Player is loading...
Round Table: Risks & Broader Societal Impact (Legal, Educational and Labor)
Lonneke van der Plas, Idiap Research Institute
Friday, March 10, 2023 · 2:07 p.m. · 57m 08s
Panel: Al Brown, Fujitsu’s Centre for Cognitive and Advanced Technologies --- Ana Andrijevic, Centre for Law, Technology and Society --- Biljana Petreska von Ritter-Zahony, University of Education Vaud --- Bertrand Audrin, EHL Hospitality Business School .......................... Erratum from Ana Andrijevic: "After verification, Open AI's terms of use provide that as a user, "You are responsible for Content, including for ensuring that it does not violate any applicable law or these Terms." (point 3(a), https://openai.com/policies/terms-of-use)"
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
if so i'm very happy to have people here from education from labour
00:00:05
from the little hand from secure i think that i think that
00:00:08
thankful all have their own perspectives on activity and and the final
00:00:12
impact on their particular field so i'd like to ask for if
00:00:16
everybody to uh introduce themselves so uh i think that the microphone
00:00:20
ah yeah so why don't you start to have it like a man
00:00:27
oh now how to find the correct is it ah okay just have fun with or
00:00:36
uh huh if it if it going on and off is that the one that's going on now
00:00:41
oh okay so i'm really happy just gotten re there and an associate
00:00:47
professor at this his university uh for education at the count on the whole
00:00:53
um i i teach teachers how to teach computer science
00:00:57
basically so i'm i'm chained as a computer scientist at the u.
00:01:01
p. s. no and i did research in competition there silence before
00:01:06
then moving onto a focusing on how to teach yeah basically
00:01:11
because i think there is a need to educate everyone but yeah
00:01:17
yep yep that's about them that well i don't know who looping through a former uh i'm uh
00:01:25
assistant professor your you'd show hospital to business school women so i'm
00:01:30
uh i have been a p. h. d. management
00:01:33
focusing on a judgement transformation and its impact on
00:01:38
human resources and human resource management and so well
00:01:42
my interests here today is also to understand how come will e. i. impact
00:01:48
in the workforce and to what skills do people need to deal with the yeah
00:01:55
think about right cell uh i maybe can have ah uh that without
00:02:01
maybe one thank you very much for having me today so that i mean did in canada are very happy to join today
00:02:07
um so i and then the h. m. o. wrists um oh
00:02:12
researcher currently a p. h. d. candidate at the university of geneva switzerland
00:02:17
at the faculty of law i work on the impact
00:02:20
of a um copyright law firms with european us perspective
00:02:24
i'm also a visiting researcher at the university of what our the centre
00:02:28
a lot technology and society hence my presence in canada and finally um unaffiliated
00:02:35
affiliated researcher at the digital law centre at the university into you know
00:02:41
yeah i thought that i thought i had i you have forgotten that you also hated yeah yeah for a couple of months right where we
00:02:47
had a show in any of the interactions regarding law and they i
00:02:52
will learn a lot from you i feel like a lot from life
00:02:55
but an okay oh yeah you already i i pray that you but maybe
00:03:00
there at a certain point that i forgot to mention yes some points that
00:03:05
no no no no no and i don't i fell i am i like this but
00:03:10
i put a question africa gap actually because that yes when i first heard about objectivity
00:03:16
most of the concerns that either but maybe that is also because we link to education anyway writing it yeah
00:03:22
was all about education how i mean students might be a user interactivity how can we evaluate them
00:03:28
can you tell us a little bit about you know what i did fears of
00:03:31
the the colleagues around you what do they say well what i didn't worry about
00:03:35
so you're absolutely right in the um it has cost quite a bit of us tour um it's
00:03:42
no we're going with this um and there was a there is no no wave of panic for sure
00:03:49
so why is that uh the big problem is a certification
00:03:54
and certification in um basically um
00:03:59
the central question is how can we make sure that um what i'm evaluating like how the violet
00:04:05
over there has been done the student that was made with the help of checked u. p. t.
00:04:11
to what extent has ted typically have him and what annihilating is is it the student work
00:04:18
or is it his prompt engineering skits what is it that right and evaluating in basically um
00:04:25
what happened is that a chat typically has passed several entry uh they like than
00:04:31
uh so there was a medical school for little school and then in da subject tragically the cotton to
00:04:38
universities we creates c. plus to be minus
00:04:43
so it's not like your best shouldn't but it's it's a pretty good student basically when he gets into university
00:04:50
and what happened next is that uh uh i think it was a public
00:04:55
school that was the first one to ban no twenty p. t. use it's cool
00:05:00
and then a slow sport followed in paris that was gonna be big news around here because it
00:05:05
was closer to us um that personally i don't think that's the way you can go about it because
00:05:12
it's difficult to then something like ted typically you know it'll become integrating in in a lot of different use you
00:05:19
using it implicitly we don't even know when using j. d. b. this it's really difficult to bend something like that
00:05:25
so the problem is is work that is done outside
00:05:30
of the class might lose its value everything that is
00:05:35
done at home or on the tail you it's it's difficult to know the one way about it is to
00:05:42
you know too well examinations more often but every at whether the energy and
00:05:47
time saving capacity of ted typically would be on the side of the student
00:05:53
put a teacher it's a nightmare it's much more work and much more time to check whether but this
00:06:00
didn't gets is the result of his own contents it
00:06:05
so what's happening is that the code oh the ontology
00:06:09
are going to change in a lot of schools like how do you regulate the
00:06:13
use of different a. i. tools in school so we need to think about that
00:06:17
no no uh i only know of institutions they're working on the subject but
00:06:23
at least rounded but no one has done the first right it's more like
00:06:27
ten teachers have thirty nine hide you can do this or that but there's nothing
00:06:31
institutional here industry the i think on that but not wrong here and um so so basically yeah that's
00:06:38
that's the big problem is like what are we going to be evaluating when we are uh uh students
00:06:45
another difficulty i would say is that i'm critical thinking is not something that is easy to teach
00:06:52
so when i tried to be these hallucinating know how how do we know how
00:06:58
do we know that the students are weird that i typically really thinking unit at
00:07:03
as authors that you need to be it's a it's a challenging task you need to be an expert to know that had to be easy listening so
00:07:10
i'm in the midst um i don't remember who mentioned really you you have mostly
00:07:15
menu in intermediate coder so for me teaching computer science if i want my students to
00:07:22
use tied to it you know how that it only helps you if you're proficient enough
00:07:27
if you use it right from the start you know you you you don't write anything
00:07:32
and basically you can just cross your fingers that it's not the end
00:07:36
and that is very difficult and finally i have a lot of mine collected they're really really free
00:07:42
uh about the practical hands on the interesting motivations of our students
00:07:48
so if you know after all this hard work you relied that you're
00:07:52
not sure isn't much slower than you know machine like what is infected on
00:07:57
on on the that's hard work and especially like why would i i just
00:08:02
didn't know what to learn how to write better we all this effort required
00:08:07
i i had such a super profusion costs right for you know what
00:08:10
i what is not nation behind yeah so no they're kind of the challenges
00:08:15
yeah yeah i i've i think than uh i i i've wondered about the the and
00:08:20
the penultimate one document if nothing about and critical thinking and figuring out whether they're having stations
00:08:27
if there was something i mean i can't think there that maybe you could turn it into a kind of
00:08:31
a you know a tool that could help you to think things sales opportunities objectivity we've talked so much about this
00:08:37
it is nice to see after the bright side of things they had some things i say well
00:08:42
the bright side is a talk to babysit tool and is
00:08:44
it tonight use it creatively and supposedly can be very very helpful
00:08:49
and poured from the teachers than they can come up with this
00:08:52
additional yet oh you can distribute your students in groups according to level
00:08:57
uh you can as as market mentioned earlier you can apply to descend
00:09:01
a systematic evaluations using a i have a lot of
00:09:05
a. i. x. factor pushing that way it's their vision uh_huh
00:09:08
and you can also you can you know making projects
00:09:12
for you that's under demonstrate it so skillfully today and basically
00:09:16
i know just type the p. t. fourth so that's such amazing and i tell that
00:09:21
are just impressing you know if i want illustrate something for my course like and just like
00:09:27
do a little no prompt and then have super illustration that an artistic
00:09:32
with taking like do your to to make and it would cost me
00:09:35
when more and um it can help me we formally text
00:09:39
they're either to their more proficient at then tend to get that and
00:09:43
where you can hardly translate official documents and things like that or for your
00:09:48
for language uh learning it has for for the students you know any they
00:09:53
can they can use it to get some simplified explanation that more complicated uh
00:09:58
uh concepts because you can say you know explain it to me is
00:10:01
a fun a ten year old reason it's me fine affecting your role
00:10:06
and then you can they can use it is an write a book they'll said like you if you get
00:10:11
in a lot of documents and they just want to just thought of it they could use it for that
00:10:15
oh they can protect languages or they can uh and uh take a
00:10:20
uh not to get responses uh for their homework are working on it
00:10:25
but they want to do this we might need to educate them and how to use it responsibly and you know what the limits are and
00:10:33
what does it mean and at what point you need to speech oh we need to have the discipline it's really difficult at that age
00:10:40
to say okay now i need to be learning so i can use it and i don't want
00:10:44
and then at one point and it's you know i'm not learning i'm just checking that there there
00:10:49
that i can use it so that you know they didn't sure enough to be able to do that
00:10:54
right that's very interesting l. activity later more about how we can educate people about check completely
00:10:59
ah anything coming right but i i thought i thought the yeah the back so i got yeah
00:11:05
we were talking earlier about chechen picking at the tool and i'm sure that you as a as a specialist later it
00:11:11
would also know uh like in human resource management what kind and how confused activity what what are your thoughts on that
00:11:18
yeah so so the first thing but by a real or discussions that we're
00:11:23
hoping to do is that we are really the current similar issues or oh
00:11:29
i see some gum or a critical thinking is the one i am also like how can use should e.
00:11:37
t. o. m. one paying him all which are you
00:11:41
or other things a menu offers nah say well maybe
00:11:46
we're getting the wrong about and uh actually a recurring of song maybe
00:11:52
one of them is really really cool products as well you know we're picking up the wrong fact we
00:11:59
are all here and not something that's come about really thinking about how can automate brings with it yeah
00:12:06
and it's actually really cool yeah i mean i read speech is likely
00:12:10
to meet my emails between emails like everyone so why not only picked
00:12:15
but what what could be thinking or is more
00:12:19
how can we admit it rather automation worked works
00:12:25
oh man preach and here that brings another chops yes i i is useful
00:12:32
but maybe you shouldn't think about a i asked something that replaces what shouldn't do
00:12:39
or what humans to but rather us copilot to humans and
00:12:45
how humans do better basically yes outsource tasks not only eagles
00:12:52
oh we find or him but our our trusty human tasks that we're not really good at it
00:12:58
and i think that's a a kind of shift of my uh what the what the what is called
00:13:03
the true trap let's not try to make yeah it's something that is just like us but let's try
00:13:10
to make a i something that we can not be and this is what really good bit shifting mindset
00:13:18
i mean that's i think something that we all have to force
00:13:23
okay if we want to help you but yeah it is useful
00:13:28
to me you have to be something else just copies us ah
00:13:33
that's a very inspiring yeah i know i but i get me thinking
00:13:38
because i shouldn't we then also changes skills completely i mean ah
00:13:43
you know isn't there how do you think it's happening how difficult
00:13:46
will that be thinks everything's moving so quickly do we have scale
00:13:49
people to be able to use a a i insertion of melted away
00:13:54
yeah that's actually that's the other fascinating part because that we
00:13:58
have to like the other ways on our overall or self right
00:14:03
if you think of yeah it not as replacing us but that's
00:14:07
one thing us then a. i. becomes a copy that we met
00:14:12
and we have to learn how to work with these women how do
00:14:15
we work with when i'm that's well we don't know if we can trust
00:14:20
so yes here are critical thinking skills but all probably work with
00:14:25
women but uh we don't know which the space bar so the
00:14:30
early church and one of us feel that they think is uh
00:14:35
well not whether schools are working one of the keys computational think
00:14:41
uh i i also write lot students uh i i
00:14:45
see how the use chuck e. l. e. a lot there
00:14:52
different qualities of answers and often becomes part
00:14:57
mm from its source to really good from the others are terrible problem we just don't know how to asks more questions
00:15:06
and i think they were really bad already visible in they are
00:15:10
not getting any better with the judge and that's the issue of how
00:15:15
we make the swing i'm useful for us and we can not
00:15:20
just expect all uh tell me how to break up with my girlfriend
00:15:25
and the judge you know we have to ask chuck e.
00:15:31
oh well formulated questions that come up with a really complex and
00:15:37
uh another points will ask you to uh but uh you mentioned uh
00:15:43
kind of reflex kind of it goes well yeah that was saying about these uh it
00:15:48
would be to to pull the plug that it's his talents that we have to fine
00:15:54
well we live on judge p. t. a. one not too i i i
00:15:59
mean i'm also became of that might sell it the other day i what cool
00:16:04
and i was like okay i'm going to ask judge p. c. but otherwise actually the pro version but i don't even know how to
00:16:14
you're not but uh it should uh and well okay well
00:16:19
get your head straight you how are you are in charge of this now ask judge
00:16:25
it to be you when uh and okay well here is what i know how to do
00:16:29
how can i put it right there that so we have to have back to uh the the final question that i need to ask
00:16:36
about like yeah how are we going to do that but we'll come to that really and i want to give the floor not to anna
00:16:41
add because we all know what we've learned today if that uh
00:16:45
check typically if they're very hungry beneath it needs a lot of data
00:16:49
and also that compute power had to get a as good as it gets and i take the elevator from the rabbit but it
00:16:56
also uses a the prompt slide with and about that so uh
00:16:59
everything you type in will be come also part of a new
00:17:03
objectivity as so i'm not from the legal perspective what kind
00:17:07
of problems the if they what if it what if eating so
00:17:12
from a cultural perspective which is the main i will talk
00:17:14
about now and maybe i will mention uh_huh well personnel that have
00:17:19
and just right after suffers from a cultural perspective you have different questions that arise
00:17:24
uh at different stages so first to do in the way you collects the data
00:17:28
in the way you copy the data that's by far one of
00:17:31
the most important one and finally the way you modify the data
00:17:35
so if we focus for instance in the way we we copy the data we know
00:17:39
that in order to train the mother wasn't point you will have to do with our
00:17:43
and this trigger is on the copyright law what we call the right of reproduction this says uh something that
00:17:48
is part of the oldest conventions and wow the copyright
00:17:52
law in spite of all copyright laws across the glow
00:17:55
and basically what it says is that you cannot copy of work that's protected under copyright
00:18:00
law unless you have the arteries authorisation over the rights holder is a very station to yeah
00:18:07
it was actually raised a very recently in three complains so in
00:18:10
general in february of this year we had three complaints against ability i
00:18:15
and all the companies that use a stable diffusion so it's more about
00:18:20
images but you'll see that it's very much less because it's also about
00:18:24
copying data and how then it will generate knew it
00:18:28
i knew in outputs and the different rights that will arise
00:18:33
uh from this process so in those cases so for instance
00:18:37
in the first case that appeared in generate a twenty twenty three
00:18:41
uh we had three artist in the united states who
00:18:44
sued a stability a i as well as me jenny and
00:18:48
uh they can quite cool the three and the three of them you stable diffusion
00:18:53
and the reason was that they had access to the database and maybe that
00:18:57
wasn't a very smart move from a stability yeah i actually to make it accessible
00:19:02
so the huge database to use with billions of data five point eights billions of data
00:19:08
um does accessible and and uh they had even created a
00:19:12
kind of search engine to try to look into a database
00:19:17
and the artist found out that the work was actually in the database and that's
00:19:22
how stability i ended up being sued as well as again enjoying it and are
00:19:27
um so again this huge for many reasons it's interesting because not only on the copyright law this
00:19:32
would as well and the unfair competition law because those tools and making money out of the work
00:19:38
uh we also have the problem that uses i using these tools to create fakes
00:19:43
so the cat the type the name of the artist and for instance you will generate a new
00:19:48
um i work in a style well this person and we can all just up
00:19:52
to make it created our little chat typically will be the same problem with text
00:19:57
it's just that in in those cases it concerned right away a images because again we have access to the database
00:20:03
uh it's not the case with company i as we know it's not the case with other tools that don't uh no
00:20:10
tables their model or a database so that was a that was interesting and in
00:20:16
two other cases getty images actually sued a stability i both in the u.
00:20:21
k. because the seeds of the company is in the u. k. stability i
00:20:25
and the in the in the united states as well so it's still ongoing but
00:20:28
it's interesting because it raises last question normally under copyright law but again unfair competition law
00:20:34
also personality right because when you use unlawfully the name
00:20:38
of an artist and you make money out of it
00:20:41
um without asking permission because in no cases that tried to ah have licenses with the
00:20:47
art is uh try to conclude a an agreement with the artist to use that work
00:20:52
i've crease issues so um it's very interesting and it's all happening right now so we'll
00:20:57
have very interesting things to to talk about as well for tragically t. will see but
00:21:02
again uh it depends on whether or not you have access to the database which
00:21:05
as far as i understand is not the case would tend to be so far
00:21:09
okay thanks an asset then about the projects right so uh yeah we've thing some thoughts
00:21:14
on and where people with high fat like we can use these tools to create things that
00:21:19
we can then you how was it for that like can i create a that a
00:21:23
knife a tax that and then pretend that i wrote it uh how did with the other
00:21:28
type of a i i'm with a with a creating for example artwork
00:21:33
and uh what i did the what the law fey in that case what can you do whatever you
00:21:37
want with and can you even pretend that you where the one creating it and felt for not mine
00:21:42
yeah so actually what you will have to look at is the terms of use and the constant policy of
00:21:47
open the eyes so the the both clients you all
00:21:50
the tools which include a deli yeah uh and uh objectivity
00:21:55
so if we look at the license the river but those provided
00:21:58
by opening yeah and says that you can do pretty much everything you
00:22:01
want a whether you have a page or not free membership up with
00:22:06
the websites but then you can do whatever you want with the output
00:22:09
however uh they say that the keeps no only the text from study also keep
00:22:14
oh the inputs that you can uh uh asked if you um if you use an image i think with
00:22:20
opening i with only actually you can yeah you can use an image of his own and then it will just
00:22:26
uh bring different variations and you can do different things with the
00:22:28
tool was only about just putting the text description i will create something
00:22:33
uh so if you have something on your own and you have rights over this text for instance
00:22:38
um when you can changes with a um tragically t. you will have the possibility
00:22:44
to um well bunny i will keep its you want people only the the tax
00:22:49
problems but also everything that's bringing on the table that you want to change through
00:22:53
that to basically and the matter again whether it is charging beating your belly et cetera
00:22:58
um then but they are the friends so you
00:23:01
can commercially seeking use it for noncommercial and commercial purpose
00:23:05
once again see can do it's true that we can do pretty much everything that you want with the text
00:23:10
however they say in their um in there yeah a consent policy
00:23:16
uh which they didn't update for chad to p. d. so the only thing i
00:23:20
found was on the website is in there as a uh i think you a page
00:23:26
is the one for opening yeah i but i'm screaming and the same applies to judge me chuck
00:23:31
d. v. d. what they say is that you cannot pretend that the text was created by you
00:23:37
uh it was created by charging p. t. so being carried shoe which doesn't say
00:23:42
that's the doesn't say much then you have to interpret what it means to tinker issue
00:23:47
uh to practically disclosed involvement of the i. model when you share
00:23:52
the the work created by a chat to it he went down here
00:23:56
and uh and they can also remove so in the
00:24:00
case of ballet for instance they can also remove the signature
00:24:04
and uh so on any work that you generate with that's who's you can remove the signature
00:24:08
o. t. i. model if you wanted you cannot mislead um and uh
00:24:13
the other people about the nature of the work so you can know
00:24:16
tell that you was entirely created by human being you have
00:24:20
to be strict worried about um the fact that was created with
00:24:24
um the eye to the only the last example level mentioned about that is a very recent
00:24:29
keys 'em discussed by the copyright office and the united states so it's not a tribunal per se
00:24:35
but is renders different decisions about works that are we just stayed
00:24:39
other copyright office it's a possibility in the united states that we don't have for instance in switzerland
00:24:44
but it's mostly a useful when you know that their chances that you will end up in court
00:24:50
so what they did is um recently on my case about
00:24:55
upcoming book that was created by a young one man who used
00:24:58
me jenny they decided that all the images created by with journey
00:25:03
we're not protected under copyright law which wasn't quite big sisters on
00:25:07
and but the all the texts that were created by her uh by the artist were protected so
00:25:13
you in the same work you have things that are protected the not detected because she used me jamie
00:25:19
and the the this ends up the the result
00:25:22
is basically that or what we understand what we
00:25:25
can infer from this decision is that uh all works that are created by a i is not protected
00:25:31
um at least these images from the journey another protected under copyright law which
00:25:36
means that even the license that is proposed by major any would be boring simply
00:25:42
because if you want to put the license on something it means that
00:25:45
is protected otherwise you cannot get a license on something that is not protected
00:25:50
but in order to really know whether it's protected or not you actually
00:25:53
have to be in front of travelling tribunals and i will tell you whether
00:25:56
or not like that yeah so there it is and what implications then for
00:26:00
this one can pushing i make less money with their comic books or how
00:26:05
either or it's a good question uh she can actually it was again she
00:26:11
can make money out of it yeah i think it's it's not a problem since
00:26:15
she used so if if you say that the work is not protected
00:26:18
on the cover of light falls into what we call the public domain
00:26:21
and the public domain is basically as it happens the same way for
00:26:25
oh works that protect the same switzerland for instance works a protected on there were
00:26:30
seventy is after the death of the author and then they fall into the public domain
00:26:35
uh_huh workload it's fifty years after the the death of the up in switzerland
00:26:40
i'm in those cases what will happen is that as soon as they are in
00:26:45
the public domain you can do whatever you want with them so if you have
00:26:47
the copyright office telling you that it's already in the public domain because it wasn't
00:26:51
created by human being that was the justification but it wasn't created by a machine
00:26:56
then the result would be that yeah you can do whatever you want with it in switzerland i don't have
00:27:00
to guess what is asked for exporting you would try
00:27:03
to identify who the author who the human arthur is
00:27:08
um which would include people who works uh for me journey and house creating mean journey
00:27:14
so maybe uh people like you who uh maybe would be a a designated as one of the authors
00:27:21
of you know the output created by uh the
00:27:24
eye so that's an ongoing discussion well very interesting also
00:27:28
into thinking stated with a different between countries i can see that companies might choose to go to a
00:27:33
company where the the law is actually more uh you know helping them to do what they wanna do right
00:27:39
yes that that but since it's and an international uh you know problem
00:27:43
and uh yeah you yeah exactly okay thanks a rifle out there are
00:27:51
not so ah and i have a question of software alan and they'll so i i think the question about mitigating that with with a ready
00:27:57
both by that was also one thing i wanted to ask you maybe there are additional things you want to mention it
00:28:02
but the other thing which i found very interesting in your talk with the menu showed us uh about uh you know that
00:28:08
the fact that in galveston models make lots amounts of data i
00:28:12
need also big computing power to be able to a generative models
00:28:16
yeah and so yeah i i as i've only a handful
00:28:19
of fall flat or in situations that can actually build these models
00:28:24
and i was wondering about yeah we think in the past that they were sometimes monopolies at being created
00:28:30
in how funny thingy here and monopoly and what do you think will be the
00:28:34
influence of based on on on power and and economics in a in a larger sense
00:28:41
our first our oh so be it
00:28:45
their distaste to degree so you know one of the approaches you can take it
00:28:50
i'm so sorry robert resulting as well so every foreigner
00:28:55
problems it leafy with these organisations is actually beneficial from security
00:29:00
o. b. q.s if they all very large mail very returns to use
00:29:05
we want to remain legal ah i may want to remind all i'm i'm if there are relatively
00:29:11
few of them you lose the resignation you have to have with them if you work of mental age
00:29:18
so you let's say we've got seven and well now i think excluding china
00:29:24
um i just don't use no actually no british school full uh
00:29:30
okay they're only relatively rich countries you go somewhere with each other external
00:29:35
it's not impossible to use condoms to get
00:29:39
all the representatives from as organisations interesting where is
00:29:43
i made them stop sharing good prior to use an isn't interested in um collectively if you use
00:29:51
if you stole impeding will be more could do so if you're
00:29:55
trying to make you deliberately on helpful use some house there is
00:29:59
there is a risk that it becomes less compressed it compares to which appears
00:30:04
and if everybody is playing with the same set of rules you will not get as much bridge
00:30:08
if you turn around so messy um you think the u. k. race
00:30:12
what are you willing to the mind just have a last time probably two different because
00:30:17
the group little money to do it on levels focused on a cool people to do it
00:30:22
are on duty rhymes in london one set of rules really
00:30:26
probably move if i remember right to market is going to do
00:30:30
but they would then release the fact that i have
00:30:33
is inconsistency really are now i am topic all of them
00:30:38
that point of integration you want has probably based around reinforced learning with u. v.
00:30:45
so you a big successive commute remote countries reduced whatever happens to be
00:30:52
one of the fundamental change use has been usability to you discussed last
00:30:57
get reinforced learning which models is behaviour whole
00:31:00
informations to talk to people or creates home
00:31:05
he creates very uh_huh functional system joints responses
00:31:12
but our user norwich resettled inputs and you'd come fifties to very few so you cam
00:31:18
you could sit machine there on wednesday right you wanna integrate pointers where it's will be
00:31:23
and you could sit there with a security experts you know was
00:31:27
also where it's a wrist rests on the whiteboard contracts and you just
00:31:32
also usually very strong will once it to g. can be easy to
00:31:36
do um there there's a risk with our you know some of the things
00:31:42
that you would know what was just to to to use you would
00:31:45
will you giving our reinforcement learning you feedback and you would walls my machine
00:31:51
most controls less colourful curing cancer or alzheimer's will will use them so
00:31:56
it's you would have to work for it yes you are you do huh
00:32:01
uh but i was only viable if you were dealing with a small number of well funded institutions you just
00:32:06
okay um so what say stephanie no one use i don't like it is interesting because i think
00:32:13
it's a challenge like an easy because i do think that he's necessarily i'm hoping to probably the most security pointed
00:32:20
and safety only with those machines in terms of the standard chart is the rest of them have not necessarily
00:32:28
thought first and foremost about possible to the nation and collaboration one piece of possible um
00:32:36
yeah a manual questionable is i think is really interesting while um so
00:32:42
you are i feel i might have much more to offer makes it somewhat
00:32:46
'cause there's such an association between preservation of interaction
00:32:50
proximal rooms here the use case we really really
00:32:55
cool um but i think you'll be genuinely really
00:32:59
useful and interesting to be just all speaking into
00:33:03
uh what's the use receiver so i don't think we're gonna be positions one standing there
00:33:08
why do you genuinely is gonna be an interesting
00:33:11
dynamic between the phone also really lost language generative models
00:33:16
arms to take shouldn't know scraping and doings uh_huh on compressed
00:33:20
advantage that gives them full of each time so there is
00:33:23
an argument to say your people's positions defensible because of the
00:33:28
taste is has accrued rather than sitting around reading but it happens
00:33:34
and if our model is true for a loss language models
00:33:37
will generative says you don't foundational e. going into the future
00:33:41
does that mean you know you're gonna have a small says extremely well funded organisations
00:33:46
sorceress when from residues you know when a channel one isn't staying where is to
00:33:53
when did you start it was nothing prearranged who was going to be the search engine
00:33:58
but it's not really hard for us to catch who just artist yeah yeah definitely yes i i was just wondering
00:34:06
about monopoly that we've seen in the past and have these were dealt with and if they can be
00:34:11
can we put them on the fine with like what we anything now happening and can we say something about
00:34:16
the future and maybe also i'm might have some i don't know if you have something to say on that
00:34:21
or an hour yet as well as to where oh he's which is interesting study um
00:34:27
so one of the effects of the nineteen eighties nineties here memo labels to change ones
00:34:34
actually decreases the standard will choose to draw it'll sounds
00:34:39
awesome presence foreign cars mark s. cummings couples espouses it with
00:34:45
really rich i regret mortimer standard was cool i'm doing is pretty no one has little use rule in the nineties
00:34:52
and it is not one of the things that change was how not lose this point as with which is all about
00:34:58
i didn't lose if you joined consumer prices which is the just a
00:35:01
whole lot okay whereas having him or is he has to be some areas
00:35:07
where is the consumer processed a lot because conceded not we all homo uniforms all
00:35:13
but i still don't also while ours might be good for the rose last no uh_huh yeah it's going to
00:35:19
be really interesting because you couldn't the role or a
00:35:23
affecting produces more when you were facts on its good place
00:35:28
but they don't work on short come there will be some
00:35:32
yeah right and i do have something to fail if if not really
00:35:37
your done a yeah it's the last made the main pressure uh
00:35:42
yeah i'm not i'm not sure that anything yeah the words that uh that i i i
00:35:47
wouldn't know exactly there's so many things to say about that we don't know exactly where to go
00:35:52
and the and uh yeah it's a a lover or not yeah that anything that my head
00:35:59
and so i think now is maybe try that i'd like to ask the audience if you have any
00:36:04
questions uh otherwise i have a few more but uh and it's much more it's much nicer you had some
00:36:10
uh for our panel any of that the people that actually i
00:36:13
i there and back the ah microphone uh mm that's right there
00:36:21
fans are quite alright to remind you mentioned um if i remember correctly raw yeah
00:36:29
uh open uh i'll quote from the pork bun yasser or what
00:36:34
uh_huh never provision for right everyone tutors e.
00:36:42
why yeah that's responsible for put all four of the reserve trial ever
00:36:54
ah response whatever your answers comes reporting question
00:37:00
the first chance to read what's uh from what
00:37:05
come to my mind why you've or whatever preferred responsible for our souls or something
00:37:14
but i think one that that that one i think
00:37:18
then a new editor for that no i think that's the the phrase this you know what
00:37:27
yeah so so i i i couldn't hear completely the question got right
00:37:31
sorry about the games that's why i was about the terms of use
00:37:34
so just summarise what i was saying again and i hope we will help answer at least part of the question so
00:37:40
what the terms of use of any i say is basically as a user you on
00:37:45
the you can do whatever you want from a couple commercial or noncommercial perspective with the output
00:37:50
and in exchange um what well what company i will do
00:37:54
is that they will keep no only the not only the output
00:37:57
but also the inputs a text problems plus anything else that you can bring to the table such as an image text et cetera
00:38:04
uh and that will help them to justify it by the fact that they will actually use it to keep training the models basically
00:38:10
um so that's uh that's that's maybe answer the question the one
00:38:15
thing i think i think if they can humidity or to come
00:38:20
so what we write terms this yeah it's responsible for
00:38:26
all put forward for your answer all put if you um
00:38:34
for for more thought well you know for expelled ah i'm all for what we want
00:38:41
mm okay so uh no i i would think that you would be responsible
00:38:46
for what a d. v. d. creates however they tell you and a constant
00:38:50
taking it now they come to the policy this say um form
00:38:55
or whatever from for remember reserve but i'll i'll put surprise huh
00:39:03
to respond to whatever to to learn in this way that kind of making sure that they can be a a you know
00:39:09
responsible for it for example creating harmful content or or by
00:39:13
if count that or that it's all user that in that case
00:39:17
so they say in the content policy that you shouldn't attempts to create upload a
00:39:21
reading from from that's to create a plough share images text as well in this case
00:39:26
um that could cool cause harm including hate harassment violent self
00:39:31
harm sexual shocking illegal activity deception legal public in recent null while
00:39:37
and spans so you have a whole list of things that
00:39:40
you you cannot you and the data you cannot do independent policy
00:39:45
so if you still do it's you have a little about violating basically their father
00:39:50
policy so no i don't think that it would be responsible for it for the
00:39:54
and also they have to be a big challenge so that that's the case all
00:39:57
these models that have a beach onto the scrapping of data that they have in
00:40:01
uh collecting is that the now and you can find it in
00:40:04
their research papers uh for instance full of any i look at the
00:40:08
the research people put the only um and you look at the conclusions they talk about the fact that they have a hard time
00:40:14
dealing with images that uh have
00:40:17
violent contents work pornographic contents because
00:40:21
i do when described it basically describes pretty much everything they could
00:40:24
so now are they doing the job to try to filter it the whole thing
00:40:29
so that's why they're selling you know so to the to the
00:40:32
solution for them rather than finalising maybe to work before making it public
00:40:37
was rather to say to create this constant policy and to say do not
00:40:40
do that but at the same time it's part of the database now so yeah
00:40:46
interesting chart about why if the question uh i gotta get about if if that or what
00:41:01
a reflection yeah the user is definitely responsible to the inputs
00:41:05
oh that's for sure that's what is provided um by the total policy yeah so
00:41:11
maybe we can come back to that uh later on that if something else on output yeah that would be interesting if
00:41:17
we can find that also have a look i haven't looked at the data that is actually ah that's another question yeah
00:41:26
and though i know in no because i'm to me in no who've
00:41:34
while i'm forever and i think that all the major problem is that
00:41:40
with the picture i thought that might actually it event and unknown in the third
00:41:49
uh no i don't want on yet
00:41:56
i don't know i think it's not that bad even if you
00:42:00
put them in the e. i. e. i. what kind of hell
00:42:08
right the panel peddled by the nine one one four
00:42:12
can let that becomes teacher but he sure definitely no
00:42:17
problem can turn a little with the lapel that line and i know they have that yet but i think
00:42:26
kathy than that if the if the fifty you think it necessary
00:42:33
i i i i i thought because i was yesterday and i'm
00:42:40
a high school where a point to find the the teachers to yeah
00:42:45
you yeah i so basically you're working night pedagogical they hand me um
00:42:51
or a heated thinking is here right in front of like all the teachers and then
00:42:55
they all met together randomly and discussed it and what implications are in how they need to
00:43:01
change their habits and so it is happening so um it's it's in in
00:43:06
the high schools i and wants to know what's happening in in high school
00:43:10
so uh in nine my school uh another c. s. teacher
00:43:14
did the same thing but it's like the me and major
00:43:18
uh does the plane with the representatives and then you can kind of hear directly so
00:43:23
there is some sensibility station that is what i'm up but there's also the top
00:43:29
bounces of the stations so basically the contender who we received the um like paul
00:43:35
what they're asking us whether we know about it and didn't even want to be
00:43:40
uh talk about it et cetera and um in the good tone i'm sure they're they're also working i thing
00:43:47
at at the political level of this nonsense of the decision
00:43:50
can thing so that teachers are weary here is so they
00:43:54
they're no it's because basically i heard is that is on the teachers
00:43:58
like uh are neither how good their students really that uh the yeah
00:44:04
yeah i mean there there is a lot of effort in a lot of movement but about and also thought that it is that it's a little
00:44:12
i i can only talk from e. sure spill general exams action
00:44:18
and what i know is that but for the school e. or
00:44:22
re you by the students that e. exams should be your own creation so uh you like
00:44:32
you are generally it's actually if you used to e. it's not huge
00:44:37
that's because i thought that was the year firefighting and i'm like oh was that michael use it
00:44:46
also try to uh get head off is strong and uh we tried to do some
00:44:54
well how to use one just that's something that no don't have guidelines everyone just
00:45:02
trying as report that i think
00:45:08
any more questions from the audience
00:45:13
mm question shall we can talk a little bit about at educating people on chapter
00:45:20
eighteen press yeah uh i guess yeah the yeah that's a question for you again
00:45:26
so no it it's a very difficult question i mean yeah if you think that
00:45:29
uh yourself if you if you look around and how many people are computer scientist
00:45:34
and how many of them are teachers and how many of them know about
00:45:38
a i and then leaves us with with the human resource problem basically that well
00:45:43
if you can help but i don't know so we don't have a lot of uh many qualified people
00:45:48
you can think about competition thinking it's not a scale this year it's just you know being over night it's it's
00:45:54
more like a professional information in computer science it's noisy getting
00:45:58
after years and years of practitioner like why don't think like
00:46:01
well that and whatever but basically i it's difficult because um
00:46:06
it's it's a long process it's a it's a costly process it takes a lot
00:46:10
of people if you know when the content of who we had this it didn't know
00:46:14
training and it's has involved a lot of people and a lot of money and
00:46:18
it's it's not easy and it's it's efficient but like it it's a difficult thing so
00:46:23
i don't know what we're doing in our lab is trying to do with like pedagogical the theories
00:46:29
gains in like making it real accessible that helps to a certain extent but then like we need to
00:46:35
it took a lot of people in the thick difficult like it even
00:46:40
i around you like not many people know works how it
00:46:44
works and it it's going so fast and can affect and so
00:46:47
strongly so quickly that it's going to be needed but it
00:46:51
the answer is is it's it's challenging huh yeah i can
00:46:56
see that would be good if i mean i i had a year starts most recently we sit in your lap for students
00:47:05
and of h. r. last year right okay but you should work for
00:47:11
us to study uh if you guys going to figure where our jobs
00:47:15
and uh i mean it's uh you know it's like students on a technically there's supposed to get
00:47:21
a job after school right and so the thing was uh i was trying to explain to them that
00:47:29
that was now of what is now we have to learn the skills
00:47:33
to work with these women and uh that's actually all have some computer provisional
00:47:40
picking classes but you also have to work a lot lot trial and error
00:47:45
and in its class we also over okay well you can use to keep them show what you do with it
00:47:51
but not only provide a judge would you answer but how can you improve
00:47:56
oh it's to show all if you actually have some i mean if you create an
00:48:01
added value because it yeah i'm mostly about but we split something oh street after yeas
00:48:08
pretty useless for me and for all will be for most people in the future
00:48:14
uh_huh yeah yeah so uh talking about you your fame like yeah
00:48:19
will they just that but i just think that job right thirty have something more to what i am factor
00:48:24
then actually to uh to that particular question i think it can really take away thoughts like that actually so
00:48:31
that's our question and uh actually really appreciated remote
00:48:36
amongst our little ducks a talking yeah when when do
00:48:40
swing much machine for them to people read them so all everybody's going to get out of a job then
00:48:47
it's been the same with every single technologically will illusion students
00:48:52
you mention of it will seem a bit multiple evolution and moving is like
00:48:58
people are more acts of free m. what many people say is
00:49:02
firstly uh we tend to overestimate what technology can actually do the work
00:49:09
it's not so even we shouldn't do so innovations like a touch your else uh
00:49:17
we tend to overestimate what we can actually the second thing is off from uh well
00:49:24
a i and who uh tools can the over tasks so parts but not fully
00:49:32
jobs uh and there is often that part uh whenever that still needs to e.
00:49:39
to some extent you and so this is like okay maybe it will scale down we will be more
00:49:44
efficient and a lot more things still some that's
00:49:49
what came to these the not be fully ability to
00:49:53
and to work to other ideas to which are used a
00:49:58
condom students uh one other one idea is that's a first uh
00:50:05
e. the next time change m. m. technology has been wrong it's
00:50:11
it for example right now is the sole scanning cashiers that uh
00:50:16
a couple people or whatever it's not because the
00:50:20
technology is there that fully replaced everyone doing his job
00:50:25
i fixed i and evolution is not from one to the next and final thing
00:50:31
also wonders new technology are linked in so that's more work that should look
00:50:38
for but not their current jobs me won't exist more jobs might be pretty
00:50:44
so yeah that's really funny or a little bit of reassurance that that
00:50:50
yeah thanks a lot uh i guess it everywhere already we sure because they're
00:50:54
on the other side right that's we don't have to worry about the fact that
00:50:58
i i get letters i fashion right i'm i'm looking at a time with i
00:51:02
will if you if somebody had a question are you have one yeah okay good
00:51:08
yeah and i think you just a final question for all uh uh so if you think about uh
00:51:18
the population distribution right and the you put complications so
00:51:21
inverted pyramid it's uh so juicy you uh technology use
00:51:28
yeah technologist changing the vulnerability of some countries in terms of demographics product
00:51:35
productivity so how do you see that on the right kind of geopolitical perspective
00:51:40
because i think it connects with that the fans narrative maybe the the product you which we used
00:51:45
so maybe you wrote this some thought on this type of risks and integration in terms of productivity
00:51:54
i saw the more security questions horses for one hour
00:52:03
uh huh
00:52:07
so i think it's gonna be very interesting in a sense for
00:52:10
how it's very difficult to music you all country without being rich country
00:52:16
um they're all different models how you approach a defence security
00:52:25
alright depending on size and scale uh so you saying
00:52:28
whole uh has the cited it calls itself the police role
00:52:35
i would use that is very small it could be called upon alternator is uh probably
00:52:40
you you try to make it so unappealing as a prospect would be just one do it
00:52:44
um we have a general other countries interesting how interesting
00:52:52
demographic challenges which problems or china is joining process and security
00:52:57
use that china has to get rich state before it gets old
00:53:01
so a joint already as a result of the one child policy recipes you you you know who's there
00:53:07
it doesn't have a white population discriminate side which is we're roughly oh gee
00:53:13
i don't so you will have a really old population problems always really possible indian
00:53:20
because has a completely different demographic read in terms of working age population has interest
00:53:26
they're really interesting counterpoint something better player spin
00:53:32
uh you can't respect to discussion of water solution will be replaced as well so
00:53:37
i'm sure most people in the room a little bit more about or power switches
00:53:41
i don't want him solar is not necessarily what machines are going to be a good asked will find easy
00:53:47
voices so for example you might well be even being engine base which is currently high status for a job
00:53:55
uh it is going to be on the vastly malls read last language models and being a girl
00:54:01
use a i'm previously society we would probably us or you would expect ritualistic a mall
00:54:07
we're going to just like you do cause he's amigo personal so much but there
00:54:12
is something easy reading digits to listen i got doubles because it's basically laurie wells
00:54:19
uh it's in bowls comments manuel hammering a teacher
00:54:23
one aldridge there's almost no engineering for which our schools
00:54:29
i'm one of those minute differences cure designs on her life between high on
00:54:34
vitamins is going to be interesting because there is this is some should but
00:54:38
future warfare my one concern right when are awesome discussions very most remote numbers will run
00:54:45
but genuinely one of the last means that will be amenable human
00:54:50
volumes towards men should use cases we even momentarily this mounted infantry course
00:54:56
use the the same things that point to go in and protect our position in waffles
00:55:02
pretty much point in the same way so do you think you'll want to see um
00:55:09
and again we can compare point newsletter status or directors with these uh
00:55:14
for some points of muslim only to be replaced so you
00:55:18
don't you will normalise mounted infantry person walking around the table
00:55:22
i will receive a series of you will never be a higher
00:55:26
talk for competition from what i'll say right her industries one by one
00:55:32
you use whatever stipulations i uh but we do the
00:55:36
best followed us for supposedly recess hosting guess people shops
00:55:40
also okay sorry i system one point no you know
00:55:43
it's almost well or binaries because you have a code affordable
00:55:48
are you mormon you which you have most production
00:55:52
hundreds issued construed cynicism arms you could basically do
00:55:55
the computer many years experience or just running time first of all i would you call do person
00:56:02
so i think most where did we what were those also mean excessive
00:56:06
operations is if you have a very large population you rent house home
00:56:12
you will have still access to those elements in volumes which is going to be people under the table ram
00:56:17
do things uniforms table people which you still going to be really fair so you i don't think we'll see
00:56:24
something like simple which is a tiny country angels can
00:56:28
page because you could use is also a recipe rich
00:56:32
or just pouring huge will that will be nice now gotten better all he wrestles was a problem
00:56:41
you will still have use license last pretty for countries um
00:56:47
i think you would be interested in terms of all the changes that we will see um what else you know
00:56:54
uh or something else or commencement probably wrestle f. f. will be good okay i think
Conference Program
The Evolution of Large Language Models that led to ChatGPT (Andre Freitas, Idiap)
Andre Freitas, Idiap Research Institute
March 10, 2023 · 8:34 a.m.
664 views
Understanding Transformers
James Henderson, Idiap Research Institute
March 10, 2023 · 8:46 a.m.
369 views
Inference using Large Language Models (Andre Freitas, Idiap)
Andre Freitas, Idiap Research Institute
March 10, 2023 · 9:19 a.m.
ChatGPT for Digital Marketing
Floris Keijser, N98 Digital Marketing
March 10, 2023 · 9:58 a.m.
Biomedical Inference & Large Language Models
Oskar Wysocki, University of Manchester
March 10, 2023 · 10:19 a.m.
Abstract Reasoning
Marco Valentino, Idiap Research Institute
March 10, 2023 · 10:38 a.m.
120 views
