AI: Google Deepmind makes breakthrough in difficult artificial intelligence test | BBC News

it is that time of the week it is time for our weekly segment AI [Music] decoded welcome to AI decoded the time of the week when we dive deep into some of the most eye-catching stories in the world of artificial intelligence this is my favorite part of the week H because it is such a fast moving story so if you have thoughts questions on issues that we covering do send them in I'm learning like you and I want to know what you think at C Fraser BBC on X so do send in your questions we'll start tonight with this story in the independent Google say their Deep Mind system has made another major breakthrough in one of the most difficult tests for AI in Geometry we'll tell you why that is such a big step forward Politico reports last minute tweaks to the eu's artificial intelligence act which would allow law enforcement to use facial recognition technology on recorded video without a judge's approval is that a U-turn the time feature an opinion from two lawyers who say AI should not be used in trial until the technology properly learns how to reason the Ft carries the latest announcements from Samsung who have new smartphones capable of running generative AI they will translate foreign languages as you go need to get myself one of those uh and remember the old joke you stand around waiting for a bus and three turn up at once well not anymore the UK operator first bus has told the BBC they will use AI to design and automatically update timetables no more no more bunching which if that if that's not the best story of the week I don't know what is um some other good news uh priia is here priia lar founder and CEO of the AI education and technology company Sentry Tech Hello nice to see you um let's start then with the independence story on uh deep mine why does it matter why should we be taking interest in a system that can do geometry it's fascinating did you see my eyes that's why I'm in words and not in numbers I've spent a good few hours reading the nature article that they published um this uh this paper in and and it is really really interesting the reason why we should all be taking interest is that so some of us have been playing with those chat Bots right and they'll come up with an answer but sometimes if you ask it a mask question even if you get the result that you were looking for the correct answer actually the proof in the middle is is a lot of nonsense right mass is a language of Truth Christian right there is a single answer but there may be many ways in which you can get to that answer so the reason why this is significant is that geometry and particularly because they competed in the well they didn't compete they took sort of former International mathematical Olympiad questions right they built a model and and this is why it's significant firstly they didn't have the training data to train an AI model to do geometry right so they essentially created synthetic data right it was very very clever it's a very Innovative thing to do we don't have the data so let's circumvent the problem and let's create synthetic data and they trained a model with half a billion uh geometric diagrams and then created a 100 million proofs so this isn't actually real data that we're used to dealing with when we talk about AI often in the past we're using real life data or we're training things on the worldwide web that exist so human created data they created synthetic data they fed it into these models and the second thing is that they combined two forms of engineering so two methods of developing AI technology right bear with me I'll take you through it okay so one is symbolic engines right where essentially they use this engine to which and this this sort of AI requires knowledge and logic it requires sets of rules okay so they said here's a geometry problem solve that with the knowledge and the logic and the rules that you know and if it solves it fine but if it doesn't they then combine that with a neural network uh a language uh Network and they trained that so the language language networks don't need the rules and guidance that a symbolic engine do right what what that can do is it can do pattern recognition and it can be quite creative actually and I'm going to get people shouting at me about creativity in AI in a minute but but bear with me it it works in a different way so let's say you've got this triangle okay and you want to solve the problem of a triangle you say two sides are the same that's an equilateral triangle and you want to say now prove that the Ang two of the angles are the same okay it can go through the symbolic engine it solves the problem you're fine if it doesn't it would go back to the other AI engine and that might say well here's an idea spit the triangles in half right so you've now got two triangles it'll come up with these creative ideas go back through the symbolic engine to try and solve that until it comes up with Sol you've got one system layered over another but this is this is where some of us you know who who don't work in your field get a little bit worry because that to me is an element of reasoning well this is the third thing exactly exactly this is the final so it's an element of reasoning so when we think about math by by doing that and coming up with the answers and solving 25 out of the 30 which is a gold medal for geometry in the Olympiad it's not a gold medal in math overall because there's other math types of problems okay um but the reason why it's really interesting is because what does mathematics require you to do right it's full of logic it's full of reasoning it's full of um actually critical analysis it's full of looking at consist consistency checking it's looking at contradictions and then it's applying it to real world world problems right and so that's why it is absolutely significant because if they can continue to do this yeah then they can demonstrate that AI can conduct reasoning although mathematicians shouldn't fear too much right because this is don't point at me this is I was being polite this is I was being inclusive so this is um but the reason why you know we're still a long long way away of things if you think about like firat theorem for example it created in the 1600s and he wot in the margins of a book right here's a theory but and he literally wrote but the margin doesn't have there's no space for me to write the proof and it took 357 odd years for someone to prove it which Andrew WS did right and it's a lot of different again it's all of that mathematics right it's very very complicated this is sort of below college level advanced mathematics so we still have a long way to go but we should be excited about this because you know engineering architecture um physics there's all sorts of areas medical research where if you can start to solve math Mass problems at a level that humans can't then we might be able to solve problems and answer questions that we haven't been able to do so before to solve some of the world's biggest issues another contentious area for AI is facial recognition yes um so I want to talk about this this this story about the EU um maybe not urning but but sort of going somewhere where we've been very critical of China we we've talked about facial recognition in life terms as people walk around town C cers it's not going as far as that because it's recorded video but it's a step in that well it's a step in that direction right yeah I mean I'm I'm really alarmed by I was so shocked by this we've analyzed the EU AI act on on on this program at the BBC before and it was very very clear uh what the view on facial recognition is so I'm actually not surprised why some members of the European Parliament are are up in arms about this and they're saying hang on why is this last minute wording change happened and it looks like a loophole now what they're saying is that we could use recorded video footage right and for for crimes we could then use facial recognition on top of that now there were exceptional circumstances in in the AI act pre- Christmas right but the thing is here you don't need to go to a judge to use it that's what's significant right so for example you may end up with some sort of pessy crime or you may end up with a protest on the street then using for example CCTV uh footage and recordings to then spot who these people are is so in for instance if you're looking for trouble makers you'd spot them in an instant how good is this technology there technolog amazing okay because I've seen it I was so there's a company called Clear View Ai and last year in La they were demonstrating the technology you can see it online it's actually it's extraordinary so the the founder CEO demonstrated the technology and explained that the feds in the US were looking for uh somebody who is a suspect in a rape so significant crime they couldn't find this suspect anywhere and then they used clear viiew AI they found the suspect image in the background of someone else's Facebook photo holding a pool queue in a bar in I don't know where maybe Ohio wherever it was right but it was nowhere near uh where the alleged crime had taken place okay they went to that area they found the suspect and then the investigation proceeded and when you look at this this image this software will spot you in the background image even with a different haircut if you have a beard at any point if you go it it will be able to spot you it's it's it's actually incredible I and I think I I get the impression that what the EU have done is they've looked at this sort of software and they thought actually this will help law enforcement right the reason I said why isn't it like real time as I know that China obviously with you know been picking weer Muslims off the street they've been telling children they've been jaywalking and they've been using mass surveillance for this but you take images on your smartphone right so do I we upload them right is it it's not instant but those images end up online somewhere to be used and to be uh scraped by these sorts of companies pretty quickly yeah so and all the yeah and all the guard rails all the guard rail privacy I think the and the key is privacy so do you want to end up in a in a situation in China where people live in fear you've got that chilling factor I think that's the key right people live in we' SE how it's being used we've seen that we don't want that in Europe um time's ticking let's try and rattle through some of these uh these other ones um look priia is a barrister I mean there is nothing that PR can do was Baris now he's an AI expert um trial by AI yeah I mean I I don't see a future where it's a really interesting article for the viewers because we're Russ you should read it we could do with it we're not going I think we're going to sit situation absolutely where it augments a lot of things that we do including judgments right because you're looking at AI what's the benefit okay can it actually be more objective than that than a judge the problem with all of this right is that actually AI is fed with data right and if it's bias in you get bias out at the same time humans also have bias right and the reason it's not going to replace judgment magistrates judgments where you've you've perhaps done something wrong in your car that's a pretty straight up and down isn't it in every single proceeding if someone represents themselves what's equally important in terms of the substance you know as well as the substance of what they say is that their demeanor how they say it their tone I've seen this so many times it's not this is not a worry for us and yeah the article is absolutely right but if it can help us in some areas great um I i' asked the viewers to look at compass in the US a parole a parole AI judgment system that went horribly wrong because of bias right um because that's super interesting but I'm sure it'll come up again in AI doodas we're just a minute and a half left so I'm going to editorialize let's do the let's let's do the buses my bus comes in trees yeah and I'm sure everybody else's does what's how are they going to do this quickly they've got a company and they take the data they create a digital twin of the logistics and the infrastructure how the is run when you create a digital tin twin you can essentially use predictive analysis to figure out which bus routes are best how can we optimize them how can we schedule how can we create great Tim taes that work for Christians so he's not getting three buses on a time how can we make him a happy man so he can come and do the context on time and not be late and um this is this is a fantastic exact use of machine learning and artificial intelligence and the more we have this in Logistics in Supply chains infrastructure Pro projects the better and that is exactly what way you're going to see the world go uh in the next couple of years yeah forget here on time we need more time on this and we never get through all the stories there's so much to say so much that is new um and maybe we'll we'll try and find some more time in weeks to come uh listen Pria thank you very much uh for that we will do this same time again at next week uh do hope that you'll join us for that

Channel: BBC News

Views: 224,459

Keywords: bbc, bbc news, news, world news, breaking news, us news, world, america, usa, usa news, india news

Id: E_a78HF4Q1c

Length: 12min 18sec (738 seconds)

Published: Thu Jan 18 2024