Transcript — Artificial Intelligence (Lieven Scheire)

⚠️ Speaker labels are absent in this transcript. The source had no per-speaker labels. The opening ~10 lines are spoken by an unnamed host/MC introducing Scheire; everything after Scheire takes the stage is him speaking (apart from brief, inaudible audience reactions). Attribute the main content to Scheire; attribute the introduction generically ("the host") unless a named participant is clear from context.

The transcript also contains numerous speech-to-text artifacts that have been preserved verbatim. Likely intended phrases for the more confusing ones:

• "she's always hard" → "Scheire is always hard"
• "dodge" / "in dodge" → "Dutch"
• "wearer's only robot" → "Where's Wally robot"
• "AGI will lead to AGI" → almost certainly "AI will lead to AGI"
• "two of trolls" → "Two Towers" (LOTR)
• "in betraying it" → "in the training data"
• "banhan" / "banham" / "Benham" → "Ben Hamm"
• "Inverness… venet" → "Inverness Caledonian Thistle"

Do not silently correct these in quoted material; you may add [sic] or a bracketed gloss.

---

Host introduction

Was it now? There? With. Awesome, how is everyone? Enjoy lunch? Have we lived up to our reputation in England for serving bad food or it was lunch? Good? Cool how's everyone enjoy the conference on getting good value out of the sessions? Who's talked to someone they don't know yet? Not good enough come on people come on people right where are going to our second keynote of the day and this is going to be a little bit different we have from from Belgium and Lieven is a big deal in Belgium and we thought we can't let these Belgians get all the best bits out of life so we've invited Lieven over to give us a deep dive into AI in the humorous way which Lieven does so it gives me great pleasure to welcome on stage oh and by the way by the way Lieven has has mocked other people on how they pronounce his name. And created YouTube videos of things like this of all the weird and wonderful ways in which Lieven is pronounced. So it gives me great pleasure in introducing Lieven.

Self-introduction & framing

Good afternoon everyone. And welcome so that is right my name is my actual name she's always hard when I speak engage [sic — "Scheire is always hard when I speak English"] because I will get on stage and say hi I'm Lieven. Confusion here and there but that's who I am it's my first name one of my hobbies is going to train stations where I stand on the platform and I shout stop the train I'm Lieven I am my last name is. How it's pronounced. We have more of the sounds in dodge [sic — "Dutch"]. As the saying goes Dutch is not a language it's a disease of the throat so that's what's going on there hardly ever pronounced right over all of course one time I was announced as Lieven Shire. Practically made me the first chapter of the hobbit so that's who I am that's my name what do I do I'm a science communicator so I studied physics now I talk about AI because when AI introduced or winning Nobel prize in physics I feel very entitled to talk about AI and for the past three years I've been traveling Belgium and to explain people what the idea is well travel is a big word when you're in Belgium you basically walk from one side of the country to the other and explain science by doing it and of course artificial intelligence has been a huge topic much interest in this topic from the interest of layman and now I speak for a room full of developers. Which makes me a bit nervous I feel pretty peer reviewed here but we'll see if you know everything already that I tell you that maybe you can it can teach you a way to explain it to other people who want to know more

AI's birthday: Dartmouth 1956

so yes what do I tell people about AI first of all that it's not a brand new technology. But it is all tech. AI is having a birthday in two weeks June 18th. Artificial intelligence is turning 70 years old. You can organize a birthday party for AI on June 18th why this date this was the start of the Dartmouth summer research project into artificial intelligence June 18 1956 this was the moment these words were used for the first time by John McCarthy so it's heavy his birthday on June 18th so AI is 70 years old this amazes people and they say what took them so long. Why if it's 70 years old why did this this lost this big boom in the eye only happened the past few years and of course the answer is is quite simple it is because only now did we get enough computing power and enough data to train good AI. If you if you try to build the eye with insufficient computing power it needs efficient data you do not end up with artificial intelligence you will end up with artificial obnoxiousness. Which we also tried. This. Everyone over 30 now thinks oh no not him. Alone on the 30 nothings is that a pokemon movie this is clicky so this was here with insufficient computing power and now we do have to compute it power and the data to build ai that does amazing things.

The one-sentence definition

And I like to I like to always be very down to earth about what it is and what it can do but still it is incredible what comes out of these neural networks and this has the effect that people sometimes start thinking that these are magical machines that can think and reason like a human being. And I'm still convinced that they are not. So I like to always get people back on earth about the eye by describing it in just one sentence. And it is this one what is artificial intelligence? It is simply a new kind of software that is good at pattern recognition. That's how down to earth I like to describe it a new kind of software that is good at pattern recognition.

Why classic software fails at pattern recognition

So pattern recognition things like recognizing faces recognizing objects and stuff like that because of course at this moment it's all about large language models and it's all about language but AI is much more than that it started with pattern recognition machines and the classic software that we have been using for so long can do a lot of things very good it's very bad at pattern recognition normally I have to explain what is written here I will not bother today. Because you all know what this is so this is software that asks five people for their age and then calculates the mean age of the group. Every one of us even people who don't write software can break this task down in simple executable steps for a computer ask five people for their age count them all together divided by five then print on the screen this is the mean age of the group it's very easy it's a very easy task for classic hard coded software turns out it is very difficult nearly impossible to code pattern recognition with this kind of software. So as an exercise which you probably maybe already did what if I gave you 1000 digital images and I told you 500 of them had a cat in them and five all of them have a dog home and you have to hard code software that can decide which pictures have a catalog which patients have a dog in it that is nearly impossible. What what hard coded rule that a computer can execute can you use for this you can't tell it to look for certain colors because these animals have all kinds of colors you could maybe tell it look for very similar two very similar dark spots and then hope that these are the eyes and then find some structure in the head but when there's only one lion picture it doesn't work. So pattern recognition with this kind of software simply never worked.

Pattern recognition as a new superpower

And that was the big change when these neural network became so good for the first time ever we had software that was good and packing up initiation [sic — "pattern recognition"] it was a new superpower something we never had. And the moment we had this new superpower in software the brightest minds of the planet started using it to solve the biggest problems of our society. Such as where is bullying [sic — "Where's Wally"]. This is a wearer's only robot [sic — "Where's Wally robot"]. We needed it. So it's a camera of an AI that will look for the pattern of what is feed [sic — "Wally"] on the pitch and then the robotic arm will say there is money [sic — "there is Wally"]. Because also our pastimes have to be automated. Now 20 years ago you could take 2 million pounds and go to a big software company and ask them make me aware of only robot [sic — "a Where's Wally robot"] they could not do it because we didn't have software that was good at pattern recognition today anyone can build this at their kitchen table with things that you can use free online so this pattern recognition machine is the new superpower and it is very strange that this one simple thing the fact that our software was never good at pattern regulation [sic — "recognition"] and now it is for that matter there's one simple change started the entire day I revolution that we see now and is actively changing the world all from this one very small conceptual thing

From pattern recognition to generative to (?) AGI

now like I said when you talk about the I two people today they always immediately assume you're talking about ChatGPT some noise AI is much bigger than large language models so it's not a renewable networks [sic — "neural networks"] that became good at pattern recognition after the pattern recognition it could not just recognize patterns that would generate patterns images and videos then it started generating language that's when the magic happened because it generated language that we get feedback into the system that's how we got a genetic [sic — "agentic"] AI and now the next step they all dream of is AGI I put a question mark. There. I am not sure that AGI will lead to AGI [sic — likely "that AI will lead to AGI"] I think it is very impressive language imitation I don't think it does it has the architecture of conceptual thinking and suffering [sic — likely "reasoning"] that we have not had things to discuss over the party this evening.

AI in your pocket: the smartphone

Fun things to do of course when I talk to regular people who are not into software I always tell them AI is everywhere it's even in your pocket. Your iPhone your smartphone uses a lot of AI well the smartphone many people curse it because it draws their attention the way it's stuff like that but it is an absolute miracle I love this machine. Well to describe how much of a miracle this thing is I like to go back to the moon landing. Well first I ask the people what historic event happened on July 21st 1969. And then of course the right answer is the first victory of any banks in the two of trolls [sic — "Two Towers"]. But also there was a moon landing. And then you. Could you could count together all the computing power that was used for Apollo 11 so they had two supercomputers in the rocket and then they had an entire server part of IBM mainframes on birth in use all this compute about used for Apollo 11 if you count it together. All that comput. Ing this smartphone is 100,000 times its power. So in our pockets we literally carry 100,000 fold of the computing power that we need for a man to move landing. And we use this to slide fruit candy in a row tenant explodes. But that's what regular people use it for you use it to fight code and the IA that reads the emails of your boss but another story no people use this for that and so there's a lot of apps that can do amazing things that would have been absolutely impossible 20 years ago because we simply didn't have we simply didn't have any software that was built at pattern recognition and I now have passive recognition that recognizes songs on the radio I have pattern recognition that recognizes bird songs when you're walking around in the woods you open the app called Merlin and it listens around and it just gives you a list of all the birds that you hear at the end I will show you a link with all the apps and websites that I use if you want to play around with them I have accepted recognized plants and animals when I take photos of them in nature

PhotoMath & education

and I have the wonderful at photo math anyone here who use this photo math. H? Okay photomatism active they simply activated the camera it starts filming the environment and just imagine you're in math class and there's a wonderful integral on the screen what do you do with your app you take a picture of the integral there you go the integral solved. That's fun. Who has kids in secondary school? Who teaches a secondary school no money. Such fun what do you do as a teacher what what can you there's a free app now that solves all the homework with one picture within the second how can you still be homework. Of course these teachers are smart they know if I give homework today I have to ask the students to write down all the logical steps they have taken to solve this. Now there's a red button here when you push it you get. Is this. A disaster for education? I don't know I don't think so but it doesn't have to be you can use this the entire year to cheat on all your homework and fail the exam because you won't be able to use your phone if during the year you want to solve something that you don't understand it you can ask the AI to solve it you can click every one of these holding steps and it will give you the theory that it used for that step this can be a very good and free math teacher. It's one of these many examples of technology that give you the choice if it's going to make you dumb or smarter. Choose wisely of course for you I can I can talk a bit more about this it's that the only AI step here is recognizing the pattern of formula and turning it into a written formula that they can put on the stats all the math of course is done by hardware and software and when I tell people this they go oh my gosh did computers become very good at mathematics well they have been formed too worried yeah

Speech, translation, voice cloning

so that's my recognition first step the fact that you can recognize faces the fact that you can recognize sound that you can recognize the pattern of sound and turn it into speech the speech to text software that in Belgium we were very fond of. For a few years so speech detects software then translation is done more and more by AI systems it used to be hard coded that's why we got this master Yoda sentences because they change words into other languages but didn't copy the grammar AI does this a lot better AI can also turn text into speech it can imitate your voice because the very specific sound by which we recognize a voice is a pattern in the sound wave. AI can analyze this pattern can imitate it AI can also edit video for us now all generative AI and then you can combine these that's what you do developers you combine all these tricks into all architecture that does things that would have been absolutely magical 20 years ago

HeyGen Italian-dub demo

so just before we came up here I filmed myself I filmed a small clip just outside I'll first show you the clip then I'll tell you what I'm gonna do with it. Truly. Where I will. Be giving my. Talk on artificial intelligence very hot topic for a science communicator because you can do so many demonstrations demonstrations that I guess were built by developers for me to use I'm gonna do a demo with this clip right here what I am going to do with it I will tell you now on stage. So yes I filled this clip and then I uploaded it to HeyGen. HeyGen is a website you upload this clip and then I asked this website could you please make me speak italian? So translate everything I said into Italian making me imitate my voice so it sounds like I'm speaking Italian and while they were at it also edit the video so it looks that the movements on my lips are more conformed to the sounds I make at the time I think it was it did some it was computed for five minutes I think and then I got this clip back in email so it's exactly the same clip just gone through an AI system to make a speed. Tag. For solid intelligence difficult because it was not a tangle most ratio demonstration image that is to video. Well. Did that make me look sexy I don't speak it was it right in his hand anymore it was all correct excellent well there now I do speak it out so these are things that would have been completely magical just 20 years ago and now because just the combination of pattern regulation and generating patterns now makes this possible.

How brains do pattern recognition

But how did we get here why didn't we suddenly find a kind of software that was good at pattern recognition we didn't have it for so long however if 30 years ago you needed a good pattern recognition machine you could easily hire one. We called it an employee. Humans are very good pattern recognition machines this is what we do the basic task of any biological brain is pattern recognition. When a baby is born that's what it does recognize faces recognizes voices recognizes situational patterns recognizes objects gives it names it's all pattern recognition. The reason that our brain is so good at it is because it works in a completely different way than this hard coded software. We do not have a list of if there's then that rules in our hand that's not how the brain works. Our brain is one big sponge of connected neurons. 86 billion neurons in one human brain so it's all just connections and every neuron has these tentacles that connect to some other neurons and through these tentacles an active neuron can activate other neurons in the network so for example a few neurons get activated at the bottom left through their connections they will activate some other neurons in the network it's a very simple system. Every neuron has a few hundred connections some are stronger some are weaker to other neurons in the network when they get activated through these connections they activate all the neurons and this very simple system. Almost automatically creates pattern recognition. For example this feeling that many people have you're walking around somewhere and suddenly you smell on the roller [sic — "an aroma"] that you remember from your childhood and a very strong memory is activated in medium. What happens in your brain at that moment is exactly this. So these neurons at the bottom left. They're connected to your nose they can activate it by this smell by these chemicals. All the neurons at the right they are the neurons that got activated by the situation where you used to smell this as a child. Who was with you what did you feel like what did it look like these are the memory rules. In your childhood. These auroma neurons and the situation neurons have been active together over and over again. And that's what we bring does neurons that get activated together build stronger and stronger connections. So later in your life when there's a roman neural activated through these trained connections they will automatically activate this memory. So a simple system that leads to good pattern recognition

Building a neural network

that's why the computer scientists thought. If this biological system is so good at pattern recognition. And our software is so valid. Why don't we mimic this structure in software. And that's how they build the neural network. Now normally I have to explain what this is but I think most of you have who has never seen the neural network in person. So yeah so you have your input digital image for example ones and zeros. You consider every one of these pixels every value of these pixels you consider these values to be neurons that are active or inactive and then you program connections some stronger some bigger and they activate neurons in the second layer and the third layer and the fourth layer at the end you want the neural network to tell you there's a cat or a dog in the picture. And of course if you just build this it doesn't work but because it hasn't been trained. How do you train it you show it the picture of a cat you clearly tell this thing this is a cat these two sides are active at the same time and based on this the computer makes some connections stronger and weaker in the neurons that's about how deep I can get a general audience if you want the entire mathematics or back propagation there's also a link to the end to the 3Blue1Brown YouTube channel that does amazing videos about it so yes one picture is not enough for training you'd show with another picture of a catch you clearly tell it this is a cat these two sides are active at the same time just like in your childhood the aroma and the memory reacted at the same time based on this it makes stronger and stronger productions you show with a picture of a dog you clearly tell it this is a dog both sides are active at the same time the computer makes some connections stronger and some weaker so every step of this training this network gets better adapted. To start reacting with the word dog to any possible picture of a dog. And with the word hat [sic — "cat"] to any possible picture of a cat. Does it know the dog or a cat is I don't think so. It's just the computer says yes. That's it that's what happens and so then when you've trained enough you get test if it works you show it in a new picture of a cat that it hasn't seen before you use these pixels to activate through the connections all the other neurons and somehow it says this is most probably a cat. You have mimicked the architecture of the human brain in software and let it adapt itself until it recognizes cats and dogs.

The black box of AI

Now how does it do this so why does this resulting network apparently react with the word cat to any possible picture of a cat and we're dog to any possible picture dog nobody knows. We have no idea. But it works through even. That's how in Belgium we deal estate politics. It somehow functional don't touch it in the fall apart. The black box of AI and this is a very intriguing difference with hard coded software before you start writing hardcoded software the human has done the thinking came up with a solution and tells the computer how to execute this solution here we didn't find the solution. We just told the computer adapt this neural network until it works and then others more. You tell the computer go figure it out yourself. And then in the end you have this black box neural network that apparently reacts with the word cat to any possible picture of a cat and the word talk to any possible picture the black box of AI.

Training-data bias: wolves vs huskies

Which of course can give you trouble every now and then neural network will do something strange and we will not know why. A few years ago a neural network was trained to see the difference between wolves and huskies. Difficult to tell apart. They trained in a lot of picture of wolves and huskies they got a neural network and it worked perfectly well. They showed a new pictures it'll always be able to right answer until suddenly they showed me the picture of a wolf and the neural network said this is 100% sure a husky. Something went wrong. You cannot ask this neural network why do you think it's a husky? Because it has no clue. But then they did something clever they wrote new software that would just feed the same picture to the neural network over and over again but make small changes in a very specific area and then change the area. Then you can see how much how much difference the result has when you change a certain area and you can make a heat map of how important certain parts of the picture are and that's how they found out that to tell wolves and hospitals apart the software was looking at one thing only. Is there snow in the background? This example you could find in a line that the snow on the background it is often used to warn people who work with AI for this snow in the background effect, the so-called bias in the training data. Humans are very bad we have a blind spot for this kind of bias. It is so obvious that we do not see it.

Training-data bias: skin cancer ruler

And wolves and huskies, it's all fun and games. A bit later they developed an app, I think and that was developed that could see if a spot on your skin was dangerous or not. So you take a picture on a spot on your skin it would tell you if it was skin cancer or it was just an innocent spot. How did it do it? They took a lot of pictures of innocent spots from volunteers and they asked a lot of pictures of confirmed skin cancer spots from doctors and the first version it went wrong because the main thing that the computer was looking for was is there a ruler of a doctor next to this part? It is so obvious that we don't see it. The bias in betraying it [sic — "in the training data"].

Hobbyist AI: Custom Vision & Teachable Machine

Now can you train your own neural network you certainly can. It is very easy. You can go to websites that trade it for you. So yeah we are now at the pattern recognition phase. You can go to websites that train it for you such as this one Custom Vision. This is not really developer stuff. This is hobbyist stuff. You can be a hobbyist maker doing stuff with ai. So what do you do? You take a lot of pictures of wolves and of huskies. You label them in the right way. You upload it to this website. You get a neural network back that recognizes these things. This is also into play with. If any of you want a camera at your front door that recognizes all your friends. And makes a different sound from one you like and what you don't like. You can build this yourself. It's also Google has made one of us saying that young people is called Teachable Machine. It can recognize images. You can train it to recognize images to recognize sounds and musical styles. It can see your stance that you can use to control games. It can recognize your voice. If you have a lot of recordings of your own voice. And if other voices you simply label them, upload them and you will get software that recognizes your voice. Any teenager that wants to walk on their own door that only opens when it hears their voice. Can build this today. It's not that difficult.

Ben Hamm's cat flap

And all over the world people are now hobbyist tinkering with these kinds of pattern recognition machines. Such as banhan. An American engineer banham [sic — "Ben Hamm"]. As a cat. And he has a small problem. It is that his cat really loves him. And it brings him present. Ben will arrive home and there will be a dead bird in his bed. Which I realize sounds strange in British English. If he was lucky he would wake up next to you and read this. Ben didn't like these birds in his bed. So what did he do? He installed a webcam on the cat flap that took pictures when the cat came in and he selected a few hundred pictures of the cat with grey. And a few thousand pictures of the cat without prey and he hand labeled them and uploaded them to Teachable Machine and got software back that could see if his cat was carrying the brain. He put this on a small computer connected it to the webcam at the cath lab bends cat flap now automatically closes locks. When the cats carry prey. He also wrote something in there that it does an automatic donation to bird protection. His talk is online. Benham you can find it.

Inverness Caledonian Thistle bald-referee AI camera

So people can build these things all over the world. People are tinkering with their own AI project. Also in Scotland football club and venet [sic — "Inverness Caledonian Thistle"]. During covet. People were not allowed inside the stadium. So they put a webcam on the field so people could watch at home. This band can fill the entire field. It was kind of boring. Somebody built an AI system that could recognize the football. And it could digitally zoom in on that part of the field. And that was a lot more fun and interesting to watch and it went perfectly well just almost home coded ai protector. It worked very well until this game. What is the problem with this game? It is that the assistant referee is bold [sic — "bald"]. And the AI is looking. For a round shiny object. So the entire game. People were looking at Patrick. They started foaming the stadium telling them please give this man a wig. The entire time. I feel you are developed roam.

Patterns easy for humans, hard for AI

This happens every now and then a pattern is easy for us to recognize and difficult for AI. There's a few very famous pictures sets that I think many of you have seen already that are used to see how good your network is and recognizing small differences. By far the most famous one is this one. It is a muffin or shiman [sic — "chihuahua"]. The classic. But there's a few more. There is apple or this is sloth or pan or shukuna. And the classic apple were born ow. L. And probably my favorite is well Afghan windhound or solemn. Difficult for AI and very easy for. Us.

Wrap-up: LLMs as language imitation, not thought

Okay I'm running out of time so I will have to wrap up and how far did we get so renewal networks that can recognize patterns and then we train them to generate patents first we train them to generate images and videos and then we train them to imitate language. And I'm still convinced that's what they do. They imitate language and when I ask a lot of language model to write me a story about a cat, I'm convinced. Doesn't have any conceptual idea of what a cat is. But the word cat is a data point that statistically follows the previous data point. S based on the training data. But then you can ask large language model to write a summit. About an animal and a certain emotion and what comes out absolutely blows my mind. So a lot of room for debate on how much conceptual thinking of using neural networks do that I don't have flavor out but we have a party this evening.

Resources

I promised you a website where you can find all the apps and all those fun things that I showed you. It is this one lieven.scanner.com/ai-links and because of course by tomorrow you will have to couple a spinning off by name. I have also registered lieven.shire. Thank you very much for your attention and see you.