Player is loading...
The Evolution of Large Language Models that led to ChatGPT (Andre Freitas, Idiap)
Andre Freitas, Idiap Research Institute
Friday, March 10, 2023 · 8:34 a.m. · 09m 52s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:05
at the end of the day so thank you for your presence so there's not much
00:00:09
and whether but a very glad that uh a lot of you would make it to
00:00:14
have this quite informal discussion on chatty peachy which certainly is getting a lot of attention
00:00:21
right so from my present my personal perspective working an l. p. so now everybody that
00:00:26
i know knows what naturally to processing is about so you have finally my parents understand
00:00:33
and there is a lot of of course opinions around and perceptions and
00:00:38
yeah question's lot of questions right that you should be
00:00:42
answered and uh we also have some facts uh that
00:00:46
i think we can state off uh yeah what you
00:00:49
yeah tragically tease about ten where this whole feud is going
00:00:55
so uh i think we could start this event thinking about how we can
00:00:59
factor out some facts right so there is a lot of perception opinion perspectives
00:01:05
but maybe we can converge on some specific facts so first
00:01:09
check to be t. definitely changed the way the general public
00:01:13
receives right uh the state of the art a i mean actually processing so
00:01:18
that was a recent and be get shipment at big never it from their initiative
00:01:24
so tragically g. also is a great demonstrator of the capabilities of
00:01:27
large language models like so i think we can agree with that
00:01:31
right and here we are going to use the more the technical term the product at every largely with the models
00:01:36
when we're talking about things like chatty beatty light but uh yeah
00:01:41
we're going to talk about the different methods and models on the line
00:01:45
uh and then check the p. t. is a product with an outstanding based go right so uh it's impressive that
00:01:51
they manage to reach a millions of users wis few months of their launch
00:01:57
a user's a user base that super letters right that's a big
00:02:00
achievement uh and that uh yeah it's a testimony to to the merritt
00:02:06
off of that specific model right and all of the the methods that uh
00:02:12
you the past week shoe actually deliver a language models with that capability so also
00:02:19
we can agree that there is a lot of hype around thirty p. t.
00:02:23
uh um and they realise also some substation investment so in the orders of beans
00:02:29
already committed to this type of technology right so i think this is just establish a
00:02:34
little bit the the bones of consensus and i think with this fact they're not too much
00:02:41
questionable right but we want to to that reflect a little bit more and make discussions
00:02:47
likely more a nuanced right so i i'm not the other facts that we can state is
00:02:54
lighting the models is one way of doing a language processing right it has many limitations
00:03:02
so when you should be critical and aware of these limitations in some
00:03:06
of these limitations will be addressed specially with this been scale invest and
00:03:11
right uh and they are already were changed to use larger
00:03:14
and more than part of discussion today it will touch this
00:03:18
inserting there are risks in implications uh of
00:03:22
using large language models especially at this early stage
00:03:26
uh in this level of maturity of the technologies so the
00:03:31
spirit and the intent behind this event is us making sense of
00:03:36
these advances so it's uh quite informal town right uh so
00:03:41
it's really with this cues that we have represented in this room
00:03:45
uh it should discuss together it is not the should not be a
00:03:49
a speaker only vital kind of this only perspective so that mean expert here
00:03:54
uh in language model so we like to your uh your expertise in her voice so
00:03:59
to have a kind of collective discussion so that we make sense and they're all or
00:04:05
uh constraints uh off what does this all about and where all
00:04:10
this is going right and we want to qualify some of this previous
00:04:15
facts we want to be a bit more new once double this x. so there's not going
00:04:19
to be complete uh attempting to be somehow balanced but of course uh we don't have all perspectives
00:04:26
represented here right so it's uh quite informal exercise but very much uh where we intend that uh
00:04:33
in on the curiosity side of things so just talking about the structure of the then should they
00:04:40
uh we start with the technical introduction session uh uh
00:04:44
where we're going to talk very messing up higher level terms
00:04:48
right not to protect ago about the principles behind this key technology right they get technology behind j. d. e. t.
00:04:55
uh that led to check to see uh then we're going to talk about an
00:05:01
imperial perspective why some people are going to talk about their experience with large thing
00:05:05
the models uh and the type of behavioural analysis that you can get so just
00:05:09
to get a very practical empirical feeling what you can do what is really far
00:05:15
from being done what are the limitations to get this practical feeling of this
00:05:19
right then we have uh some related work it out uh expressed in the
00:05:24
poster session that will be doing lunch and then we're going to discuss
00:05:30
uh industrial challenges and opportunities in the context of a panel invited panel
00:05:36
many companies uh well represented risks implications separate panel and
00:05:43
there's more those just be very clear light so they are
00:05:46
they're quite expensive tribute there is a a massive heart infrastructure
00:05:50
required should make them work uh and they're going to have that uh perspective represented as well
00:05:56
so that even if you prefer not experience with that you can get this
00:06:00
feeling on what is the type of investment need right needed for for this in
00:06:04
what are the options available in terms of infrastructure so i just like to uh start as well
00:06:12
saint many thanks to all the local organise this so people who really did they have lifting
00:06:17
in managing all the details behind organising an event like this so you steamed once uh too much seems
00:06:25
well uh to the many speakers kind of this
00:06:29
who really made the time some of them travelled overseas
00:06:33
to be here so many things for for your
00:06:37
availability in for your input uh yes and also
00:06:43
participation thank you so we have a i think a good number of posters and make their own pen
00:06:48
so thank you for corporate speaking the discussion as well
00:06:52
uh and i also like to uh think my scientific organisers
00:06:58
oscar this artsy and uh one okay on the bus for for their
00:07:02
uh organisation as well on the scientific a side and you right so thanks for for
00:07:10
can hear things for spending the time having this discussion and you're looking for to to start
00:07:17
a a day off intense interactions so uh before we we we go to the
00:07:24
the core so some small announcements so easy
00:07:28
up has a animal uh international great challenge
00:07:33
that uh it seems to be organising at night
00:07:36
they happen to explore new technologies a. i. these technologies
00:07:41
uh and yeah i participated myself in the previews i think it's
00:07:45
it's the organisational for this type of event is stellar so i really recommend it's impressive how much people can do
00:07:51
in a time constraint and under the right section so i really invite everybody was interested
00:07:57
in experimenting with a i to participate in share this with people that you think
00:08:03
would be of interest and maybe another announcement that we also have a continues uh
00:08:09
master yeah program and will starting a at in partnership with when
00:08:15
the stands are in this particular tailored for industry collaboration so people
00:08:20
uh looking for a experimenting with products in an industrial setting so uh quite convenient to explore
00:08:27
technologies for example that we're talking about today so
00:08:32
final announcement this event is recorded so yeah he's uh
00:08:37
you like in closer to try will be recorded at the aware of this uh so questions are recorded
00:08:43
and is going to be available afterwards in the level platform
00:08:49
so i think that's it uh from as a welcome message
00:08:54
uh i think you now let's move to the first section
00:08:59
and so that's talking about really the under the hood what's behind
00:09:03
uh this technology so i'd like to invite james henderson a
00:09:08
corrugated yep so research group leader from the natural language understanding group
00:09:13
uh uh yeah james is uh i think the person that i know that knows more about
00:09:18
transformers uh at and is most excited about transformers
00:09:22
as we also super enthusiastic about as far as
00:09:26
so yeah his background uh yes is essentially long a set of a
00:09:33
different research positions from them all the jean yves a u. p. as
00:09:38
we all right so we have been really in in different places uh
00:09:42
yes in l. a. d. d. out really focusing on transform these models
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
