Player is loading...
Q&A: The Risks Behind Large Language Models (Al Brown, Fujitsu)
Al Brown, Fujitsu
Friday, March 10, 2023 · 2:01 p.m. · 05m 07s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
and then on this and i don't know what's that known huh
00:00:08
i have lived in a few minutes though uh i mean it's you questions or something if somebody has a question like now
00:00:16
yes and you know the number
00:00:21
thanks so great and talk a good question on uh something that should touch to write this
00:00:26
imbalance in terms of the access should infrastructure but maybe in the a. i. context so do
00:00:32
you see the um the projection of this imbalance in terms of the text whole what you
00:00:38
see as the major risks uh involved in that
00:00:42
in dallas uh across different countries uh across different
00:00:46
types of corporations institutions government versus private sector like to hear your thought on
00:00:52
that and if there are risks what would you consider mitigating uh yeah strategies there
00:01:00
and uh so it's a very very complex failed and i think one of the challenges that we have
00:01:06
in this space is that um it's very difficult to
00:01:11
find previous parallels that much perfectly uh on the yeah
00:01:17
interesting contrast for example is the manhattan project say you
00:01:21
have a sensual changing technology the last time that happened
00:01:27
you know within the space of a decade or so you got smart and projects and you
00:01:32
discover that you require huge measles which meant
00:01:35
it was e. o. you will only the richest
00:01:39
but in this case it was the riches states um ah and in this
00:01:45
instance where i think we have the deviation is that's that's a lot inaudible
00:01:49
so the only place well that is the change really different well there is a different dynamic is in china
00:01:57
whether blurring between uh stay academia and industry
00:02:01
is deliberately enforced by state figure out um
00:02:06
but uh the challenge in other places we the
00:02:11
other also to construct a society lost in west
00:02:15
um as that control as well as legal determination that
00:02:20
you were last when one of these elements like here in sideways case where there's a crying out
00:02:26
is that we see ah as i'm getting ourselves that's
00:02:30
even as the basis of compressions and sees um the you
00:02:35
some of us are expecting to okay 'cause of the last language
00:02:38
model allows one of the areas that they are interested in use
00:02:43
um there are existing holes in mind already been
00:02:47
found also if you accidentally crazy your loss language model
00:02:51
and elements of copyright that cover something like three uh
00:02:56
have this dust capabilities for next generation point is you find that
00:03:01
your model is immediately subjects the lord points all cannot be exploded
00:03:06
how many subjects also sitters one season you all get it and closed and one hundred years ago
00:03:13
so there was an interesting tensions that are going to play out
00:03:17
i do not know how that will work and also the issue that we that
00:03:23
it will be how you have that discovery mechanism becomes really difficult
00:03:27
home general i think it's really different from previous model so be
00:03:33
there is a divergence between creation hosting infrastructure
00:03:39
uh understanding what is in the code itself and use
00:03:44
because the thing that is different no chance t. v. t. v. and say going
00:03:48
back to compare it with a a day you're the colossal effort involved immigration you'd reruns
00:03:55
is the the user was a single instrumental e. i mean this case actually uses
00:04:01
the bigger anyone can do right now so i i think there's some really fascinating aries
00:04:06
um one of the things i think that would also be interesting use
00:04:11
chan g. p. c. you may also be one of the great he's data gathering exercises at the same time
00:04:18
because if you want a mobile home he uses engage with loss language models
00:04:23
what's everybody is putting it is family going back to the creators of those
00:04:26
last like one so the behavioural aspects of the international space okay thank you
00:04:31
he usually impose also days uh and say well there is a uniform as
00:04:35
he uses what the most it uses across anyone has access the ins mass
00:04:39
access to the database generated in terms of the behaviour between humans and machines
00:04:44
in terms of addressing a excels responses to questions is still being eminently instinct sorry miss
00:04:50
there's loads a i i'm only scratches the isn't probably giving you very ugly i'm sorry
00:04:57
thanks very much
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
