot.
(06:30):
Oh, yeah, it's a really great resource.
I shared it to my team because we were grappling with, we did our first machine learning curriculum
concept back in 2019.
And we were trying to predict which students would fail a course during the course.
Oh, wow.
Yeah.
Because it's highly correlated, students fell in courses means they failed degrees
(06:52):
and it's highly correlated to them self harming.
So we actually wanted to find out who might be on a trajectory towards self harm.
Got it.
So the course failure was like the leading indicator for the thing that you were really
concerned about that was more rare.
Yeah.
(07:13):
And the thing was, you know, we had data on who was in the counseling service and was
getting treatment, but all the people who weren't were the ones who would go on to self
harm.
So, so it was kind of a real, really big problem that we wanted to solve.
So and we was efforts for a intuit.
And we're on the Microsoft platform and they were really helpful to us and we did a great
(07:37):
proof of concept and took six weeks to build it, build this model.
It's still running in production now.
Oh, wow.
And it just went on, which is great.
That is, that is such a good story because, you know, it's rare that someone actually
uses machine learning to make the world a better place.
(07:59):
But it sounds like actually the case here.
I am curious.
Did you have to stretch to like figure out what was important to measure or to include
in the model?
It was hilarious.
So, so the teachers had five factors that they thought were important and we tested
(08:21):
all sorts of factors because we had all the data and it turned out to be the five factors
that the teachers thought.
So they were right.
So they were very chuffed that they were right.
But there's really basic things.
Has the student logged on to the learning management system?
You know, have they done their first assignment?
(08:42):
All of those sorts of things that you would think are really predictive are and it's been
running at 95 plus percent accuracy since since then.
So it's pretty good.
That was that was our first for it and I had my development team were like, what?
And I found you found your website and showed it to them and they were like, oh my God,
this is so helpful.
(09:03):
I'm so glad.
It's funny because a lot of times like, I mean, I did a deep dive into transformers because
I wanted to understand about transformers and LLMs and all of that.
And at the very end, in my own mind, the punch line after doing these deep dives usually
(09:24):
comes down to, you know what?
It really comes down to how good the data is going in.
And you can have these huge models and dump in all of these unrelated features and the
results are met.
But if you have even a very simple model and you have domain experts like the teachers
saying like, you know what, my gut says it's this handful of things.
(09:47):
Let's look at it.
That wins every time.
Well, the interesting thing is, you know, every organization needs to get better at
data governance, which includes data quality.
Because you know, if you're working with poor quality data, you're going to have some standard
outcomes, especially if you can be driving actual decision making by these technologies.
Yeah.
(10:08):
I could not agree more.
It's easy for algorithms to get a lot of airtime.
But issues that really get my attention are, get my positive attention when someone's collecting
more data about something or higher quality data or more careful data and gets my negative
attention sometimes when someone's making data harder to get, less visible or somehow
(10:33):
corrupting it or calling it into question.
Like that, at that point, like there's not really anything that you can do.
I'm thinking about like government, public entities, you know, the environment.
In Australia, we had a particularly bad thing called RoboDat, which was the government using,
(10:58):
and it wasn't machine learning algorithm or anything.
It was just Excel basically, but it was still using a formula to determine if people owed
it money.
They were doing really egregious things.
Some people committed suicide over being harassed by the government saying that the government
owed them money.
It was really an evil program.
(11:22):
The fascinating thing about that was the public servants were just saying, oh, we're just
impersonally executing government policy, but it was a really bad policy, but it was
driven by this underlying algorithm and you couldn't prove that you didn't have a debt
for the most part.
If you had a lawyer, if you could lawyer up and just go, my lawyer will be in touch,
(11:43):
you got off, and if you were a poor person, you would end up having your pay garnished
to that government debt.
Wow.
That sounds like exhibit A for like how not to use automation and data to make the world
a better place.
Yeah, it is.
It is.
(12:04):
I spoke out about it when it was running and people were like, oh my God, you're saying
it in public and I'm like, yeah, I'm saying it in public.
It's really bad.
It's not good for people.
It's so, so bad.
Yeah.
The ideal use case for machine learning is when there's a lot of decisions that are
(12:25):
small and there's not a very high penalty for getting them wrong, but there are just
so many that a human either gets bored or just can't move fast enough to do them, or
if the information is just like so broadly dispersed to be able to pull it in.
That is the absolute sweet spot.
(12:47):
But the whole, the book, Weapons of Mouth Destruction, Cathy O'Neill, some great case
studies in there about how similar to this, it's like when you have an algorithm, there's
going to be mistakes and when there are people, it's making important decisions about people
on the other end.
(13:08):
Those mistakes are unacceptable.
If those decisions have to be made, you have a person make them.
Always like you just get out of it.
You can have the query it.
So, I mean, the whole thing with the advent of generative AI, and you know, it's been
a kind of a thing for a while, but it's sort of popped into the public perception back
in 2021-22.
(13:30):
And I think the interesting thing about that is ordinary people could understand what AI
could do for the first time.
Like people like my husband, who's a high school maths teacher who is not technology
driven at all.
He uses MATLAB and stuff, but he doesn't know any programming.
(13:50):
But he could see it and see what it could do for the first time.
And I think that sort of captured people's imagination.
But you know, the problem of hallucinations is an interesting one.
I think imagination is the key word there.
It's like in our household, we name our automobile, we name our devices.
(14:15):
It's just a cute thing and it's fun.
But it's easy to personify an inanimate object no matter how dumb it is.
And to start to ascribe motivations to it and ascribe internal processes to it that
simply aren't there.
Yeah, if you like the classic ELISA computer program from I think it was the 60s.
(14:42):
Where it just followed some kind of rope.
You say something, it asks you a follow up question based on that and it sounded plausibly
human.
That was the original blow people's minds.
It's like, oh my goodness, computers are thinking and it was in this first round of
AI sci-fi, it's going to take over the world.
(15:05):
But now with so much more data at its disposal, like LLMs are so much more plausible, even
people who are familiar with how they work on the inside have to be really, we have to
remind ourselves, there is no ghost in the machine.
This is really, really good autocomplete.
(15:28):
And yeah, if you give it enough data, it will give you very good answers and maybe even
the right answers in some fraction of the cases, which it does.
It's great for a lot of things, but it doesn't have this rich.
It shouldn't be like those lawyers who just took the generative AI, created case summary
(15:48):
or whatever it was and sent it to the judge and had it completely invented, completely
hallucinated citations.
Yeah.
Yeah, I mean, so my wife is in the education world.
She is a PhD candidate.
So she gets to see students and professors working, dealing with generative models.
(16:14):
And they're really good for cases where you have to produce words and you don't care what
they are.
And there are, sadly, like that is a common use case in even in education.
And it's great for that.
It can be quick.
It's also really good.
I've seen really good use cases where you wanted to hallucinate.
(16:38):
You're like, hey, I'm thinking about this thing.
I just, can you give me some suggestions, some ideas, and then I will filter through them.
Yeah.
It's great for that.
I think the process is an important concept I think people need to get.
Yeah.
I was just going to ask you, though, do you think that the RAG models, so retrieval or
(16:58):
augmented generated models are going to be able to help reduce the amount of the hallucinations?
I think so.
And my kind of like one sentence summary of RAG models is they're really good at taking
what you say and then doing like a really quick local search and pulling in actual real
(17:21):
texts that are related to it and using that.
So it's kind of like some kind of a hybrid between an internet search and generative.
But I think they're effective because they are a step closer to sticking with the raw
data.
And so they, I am still waiting to see some mechanism by which they could say like, well,
(17:49):
even with what I've been able to retrieve, I'm not confident that this fact is the case
or that this legal precedent has occurred.
I'm just going to sit back.
I'm just going to say nothing.
LLNs are really good at nothing.
I genuinely think that they're actually throwing into disarray what we've conceptualized as
(18:12):
intelligence because a lot of what we've thought of intelligence is just goodness.
Yeah.
So 15 years ago, I was because of a line of work I was in, I was attending these AGI
(18:34):
conferences and this is back when it was very much a fringe activity.
It would like co-locate with these AI conferences but be like in a little venue, like not even
co-located with me.
Yeah, like two days before the real conference and you kind of said you were going there
(18:55):
and like winged and be like, you do it ironically.
But yeah, even then it was hard to pin down what we meant when we were talking about intelligence.
And the thing that really drove this home for me is there was one big Q&A session where
(19:17):
we were getting into it and I just stopped.
There's room of like 100 people.
I was like, so I feel like I'm hearing two things.
Intelligence is it operates with the same mechanisms as the human brain or intelligence
is it does approximately what a human can do, like functional.
And if you could have just one or the other, which would it be?
(19:41):
What's core to you?
And the room was split 50-50.
So like this gathering, like we were all looking for something different.
We all meant something different when we were talking about intelligence.
And if there is a place where we had a hope of defining it, it was there.
And we weren't even close.
Yeah, but I'm just fascinated because you know, generally AI is capable of string together
(20:08):
really good words that sound really good.
And we I think a lot of us have mistaken that for intelligence for a very long time.
And we probably need to dig a bit deeper because we now know that generally AI is often quite
wrong.
Yeah.
(20:29):
Yeah.
I mean, there are the there are the jokes about, you know, there are there are humans
who are very good at stringing together words that sound great.
And then when you pull them apart mean nothing or contradict themselves or have no grounding
in reality, in which case, LLMs are extremely human.
Yeah.
Yeah, yeah, they are very intelligent.
(20:52):
Whether they're intelligent is a separate question.
But I think that that's a big question for our time.
What do we really mean by intelligence?
And when we're probably going to be, like you said, 50 50 split on what it is.
Yeah.
And any room that we're walking to.
This is something that has has been in my mind for a while because even going back.
(21:17):
So set aside reasoning, production of words, you know, playing chess, things like that.
Even if you go down to something very mechanical, like moving an arm.
So back when I was doing human movement, stroke rehabilitation, there were some very large,
(21:37):
ambiguous, unresolved questions about how the human brain even moves our limbs.
And it's just so, I mean, it's just so hard to measure what's going on.
We don't even know what to look for.
There was even a debate, like, are we looking to measure like one neuron that controls something
or somehow it's in the distributed set of neurons that are doing something, or maybe
(22:01):
it's in how those patterns of neural activity change over time that's doing something.
And there's so much going on and it's hard to measure it accurately.
Like no one could agree even on what we were even looking for.
And this is something where we can measure movement very, very accurately.
We can all see it and agree on it.
(22:21):
There's no disagreement there, no confusion there.
And we still don't know what what motor control is at its core.
And so abstract that out a little bit to like reasoning and empathy and model building and
like, you know, building metaphors and analogy and communicating, building theory of mind.
(22:49):
And it's like, I love it because we are so utterly clueless.
Like it is in my mind, the frontier.
We all do it.
We all see other people do it.
And we don't know how.
Yeah.
Yeah.
I only recently became aware of how difficult it is to build bipedal robots.
(23:15):
Oh, yeah.
Yeah.
So I had friends in the MIT Lake Lab, which has done some great work with bipedal robots
going back into the 80s.
But back then it was fun because I was fascinated with robots and they're doing this great research.
(23:38):
But even then it was like their their failure, their B-roll failure reel was hilarious because
these robots would just just like fall over or jump and land on their head or do a flip,
but but not quite clear it or just stumble over a chord.
And it's easy to forget that now because that line of research has since evolved into
(24:00):
Boston Dynamics.
And we have, you know, the big dog and we have Atlas and some robots that do some just
amazing things.
But that's decades of focused research from some very dedicated people, which I have immense
respect for.
(24:20):
And even then my understanding of the current state of the art is you change that hardware
a little bit and the policy that they have is not very helpful anymore.
Like it's very specific.
So things that humans do without even thinking about it are really, really hard to program.
(24:43):
Like we're really complicated.
Yeah.
Yeah.
It's kind of funny too, because we stack the deck in our favor when, you know, chess,
for instance, for a human, that's hard.
Like you have to study and practice for a long time.
That's a great problem for a computer.
Like just small finite set of moves and rules and finite possibilities.
(25:08):
The three heavily rules based world.
Yeah.
Yeah.
So when we initially set up checkers and chess as measures of machine intelligence,
like we were really stacking the deck in their favor.
But things that a five year old does, you know, like stumbles around on a shag carpet
(25:28):
with sticky sold shoes.
Like, oh my gosh, that's a nightmare for a robot being controlled by a computer.
I mean, one thing I did want to talk to you about is this idea of where you think Generative
AI is going.
Because you've heard people talk about a genetic AI, which is the idea of the next generation
(25:54):
of Generative AI with agents that can daisy chain their own prompts.
And so I've been talking at conferences about how I think things like AI, Internet of Things,
robotic process automation are all joining together to make autonomous systems, which
now people seem to be calling a genetic AI.
(26:17):
And I sort of see that as the next interesting thing that people are going to do things with.
I like that line of research.
I think that has a lot of possibility.
Let's see a couple of thoughts.
So LLMs are not perfect, but they are a tool that we have not had before.
(26:40):
So like any tool, there's some things that they are very good at and some use cases that
they just really nail.
But like any tool, if you use it places where you shouldn't, it's going to look silly.
So we're kind of exploring where that boundary is right now.
But to your comment before about, you know, a lot of us, a lot of us, a lot of the times
(27:01):
are just stringing words together in a way that makes sense given the context.
And so for a lot of non-critical situations, like if there's a computer program that can
do that, like that can meet certain needs.
Like if you're going to an information desk and asking for help, like how great to be
able to do that in natural language and get a reasonable answer.
(27:24):
So I think we'll get better at helping LLMs, like not say dumb stuff.
I really think that being able to train an LLM on the data that's specific to what we
want it to do, as opposed to having like giant.
General purpose LLMs.
(27:45):
So we think that we're going to get away from general purpose to specific purpose ones.
Yeah.
And it may not even have so much to do as a difference in architecture.
It's just the training, training data.
But yeah, like domain specific.
And if we can get a better at being confident in what it produces and not producing things
(28:07):
in time.
There was a smaller training data set with a more constrained set of questions that it
can answer confidently.
Yeah.
Yeah.
Then we could treat it like any other human where it's like, you know, I meet someone new
and I interact with them and I'm building my own assessment of whether to trust what
they say, whether they know what they're talking about.
(28:29):
And yeah, I think there's some really great opportunities there.
Yeah.
I see some really good use cases there.
It's also really fun.
Like on the data side, like right now, the easiest thing is to train on words, characters
(28:52):
from the internet.
That looks, you know, different depending on like, if you limit yourself to one language
versus considering lots of different languages versus considering like very localized slang
and including emojis and kind of like making it broader.
But it's still just like a sequence of symbols to a sequence of symbols.
(29:16):
There now they're including like steps to expand that to audio, some great demos on
the most recent GPT I have to give a nod to like a great job doing audio to audio and
text to speech and speech to text.
(29:38):
And then there is the promise of doing this with video too.
The trick being that with printed text is already symbolic.
It's already reduced to ASCII or Unicode.
And with speech and video, there's this extra step of how do you take this high dimensional
analog continuously variable input and reduce it to symbols and then get it in there.
(30:03):
So I expect that.
Yeah, I was talking with a startup that's trying to do some stuff in that space at the
moment and they're just trying to try to automatically translate it to from video spoken from language
to language.
And that's already a pretty hard problem at the moment.
(30:23):
Yeah, yeah, it's super hard, which coming back to robots.
It's fun.
Like a lot of early robot work was done in simulation where you could chunk the world,
you can make the world symbolic and make it so therefore straightforward, more like a
chess game.
(30:45):
When you're dealing with physical robots, you're dealing with actual continuous.
There is audio, there is video, there is these weird continually varying sensors.
And you have to figure out how to get them into this symbolic state, how to make them
symbols so that then you can get them into the computers native language like the computers
(31:05):
playing field.
But that's to me like those data issues, data quality, data transformation.
Those are hard problems.
That's not something that a company is going to plow through to get the next feature out.
Those are things that will take years to sort through and trial and error before we really
(31:27):
make much headway there.
Which kind of goes back to something that I used to keep saying at work is that data
engineers aren't going anywhere in the near future.
We're still going to need all the data engineers to do all the data munging to make the AI
work.
Yeah.
The form of that that I hear a lot is like, well, we have co-pilot, we have coding assistants,
(31:53):
coders are going to be out of a job soon.
And another person that I follow on Mastodon, Jason Gorman, he's like, typing is the easiest
part.
Like actually deciding what, yeah, writing the code.
That's the easy part.
Understanding the underlying data.
If you've got a data set that you don't understand, you can't do anything with it.
(32:17):
You need to actually understand that data, understand what drives what behavior.
You need a human being in the loop for that too.
So it's a real challenge for organizations.
Like I told my developers that if they were thinking that they would be building Power BI
dashboards in the future, they need to get that out of their head that people would be
(32:38):
able to build those using co-pilot.
That'll be.
If you're seeing your value in that, you need to get that out of your head.
The real stuff is the value add of understanding the data, getting the data over into our world,
transforming it and translating it.
So it's ready to use FAI.
That's the real value add.
And that's where we want to play going forward.
(32:59):
Yeah.
The domain expert, the person who has a gut feel that measuring this and discretizing
it a certain way and combining it with this other thing, like the teachers who knew the
five things that they were already kind of watching for.
Like that value is we don't have any kind of a computer that can do that.
(33:21):
Humans are where that comes from.
And that's what takes any problem and changes it from intractable to all of a sudden just
a challenge.
Yeah.
And that was the best thing.
So we had some expertise from outside helping us with that proof of concept.
We had my team learning on the job.
And then we had the people in the education space working with us.
(33:45):
And then we had the teachers input.
So it was really a collaborative thing.
So one of the things that I'm really seeing will drive a lot of this in the future is
people working together collaboratively, bringing different bits to the table.
Yeah.
The projects that I've worked on that I've enjoyed the most have had these in the same
(34:06):
room.
You have the data engineers and you have the algorithm person and you have the person who
has spent a lot of time with users and has a good sense of kind of like what they value.
And speaking different languages, valuing different things coming from backgrounds where
there's totally different incentive systems, but taking the time to build the bridges to
(34:30):
understand like, okay, here's where the real sticky points are and what we're trying to
do.
Here's the hard parts.
Here's what, and here's what it'll take to make it good.
Let's kind of do this.
And it's confusing and it's messy and it doesn't look like a homework problem that you get
a thrill out of like just nailing the right answer.
(34:53):
But when you're done, you feel like you have made something, like you've created something
real.
And that feeling is like, that's one of my favorite feelings.
Yeah.
I love that too.
That's one of the things that is really exciting me about being able to build things.
I'm still a techno utopian at heart, even though I've been crushed many times.
(35:16):
So I still hold out hope that AI can make the world a better place.
I am fully confident that AI can make the world a better place.
It all depends on the hands that's holding it.
And that seems like a good place to draw this lovely conversation to a close.
(35:36):
Thanks, Brandon.
Really enjoyed our chat.
Thank you so much for having me on, Kate.
It has been an absolute delight.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.