00:12
Hey, Asami.
How's it going?
I'm exhausted, but I guess it's going well.
I am wondering, what is the legal limit of usingjet lag as a way of excusing oneself to perform
less or show up less as a human being to thisworld?
Three months.
What do you think?
Three months.
That sounds really great.
Let's just hope that my boss agrees with that.
I know that we probably just published my chatabout my summer travels that's already over.
Yes.
Exactly all of my summer fun travel is done.
And it's not even end of July, so, you know, haveto figure out what to do for the rest of the
summer.
But I also just came back from another travel fora week.
This time it was in Switzerland.
But this time it was for work.
It was for a conference.
Because my DM is plagued by your AI rage content.
What is wrong with people?
What is wrong with people who think AI can solvetheir problem and save them from their misery type
of shit?
Of course, you know, these things are just now inthe background of my baseline thinking.
I guess I'm not the only person.
You know, it's pretty hard to ignore AI discussionin this day and age.
If you are even a little bit participating in thesociety, I guess.
Yeah, I think that's a fair observation.
I say this to my students a lot, right?
I say it to faculty and others that I work with.
Because of the way that the tool functions, andbecause of the resultant capabilities it has, and
the way that it's designed, and the sort ofresources it takes, and the people power it takes,
right?
It just kind of touches everything.
And it causes a lot of uncertainty.
And that level of uncertainty is differentflavors, right?
It could be distress.
It could be thrill.
It could be, you know, pick something, right?
Yeah.
Yeah.
Well, not because of just that, but also probablybecause it's been two years since the same, not
03:00
the same group of people, but the same conferencewas hosted.
This only happens every other year.
And sort of I would put this conference as one ofthe more important international conference for my
field of conservation science.
It's important, not necessarily because of, like,the high profile names that come to these
conferences.
They do.
But because we talk about sort of more like a bigpicture stuff of, like, the field as a whole,
like, where are we heading?
What's going to be the important problem to solvefor the next decades?
And because of that, it's very small.
It's, like, capped at, like, 150 people in remotelocations.
It's really, really hard to get to.
And, you know, because there's nothing else in thevillage, we eat breakfast, lunch and dinner all in
the same dining hall.
And, you know, we all eat at the same time.
So we have to converse with one another.
We will actually get to know each other.
Oh, no.
Human interaction.
I know.
The only sort of problem is that maybe becauseit's set in Europe, it's way more represented by
European, like, populations of this field, whichis already overrepresented by Europeans anyway.
So I was a proud handful of Asians working inAsian institution to show up there and represent.
Make my voice heard?
No, not really.
I was just more of an observant situation thistime around.
Fair enough.
Anyway.
So imagine this, right?
Like 100, 150 scientists working in extremelyniche field, locked up in the mountains of
Switzerland, talking about what we do, but alsowhat we should do for the next decade or two.
And two years ago was my first time participatingin this conference.
And I think during that time, there was exactlyabout one talk that touched anything remotely
machine learning or AI, really.
And from what I remember, even then, I felt likeit was a very sort of crude step and foray into
machine learning.
I felt like their application was too narrow forsort of the expense of training and the model
development.
And even then they were doing like a very sort ofpretty basic architecture.
06:07
Like they didn't really invent a new architectureor model training method.
It was more about, oh, we took this verystraightforward convolutional neural network into
the stuff that we use.
And it was like, okay, novelty on the combinationof the application, but no novelty on the science
front was my impression.
Yeah.
Did we talk about that?
I don't think so.
Okay.
But famously, my field is kind of like highlyinterdisciplinary and that's why it's exciting for
me personally.
But it is also kind of easy to quote unquote fakeinterdisciplinary.
That's a talk for another day, by the way.
All right.
But sort of like for me, there's not much point.
Like I don't think you can call it reallyinterdisciplinary when you just like took another
field's established technique and apply it to yourown sample.
That's like good job on spotting that you could dothat.
But I think we're just going to dilute the meaningof interdisciplinary if we keep using that word
for these kinds of investigation.
Anywho, long story short, two years ago, prettycrude.
This year, a lot more student poster presentationinvolved AI and machine learning stuff in somewhat
more innovative use.
And also sort of like addressing the specificchallenges that my field face, which is usually in
my field of conservation science, because we'redealing with like uber precious sample.
It's really hard to find a situation where you canfind or have a ground truth.
So that's sort of like a fundamental to any AImachine learning investigation.
You know, where do you stop? Right.
Like, where do you have your stopping point, whichis your ground truth?
We often don't have that.
So we always have to kind of negotiate how usefulthe model is based on some other metric.
And I think this time around, people are using itfor more clever and measured way where it's
usually in a combination with other skills andother techniques of analysis.
So felt more encouraged that they're using it.
And I think they're using it more and moreprojects seem to be using it for stuff that AI is
09:02
actually good at.
Like machine learning is actually good at.
Okay. Yeah.
Which is to say, you know, I feel like sometimes Isee papers where you could have solved this
problem just classically, like just using thescience we already know.
It's like, you don't have to machine learnifyeverything.
There's, okay, very small sidebar.
Yeah, I just want to add to this, in the sense ofthere was a recent article, which I won't pin, we
can maybe save for another chat.
But that used, essentially, a very fundamental,they use like alpha fold, right?
For protein folding type stuff, right?
Yeah, yeah, yeah.
That's not a large language model, unless I amvery mistaken, right?
But their introduction, their introduction writesabout large language models first.
And so they essentially, they're tying it togetherwith this idea of AI.
They tied it.
I respect some of what they did, where there was abit of like, they did use sort of the alpha fold.
And they seem to have played around with like,maybe a language model.
I'm not really clear whether they did or didn't.
For like the, what did they call it?
Like the sort of ecologically relevant, like somesort of little bit of tidbit and maybe, I don't
know, a bioinformatics angle or something.
But it was, to me, it just screamed, this is notthe important part, but you put it here because
you are attempting to, you know, situate it withinthe gen AI discussion, right?
I think it's really boring and claims sort of lackof creativity on our part, scientists' part.
When you take something like Alpha 4 and your ideaof making it user friendly, quote unquote, is to
slap some sloppy AI agent on it.
So that people can search using words instead oflike, protein structures that they already know.
I don't know.
Like, this is stupid.
I don't know if that's what they're doing.
But you know.
There is a place for that type of like discussionand sort of accessibility, if you want to call it
accessibility feature.
Right, right, right.
But that's like a short draft or like a quickworkup.
You know, maybe not a working feature.
It's such a distraction.
It's just like a, hey, okay.
All right.
But now I have to, yeah.
Perfect.
A distraction.
Now you have to deal with it.
It pops up in front of you.
And like, yeah.
So, oh, yeah.
Anyway.
So like, I was pleased to find that there weremore projects involving machine learning, AI,
12:05
various other form of, you know, non-mathematicalmodel based modeling.
That utilized it in a way that is like kind ofsensible to like bridge the gap and address the
problem that we often have in cultural heritage,which is that we don't have ground truth.
We don't have a lot of sample.
And we don't often have access to many, many, manydata.
Because gatekeeping in museum world is a thing.
But I was interested when sort of like one of thebig keynote speaker talked about like, hey, so
like, you know, I don't think our field can escapefrom this wave of adopting gen AI and other forms
of machine learning.
Just like nobody can really escape anymore.
But what is sort of how should we as a field thinkabout it, deal about it?
And I was sort of like, I felt like there needs tobe a lot more discussion about this.
Because after he mentioned several concerns, oneof which was really hit like a like a big point
for me, which was the environmental cost of usingand training machine learning models.
Like, in museums nowadays, we are super concernedabout sustainability policy and keeping the
cultural heritage alive, because many of them aresuffering from climate change.
Many of them are suffering from increasing energycost to keep their collection safe.
And like, these are more dire concerns thanwhether we can keep up with the AI wave, in my
opinion.
But like, it was just also really hypocriticalthat like, like the the very next talk after he
mentioned this was like, so we like chuck this allof data into like our machine learning models.
Oh, great.
Right, right, right, right.
Here we go. You know?
Oh, boy.
It really is. I don't know, I guess this is stillapplicable to larger society, too.
But especially kind of concerning in science,where, you know, I still naively maybe believe
that we are after uncovering something that isscientific, true, scientifically true, or what we
think it is scientifically true.
But like, when it comes to just kind of like bruteforcing everything, all of your research on some
15:06
five coded machine learning model raises a lot ofquestions.
And, you know, it becomes the problem of like, whobrute forced it with the most amount of resource
and the fastest possible way.
And it becomes so, so far away from empiricaldata, like based chemistry and physics that we
have developed over decades of this, you know,fields existence.
I'm like, I'm far more interested in that.
And like seeing what the data has to say.
And if machine learning can help us uncover it,because we come back with like terabytes of data
when we don't know what to do with it, or where toattack it, or something, right?
Like, if there's some sensible way of using it,and if we're really careful and critical about how
we're using it, I think there is a room for it.
But I'm worried that this tiny, tiny little nonmoney making field that is cultural heritage
science is also following the same economic modelof sort of like pharmaceutical or other, you know,
more money forward industry, where there's like,gen AI the fuck out of it and like see what
sticks.
And I am like, like, no, we don't have to do that.
And, and, like, can we not?
Can we not is like the phrase that I cannotprobably count how many times I have said aloud
and in messages over the past, like five years.
And I think, you know, I'm saying five, that'snot, you know, GenGBD technically came out later,
right?
There were a lot of can we not for a lot of thingshappening, you know, during a particular pandemic
time as well.
But the sort of, you know, gen AI explosion,craze, age of whatever we want to call it, it's,
it's a lot of can we not? Can we ask those reallymore critically focused, more intentionally
focused questions around whether we ought to bedoing it first?
Right, right, right.
I don't think I'm a closed minded person.
I think I'm all for new ideas and new methods.
But this is cheap.
Like, this is not, this is just like, riding on asame bandwagon of like, everyone else is doing
like, what, look at all the cool kids are doing,like, you know, whoever gets the most amount of
18:02
cluster space wins the race, like, that's soboring.
It's a science.
And yeah, but making ideas that are unique, or atleast, you know, come from one's dedicated
experience is hard and challenging.
It's much easier to pick up a thing that lookslike it can just do the work for you.
And people are like, ooh, shiny toy.
Didn't we decide to do science because it'sfucking hard and interesting? Like, is that not
the case for people?
Some people, yes. Other people, no, they were toldto do science because they were told, for
instance, it's the place where you make money.
Nobody fucking told me it's easy. I don't intendit to be easy.
I don't think people, that's, I mean, it's a goodpoint.
I don't think people tell us it's easy, right?
But in the same lens that we see shifting of wherepeople spend their, you know, youth training,
right?
Like, do they go into university?
Do they pick a philosophy based degree?
Do they pick a STEM based degree, right?
The motivations behind them are never as pure asthe one you just described, which we might hope
for, right?
This we go into, it's hard, but it's interesting.
It's questions to understand better the worldaround us, the importance of that thing, you know,
to try and recognize that we are doing ourdamnedest to follow that.
I mean, I don't know, though.
I mean, that's probably not the case.
You know, this whole, like, idea of people aresort of losing the initial motivation type of
thing.
Like, that might be true for, you know, many, manyother scientific fields.
I'm particularly disappointed if it's my field ofpeople doing that, because we did not do this
because it makes money.
We do this, what some other scientists might callacute science, because, you know, we work with
artworks.
We work with cultural heritage, like usinganalytical chemistry technology that was like
invented decades ago, still, because that's all wehave in a field.
And we're still trying to do cool science and newresearch to better understand cultural heritage,
because, you know, by the time it's lost, it'slost forever.
And like, if we stop caring about it, it's goneforever.
And so many of the art history and culturalartifacts is already lost as we speak.
Like, you know, war happening.
Have you heard of that?
I will not be sarcastic about that.
Yes, I have obviously heard about it.
Like, there is real global massive changes anddestruction of people and the creations they come
from.
Yeah, exactly.
And this is sort of like, yes, it's a little bitrosy gloss and idealistic view of the world.
21:02
And I do find it extremely privileged to be ableto make a living doing this.
But like, I think, like, from my people, you know,it's just like, come on, like, do we are not gonna
have open AI money, we are not gonna have EliLilly money.
And like, so, like, we don't have to hop on thisthing.
Like we can we actually don't have to compete witha whole lot of people.
You know, we don't have the same kind of publishor perish, like, pressure the way if you were
doing like Alzheimer's disease research, or likediabetes research, or, you know, other things like
that.
Like, I mean, we do still have to publish, butit's not like, oftentimes, it's, there's only so
many of us in this field, you know.
So it's like, we can afford to like, take timeinnovating.
And I think it's like, that was sort of like mymain takeaway from this like AI discussion that I
had in this conference with like, people who havea lot of power over what the museum industry does
to the larger population, right? Like, like, theseare people from the top museum of the world, or
like universities that collaborate with, like,these people.
And even though museum industry, like we onlyoccupy a tiny sliver of it, not the whole thing,
but like, we still have access to this very publicplatform that is called museum.
And it's like, I would be kind of sad if, like,all we have to show for is just like, some cheap,
not even innovative use of AI.
Right, just, we tossed, we tossed a bunch oftextual descriptions of, you know, 17th century
art into an LLM, and it produced a bunch ofanalyses that we have now published in a paper.
Yeah, like, that will be boring.
Yeah.
So, I don't know, just food for thought. I don'thave much more to expand on, honestly.
No, and I appreciate you sharing it, though,because this is, the more that I talk about this,
right, because I am the one filling a lot of yourinbox and feed with the things that I am paying
attention to.
With that said, you know, a lot of theconversations continue to revolve around the same
types of things, right?
This is why I will get regularly, I got upset justbefore we started this recording, about people
being surprised about, you know, what happens inthe field right now with these.
24:03
And we won't get into that right now. But, like,with your field, and with this sort of, I would
say, parental disappointment that you have towardsthe people in the field using the large language
models, do you have a feeling for like, maybe whythey're leaning into it?
Is it just a, oh, we're trying to get money, butthey don't seem like they should be rushed to do
publications?
That's it for the show today. Thanks forlistening, and find us on X at Eigo de Science.
That is E-I-G-O-D-E-S-C-I-E-N-C-E. See you nexttime!
