Player is loading...
The Risks Behind Large Language Models (Al Brown, Fujitsu)
Al Brown, Fujitsu
Friday, March 10, 2023 · 1:42 p.m. · 18m 35s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
i come back to this i that's that's enough of today sets
00:00:07
in this is that the facts off ah tools insists activity and society
00:00:14
this it isn't the talk uh and it's it's a centre for cognitive an
00:00:20
advanced technologies here because that is going south it's it's you know it's the
00:00:28
insecurity says that that's this and i think the contents to to tell us
00:00:32
that that is uh it's getting this stuff last activity still i one example either
00:00:38
he was this is the the u. k. the minutes is descends on the
00:00:42
study of those cited trends but it's an artificial intelligence and its impact on conflicts
00:00:48
ah the other think that yeah but he did a lot of flexing ah
00:00:51
on these subjects absolutely citizens what's that i don't think it's it's it for example
00:00:58
what is is a very interesting i think for this that bias it read
00:01:01
it is also the former army officer and he commended upon disposal in afghanistan
00:01:08
some interesting have you yeah excel and
00:01:13
that's not an l. and where i was also the vitality of the asbestos in different ways yes
00:01:23
that remote or so no magnificently i'm hopefully you should be able to see uh some slides there but haven't
00:01:30
we require i don't know if you're looking at some images or not may yes it's for him to protect
00:01:43
yeah right so raw on guys talk about oh it's a lit solid here
00:01:52
um so exams risks was deciding what we're is we're gonna talk about um
00:01:59
like um some is most was a little bit say yes or whatever it
00:02:04
is you want to think about and risks can probably does only be disappointing to
00:02:08
risks to life a prophecy unlimited so life
00:02:11
comes to g. standards associated with violence fayette injury
00:02:15
prophecy has to do with the protection of property chronology i don't know unless he is associated with the
00:02:22
changes you'd expect to see to runs society the freedoms you normally enjoy so
00:02:27
seven digits a pacifist our life um so one of the uh really thought through
00:02:35
uh your uh projections the society enjoy home or actually idiocy and slot on okay just
00:02:44
a laziness and people being daft all to the craziest on some heroes in prevention of
00:02:51
threats of violence against ah society um it is one of the issues you will always
00:02:58
have that it was very difficult to do with his brains determined people one schools man
00:03:03
but there was was basically a subset of the population you comfortable it would you know have the way to do
00:03:10
it will most commonly combination of the two it's actually
00:03:13
very difficult to stop somebody is willing to trade paranoia determination
00:03:18
and has the equivalent of a p. h. d. in biochemistry going out doing something truly will start to
00:03:23
proportionally most people are very runs a very determined also usually
00:03:27
wise and sensible enough to know that they're vested interest to
00:03:30
the things that we wanted the world and that which is good um uh how how usually long bus was thinking anyways if
00:03:36
you run retired seventy your city you've been a good lead me some topics are are actually pretty low you know in
00:03:44
the last well we used to have to worry about was in it before the internet era would go no go better knowledge
00:03:50
we're going to get a things like having his book of work and would you put a lot and an older for
00:03:55
yeah and insecurities as it would take don't because you work with the time money and effort to get ah and that
00:04:01
would tell you how to go create problems i see you're
00:04:05
trying to also some other little it's associated with oliver test
00:04:09
oh yes obviously change time essentially load of all so well last
00:04:13
language models along with these change anything is basically the highest ends
00:04:18
all been somewhat uh sorry isn't going to be particularly good increasing use does points uh
00:04:23
what am i saying what is the source of the population has previously not been i'm
00:04:30
intelligent enough to have the background knowledge wall in particular the effort required to get past
00:04:36
a base perry is going into something anymore i guess slightly unusual but the first studies on the effects
00:04:42
of los language models on working behaviour comes really interesting things it takes less time and efforts to do stuff
00:04:49
uh los abilities work is all those who have benefited by
00:04:52
just um hacks me you know bring deleting couch potato potential terrorists
00:04:59
b. is the increased enjoyment of the task as well that makes a difference no
00:05:03
uh whatever work with uh so the last language models uh it's a
00:05:09
bit of developing those because they genuinely already gone down to the fact that
00:05:13
being helpful to people who you do not wish to have is a challenge so you start since we are at the moment is
00:05:20
that if you want to care less willing to buy weapons particular
00:05:23
you wants to join from ground zero i don't really know enough
00:05:28
yeah i think race unaffected by the uh the status we
00:05:31
are right now is that it is no significantly easier at
00:05:36
all then we use the exact same last language models don't
00:05:40
hold an off right right one now they are really really
00:05:44
big threat or problem potentially is that we on my way
00:05:48
may limits we don't know what it's all features skating mine
00:05:52
to to these models of become more helpful as they scaled
00:05:57
and that has happened between every six twenty one or it involves
00:06:00
so this genuinely question we need to look at now to work out how we engage with some of these
00:06:05
organisations to make sure nobody really hopeful known but that
00:06:09
isn't the stalled at the problem so this is the
00:06:12
um uh the internet research association was it's chilly goal which is to pieces but ah which is a troll
00:06:20
fall i was language models already to now is accelerating
00:06:25
capacity of people to engage in the production of trolling material
00:06:30
and you will see from the news or oscillate yours coming engaging
00:06:34
how people can engage a if you know it's just it's a consolation
00:06:38
effectively the turing test has been passed um i saw somebody recently
00:06:43
uh in cajun catfish so tending to be somebody else on the internet
00:06:47
used to take time and effort on i have seen somebody
00:06:50
you as a navy hourly wage tax laws language model information
00:06:55
tends to be attractive women and he has been using a
00:06:57
stroll co workers have work ah on on with especially automated
00:07:03
and it's also when a numbers game because i once people would hold
00:07:08
obviously it's no longer a threat but the issue is if you can do that us cable but it's
00:07:13
not going to point you might uh you can do not what's it ah ah to be honest it's going
00:07:20
to be an issue in some ways that's also how you affect spiegel change behaviour is how you could
00:07:27
four seasons to society changing what people want to engage
00:07:31
on and what drives people to was violence and anger
00:07:36
uh it was gonna change the one of the things is also religion in los angles he's even before you do anything like that
00:07:42
it turns out that they are really good yeah catch a must be i mean we'll send it won't you
00:07:47
oh but haven't have we have the tendency are
00:07:49
previously to drive a vehicle to withstand echo chamber behaviour
00:07:55
which has seen it only be up to the city all um and
00:08:00
so there's a discount basically i'm sorry advanced is an interesting one so
00:08:06
the most obvious when you could think of would be you think would
00:08:09
be using gently peachy already to create
00:08:12
code and post exist rooms and actually
00:08:17
sort of get caught code on it but the more advanced why ah
00:08:20
the listings icing on a suggested to this was really of most used
00:08:25
to the intermediate level coders to record and the coding it wasn't hugely
00:08:29
helpful if you're really really good coding you kind of you to do it
00:08:33
uh and then on posted to cover the code is you work for right now i am tropic
00:08:38
and it turns out that they are very hardly using
00:08:40
or even especially especially systems to accelerate their encoding speak
00:08:45
so only you could guess it's around where the road which saliva
00:08:50
threats can be developed would increase but i think it's more likely
00:08:54
that you'll see things like this um which it's it's older people
00:08:59
creating to push happens well what it really is primaries no advanced techniques
00:09:04
bob morris that social something say that social engineering that you would say say if she attacks
00:09:10
proxies ah well the last one of our dogs to read just to get the frame discussion use
00:09:16
the first element in this is all giggling was the more difficult ones to get my
00:09:20
head around and think out because first of all it is a long that's an issue
00:09:25
and there is also a male lawrence's cool p. i. fantasy a male voices
00:09:30
description of yeah sauces this assumption that is the first technology that will rowing
00:09:34
our society's lines that will model itself to us in well that
00:09:39
that too was rather than all other forms of technology in the past
00:09:43
which have been symbiotic so they have changed society unchanged us the same time as we have
00:09:49
incorporated them into what we want to use um so if you're thinking about rights and freedoms we
00:09:55
technically have them on the rights of freedom you know hey that looks like this they are established point
00:10:00
social argument protest and oppression and and political moves
00:10:06
but there is also correlation between economic value lately but
00:10:10
and what happens in the rights and freedoms we have given pigments society say this is model big thing called and in the
00:10:17
u. k. particularly in terms of stuff project movement and once when
00:10:21
it uh that protested home for a very very long time and
00:10:25
the point which writes the voting for women arrive was noticeable
00:10:29
will always point one minute become immensely economically valuable as part of
00:10:34
like princess hockey and russell one off so the others on here
00:10:39
but i'm i'm a cancer isn't handled wasn't present privileges is hated
00:10:44
and we know that is sort of hostile recall precedence that parents documents to much
00:10:50
of what annoyances society uh all really heavily correlated with things like that like that
00:10:55
so when hasn't numbers increased hugely they suddenly became economically much more valuable
00:11:02
and they change their status and their ability to move around the country
00:11:05
work out that those are examples of progress where is going in one direction
00:11:11
but we also in that in a language phrase in english
00:11:16
uh the us laws on it which we have typically now
00:11:19
associated with and those who were against technology the actual allowance
00:11:24
themselves represent something that was in the other direction which was
00:11:27
included in the meeting industry when the machines turned up
00:11:32
discovers the then wages and then once had diminished usually
00:11:36
is why they engaged in a violent protests that get um and there
00:11:41
will always embrace partial uh if you move to the mall near you
00:11:45
i think of the resource resource curtis ah and only the worst situation
00:11:51
you can be in rice and privilege position with respect to the value later
00:11:56
is if you can be made to date little diamonds handles the
00:11:59
ground and be going when you're doing it but still be economically valuable
00:12:04
you were one of the run very few economic pressures to preserve your most you know most of all based in an era when the common
00:12:11
colossal unsubstantiated full physical possible so you had passed vessel technological trend that
00:12:19
and we are interesting interesting answering this you know where it is intellectual muscle
00:12:24
that is now the feature about lawyers um hum was really good discussions actually could say oh i
00:12:29
um uh one of the professors that they don't really it's true yeah and he said that uh
00:12:35
chan g. p. g. is only the first thing that as a
00:12:38
write in society that feels like we expected i owing to feel
00:12:43
it's probably the first time that people really do you know a isolate the world when trent david what
00:12:48
are you shopping and everything else is on a
00:12:50
lot but it's invisible recesses things where intellectual muscle house
00:12:56
so uh substitutions also connected universal have already have sex
00:13:01
so one of the examples i'll give you his the black advised us is the usual
00:13:06
yeah we learned that cost little being a good citizen of
00:13:10
of any with your it's also to warrant streets to develop but
00:13:14
i wasn't used to this case i mean it's it's not a very good schools
00:13:18
ah intellectual muscle was what gave the economic with only to
00:13:22
clark can't drive us and they were able to how one yes
00:13:26
um associated with that i worked in the strong when the analogy to navigate
00:13:31
the world is here and and i just don't buy any with a fine
00:13:36
you will see in all of the news that is
00:13:38
associated with the legal challenges associated with whether people try those
00:13:43
uh it is uh holidays for example um uh i think i will talk about
00:13:48
the smallest when donald all covered really it's a particularly interesting field in in terms of
00:13:55
a cocky going to say the image on the left there is a lesser when uh it was a test case going through
00:14:01
somebody at all he that bad he's of generative models to create all talk
00:14:06
to them all the machines call you want which us copy mobile disputed which interestingly
00:14:13
does not necessarily ugly matchup with the terms and conditions you only um and the
00:14:20
other relevant what is quite a and i think this has to cost it is
00:14:24
if we're looking at some of those latest this is actually
00:14:26
close well there is a a a science fiction uh my busy
00:14:33
but has a reputation for being particularly good dealing probably unofficially with
00:14:37
all the submissions and how's temporarily closed it's doors to new additions
00:14:43
because of the volume of material that it is getting and it's inability to work
00:14:48
out from that all you was always beyond a constant care it but secondly holes i
00:14:54
gives them nightmares about what kind and how old age because they do not know whether i stay with the spectacle liked
00:15:01
the other no i think that is interesting in this how we might
00:15:04
change society is going back to those ideas all the internet research agency
00:15:11
uh and it's ability to account how many volumes of nonsense into the internet
00:15:16
because it was just amendments opinion um or even know
00:15:20
whether the asians about where china g. p. t. material goes
00:15:24
out on the internet and is recorded as the next corpus
00:15:28
of material is the potential for corruption of the inverse yeah
00:15:32
so the place that we currently mowing for information has the risk of being polluted
00:15:38
with information that may not be validated because it is not embodied in the real world
00:15:45
but is a creation not sense um there the other other than onto quite in terms that that
00:15:51
relationship between industry economic like that and him down
00:15:56
to you and you can also see in this car
00:15:58
which i think is interesting so undergrad formalised uh it's the colour the masses i agree that's coming
00:16:05
through uh which is looking at what is available
00:16:09
to you research organisations like everything that is using now
00:16:13
what is actually the the view of the news
00:16:17
extremely well funded lands you then have the resource
00:16:21
to create a computation and the database that and
00:16:25
volume that is needed to create these last language models
00:16:29
and you'll see that as usual get tools rise of a graph it is predominantly and only would say less than six laps
00:16:36
and well and you can see on the right hand side the
00:16:40
scale of uh i launching on results that are coming from academia
00:16:45
i'm rather than other instance um and i think the other thing that
00:16:49
is interesting and worth thinking about now is we will have a tendency
00:16:53
to think about all of these issues in the what is possible point
00:16:57
now but it is also worth being very aware the scale of progress
00:17:02
um as a speech result all in this um because the brain to change
00:17:09
uh is going to be difficult to assess it could carry overs current trajectory
00:17:14
if somebody's you'll see if they're looking to pass and you should really like to know
00:17:18
what we haven't spent all that is um and if we look at one of the relevant
00:17:25
i would say that we have an interesting being engaged in these uh decisions so so
00:17:30
that will be really boring like copyright but sorting them out in advance will be extremely important
00:17:36
and we're example give us that going back to that concept of substitution of physical muscle
00:17:42
for kinda colossal that wasn't something that just
00:17:45
infected humans are obviously we medication doing manual farming
00:17:50
ah but we also have beasts of burden um the population of
00:17:55
holes is actually teaching about on t. i. t. nineteen nineteen and diminish
00:18:00
but if you look at what happens to them only murals is life
00:18:04
is much better now but well i think this is genuine questions about
00:18:10
well we're gonna be the holes on the left rather wholesome around
00:18:13
how we engage with the risks challenges freedoms that we already enjoying and
00:18:18
unnecessary things with that that should i see those those things with a
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.
667 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
