• 2 days ago

Category

🗞
News
Transcript
00:00Good evening everyone and welcome to the latest virtual event from The Independent.
00:03Thank you for joining us as we reflect on the significant AI developments from the last couple
00:07of years and how they're shaping the future. It's been two years or so since ChatGPT first
00:11launched which sparked a worldwide debate over the potential and opportunities of artificial
00:16intelligence and over the next hour we'll discuss what tangible progress we've made since that
00:19moment and examine how AI is already changing our daily lives. My name is Andrew Griffin,
00:24I'm the technology editor here and I'm joined by an expert panel today including my colleague and
00:29deputy tech editor Anthony Cuthbertson, Dr Catherine Breslin who's also with us today.
00:34She's a machine learning scientist and best known for her deep expertise in AI and voice
00:38recognition technologies and our third panelist is Neil Maiden, Professor of Creativity at Bayes
00:43Business School and Director of the Institute for Creativity and AI at City St George's.
00:47There he works closely with funding bodies, businesses and other types of organizations
00:50to understand the new challenges posed by AI technologies and offer novel solutions to improve
00:55creative work. As you know these events are totally free to attend but you can donate to
00:59the independent using the link in the chat box. Donating allows us to keep our journalism truly
01:03independent and helps us to organize more important and insightful virtual events like this.
01:08Many readers submitted their questions ahead of time and I'm pleased to say that we've incorporated
01:12as many as we can so let's get started. To set the scene let's start with a basic but I think
01:16important question. Chris Basson asks is AI actually intelligent in any meaningful way?
01:22Isn't what's called AI simply machine learning by another name? I think that drills into something
01:26we probably want to think about quite a lot tonight which is what we actually mean when
01:30we use the word AI, whether it's different, usefully different from machine learning
01:36and what we're talking about when we use those words and those phrases. So
01:40does anyone have any particularly strong feelings about what AI means at a simple level?
01:45Well firstly machine learning is a sort of branch of AI. All machine learning is AI but not all AI
01:51is machine learning and in terms of intelligence we'll get onto that a lot deeper I think but
01:59things like chat GPT aren't really intelligent, they're prediction models and let alone getting
02:05onto subjects like sentience and things like that it's not what we would consider human intelligence.
02:12Yeah exactly I think machine learning is a part of AI but I think of AI as something of an idea
02:18that sort of encapsulates where we think we can take technology that might mimic some of human
02:24intelligence in the world and how we can go about building that and historically it's included a lot
02:29of different technologies not just machine learning. Machine learning is this set of algorithms, set of
02:36ways of using computers that take lots of data, ingest lots of data and learn some of the patterns
02:41in it and then can repeat some of those patterns back to you and a lot of the progress that we've
02:46seen particularly over the past decade has really been in the field of machine learning and so when
02:50people are talking about the progress in AI right now a lot of the time what they're thinking about
02:56is machine learning technology but like you said AI is a broader term and I think it's not as a
03:01stricter definition as something like machine learning where we have a very good definition
03:05of what that is. Yeah I think actually we've got another question here from Kenneth Toomes
03:11that gets to some of that separation between where we're going and maybe where we are.
03:17He asks why are so many people talking about AI yet doing so little with it in real terms
03:21and I wonder sort of with that in mind if you could share any examples of either in your own
03:25daily lives or the people you're working with how what we might call AI as discussed what we might
03:32not want to call AI is being used by people in interesting, helpful, transformative ways or
03:39indeed not if you think that's not happened yet. I think it's beginning to happen it may be hard
03:43for people who are not in particular professions to see the change they're still using chat gpt and
03:48so on for sort of relatively trivial tasks but my team are engaged in a lot of research looking at
03:54how particular gen ai is being used in different professions so in the design industry the tools
03:59are being used increasingly extensively there is first of all the resistance for many designers
04:05to understand that the tool could replicate at least some of the outcomes that they do
04:09there's a degree of denial about the competence of the tool but now this needs to be a growing
04:15acceptance what this means is it's fundamentally changing the nature of professions
04:20you see this in marketing you see it's in branding many many of the disciplines where
04:24there's content creation the the uptake of the tools is widespread the consequences for what
04:30it means to be a professional in this day and age are quite profound we're only beginning just to
04:34understand those changes so i think there is use happening and not only is it superficial it's
04:40quite profound in the nature of work we have a question actually here uh from laran joana
04:45hopefully pronouncing it right um how does ai affect journalism so a version of that question
04:51to you antony as someone who both writes about ai and is kind of both studying it and also
04:58maybe under threat from it i write a lot more about ai now than i did 10 years ago i've been
05:03a tech journalist that whole time but 10 years ago i'd have written one article a week about ai
05:08now it's about one a day almost um but in terms of the tools i use it has changed a lot um every
05:16journalist i know i think uses something called otter ai which is voice transcription or a version
05:20of that there are others available but it means that really laborious task of writing out an
05:25interview or um someone's speech goes from hours to its instant and then you can search it for key
05:32words and um yeah that that's made a real big difference as well as um chat gpt i find and
05:40other generative chatbots are becoming more integrated into just my daily online routine
05:49in the same way that i'd use google to search for information about a subject i'd now use an
05:55ai chatbot to sort of see what other ideas they might have about it um so yeah i feel it's seeping
06:02in more and more and um the i can't remember which reader said that they don't see it anywhere
06:08but i feel like it's sort of in areas that people might not even notice um like talking to a they
06:18might be talking to a person a company you know in a little chat box and it's actually an ai
06:24chatbot they're talking to um or their algorithms on social media a lot of that now is being sort
06:30of done through ai to get their exact uh personality down to a t so they know what
06:36what to feed them um yeah those are the uh the ways that it's been coming in i think
06:41yeah it's interesting uh anthony obviously mentioned otto speech recognition tool katherine
06:46you've you've done a lot of work on on that technology i think that's probably as we're
06:50thinking about the integration of ai and daily life that's probably a lot of the ways that people
06:55whether they know it or not are interacting with something like ai at least in alexa or
07:00wherever else could you talk a little about how you think that has been integrated into people's
07:05daily lives and whether it's been a success uh and how you think that's gone yeah absolutely
07:10and i think you've given a great example of how voice technology has really changed over the past
07:14decade or so from this technology that really wasn't all that good and it worked well for just
07:20a handful of people on a if you were talking about very common vocabulary terms in your speech
07:26and to now we've got these systems that are much more capable and we've integrated them into these
07:30tools that you can just use in that on a day-to-day basis so i think we've seen a real shift over the
07:35past past decade and if you go back sort of 10 15 20 years voice technology it wasn't hugely
07:43accurate and it took a lot of effort to build a voice recognition system that worked in a
07:48particular language for example you want something to work in english or you want it to work in
07:52spanish or you want it to work in german you had to put a lot of effort into building those separate
07:56speech recognition models and so over the past years what we've seen is technology called neural
08:02networks which is the foundation of chat gpd and some of the other technology you'll have heard
08:08about they've impacted all of the field of ai i think and neural networks have proven to be
08:15much better models than some of the ones we have especially when they're combined with
08:19huge amounts of data which we can collect now because we have the internet we have much better
08:24data storage so we've been able to combine lots of data these powerful models and make voice systems
08:30that are much more accurate which perform much better across you know multiple languages which
08:36is something we didn't have in the past and then we started to integrate them into tools like otter
08:40or some of the other note taker and transcription apps that exist out there or in another way maybe
08:46sort of subtitling on youtube things like that where you can automate the captions and you can
08:52search within the audio of your videos now and you can search a podcast or you can find the right bit
08:59in that podcast based on on the transcription all these things are sort of much easier so
09:03and live translations i'm saying live translation again that's been around for for quite a while i
09:09think we've seen early systems but we're starting to see again more and more capable systems that
09:14can translate between multiple languages in real time so you can actually yeah have a conversation
09:20with other people so i think there's there's two things maybe that the fundamental technology has
09:24really improved underneath the hood and also it's got easier to integrate them into the sort of
09:28products that you or i can use on a daily basis yeah yeah and it's it is interesting so many of
09:35the technologies you mentioned that obviously predate that great moment of awareness with
09:39chachi bt and everything a couple of years or so ago we're obviously having this conversation in
09:43that context but you know as you sort of all gesture that it does feel like a lot of what
09:49we're calling ai is technologies that you you may well have been using for you know years maybe
09:55approaching decades uh but the big move that obviously this happened over the last couple
10:01of years and i think got people excited and terrified and whatever else is is is the move
10:07into generative ai neil you work obviously very closely with creative industries and thinking
10:12about how these technologies impact uh creative people and creative work and all those things
10:18could you talk a little um i mean with kind of all that context in mind uh the the previous
10:24technologies but also this explosion of generative ai could you talk a little about you know your
10:29work and your research but also how you think um creative creative life is changing um given
10:36these new technologies as well absolutely yeah i mean it was interesting we built a startup a few
10:41years ago pre gpt3 arriving to help journalists be more creative with ai find new angles be find
10:47more diverse voices and so on and unfortunately some of that our ideas got swept away by the easy
10:52access of gpt and the other models yeah to me and our team are fascinated about how to make people
10:58more creative every day it's not about sort of big ideas that most people associate with
11:03creativity if you ask people in the public it's about this everyday creativity little directions
11:07little changes little new angles new ideas and that sort of culminates up to something more
11:13substantial in my work we've tended to find probably more resistance within the creative
11:18industries where professionals such as yourself identify as being creative like the designers i
11:24talked about before but one of our missions is to reframe a lot of work that happens outside of the
11:29creative industries as creative because there's problem solving every day needing to come up with
11:34new and useful solutions so we've explored tools pre the gpt explosion in social care how do you
11:42care more creatively for older people with alzheimer's health and safety manufacturing
11:46and attractive production line that doesn't sound very creative but there's a lot of creative
11:50potential there business modeling for smaller medium-sized businesses we're currently looking
11:55at how lawyers can be more creative in their problem solving policy development or innovation
12:00innovative policy development and so by reframing all of this work as creativity it enables us to
12:06bring in new approaches to to ai and again going back i'm part of a conference a group that's been
12:13around for 30 years researching this idea of digital creativity tools clearly the emergence of accessible
12:19large anguish models and gpts has given the technology a boost but there's still underlying
12:24principles there around human-centered ai out of the tool augments human creativity rather than
12:31seek to automate the humans out that's what i think where most of the action is ultimately creative
12:37thinking is a skill in most domains it's a skill that people don't have the knowledge or the
12:40resources the time for example with journalists i was talking to journalists i try eight stories
12:45a day as an academic that's a terrifying problem i can tell you and uh you know they just didn't
12:50have the time and the resources the tool has to be there to recommend to nudge to to guide and
12:56you can take that principle into other other problem domains and that's really where i think
13:01a lot of interesting activity will be yeah is there anything kind of in that work could you
13:06share maybe a more concrete example of how some of those transformations might have looked and
13:12whatever the problem um certainly i mean we're working in particular at the moment with a lot
13:17of businesses so uh i think there's an interesting concept that when when certainly small smaller
13:23organizations are looking to establish their business model or business models there's a
13:28tendency to reach off the shelf what is the business model i have and of course that is
13:32copying probably your competitors so tools that are provoking businesses to think more creatively
13:38about their business models is something we're seeing traction in um likewise in a lot of
13:42consulting work there's opportunities for ai technologies that have co-creative design and
13:49algorithms built into them to help the the innovation specialists search very large
13:54amounts of information in order to pinpoint the potential um creative opportunities for invention
14:02and this is good in terms of technology it has to go beyond machine learning because we've you have
14:09to understand there's a lot of codified creativity knowledge and creativity theories and often talk
14:14about creativity is trying to shine a light on those spaces in between what we know now so it
14:18may not actually be codified and it can't even be combined through traditional uh large language
14:23models there's something else that needs to find the gaps and so it's combining all of these
14:27technologies together into into genuinely creative search tools yeah yeah i've got a live
14:34question which gestures to some of the stuff you you've talked about there it's important i think
14:38to put this down to me because of something you mentioned uh lucy van der steeg uh says do you
14:43always trust the ai answer or do you are you fact checking too no i definitely don't always trust it
14:48there's um uh obviously ai hallucinations has become a big thing um where it just felt gibberish
14:56um and other people i know in completely different jobs to mine who use chat gpt for similar things
15:03they say they feel it's almost like they have a personal assistant but they've become like a
15:07quality control manager for them and they have to you know if there is something they're not sure
15:14about they have to then fact check it they use by the usual methods so no definitely at this stage
15:19we cannot trust it yeah and i suppose there's a there's a there's a broader question about
15:24productivity and creativity is in there too which is that it does feel like at least at the
15:30moment we're at this such a point of explosion but an early point of explosion that actually
15:35they may all these systems might end up just making people do different kinds of work rather
15:40than less of it didn't that's fair i mean do you think that's where we are yeah there's that old
15:47saying well not really old saying but people say ai won't replace your job but someone who knows
15:52how to use ai will um so as neil was saying there's so many more opportunities now with ai
16:01as this tool you can use um to create things or to like come up with new ideas or ways of creating
16:08things uh creating podcasts on the fly or like little short films or so many things you can do
16:15with it and it's evolving so quickly that we're not even sure what i'm not really sure what my job
16:22will look like in a year's time in terms of my day-to-day how i'll be um interacting with these
16:27tools i think what i see is most of the use cases at the moment are around efficiency and productivity
16:33that's where most people are trying to drill down the cost but there's only so far you can go with
16:37this so we then need to start exploring what are the the real value-added use cases for ai
16:42technologies and that's where my interest in creativity i think has a role to play but there
16:46are other types of use case i remember when we were working with lots of news organizations
16:51trying to get our tools in six seven years ago the journalists were really interested in the
16:54creativity angle the news business owners were interested in the operational and the efficiency
17:00side and there's a quite a core tension there that actually impeded our ability to run out
17:03the tools in newsrooms with the support for management yeah yeah i mean generally that must
17:08be it's interesting we had a question about what jobs and industries are safe from automation that
17:12there does seem to be this at least among kind of the the people lower down the chain this this
17:18great fear that it's it's it's you know as this question kind of the the context of the question
17:24makes clear it's it's about safety as if you're hiding from a threat rather than anything uh
17:29positive and about potential do you think that's still as i say we're kind of two years into this
17:34into this great explosion that through those two years we've had moments of great excitement and
17:41then moments of great terror both of which i would say are almost at the most extreme possible they
17:46can be you know we've heard that it's gonna agi is about to hear it be here and we're going to be
17:50ruled by ai gods or everything you know ai powered weapons are going to kill us all so
17:58um on that kind of terror to to glee spectrum where where do you think we are and where do
18:05you think the kind of general consensus has has found itself well in terms of uh what you first
18:13touched on about on about um which jobs might be completely automated there's things like taxi
18:18drivers or long distance lorry drivers you can imagine in the not too distant future it could
18:23be replaced by a self-driving car for example there's others that you think are fairly untouchable
18:27at the moment unless you get very advanced humanoid robots that were nowhere near
18:32like trades people plumbers electricians that's going to be very very difficult uh there's others
18:37that are a bit more um like people want a human connection like a priest or a spiritual leader
18:44people don't necessarily there was an ai priest that they've been trialing ai jesus you know ai
18:48jesus yeah not even a priest in switzerland they trialed it in a church for a few months
18:54um and some people were sort of you know open to it but others thought it was you know a horrible
19:00thing um and there's other little things um like a nick bostrom writes in his latest book about how
19:07um no matter how good ai gets a drawing would never be able to replace a child's doing a
19:12drawing for their parents or something like that that kind of emotive um like efforts could never
19:20be um automated um so in that sense i think what it means to be human for them for now is very safe
19:26it's more that we're going to incorporate it but there is a big risk massive job displacement if
19:31it's not done in the right way yeah and i think that the thing is we can't really predict necessarily
19:36what fields ai is really going to be good at because we've talked about llms and chat gpt but
19:41another thing that's happened about two years ago as well are these image generation models and video
19:46generation models and go back further than that and people thought that sort of art and creativity
19:53was safe in a way from from computer intervention and now we have these quite capable models that
20:00can generate images that work really well in a lot of different situations and people wouldn't
20:05have seen that coming necessarily so i think we shouldn't really predict too and at the same time
20:11i guess people were talking about self-driving cars sort of being finished by 2020 think back
20:18to that time so the progress doesn't necessarily follow the lines that we believe it will because
20:23we we're still figuring out what these models can do and where they will have an impact and so i
20:28think we there's a lot of also um careers where like you said earlier i think the computer will
20:33change the job that you're doing rather than replace it entirely so there was a nice example i
20:38was reading in in the field of material science where you have people whose job it is to sort of
20:44come up with new ideas for materials and a lot of their job is quite creative to come up with and
20:49suggest these new materials but then there are ai models now that are getting better and better at
20:55suggesting new materials and so those inventors and scientists jobs then change from being the
21:00ones coming up with the materials to being the ones to vet them and then decide to do the quality
21:05control rather than that that initial stage so it can change some jobs as well as we're not
21:11necessarily talking about sort of automating a lot of these things away but but changing what
21:15people are doing in their jobs there's actually one more job category that um i just thought of
21:21that is seems to be fully immune which is sports or kind of entertainment um like uh computers got
21:28better at chess um back in 94 or whenever deep blue beat kasparov and yet no one's watching
21:35computers play against computers yet the sport of chess is as popular as ever um and there's i was
21:43in japan a few weeks ago watching an ai car race against a human xf1 driver uh the ai car
21:50unfortunately crashed on the way to the start line but it's it's part of a autonomous racing
21:55league that they're trying to start um and there's a feeling that this might not be people like the
22:03human element of these um these sports and like deep mind trying to make humanoid robots that can
22:09play football people still want the drama off the field and on the field um so i think that and maybe
22:15acting and other creative industries will be immune from from the incoming ai wave yeah because
22:22it feels like i mean especially in the context of the image generation stuff that you mentioned
22:26katherine we hear so much about how great the outputs are but it doesn't feel like
22:33as you say it's often that human connection it's the child drawing the picture of the parent
22:37or indeed you know the other end of the spectrum the greatest works of painting aren't necessarily
22:43about how well you've realized something that's come before it's it's quite specifically about
22:49a creative act and i feel like a lot of this conversation is kind of skirting around what
22:54exactly that means you know i expect you have a very good answer to it given your job title but um
23:00do we think that we we spent an awful lot of time uh both talking about and building
23:06um these systems and also we've spent an awful lot of uh energy um both kind of human and um
23:12the other kind uh which is something we should come on to the kind of environmental cost of all
23:16of this building these systems for creativity that as antony kind of gestures that maybe
23:22that's exactly what we don't want because these systems are doing something that we actually quite
23:27like about being alive um and and there are other things that we could more happily automate a way
23:32that we're maybe not putting as much effort into i think with with art part of the value that we
23:37attribute to is the the process that the artist goes through and you see this going back to our
23:42studies with designers they they they can see that the mid-journey is producing an outcome that they
23:48which is very similar or even superior to what they produce but they they know it's not going
23:52through the same process so they're going through a learning process they're going through an
23:56emotional journey that creates an attachment one of the most fascinating views around creativity
24:01is this idea of an extension of ourselves so often really brainstorming group will hold on
24:06to the value of our own ideas well beyond the time that they've been shown to be invalid because
24:11they're an extension of ourselves if we decouple the self from the idea then there's something
24:17changing in the nature of the output i watched a program around michelangelo and his production of
24:22the uh the pieta the other day and they only really only really value it it's a wonderful
24:25piece of work but you only really value it when you understand where he was and what he was doing
24:29and what he was trying to do at that point in life um in the production of it so that's where the
24:34value what is part of the value comes from machines can't contribute that in the same way
24:40there isn't a human story there i think that's important i neil i'd be interested to hear your
24:45thoughts actually on art specifically because impressionism you could say was a reaction to
24:51technology cameras came along and rather than doing hyper realistic portraits they monet and
24:57everybody else started making how their sort of feelings and that a whole new movement was birthed
25:02pretty much from the arrival of cameras and photography do you think are you seeing it that
25:08artists are now sort of coming up with new ways do you think there will be a similar kind of reaction
25:15i think so i mean just in case the impression is they were also enabled by new technologies such
25:19as the um the oil paints that they could take outside for the first time so strangely it was a
25:23technology driven movement uh not just a reaction to photography yeah i i think i think a society
25:30we're going to see a return to human made you see this already people going back to digital cameras
25:3620 years ago you see this with the young generation rushing towards analog um and i i think it all
25:44comes down to how we attribute value as a society and individuals and people are what we always be
25:50searching for something new and the technology might be leading us into some kind of cul-de-sac
25:55of how we can attribute new types of value and interest as something so i suspect we will be
26:02seeing reactions in the same way that you could argue the impressionist was a reaction i don't
26:06know i think prediction in this game is extremely difficult and that's not my specialty i was at
26:12lunch actually with the um of all people the creative director of beaver town here at the
26:16brewery yesterday um and he is obviously if you know the brewery it's it has a real look to it
26:22it's kind of comic book inspired and he was saying he's not worried at all because nothing much
26:28looked like their stuff before you can you can now ask mid-journey or whatever to create a
26:34beaver town style beer can but when he drew it you could never have done that and that he and
26:40he was almost suggesting it's kind of all of the stuff is almost definitionally boring because it's
26:46it's an amalgamation of everything else yeah there's homogeneity now that's emerging in
26:50certain design sectors and people will seek the human input to reinsert that that novelty i suppose
26:57that is arising yeah from technology over technology use there's a couple of questions
27:01that kind of that that sense of how you manage that change shone through i think in a lot of
27:06the questions here that this this feeling that people are finding themselves i think increasingly
27:11set against this technology that i feel is being imposed on them from outside and it's like
27:17we're having to deal with i i think i mean as a as a as both a news consumer and a news producer
27:23i feel like most of the headlines i see around ai most of the discussion i see around ai
27:28is people saying i never really wanted this in the first place and the sense of imposition do you have
27:34any feelings about how how we should manage that change how we can better i mean there's one example
27:40here moshe friedman asks given that ai is a new technology what are the greatest challenges in
27:44improving the way that humans interact with it do you think how do you think we should better be
27:49making friends with these machines and making maybe people feel like they get on with the
27:54better and they're not kind of enemies that are coming to change their lives in bad ways
27:59well one an interesting thing that happened actually this week with chat gpt was that
28:04people began to notice that it can't say certain names you ask it to say the name david mayer
28:09and it'll you say can you write the name david mayer and i'll say yes i can write david and then
28:14it crashes and says there's been an error and no matter how you do it you say can you um say the
28:19name david may david's layer but with the surname rhyming with them beginning with rhyming with
28:26layer beginning with them and it will try and then crash and there's five or six names people
28:31discovered and they think it's due to sort of gdpr requests people have made to be sort of
28:36deleted from these systems but these small things like erode trust in in the platforms but also in
28:43the companies that are running them so i think there needs to be a lot more and i contacted
28:47openai several times saying what's going on why is this name not allowed uh why are you removing it
28:53it's since been reinstated but there's about five more that haven't but um they didn't reply to me
28:58and if they're not replying to me then they're they're not replying to the general public i
29:02can't imagine they're not very active on on the social media or anything so um yeah they need to
29:07be held accountable i think a lot more and whether that's through regulation or just
29:12um people speaking up um that'll be one of the big challenges i think yeah emily twist actually
29:18asks here i think this is something else that we can um maybe stick the boot into opening up this
29:24pr machine slightly because she asked how do you navigate the differences between expectations of
29:28what ai can do versus the realities of what it can currently achieve and i know katherine like
29:33when we had a previous version of this this conversation it was kind of a year year and a
29:37half ago i was really amid the chat gpt excitement kind of festival which was very much um helped if
29:46not led by open ai saying you know in x amount of months we're going to be achieving this in
29:53in years and you better regulate us because it's so terrifyingly good and all those sorts of things
29:58what's your sense of both you know the two parts of this question now how do you navigate the
30:01difference between expectations of what i can do versus the realities of what it can achieve
30:06do you think those two things are closer together than they were maybe when we last spoke or further
30:11apart or i think maybe one of the problems is that this technology ai is sort of shrouded in
30:18some sort of mystery and hype and it does a lot of people good to keep that image going that this is
30:24really difficult to comprehend technology that you have to be really expert to understand and i don't
30:30think that's the case at all and i think technologists should be doing a better job of
30:35demystifying some of this technology and helping to helping people to understand how it works and
30:40what it is and what it isn't because it's like what it is and what it isn't thing there are people
30:45um i read sort of treating chat gpt as a search engine and as you were saying it's not a search
30:50engine because it has these hallucinations and it's not necessarily built and designed as a
30:55search engine but given the little amount of information out there and the difficulty to
31:01pick through it if you're not an expert in the field it's not surprising that a lot of people
31:05don't know that difference there so i think there's definitely something in the narrative
31:09that's being perpetuated about technology and who that benefits and who that doesn't benefit are
31:14there any particular misconceptions you think that that it'd be good to to address now so do you
31:20think people maybe on the stream might be help benefit from from hearing the truth about um i
31:27don't know if i have anything specific in concrete um but i do think there is um
31:34this technology does something it's not simple but it's easy to understand that what you're doing is
31:39you're if we're talking about llms you're feeding in language and it's trying to predict a good
31:44response to that based on what it understands of the data the input data that it's seen from the
31:50internet and so conceptually it's not a very difficult thing to understand you've just got
31:56this machine which is trying to predict what might be a good answer to your question
32:00and i think that's where a lot of people fall over because they think it might be doing something
32:04more than that and there's a lot of different types of methods that companies use to train
32:10the models to do very well at this and they have three or four different ways that they three or
32:15four different approaches to building good models that can come up with these good responses
32:20and so i think that the idea that it is this sort of really smart sort of text completion system
32:26really smart sort of text completion system which is smart enough to um you know be able to complete
32:34huge amounts of text and some of these um capabilities are really quite unexpected
32:40um i always think of machine translation translating from one language to another
32:45which some of these models can do but that is purely because they've been trained they've
32:50ingested such a lot of different language data from across the internet that they've somehow
32:55internally figured out how how they can do that and what it should look like without explicitly
33:00being told you're going to translate this language from one to another so i think understanding some
33:05of these nuances about what they're designed to do and what they're they're not designed to do i
33:09think it's a really useful thing to get to the bottom of and we often talk about a lot of these
33:14approaches and models as being black box and i think there's a degree of truth in that but the
33:18way that they're put together for tools that most people use it becomes grey box or even white box
33:24because you can see the different parts and quite i've been fascinated by a lot of the more
33:29substantial applications of gen ai it's not just a gen ai solution they've bolted on old
33:36older symbolic ai approaches such as rule-based reasoning and so on to direct for example the
33:41prompt engineering the prompt generation so you actually pick apart various insurance applications
33:46and so on they've got all kinds of technologies in there you can draw a little box and lines
33:50diagram and that helps to understand the information that's been flowing and the control
33:56and so on and once you start to understand these are components that people are increasingly putting
34:01together you know different different llms different types of model different types of
34:05technologies becomes more transparent yeah um okay people might argue i need some sort of
34:11engineering background but it starts i think this domestic use the term demystify and i think
34:16we have a duty to demystify what these technologies do and what they don't do more importantly yeah
34:21yeah yeah because i think like as we've sort of all gestured out you there are an awful lot of
34:25people in this industry i would suggest who it is interested is to mystify it to the greatest degree
34:31possible that it's kind of very sexy very powerful monster and it's it's a machine which is one of
34:37the things i think i sound like i'm banging around about creativity but one of the arguments i make
34:41is that we've been researching creativity for over 100 years one of the things that often winds
34:46up my computer science colleagues has argued there's been more publications in creativity
34:50science than in computer science but um you know it's provocative the body of creativity research
34:57knowledge is pretty substantial and most of that is not accessible yet to llms that most people can
35:03access online because the the engines haven't reached them but we've got all this codified
35:07knowledge about how to be creative it's the sort of thing i teach your master's courses and we we
35:12train people on you can put that kind of knowledge very explicitly into other kinds of technology
35:17that then drive for example a machine learning engine or a generative ai component and once you
35:24start to understand that you can actually make white box codified knowledge interacting with
35:29this it suddenly becomes more understandable i think we have a duty to demonstrate that not
35:34all knowledge exists within a machine learning model resulting from machine learning there are
35:39other ways of sharing and communicating content the reasoning with it i should say yeah and it
35:46may be there may be some more simple systems than we realize they're but they are very expensive
35:51systems and we had an awful lot of questions about climate change which again i think is not something
35:56that came up last time we had this conversation so much uh so the sort of most general level
36:01how do you feel about just how you know energetically expensive and computationally
36:07expensive all of this stuff is and and whether that gives you any kind of pause for thought as
36:12we think about how and how often we use this stuff i agree it makes me i feel awkward working in a in
36:20a in a sector that is potentially doing a substantial amount of damage to the planet
36:24there is a movement around approval ai i think also going back to my argument there are many other
36:30more lightweight and established ai approaches we're often sort of using sledgehammers to crack
36:35nuts you know do me a shopping list on gen ai drop down the tree it's ridiculous there are other ways
36:40in which you you can do this so it might be that we might see more hybrid approaches to ai machine
36:46reasoning which could substantially reduce the cost that wouldn't be one more approach i would
36:51argue for obviously some of these big models have taken huge amounts of compute to train and
36:57cost a lot of money and to use a lot of carbon and there is work around shrinking those models down
37:05which is a lot of research has gone into that so either starting off and building so-called small
37:10language models instead of these large language models and from the start or taking a large
37:15language model and compressing it down like you would compress a picture into a jpeg for example
37:20you can kind of compress a large model into a much smaller model and there are other things
37:24you can do to sort of cut chip away at it to make it smaller as well and so a lot of these techniques
37:30are being used and i know these these sort of large foundation models are big to train but one
37:35of the things that we've seen happen off the back of this is that a lot of people are taking those
37:40foundation models and building on top of them rather than then having to train them from scratch
37:45so sort of a new way of working that's emerging as well where not everybody is training these
37:50foundation models from scratch but a few of them exist and a few of them have been trained so i
37:56don't know how it weighs up against you know the the other possibility where everybody's sort of
38:00training their own models but there's definitely a difference in how people are using them versus
38:05how they're originally being trained that changes the equation. Yes some of the big tech companies
38:10are already talking and trying to adopt on-device AI so that rather than sending the queries out to
38:17big server farms they've got a powerful enough chip in their iphone or smartphone that they can
38:23handle the query there and then but yet the fact that many big tech companies are kind of going
38:31back on these climate pledges they made a few years ago because of the arrival of AI that is
38:38there needs to be solutions to that and i think what you were saying both of you were saying
38:43i think we as consumers can make choices in the same way we can make choices about the kind of
38:47cars we drive and how we drive and flights we take i think it'll become more transparent
38:52about what kinds of models we use and indeed when we should be using GPT you know just you know
38:57using it for trivial tasks we've got to accept that we are contributing to the impact by doing
39:02that so yes there's probably an emerging consumer movement to be seen around
39:09more selective and informed choices around AI use is there anything you'd say to you know i mean
39:15you could say it to me because i'm actually in this sort of situation i'm deeply concerned about
39:19the the energy it uses but also fairly um uninformed about which which models and which
39:27companies are doing that in the most responsible way um it feels very difficult often to find
39:32information we know that some big tech companies are being kind of quite uh consciously opaque in
39:39the ways that they report the energy use of these systems which how how do you think people should
39:44responsibly think about the kinds of companies that they're operating with and the systems that
39:49they're using i think there has to be a push for transparency here right so um if you take a look
39:56at some of the academic research that's being published there is a lot of work to quantify
40:00sort of the computing power and hence the environmental cost of building these things
40:05um so it can be done um and i think we need to push companies because as the individual consumer
40:10you don't have the information available to you right so i think pushing for transparency has to
40:15be the way in like you said evolving movement information is power yeah yeah and to take you
40:22on a kind of brief tour through the the other uh big ethical uh concerns that that these systems
40:28bring up um one thing again i think that we didn't speak about so much maybe last time and
40:32wasn't such a least discussed concern is obviously having these systems having access to them is a is
40:39confers great power on the companies that run them and the people that that can can use them
40:44and we've seen an awful lot of more summits than i can count about how we regulate that power and
40:49how we decide who holds it and more conversations than could ever be counted um we've got an
40:56anonymous question here about uh what issues surround power control and accountability and
41:00how might governments or big tech misuse them do you do we have a sense do you think of um
41:06how responsible governments and companies are being with this new power that they've
41:11in some cases sort of stumbled upon uh in other cases being active in in building and
41:16kind of harnessing do you think people are behaving responsibly are governments and companies
41:22being good citizens i would say not responsibly enough might be concerned i'm less i mean ethics
41:28is a huge issue and of course i recognize my biggest fear is the related issue of bad actors
41:32you know russia china i think that's you know we could we're getting to the point quite soon where
41:38we're not going to be able to distinguish very effectively from the evidence in front of us
41:42whether something is real or not you know not just an image but five ten minute video
41:48um and at that point there's some fundamental questions for the kind of societies that we live
41:52in at the moment um and we're already seeing uh fragmentation of our societies in north america
41:59and europe um so in that context i think we have to argue that the the technology companies are
42:04not taking the responsibilities um sufficiently um yeah let me leave it at that i think it's
42:11quite great there's a counter argument that if you were to regulate these big u.s companies
42:16that companies in china are just going to do it anyway so the big fear of say having a
42:22completely deep fate um video of say macron announcing something very terrible that could
42:29spark a geopolitical event perfectly like synced up to his lips perfect audio everything
42:36it's going to happen what people need to be aware of and the media and everything has been like be
42:43very careful of what you're actually where did this actually come from um so it's more also a
42:49question of education for people of how to how to sort of respond to this technology that's coming
42:54through yeah yeah at base we've made a big point of teaching critical thinking now to all of our
43:01undergraduates first year first semester entire courses on critical thinking yeah i know there
43:07are only five six hundred of the population that's coming through but yes we we need to have
43:12a more critical population in terms of being able to judge and make informed decisions about what
43:18they see and hear because again to take it to just like a more practical level maybe you can
43:24give us some of the lessons that you're giving to them how do you you know as a as a kind of
43:28media consumer in in this newly ai filled hopefully not filled just yet but certainly populated uh
43:36world what are the things you need to be thinking about as someone who's spending time on the
43:40internet spending time reading social networks how do you how do you deal with that environment
43:48it's a big question i would one of my facetious answers is don't spend so much time on social
43:52media and the internet there's the real world out there you can observe think think long and hard
43:57about this your source and what they might what the motivations for them putting that information
44:01there might be but for cross cross referencing you know this this journalistic practice of at
44:06least two sources those kinds of those kinds of activities can at least lead you to being
44:11more critical of what you're seeing hearing and thinking about but any new technology people are
44:16going to find ways to misuse it and one of the biggest real world uses of generative ai right
44:24now is in phishing attacks through like cyber criminals using these tools to
44:32elicit personal information out of people and it's become a big an increasingly big problem because
44:38before these phishing attacks you might be able to tell the english is not quite right or there's
44:42something a bit off about them whereas now you can have an extended chat with them without
44:48realizing that it's just an ai tool that's been programmed to get all your personal information
44:54so um yeah you've just got to be recognize those threats and then publicize that this is something
45:00that criminals are doing you should expect it i think that's that's the reality now that is your
45:04default you know it's sad to say that we need need to be more suspicious but we are in a period in
45:10which we under under attack at many different levels individually and as societies and people
45:16need to be much more on their guard so it's not actually i don't think there's a technology
45:21solution to this or a regulation solution it's actually about the population having to change
45:26how it treats social media and digitally delivered content exactly it's not just one solution either
45:33so i'm thinking back to covid times you know we had lots of prevention preventative measures so
45:38we'd wear masks and distance and wash hands and you know all of those things layered up meant that
45:42we could protect ourselves better and it's the same with you know misinformation and disinformation
45:47i don't think there's one technical solution or one societal solution to this a series of
45:52different measures so critical thinking is one and there are you know initiatives about labeling
45:58images if they are ai generated and then verifying the source of the data and potentially relying on
46:05trusted sources like journalists to do a lot of fact checking as their their day job to add that
46:11extra layer of trust and so i think layering these things up is is the way we have to think about it
46:16rather than looking for the one solution which is going to solve this problem yeah it's interesting
46:21actually we have a question here from from brian ratner a live question can ai be used to spot
46:26fake photos and videos can it kind of eat itself is are we seeing ai to stop good ai to stop bad
46:32ai just yet with llms language models we know that the detectors of ai generated text are not very
46:40accurate um and so i think maybe the same as pictures as well you know detecting it is a lot
46:48harder yeah i saw a study in my emails yesterday i think i didn't write about it but it said
46:55something like 93 of ai generated content in college level essays goes undetected i don't
47:04it seems like a very difficult problem open ai claims to have tools to catch when something's
47:10been um uh ai generated and there's also very obvious tells now but they're going to probably
47:17get better the thing like with hands if you see an image and their hands are a bit weird i think
47:21we'd be on that already we'd be on that yeah potentially uh videos again though um can go
47:26very weird very quickly but these are i think those that's what we're talking about last year
47:31when we're having this i think yeah like you said that's already changing and in another year
47:35the videos will get more realistic we'll have yeah you can kind of tell if you have lots of
47:42students do the same essay and they all give you a very similarly structured sort of essay you can
47:47start to see that maybe they've all used a generated system to create it but if you were
47:52to just look at one of those in isolation you may never spot it so there's something as well about
47:57seeing patterns of how people are delivering it comes down to having to change and i think
48:01many people in higher education are accepting that the old-fashioned right 500 words on x
48:06which is largely reporting to clarity knowledge is kind of dead as a valid form of evaluation
48:13going back to ai as kind of gatekeeper it certainly can't help i mean there's that
48:16famous case a few weeks ago of daisy the uh daisy the grandma who's there to stop the scammers have
48:21you seen that yeah yeah i mean that's that's a really creative way and i've saw you know many
48:26people saying it's one of the best applications of gen ai yeah yeah that's come out um so there
48:31are more creative uses of technologies including gen ai yeah that can be applied to stop this
48:37process it might not be about detection it's more probably on the front foot yeah yeah i feel like
48:42that because so daisy was she was a kind of fake grandma wasn't she kept scammers talking for as
48:48long as possible you see on twitter as well i feel like you have often scam bots accidentally talking
48:54to other scam bots and just kind of keeping each other busy which i mean for the environment is
48:59terrible no doubt because every one of those is okay but uh at least for keeping them occupied
49:05is maybe good and actually i mean we've we've we've as i say we've gone through a quick tour
49:11there of all the the the bad things um the team have asked the attendees on this stream whether
49:16they felt overall they feel positive or worried worried about ai the results are 46 percent
49:22we're in a 54 percent positive and i think we both in this conversation and in the world we've
49:27probably spent a lot more than 46 percent of our time talking about how terrified we are and how
49:32terrified we should be we've got 10 minutes or so left of this stream so i'd like to represent
49:36those 54 positive people and hopefully send people away feeling a bit more happy or at least
49:42a bit less depressed about things and is there anything you know um you know as i say we're
49:48having this conversation a couple of years into this chat gpt revolution is there anything either
49:53in that field of gpts of llms or maybe in ai generally that's making you especially excited
49:59and especially optimistic about where it could go even if not where it's going yeah i've got one
50:05that's not llms but it is ai generally um would be deep mind's alpha fold which was um a protein
50:13um mapping tool thing that's one uh demis has said this and john jumper i think the
50:21nobel prize this year and that is transforming um it's used in almost every sector every field
50:28of biology but it's transforming everything from drug discovery to vaccine development
50:33to coming up with new treatments um or understanding like neurodegenerative diseases
50:39um and that that i feel like we're only just the tool is there and we're now finally getting the
50:47practical real world use cases out of it and that could be so transformative for so many people's
50:53lives and that's something in the background that people don't really have to worry about
50:56in their day-to-day lives but that's happening in labs around the world um on computer screens
51:01around the world um our understanding of science through tools like that and others like that um
51:07which could transform society in a very positive way yeah and that feels often with the good stuff
51:15it's often slightly more complex slightly more deep science slightly more abstracted than the
51:22various horrible things that we see the misinformation and stuff we've talked about
51:25but it it does feel really important also to point to that even if it is like slightly less
51:30kind of exciting yeah i was also going to talk about ai for science because i think it's a sort
51:36of blooming area right now and there's a lot of interest in it as before it's one great model that
51:41contributes to this but i spent the past year or so working with a lot of researchers at cambridge
51:46and figuring out how they can use ai in their research and there's some really great applications
51:52that people are doing so one of the things that's happening in science at the moment is that people
51:58are starting to gather lots more data so whether that's data they've collected from from people in
52:03medicine or whether it's sensors in biology able to sort of look at um people's bodies on a cellular
52:09level a little bit clearer than we could decades ago or whether you're collecting written data sort
52:15of in law maybe there's a lot of written data which is now accessible that we can we can do
52:19science with so we're starting to find there's lots more data and people scientists can't manually
52:25sift through all the data anymore and so ai is getting used increasingly across the research
52:32world to help make sense of that data so to try and look at scientific data we're gathering to
52:39see if there's any patterns that they can spot whether we can look at the cells in our body and
52:44try and identify you know particular types of cell that we could then try and see if that's something
52:50that is related to a particular disease that we could then go and test in a lab and you know
52:55potentially find a cure for that disease or a treatment for that disease or whether it's in
53:00medicine and we're trying to understand which blood test results are related to which conditions
53:08or which symptoms you have in much more detail so that we can try and find a diagnostic test for
53:14something or you know another treatment that's going to work and in law i was talking to somebody
53:20today about trying to increase access to justice by allowing people to you know use llms to access
53:27legal texts and understand them in a lot more detail all of these things we couldn't do a few
53:32years ago we didn't have the data we didn't have the tools and so i think there's lots of places
53:36and like you said it's not the kind of thing that people see in their everyday lives this takes
53:40time to make its way through the system because you have to go on and do experiments you have to
53:46do tests in the in the lab you have to do clinical trials and that that takes years to do so we're
53:51at the very early stages of seeing how this is going to play out but i think there's lots of
53:55exciting ways that people are using ai to help advance our knowledge of science yeah that last
54:01point is really powerful um there is there is a huge body of knowledge and information that we as
54:07society as individuals can use to manage and improve our lives most of it's not being accessible
54:13to most people because it hits it's behind publisher paywalls and even if they've got
54:17access to it they've got time to read through all of the information they can't afford to pay for
54:22lawyers and so on or other kinds of professionals all of a sudden ai has the capability to
54:28manipulate that information and provide it to you in a form that you can digest and use when you
54:34might need it and then you know to me one of my fascinations is how can we improve quality of
54:39life certainly in the last third of life and all of a sudden these ai technologies have that
54:44capability there's a reasonably good understanding of what quality of life is you can model it
54:49um but how do you get that from some sort of rather conceptual framework academic literature
54:55to you know individuals caring for someone with alzheimer's in their own home in the last days of
55:00life you can't do that by any other way i can think of a world design technology can do that
55:05so i think we have the opportunity through ai to become a far better informed society
55:11and that can lead to numerous benefits and that doesn't tend to get talked about at the moment
55:16yeah there's a slightly morbid way of thinking about the same thing but i've read about this app
55:20this week death clock did you see that no it's uh you uh give it a load of information through
55:27a questionnaire about you know your lifestyle so it works out with that but then it then it says
55:34and we've also crunched all these um you know proper scientific work on longevity if you were
55:42to change these things you'd get two and a half years sounds like you maybe you don't like that
55:47idea it feels like my eyes are simplistic yeah so yeah there's probably could be benefit i'm doing
55:54the developers a disservice but uh yes i think we we can make choices i mean most of us can
55:59prolong our lives by making informed choices about nutrition exercise lifestyle stress and so on
56:05and there isn't really a concerted and regulated set of of tools to enable us to do that and ai
56:13is one vehicle by which those tools can be delivered yeah yeah brilliant thank you well
56:18that feels like a a very positive and optimistic way to end after uh maybe not quite as positive
56:24but not quite as optimistic a couple of years i think and also you know we've we've also talked
56:29about some of those threats and dangers and terrors and they're all very real i think because
56:34we've kind of gestured out that there's good stuff happening too so that's
56:37yeah exciting place to end thank you ever so much out there for for joining us this evening thank you
56:42to everyone for your insightful comments and uh information all those sorts of things um if you
56:48enjoyed tonight's event and would like to support the independent please use the link in the chat
56:52box to support our journalism with a donation uh our next virtual event is next week actually it's
56:57december the 11th focusing on domestic abuse and the role of women's refuges in partnership
57:02with refuge the uk's leading charity the panel includes refuge interim ceo abigaila ponfe
57:10radio presenter and broadcaster nikita kanda and our own women's correspondent myra oppenheim
57:16the link for tickets is also in the chat box so make sure you grab those there
57:19um that's part of our campaign brick by brick with refuge
57:24thank you ever so much for joining it's been great thank you

Recommended