AI利用における懸念と分野の現状
Though because this is the more that I talk aboutthis right because I am the one filling a lot of
your inbox and feed with the things that I'mpaying attention to um with with that said youknow a lot
of the conversations continue to revolve aroundthe same types of things right it is this is why
I will get regularly I got upset just before westarted this recording um about people being
surprised about you know what happens in the fieldright now with these and we won't get into
that right now but um like with your field andwith this sort of I would say uh parental
disappointment that you have towards the people inthe field using using the large language models
do you have a feeling for like maybe why they'releaning into it is it just a oh we're we're trying
to get money but but they don't seem like theyshould be rushed to do publications so is there a
weird tension here like what's the um no I thinkit's just that there are some like there's always
gonna be you know people who are more curious andnew adopters of the new technologies all right
versus people who are a little bit more laterjoiner of the trend and like remains conservative
I would say majority of our field not because it'scalled conservation science
tend to be more conservative like conservativeabout technically different meanings right
very different meanings yeah but same word yeahhomonyms blame them serving but in a different
sense so like we tend to be more old school Ithink in this field and um and that is also not
the way I think like it's not really like stickingto the old school science is not the way forward
either but um so I appreciate that there are a fewplayers in a field who are like a bit more
forward and curious about adopting this technologyto the sort of problems that we have so I I'm not
against that what I am against is what seems to besort of like careless adoption of like
the technology right yep like they like when youhave this wealth of knowledge that is
not digitizable in the field you know that is notexactly feedable to the learning model
I think that's treasure I think that we shouldtreat it like a treasure and we like that's that's
that's the giant shoulder that we all stand onright and I think some people have so little
respect for that and just like oh look what we cando in like three hours what a concert
a conservator takes three months doing kind ofthing efficiency right that's the thing
our field was never about efficiency like no itwas never about making it happen faster
you know or like making it happen uh parallelthings because cultural heritage by definition
is like unique on its own like not everyone hasaccess to Mona Lisa so like people who have access
to Mona Lisa can take all of their time dealingwith Mona Lisa you know yeah yeah I and still
publish in nature probably but like it's like it'sjust like it irks me so much that because okay
this is sort of like a field-wide problem most ofus are trained in sort of wet chemistry like
organic inorganic but like right chemistry yeah Iwould say that's the majority of scientists
few of us are chemistry physics like me and moreemerging group of people are sort of engineering
science type of like bridging between people whoare good at imaging computations like that sort
of things yeah that's like entering the field butlike not quite a big part of the field'spopulation
yet so when if you think about a field that'smostly occupied by traditionally trained organicchemists
they are not proficient most of them in this newlanguage of
like machine learning or yeah even even anythingthat's sort of like data-driven type of modeling
you know and um like we still like to do GCMS westill like to do like FTIR like that's sort of
the vibe here right and and it's these people whohave like sort of it's it's better if they claim
文化遺産とAI利用の倫理的課題
that they just like don't really get it but thereare some of them who are like they feel
like they sort of get it because they read enoughtech bro blogs or like um like youtubers who talk
about um the latest headlines regarding AI andthey sort of feel like they know the cool bits
of this technology and but without like the fullcontext or like full ethical consideration of like
what does it really mean to put these spectraldata on a public learning model that is like
you have no security over this and like what doesthat mean right like like information ethics like
these are people who are not who are not aware ofthis i don't claim any expertise either but like
i at least worry about it you know and and talk topeople who i think know more about it than i do
exactly which that's i mean that's a greatopportunity to have a human human collaboration
and questions about these things right like yeahto to do the science right to think about
if you're gonna if you're gonna play with itplaying with it is one thing right like tooling
around with a new a new tech related thing a newtool a new system yeah um but as you're describing
here this like i think i understand it enough butwhat they've understood is the messaging about
the object right they've accepted one version ofthe messaging about the object and they have
which is i think i've said this with many thingsbefore it's a very human thing to do right you you
get this information and you try to attach it toyour scaffolding right you attach it to the
organic chemist scaffolding and you start seeingoh if we could like you know train it on you know
the the organic chemistry structures that i'mfamiliar with then like i wouldn't have to go
through you know two thousand old-fashioned youknow synthetic uh synthesis reactions you know
defines i could just describe something i'mthinking about and you even that is like that's
a that's a fair thought right like maybe there's away to to think about this but as you mentioned
things like information ethics sort of security ofdata um yeah not recognizing that these tools
are owned by companies that are using whateverthey can get to train them so yeah like at any
moment charge gpt can shut down and be like nownow it's not available right like and to put on
a data that has to do with like humanity'sheritage artifacts you know like these are
something that like we cannot replicate we cannotlike print out the copies of and um
i don't know it's like to me it doesn't take thatmuch leaps of imagination to like get worriedabout
and like try to think it through right before i doanything about it which is why i honestly
prefer to work with like what we call mock-ups soit's like fake painting or like um you know
things that are made of the same material but notthe artwork so that i can understand about
the material but not necessarily have any damagesdone on the artwork like i honestly
much much prefer that over like and i would onlytry to interact with the artwork itself if it's
like the whole point of doing that right um yeahbut like i guess there are other people who are
like this is a cool new thing to do and we shouldtry it before others get to it yeah i know trying
to be ahead of the the curve this is right becauseunfortunately publishing culture rewards the
person who reports it the first so yeah like itcomes to like you know who has the most amount
of money to chuck it into this new project or likewho has the most amount of access to these
precious data and um and it just seems like beingable to work in this like very unique hybrid of a
AI研究における貢献と批判的思考の必要性
space of art and science and and this is what wecome up with like seriously can we not do
something a little bit more interesting with sortof the tools that we have and access that we have
i don't know i think you do know i think you knowthat we could do something a lot more interesting
like then and that's the thing because like youknow at least at this point what we're doing in
a field is like not so groundbreaking to the pointthat other people in other field haven't come up
with it um more often than not we're kind ofborrowing what is already tried and tested
using like medical imaging data for instance toapply to our imaging technique sometimes it's like
a good adoption of you know why in why like whyyou know don't reinvent the wheel kind of thing
right sometimes that's clever sometimes it's justlike okay so where was your contribution in this
you know like yes yeah where where is it that youcame in you know i think it's really easy
to fall into this trap of like oh i applied thelatest language model to this unique set of data
that we have and look what it did and like youknow the end result is like did we even want this
like does anybody want this that doesn't matterwhen the point is just about doing it fast and
doing it wild and sending it out there exactlywhen when the conclusion is like oh you don't
have to take it but i wanted to publish it becauseit's interesting and it's like new thing
like like i think we more than other scientistsshould be aware of sort of like the historical
ramification of bad science right we deal withhistorical objects we deal with historical
attempts of bad conservation like people who likethought they knew how to conserve an artwork and
they really didn't and we constantly have to dealwith that frustrating and more than other
scientists we know that one step towards bad thingis very costly down the road
and like still do this bullshit isn't like yeahyou know yeah so yeah i think that's just like
i'm still interested and i still want to learnmore about machine learning and how like i can
use that to address some problems that we have inour field that is like very specific to our field
i want to understand it enough so that i can useit as a tool rather than like
just copy and pasting an existing function into myecosystem and like let it run and be like
look it's been a cool new thing yeah i don't knowwhat it means but it's like cool you know yeah i
don't know what it means how i got there orwhether it has some sort of relevance to what'sgoing on
but it looks cool right um like i i don't want todo that kind of research but even if that
gets me like publication every year or somethingyou know like i think it's boring i don't i think
it's just like i will i can easily see howirrelevant they become like so quickly you know
and um i would rather take my time um eitherunderstanding what the model is doing or come
up with ways that doesn't rely on machine learningto like you know like we have physics
we have chemistry like that's a thing it's notdead yet you know it's and it doesn't it can't be
it's not like these tools are some sort ofreplacement for you know physics or chemistry
right it's right it relies on this knowledge thatwe have and we are very privileged to have
the education to know so like you know i wouldeven if that takes me like three times longer
to solve this problem in a traditional way quote-unquote then chuck it in a black box
again much rather do that and feel confident in myscience than i think my model did and
i tweaked it enough hard enough so that it matchesthe ground truth or something you know
AIとの向き合い方と研究者の責任
yeah the to me sounds a lot like a a sense ofcomplete dissatisfaction and disinterest
with that type of work recognizing that the valuethat it is to bring if i were to give it
even the most sort of not benefit of the doubt butthe most value i could give these types of
explorations is that they should be you knowcouched in or they should be surrounded by
language and discussion of it being exploratorytesting and sort of just playing with it yes
that these things do not there there are thingschanging and these are not going to be the same
and this is not definitive on this thing rightuntil you as the researcher have done work
to build the fully constructed story between yourdata the analysis and the conclusions exactly
exactly exactly yeah i don't need you just said iteverything perfect i'm glad it's because but
as i said all this stuff is both connected andit's not as if these are unfamiliar discourses
educational papers that were related to ai youknow like ed tech stuff you know especially
in the first few years it was like reading aproduct review for an academic paper right like
a paper where you expect to be thinking andinstead they were like this could have been a
blog post which no offense to the blog post pleaseshare your explanations this is where i tellpeople
to keep an eye on you know people playing with thetools that's where you're going to learn what
people are doing with them the academic literaturethough you think you'd be a little
bit more sort of critical but there is thathunting for maybe it's papers maybe it's just
you know exactly either because not enough peoplein the field understand it or care about it uh
the critical engagement that we often you knowsubject all of our you know reviews and papers
on like it kind of goes out of the window as soonas the word like ai and machine learning
comes they're like oh i don't get it and then likeyou know yeah gloss this over i'm like no no no
like we're gonna have to live with this whether welike it or not so the only way forward as
especially as scientists but to everyone really isto like engage and try and understand it in a
way that you can describe it to yourself andconvince yourself how you want to interact
with it right and i have decided that i don't wantto use machine learning and ai in this
cheap way that i've been describing for the past30 minutes yeah uh if that's if your conclusion
disagrees with me that's fine that's yourconclusion as long as it's your conclusion
and not just like some lazy like oh i wish i couldfinish this analysis by tomorrow kind of thing
yeah not just like i wish it was done you knowi've had we for those that play with it it's it's
tempting right these tools and the way they'remessaged are designed in in a a helpful assistant
framing right and they can do all of this work foryou in there they're just like having a grad
student at your disposal none of which i would sayis true um i i think that the ways in which
you can get super specific about their use casesyou see how people will attempt to replace
those types of steps because they feel rushedbecause they feel yeah uh unable to spend time
on it because they are exhausted or worn out bythe process themselves right um yeah all very
reasonable like human responses to trying to finda way through it but yeah i would i'd prefer you
be thinking about it at least in making yourdecision not just being like well we'll just
have it right like not just be kind of reactionaryabout it and actually like decide what you do with
it and yeah so i think better than two years agoin terms of the adoption of machine learning and
ai discourse i think uh there needs to be a lotmore now that the tools are becoming more and
more available to people now it's important tothink about like do we really need this for it
and do we really like like is this is this goingto be better than if we diligently collected data
for like two years you know and yeah it's justlike more scientists need to educate ourselves
to be like for your specific field how do you wantto interact with it yeah and i'm i'm still
forming you know what it means exactly but i havea strong intuition on you know what i don't want
to do yeah right like there are things that don'twant to be done here um uh which is good the
starting point sometimes is things you don't likethings you don't want to do it's the negations
right that often come first because it's almosteasier to say no well sometimes it's harder to
say no but it's easier to say no not that than itis to identify exactly what you're looking for
right yeah yeah yeah all right uh everyone beintentional think critically you know play around
with stuff that's fair too you gotta learn um butkeep these things in mind uh and maybe yeah don't
get lazy on thinking yeah don't don't get lazydon't get lazy on thinking i know there's all
the you know social media and everything else outthere trying to eat your attention away
um but don't don't do it just don't don't let ithappen you know just like everything else just
don't just don't do it we know that it's hard i'mnot gonna i'm not pretending that it's easy right
in a lot of these cases but it's you know justdon't do it go out there care about some culture
care about some people um do it before the aicompanies destroy all of the old books because
they're turning them into more training data uh sosorry more more things to think on um but yeah
so thank you for sharing that i feel i feel likeokay the there are fields still changing but
everybody's moving at their own sort of pace hereit's very interesting to yeah yeah see how they
but all right all right
that's it for the show today thanks for listeningand find us on x at
egode science that is e-i-g-o-d-e s-c-i-e-n-c-esee you next time
