AI wins first gold at maths Olympic games; How mitochondria are linked to sleep; Famous psychology trick works on octopuses too

0:28This episode is brought to you by Accenture. This episode is sponsored by Wellcome's podcast, When Science Finds A Way. Hello and welcome to The World, The Universe and Us, the weekly news podcast from New Scientist. I'm Dr Penny Sarshe. And I'm Dr Rowan Hooper. On today's show we reveal the link between mitochondria and sleep and the link between mitochondria and Ozzy Osbourne. Oh really? Yes. We're also hearing about a famous human psychology experiment that's been reproduced in octopuses and we're going to get into what that tells us about octopus cognition. That all sounds great. And we're going to start with the week's big AI news.

1:03This week at the Maths Olympic Games in Australia, or the International Maths Olympiad, to give it its official name, two artificial intelligences won a gold medal for the first time. It's the first time that AI has got to this standard. And given the strides that we've seen over the last year or so in AI across many aspects of our lives, this seems like quite a big deal. maybe it heralds the the first whispers of artificial general intelligence this holy grail of ai alex wilkins is here alex fill us in on this what's happened what tell us first what the imo is for people who haven't heard of it yeah so the imo is one of the most prestigious maths competitions in the world it's only open to 16 to 18 year olds and to enter the kind of the full thing you basically have to drop out of school train for it it's fiendishly difficult and it consists of about six questions and they come from geometry algebra combinatorics and it's so difficult that only a few contestants each year out of hundreds of thousands of kids all across the world get gold medals and that's because it requires some really difficult reasoning and kind of complex analysis in the form of a long proof to answer it correctly yeah so before we came on And I had a look at some of the questions that were in this year's Olympiad.

2:22And what's funny is you can choose the language you want it to download it in. But I found it basically the same. Whatever language I looked at, I was like, even in English, I was like, what is this? Still incomprehensible, yeah. Yeah, exactly. These questions are really hard to answer. And actually, it's been really hard for AI to answer it traditionally as well. If you ask ChatGPT to help you on these problems, it will be kind of hopeless at it. And so lots of researchers saw that the IMO was like a really good benchmark to say, if we can use AI to answer these questions, then that will be a sign that somehow the AI is doing some more complex or general reasoning than it's been able to before.

2:59And so when you say general reasoning, that is basically what we're getting at with AGI, artificial general intelligence, right? Yeah. So it's a hotly debated term, AGI. And the companies, OpenAI, Google DeepMind, they use it as this kind of marketing term. There's no good scientific definition for what it means but you can argue that being able to answer these kind of maths questions would give you some sort of more general reasoning than they're able to do at the moment. So to give us an idea of where we are if we're looking at sort of the current models that we're more familiar with how bad are they at these kinds of maths problems what happens if you ask ChatGPT say to solve one of these sorts of problems?

3:34So there's been tests kind of testing as you say ChatGPT or Anthropics Claude or any other kind of large language model and they do pretty universally badly they kind of yes fail to answer any one question correctly uh definitely don't get any sort of medal performance we have seen examples of kind of specialized systems doing well at the imo so last year google deep mind had two models one called alpha geometry another one called alpha proof and together these models managed to answer enough questions to get a silver medal level performance but these were kind of very specialized systems they had elements of large language models like chat gpt but they also had what's called neurosymbolic systems which are basically using kind of symbols and code and they were trained on specific imo problems so they weren't really general in the way that you could just use these to answer any math question you wanted yeah so they were written in a in a special language weren't they uh the programs themselves were written in lean so this proper language that um produces a proof that um i am told it's relatively easy to check the proof but to untrained the untrained eyes but you know it's basically impossible to decipher so it does feel a little bit like cheating when you use a specialized language but this year as you say it's more like an llm has been let loose on the questions exactly and they've done really surprisingly well so we heard last week that both open ai and google deep mind have trained models that are working natural language kind of large language models and they've both got gold medal level performances and the fact that they're in natural language and have done so well at the IMO are things that the companies have really trumpeted as evidence that these systems will be able to reason more generally and more flexibly and that they will get us closer to AI that can solve complex and advanced mathematics as well as just be more generally intelligent which is if you believe in something like AGI as we said that's something that you want these systems to be able to do.

5:32It does remind me a bit of back in the day when Deep Blue, IBM's computer, beat Gary Kasparov at chess. It was obviously a landmark thing that happened then. But even then, at the time, people said it was a bit like cheating because Deep Blue had been really specialised just to play chess. It was a real number cruncher. And it felt like last year's Olympiad, the maths Olympiad, was a bit like that in that sense. But going one better this year and getting gold with a sort of large language model basis does feel like it's gone up a level. So how have they managed to do it? So it's a very good question, and we don't know.

6:14And because we don't know, lots of people are very frustrated. So mathematicians that have seen this result kind of say, yeah, it's impressive, but we can't look behind the scenes. We don't know what exactly it's doing. The information that both OpenAI and Google DeepMind have given us is very, very scant. they've said both of them have said that they've used reinforcement learning in some way to train these models and reinforcement learning if you're not familiar is the system that basically tells the ai what success looks like and then lets it kind of figure out the rules of how to win whether that's a game or something like protein folding and we've seen it deployed with really success with google deep mind for things like alpha fold and alpha go so the fact that we can't find out how these systems are working means that it's really hard to evaluate exactly how much of a deal this is.

7:02I spoke with Terence Tao, who's one of the world's kind of most respected mathematicians and all around very smart guy. And he told me that while it sounds promising, the fact it's not been done in a controlled scientific fashion means he can't really assess whether it's a big deal or not. So it's kind of unclear how much of an AI victory this is. And I guess we can only speculate, but does this just suggest that maybe these AI firms have like brute forced it and just set that particular LLMs like completely focused on this one kind of competition they've just found a way to get them better at that does it really mean anything so it's really hard to answer as you're saying without knowing how this doesn't work but it is worth saying that to answer an IMO question correctly you have to come up with a proof that's at least a page long that proof has to have kind of jumps of logic and reasoning and really construct an argument that we can understand in sort of human terms.

7:55So it's not as simple as just answering a question that gives an answer like 52 or 42. Does it take them four and a half hours to do it as well, or do they just spit the answer out? So they gave the systems four and a half hours of time like the human contestants. And they needed that time? I think so, yeah. So they were real, like, just sat there pondering. Yeah, it would say thinking, exactly. and we also don't know if the systems will run for longer if they might have been able to do even better because they couldn't answer the sixth question which some mathematicians are seeing as evidence of okay well there's still some progress to go and something's not quite working there so did they both struggle to answer the same question?

8:33Neither of them answered the sixth question oh intriguing do we know what it is about that question? it was about combinatorics I am unable to tell you exactly what the question is about to answer that question yeah i think there were about five five kids who got it correct so we've still got the edge on combinatorics great yeah and it's also worth saying that there's real growing distrust around what ai companies say and how that lines up with reality so we've seen multiple examples now of how a company like open ai or google deep mind will come out with a discovery or finding and then when scientists have gone back and looked at that it's failed to hold up to scrutiny so we saw for example last year deep mind released a ai model that claimed to predict inorganic crystal structure and then when scientists went back and saw if these were actually new they found that most of them already existed.

9:20There's also been a bit of chat in the New Scientist newsroom hasn't there about like anytime you set a benchmark you can fully just go for that benchmark but it doesn't necessarily mean that you've nailed the entire piece it just means you've managed to hit that target that you set yourself. Exactly yeah these benchmarks it's really unclear how much they apply to kind of real world use. And also it's a bit frustrating isn't it if they're dangling this advance or this achievement in front of us but we can't use it is there any hint that we're gonna be able to have a look at this soon or is this just about kind of getting investors excited surely not so on the bright side google has said that it will incorporate this model into their gemini system at some point this year open ai have been a little less specific about when we might get to use this closed ai so there is a chance that sooner rather than later we'll actually get to find out whether these models really can stack up to what they appear to be able to do and when i spoke with one of the google researchers he said that he hoped we'd soon be able to use these on real scientific and mathematical problems he even said that maybe soon we'd be able to tackle the reeman hypothesis which is one of these millennium prize problems that's basically impossible to solve so far.

10:34In that interview, there was another deep mine researcher who said to him, that's a bit of a wild claim. Maybe we should not say that and just kind of walk it back to, say, some lower mathematical problems. So there's clearly a difference of opinion in how big, even within these companies, they think this claim is. Okay. And just to come back to the holy grail of AGI, I mean, if we take them at their word, let's just try to for a minute, then you know you could argue that these gold medals are an important step forward but I spoke with a mathematician who's involved in in running them setting up their IMO and he pointed out that this increment that they've got this year from like gold medal threshold you know just under to gold medal isn't actually that impressive for a whole year's work of in in AI and he says it could be more to do with just sort of the vagaries of boundaries that you set on on medal on how to get a medal and the questions um and and as we've been saying no one no ai um solved question six um so that would be the big deal so we do want to celebrate these kids who they are like athletes dropping out of school they have to train you know um and uh and some of them managed to to get question six so go go humans yes one for the win exactly and we've actually seen another win for humanity recently and there was a world coding competition held in japan last week and open ai again entered one of their models it's not clear if it's the same model as the imo model and that model came second there was one human who beat it and he said after three days of very little sleep he was absolutely exhausted but he was kind of proud to be holding the line for humanity go him now time for a word from our sponsor the best science isn't just a laboratory endeavor it changes lives, but that link between researchers and the impact they have on people is often forgotten.

12:25That's the gap filled by Wellcome's podcast When Science Finds a Way. Host Alicia Wainwright is a botanist turned Hollywood actor who invites researchers at the cutting edge to tell their inspiring stories alongside the people in communities around the world who are working with science to improve lives and shape their own futures. Episodes cover topics ranging from cholera prevention in Kenya to wildfire management approaches rooted in indigenous practices. When Science Finds a Way is available now wherever you're listening. This episode is brought to you by Accenture. When your advertising operations fall out of sync, everything else follows.

13:02Spotify and Accenture are working together to reinvent the rhythm of ad sales, using automation, analytics, and smarter workflows to simplify campaign delivery and access better data across the business. The result? Less time spent on operations, more time connecting brands with the moments and fandoms that matter most. Learn more at Accenture.com slash Spotify. Now we're going to link sleep, which we all do, I believe, with mitochondria, which we all have. Yay! Two of my favourite things. Indeed. So famously, mitochondria are... Former bacteria. The powerhouse of the south. Have you not caught this meme.

13:41Yeah, I have not. It's a joke that everybody knows that mitochondria are the power house of the cell. They are. And not much more than that. Right. But what they really are are these specialist little organs essentially inside our cells that do chemical reactions to provide ourselves with energy. And now a study suggests that mitochondria in the brain actually help to trigger sleep when you've been awake too long by building our sleep pressure. Yeah, it's a fascinating story this. And to add, well, it's already brimming with interest, this story, but add, I will say that I'm going to link this very feasibly to Ozzy Osbourne this week.

14:14Okay, okay. But before we get to Ozzy, let's learn more about the study. To do that, we welcome Alexandra Thompson back to the pod. Alex, look, the first thing we should say is this new study, it's on Drosophila, but there are applications or, you know, implications for humans as well, aren't there? Yeah, so this is in fruit flies, which as a general rule, an animal study we take with a bit of a pinch of salt particularly if it's not in mammals or preferably primates but mitochondria are so fundamental to how cells work we can quite confidently make the leap that this could also apply to people.

14:49Okay Alex tell us more. So we have some understanding of how the brain reacts to sleep deprivation. I think emphasis on some because sleep is the biological enigma but we know that there are changes to neuronal firing, the structural shapes inside cells and how genes are expressed. We also know about specific neurons in the brain that switch on when sleep begins, but we're not so sure what tells those neurons to fire. Well, so there are neurons that switch on when you go to sleep. I guess I always thought it was about winding down and, you know, you forget. But I guess the brain is really active, even when you're asleep, isn't it?

15:26It doesn't just turn off completely. No, the brain's very active during sleep. You obviously have much less vivid dreams than me. but yeah it's a time of memory consolidation waste removal processing information and there are several structures that are involved so in people there's the hypothalamus which is sort of a peanut sized structure deep inside the brain that contains groups of neurons that act as control centers affecting sleep and wakefulness partly in response to light and it sort of triggers a chain reaction that involves lots of different parts of the brain now flies don't have a hypothalamus but they do have a brain region called the here we go pars intercerebralis that performs similar function and houses their sleep neurons so to better understand sort of the process of this researchers use sequencing and fluorescent markers to study the genes expressed by sleep center neurons in about 1000 female fruit flies which usually get about 13 to 16 hours of sleep mostly at night.

16:24That's a lot of sleep. So these young flies sleep more than a child, about the same as a baby. Well, yeah, you're happy if a baby sleeps for 12 hours. Yeah, very happy. But babies, you know, they've got massive amounts of brains and they're pruning and all these problems they're doing, you know, figuring out their brains. But a fly has just got this tiny little pinhead of a bunch of neurons. Why do they need that much sleep? I just think it's exhausting being a nuisance. But it seems to be more about quiescence, sort of a period of dormancy rather than what we consider to be sleep. So there's some evidence they sleep more if they've had a busy day, lots of social interactions.

17:05Also, flies that can't fly sleep longer, maybe because sleep helps them adapt to challenging situations. But anyway, back to the study. The team let roughly half the flies get a full night's sleep. the others were kept awake for 12 hours either by gently shaking the tubes they were in or by genetically engineering them so their wake promoting neurons were switched on by a rise in temperature and what they found was among the sleep deprived flies the sleep inducing neurons ramped up the activity of genes involved in running and maintaining their mitochondria the mitochondria also showed signs of being under stress so they broke into smaller pieces they cleared out damaged parts and they formed contact points with nearby structures that help with repairs.

17:50So this stress may stem from the fact that the mitochondria keep producing energy even when the neurons are sort of otherwise inactive. The researchers observed that this led to a build up of electrons that leak out, generate free radicals and ultimately trigger sleep pressure. Wow. This is fascinating because, okay, you know, you said the meme about them being, we think of mitochondria just as these batteries or power packs of the cell but i think that obscures the what incredible things they are so you know across the membrane they they're pumping protons across the membrane multiple times like thousands of times a second and it generates over just a few nanometers enough basically about a bolt of lightning's worth of electricity and that power the mitochondria harness that to make ATP the universal energy currency of the cell and that's just going on all the time so you can imagine with that much you know electrical charge going on across a membrane how stressed and damaged they can get over their lifetimes.

18:55Yeah and when the flies were finally allowed to sleep the mitochondrial damage was repaired. The researchers also found that flies with fragmented mitochondria in their sleep neurons slept less than normal and they didn't catch up on it after being kept awake. By contrast, flies whose mitochondria were engineered to fuse more readily, which suggests better repair mechanisms, they slept more than usual and showed a stronger rebound after that deprivation. So this all sort of supports the idea that mitochondria is involved in the pressure to sleep. We've not cracked the code of sleep. It's very complex, but it gives us a clearer picture.

19:28So this is amazing, isn't it? Because it makes me think of we already knew that in people, if you have mitochondria that aren't working very well you're more likely to feel tired and fatigued and that makes sense you're producing less energy but this idea is that the mitochondria in your sleep neurons are directly involved in the sensation of sleep pressure building that feeling of sleep sleepiness and it just makes you think what else are mitochondria at the core of yeah well this is it they are at the core of lots of other things it's really and and when you stop to think about it then why wouldn't they be because they are at the core of what we are as eukaryotes meaning you know mushrooms trees dolphins us any anything any complex life form we owe it to them yeah because mitochondria of course were once free living bacteria about two billion years ago they were engulfed by another life form a single-celled archaea and this led to the evolution of eukaryotes that's all complex life like you mentioned plants fungi animals yeah Yeah, so yes, exactly.

20:28Perhaps no surprise they have this kind of massive impact. Yeah, and just to add to the impressiveness, there's another part of the experiment where flies were engineered to have raised mitochondrial activity in response to light, and the team found that one hour of artificial light caused sleep duration to rise by as much as 25 % compared with control flies. So yeah, they investigated the brains of flies. They're not people, but as you say, mitochondria are in all complex things. And it also supports the idea that aerobic metabolism which is the production of energy from nutrients and oxygen, which takes place in the mitochondria of most animals, that can drive sleep pressure.

21:04And, you know, always looking for the news you can use. Does this tell us anything about how we can get more sleep, get better sleep? I don't know if it tells us how we can get better sleep right now. I think hopefully just improving your understanding of anything leads to better treatments. And sleep conditions are really debilitating. They're also very common. So one particular person we spoke to said, this provides us with novel opportunities to target these pathways and come up with new efficacious ways to treat people who have sleep problems. And also the team used various ways to keep the flies awake.

21:35So, you know, gene editing via temperature changes that were normal for the fly and non-stressful for them. And they all had the same effects on mitochondria. So I think it's quite compelling evidence here. Actually, another person said that what the study has revealed is the sleep homeostat is actually looking at its own mitochondria to estimate the need for sleep. So what's the link then to Ozzy Osbourne who sadly died this week in 76? Ozzy had Parkinson's disease. I think that was announced in 2020 announced it. It obviously had it for a long time but he had a particular form of it that's linked to this gene called PRKN Parkin.

22:13It produces a protein called Parkin and that gene is involved in maintaining mitochondrial function. So it's not a simple disease at all. And there's no simple link like there is in a few diseases between, you know, one broken gene and the disease. I mean, Parkinson's has got many complex features, but we know more than 200 mutations in this Parkinson's gene that have been linked to Parkinson's. So it does really seem like the mitochondria has a big role to play, certainly in the form of it that Ozzy Osbourne had. It's so important that we continue to look into that because it's the fastest growing neurological condition in the world.

22:50Some people are even talking about Parkinson's pandemic. And a lot of the research I come across really focuses on the environmental potential causes, particularly pesticides. There was a study recently linking how close you live to a golf course to a high risk of Parkinson's because they've got their sprain on it. And that's all important to study, but we mustn't overlook the genetic element either. all right hold on to your hats octopus fans of which there are many octopuses hold on to your metaphorical hat anyway the reason you're holding on is because this week we had the brilliant news that an iconic psychology experiment works with octopuses too and this is the famous rubber hand illusion where you trick someone into thinking a fake hand is their own wow love it um but it works with a tentacle that is what we've now that is what we now know.

23:39So for those not familiar, this was first done in the 1990s in people. It's this experiment that showed that you can be fooled into thinking that a fake hand, a rubber hand is your own. I've always wanted to experience it, but never have. But you have, haven't you? Yeah, yeah. Have you done it? Yeah, I've done it. It's great. It really messes your mind. So what you do is, yeah, you're sat at a table and your arm goes under a curtain or a box and next to it they put a rubber arm that's you know matched to your skin tone that you can see yeah you can't see your arm but you see this fake and what happens is um someone strokes the fake hand with a brush at the same time that they stroke your real arm that's hidden away and after a little while you you get this weird transference thing going on in your brain where you you understand that the fake hand you're looking at is your real hand and it's really bizarre isn't it It is, and it has really important implications for helping people with amputations recognise their limbs.

24:36Yeah, because it's not just a party trick. People have also looked into it like, can we use this kind of illusion to make us feel like we embody robots, for example? It can also make you feel like you've teleported into another body. There have been loads of variations on it. And now they've done it in octopuses. So our reporter, James Woodford, he says that we already knew that some other animals can be fooled by this rubber hand illusion. Wow. like mammals so mice for example but now a team in japan have found that you can also trick octopuses this way too they carried out the experiment with captive octopuses that were each given a fake arm or tentacle and made out of a soft gel and then they used like an opaque partition to block their view of one of their real arms yeah so then they did basically what you guys have done they used plastic calipers to stroke both the fake and the hidden real arm and then they moved to pinch the fake arm with tweezers.

25:30So it's a bit like sometimes in the rubber hand, they make you feel like it's your hand and then they bash it with a hammer and you freak out even though you're not actually in pain. And the octopuses did similar. When the arm was sort of aggressively jabbed, the fake arm was aggressively jabbed. The octopuses then kind of reacted with a variety of different responses. It doesn't say anything about ink, but some of them changed colour, some of them rapidly retracted their arm and some of them just got out of there. I have to applaud the biologists here for what they've done. Can you imagine trying to set this experiment up?

26:02I mean, it must be bad enough with a mouse almost. I can't imagine how they've done it. But an octopus, you've got to do it underwater and get an octopus to leave its tentacle somewhere. Yeah. While you put a fake tentacle. How did they manage to do that? So how did they know, though, that the octopus thought that the fake arm that was embodied had embodied the fake arm? Yeah, so they did various controls. So they did the same test without first doing this kind of important step of simultaneously stroking both the fake one and then it didn't work. Wow. It's amazing, isn't it? It just really suggests that, well, that octopuses have a body image like we do.

26:45Yeah, this feeling of embodiment. Yeah. And I guess you think of how good they are at adapting themselves to their environment, like camouflaging themselves, they can do and they can mimic a whole range of different things like in shape as well as colour. I always think of the floating coconut one. Well, they've got so they have an amazing self-awareness. Perhaps it's not surprising that they can do that. Yes, although I was a bit surprised because we always talk about them having in their tentacles these mini brains. right they have this sort of uh more distributed nervous system so i thought perhaps if they have these mini brains in each of their tentacles they'd be less susceptible to an illusion like this yeah i thought exactly the same and then i guess i think maybe you know the central brain has the eyes and that's what's being fooled by the illusion in this point and and so you know the central brain is still is still the central brain and and has to process this illusion and So maybe that's why it happens.

27:42Yeah. But amazing stuff. That's all for this week. You've been listening to the world, the universe and us. Do go ahead and subscribe. Spread the word. Thanks to all our guests and thanks to you for listening. See you next week. Bye. This episode was sponsored by Wellcome's podcast, When Science Finds a Way, available on all podcast platforms. This episode is brought to you by ChatGPT. Hey, it's Bill Simmons from the Bill Simmons podcast. Have you guys heard about ChatGPT work? It's the new way to use ChatGPT for bigger multi-step projects. And when you need more than just answers, give ChatGPT work access to your apps and files, and it can create real work documents like spreadsheets, slides, and structured reports.

28:27Get started at ChatGPT.com by selecting work mode available on Plus and Pro plans.