00:00:04: The Biorevolution Podcast.

00:00:06: Your hosts,

00:00:07: Luise von Stechhoff

00:00:08: and Andreas Roichler Welcome to the BioreVolution podcast this time.

00:00:15: AI-The Brain and the Biology of Intelligence.

00:00:19: Hello Izzy!

00:00:20: Do we know a lot about human intelligence?

00:00:24: I

00:00:24: think

00:00:26: we know it when you see it.

00:00:28: We KNOW IT WHEN WE SEE

00:00:29: IT.

00:00:35: So...you brought quotes as always.

00:00:38: for starters

00:00:39: One is from Sam Altman.

00:00:41: He's the CEO of OpenAI, and I think we quoted him a couple times already when you talk about AI because he says there are lot things about AI And he says We see our future with intelligence as utility like electricity or water?

00:00:55: People buy it for us on meter.

00:00:57: Oh my god!

00:00:58: Yeah

00:00:59: Okay It's scary.

00:01:00: Maybe.

00:01:02: Another one is from the Google CEO, Sunir Pichai.

00:01:06: And I think he said that on the Alex Riedman podcast and he talks about something called artificial jagged intelligence.

00:01:13: That's something we'll dive into in this episode as well.

00:01:17: It sometimes feels like you see what these models can do and then you can trivially find them making a numerical error or counting roses seemingly super intelligent capacity of these models and the idiocy on the other side that baffles us here.

00:01:39: And I want to finish with a third quote, which is from the physicist Richard Feynman.

00:01:46: scientific founder of many smart ideas across fields, even though he was I think mostly shaping the field of physics quantum physics and what he apparently wrote on a blackboard in The Day Of His Death or it was written there.

00:02:02: What can not create?

00:02:05: This is really telling i think.

00:02:07: let's dive into intelligence.

00:02:10: while we know about that how we look at intelligence in the first place, because it's very clear that in order to create artificial intelligence.

00:02:21: We have to know about the human intelligence.

00:02:24: so How do we look?

00:02:25: At intelligence?

00:02:26: what are The ways to tackle intelligence?

00:02:29: or Do we Have To Know So much About It?

00:02:31: That's I think a Key Question In This Field.

00:02:34: Yeah!

00:02:35: How Do We Define Intelligence?

00:02:36: i mean thats What I Started With Right Here.

00:02:38: know it when you see how would I mean.

00:02:40: everyone i think has a feeling of what intelligence means.

00:02:43: What do you say is intelligence for your?

00:02:45: if u said this person is

00:02:47: really hard to save the people from me with this?

00:02:50: absolutely and i wouldn't have quick answer quite frankly so.

00:02:56: i think from cognitive science and from psychology there are certain definitions that come up Very frequently, but there is not this one definition of like the one principle that everyone can agree upon.

00:03:09: That is intelligence.

00:03:10: But I think they are a few things we do agree would classify A person or maybe system Or an animal as intelligent which for example problem solving Which learning reasoning, all these kind of capacities and many more would be counted under the umbrella of intelligence.

00:03:31: For human intelligence I think there are also different schools of thought into is they're actually one intelligence which sometimes it's called like the factor G that this underlying intelligence shapes types of cognitive ability that we would classify as intelligent.

00:03:50: Another concept is there are actually multiple intelligences, which fall into different classes of capacity like musical intelligence, mathematical intelligence and many other specializations – let's put it this way – which is something that Howard Gardner postulated in his idea influential in education and something that I think, don't know if it's still there prevalent in the school system today.

00:04:19: But i remember when I was In primary school but also high school That was definitely a thing.

00:04:25: people were being classified by teachers By their peers as someone who is mathematically intelligent And this is someone who musically intelligent Right!

00:04:38: cognitive science standpoint, this is actually kind of refuted.

00:04:41: This theory that these are independent factors of intelligence.

00:04:46: I think intelligence often connected across different domains and i think like a famous example would be Einstein who was definitely very mathematically intelligent doesn't mean that if you're good at one thing, your necessarily bad.

00:05:01: At another things even though I mean there are definitely people who excel in a certain area and not as intelligent than others.

00:05:08: but generally different domains of intelligence is connected

00:05:12: Looks like inclusive intelligence

00:05:17: And an interesting part is that there's for human intelligence, at least also this idea of embodied intelligence.

00:05:24: That might also spread across different people like kind of crowdsourcing of intelligence again?

00:05:30: This is something that is debated across I think psychological science and also cognitive science but they seem to be at least for primate evolution a great benefit or social intelligence.

00:05:41: so working in groups solving problems together has brought us to excel at different tasks and outperform many other species.

00:05:51: It's put it mildly and friendly like that, so say a little bit worse.

00:05:55: but the social intelligence of course also means we share intelligence across individuals in group.

00:06:02: doesn't again mean We have like a clearer, you know?

00:06:05: Like everyone has their own domain of intelligence.

00:06:08: Not on the Smurfs where everybody has its own specialization and can only do that one thing.

00:06:14: but I still think there is definitely at least co-evolutionary component of intelligence across group because also symbolic nature of human intelligence and I think that is something that makes us quite unique.

00:06:29: so the language but also.

00:06:31: The symbolic thought that allows as to share intelligence problem solving learning across people within the group and across generations, So actually our intelligence has become storable in a way that it has not been any type of organism before?

00:06:52: sums up there are many different ways to think about human intelligence and when we think about artificial intelligence, We might have to think of some other things.

00:07:19: The expression Artificial Intelligence I Think we've been talking that before in the other episodes has been around for a while.

00:07:27: but i guess In order To get both together like Human Intelligence, Artificial intelligence And what is possible along the way in the pipeline?

00:07:38: to discuss the history of artificial intelligence a little bit?

00:07:41: Yeah, at least I think what would be important is to understand What were people actually aiming for when they developed this so-called Artificial Intelligence.

00:07:52: This idea of trying to build artificial life The Golem.

00:07:56: there's examples in that already of robots being built

00:08:01: Two and half thousand years old.

00:08:02: So

00:08:04: i mean There are definitely thoughts going around of building some type of artificial life for a long time already.

00:08:11: It seems to be fascinating to humans and it seems to me very something that we care about, I think part why we care are two things – this is also what the field of Artificial Intelligence is going for.

00:08:24: one as we want to understand ourselves better through an artificially created intelligent being.

00:08:32: The other thing A smart robot, a smart machine.

00:08:36: A smart servant who can take over tasks from us and I think both are equally valid.

00:08:42: And the term artificial intelligence was coined in The Proposal for the Dartmouth Conference From John McCarthy in fifty-five, nineteen fifty five where he proposed For the first time this term Artificial Intelligence.

00:08:56: done that.

00:08:57: But this was the time when it got this catchy name, I think he himself wasn't super happy with it but said we need to call it something.

00:09:06: and i think back then there was at least...i just listened to one of the great courses..and there was this idea proposed ..and i think that's very nicely done.

00:09:14: ...of this dual nature how to approach artificial intelligence.

00:09:20: school of scientists who want to understand how the human brain works through artificial neural networks.

00:09:27: And the other one would just wants to build smart machines, who cares why they're smart as long as they can do their job and help us out?

00:09:35: I think both coexisted for a while in that just built smart machines In this sense off.

00:09:41: there's so called good old fashioned AI symbolic AI or not cold go fi One out for awhile until There was a slope of disillusionment in the AI winter.

00:09:52: And then after some time, the neural nets took over and these are modeled very loosely based on how the human brain works.

00:10:02: but I think there is quite significant differences.

00:10:05: Interestingly when we look at modern neural nets they do work like AI maybe in similar fashion as humans brains or human neural nets.

00:10:19: But that said, in the end we don't know how they derive any of their conclusions and your decisions because it's all a black box.

00:10:27: so basically what other branch wanted is this smart machine doesn't really tell us anything about our brain Exactly.

00:10:35: Seventy-one years fast forward from the coining of the expression and we still, in a sense even though... Of course we see majestic steps going forward into development artificial intelligence first and foremost due to storage computer capacity and lightning speed.

00:10:55: which mathematical transactions can be processed today?

00:11:00: Yeah absolutely but You were referring to this expression in the beginning of this show, The Jagged Frontier.

00:11:07: What are we dealing with here?

00:11:10: Yeah so I think that is that We have this expectation That a smart AI Is smart at everything

00:11:18: Or

00:11:19: which it is not or that It's smart In a similar way than a humanist.

00:11:23: So it has This more rounded type Of intelligence like perfect at most things, but we're pretty good at most thing.

00:11:34: So that's like this where good generalists could problem solvers and AI is often very good at a very unique specified task, but it's not a good generalist.

00:11:47: And if the AI was trained to play chess then It might not be very good playing.

00:11:53: I don't know Tetris or something.

00:11:55: The modern large language models give assemblance of being good everything.

00:12:00: They are also not good at anything that requires world knowledge.

00:12:04: And this is something these models lack, so they're trained based on large data amounts and then kind of draw their conclusions from what the world looks like.

00:12:15: So very famous Alex Knapp in two thousand twelve was trained on cat pictures for internet and could recognize a cat.

00:12:22: I mean there's huge success.

00:12:23: somehow it got cat knowledge.

00:12:26: But does it know what a cat is?

00:12:28: probably not.

00:12:29: Probably, not exactly.

00:12:30: and this the problem because model has dissemblance that has been trained on What could look like but doesn't actually understand rules of world in same way human or another biological organism does.

00:12:48: I think there are different reasons for that can lead to this so-called jagged intelligence and jagged frontier.

00:12:54: That means that while the models excel at certain tasks, they are much better than humans for example... I mean we have these famous deep mind examples of AlphaGo where the AlphaGo played a move no human had ever made before because it could understand rules in different ways or explore the option space on a different way.

00:13:15: We have alpha fold with protein folding problem in my opinion is not solved because we still don't know how proteins fold, but we can approximate.

00:13:24: How they do it in an excellent way?

00:13:26: Because this huge space of possibilities the AI can scan through at much faster than any human ever could and give us approximations on all that.

00:13:36: good work.

00:13:37: But at the same time, if it's about these and that was the example that we heard from Google CEO of distinguishing roses in strawberries or we have this famous example off techie PT which model?

00:13:48: It was but one of the large language models asked for advice.

00:13:51: If you go to the car watch should you take the car?

00:13:53: or going foot and it says like yeah just walk is better for the environment are seeing the night sky from the basement and it doesn't understand that there's a ceiling in between because I think it lacks these fundamentals of world knowledge.

00:14:06: And, of course, it gets closer to simulating that but that it can, at the current stage not solve because it does not understand.

00:14:18: And I think anyone for example who tries to use AI in a business context and wants to make for example PowerPoint slides which should be easy and there are many solutions that promise that it does that well but It's basically not usable Because it cannot...it doesn't really makes this next step of understanding what really The Next Level Of Complexity Should Look Like is something that we could consider the jagged frontier, which not per se means that AI is extremely useful.

00:14:48: We just need to understand where can pointed at this current stage and I mean these are models of today.

00:14:54: there are new types of models in development That might be better or maybe other things Or being taught specific world knowledge.

00:15:03: For example for math.

00:15:04: some have been more specifically trained on mathematical data And also get better at solving mathematical equations based on that, but do they understand math?

00:15:14: Probably

00:15:14: not.

00:15:25: It's like a mix of genius when I think about the goal story which is somewhat old by now... But also at the same time i see if i think about The Car Wash and many other examples i think of helpful idiots even if i do general prompting Friendly and so flat, it depends on what you put into the machine.

00:15:53: I understand that but at the same time It appears like its still early stages And maybe one aspect of this is really understanding Of human intelligence.

00:16:07: The basic question or fascinating question for me Is also why we know so much as humans?

00:16:14: Why's it hard to understand Human Intelligence?

00:16:18: That's a very good question.

00:16:19: But I just to the point that you had, I think also these models are quite different ones.

00:16:23: so i think goal model was trained very specifically on the rules of go.

00:16:27: where is large language models or as you say like helpful idiots that our goods at little bit off everything and I mean this training this persona that they impersonate?

00:16:37: This really pure...that what developers put in reflection of human need for subservience maybe by the model because it could be more controversial them on, could have different very different types of personas that yeah we don't have.

00:16:52: Why is it so difficult to understand human intelligence?

00:16:55: I think the problem is that we are trying to study our brain with our brain.

00:17:00: That is like understanding ourselves as they're difficult and i think intelligence has only one part off picture.

00:17:07: So what we now discuss?

00:17:09: this really a thing type of intelligence than you would consider residing in our prefrontal cortex like there's cognitive problem-solving intelligence, this higher function.

00:17:21: And logical

00:17:22: cold and yes but also the creativity I would say it was out there.

00:17:26: so any like kind of what we would consider maybe higher functioning kind of cognitive ability?

00:17:32: But I mean This is only a very small part of What makes us human or what makes as function at every point Of The day.

00:17:40: There's many more processes ongoing that are below the level of consciousness, so below all what transmits into our brains.

00:17:49: And these processes also matter and they matter tremendously for example to decision making or moods and feelings.

00:17:58: this is also the rest body sensory motor systems playing a role of course transmitted through brain but many things will never make it level understanding, but still will.

00:18:15: for example I mean there are these famous studies from the nineteen sixties that show that we often make decisions before we consciously start.

00:18:24: Before we know that We made a decision.

00:18:26: and this is probably not true For any type of decision.

00:18:29: But four at least let's say i want to move This arm?

00:18:42: prefrontal cortex will tell us the story afterwards that we made the decision, but we had made a decision before already.

00:18:48: This of course tells us what we consider our intelligence is also a little bit storytelling by the brain.

00:18:54: That makes sense in our surrounding and of course We have very filtered view on reality which we need because otherwise it would be very stressed if all things going around us were constantly bombard or system so filter out a lot of thought, but we filter our sensory expression.

00:19:12: We are not aware every movement that makes in conscious way and which is very nice because it would be stressful to do all this.

00:19:20: so I think the reality that we're perceiving as a filtered one also true for intelligence.

00:19:30: what we consider intelligence is small part Which also means that whatever we're trying to replicate with the AI is only a very small part of what goes into our functioning and into our so-called world knowledge.

00:19:44: So why do we understand how The World Works?

00:19:46: Because, We Also have a somatic level of understanding Of How The World works.

00:19:50: That might not be easy To encode Into an AI system.

00:20:07: I will love to hear a little bit about what in your view are the limits.

00:20:14: I mean, we're also talking but by revolution and use cases in pharma development drug development for instance and all that where AI even if it's the dumb model as I might call it still is so helpful because does all the math in lightning speed.

00:20:32: But of course talk-of-the-town these days is the general AI.

00:20:37: and after what you just said, you know intelligence is not playing in the frontal cortex solely or exclusively but in other parts of our bodies as well!

00:20:48: And we don't even really recognize it.

00:20:50: so that puts limits to a I in comparison to real or human intelligence, even now where we have so many landmark developments happening as we speak maybe.

00:21:07: I would say so biological intelligence functions differently from electronic intelligence because first of all electronic intelligence artificial neural nets are simple.

00:21:20: I mean, they are purely digital.

00:21:21: there are ones and zeros but we're much more right?

00:21:25: We have analog layers And we have digital layers in our neural processing which already give a much more complex way of signal transmission.

00:21:34: We have more parallelism or synergies on our neural net.

00:21:38: so just from the get go i think that the system is designed In such A different Way That drawing conclusion From one to The other Is Just it doesn't make

00:21:47: too much sense.

00:21:48: Can we think of a combination?

00:21:50: That is one thing, I think that's really interesting parts.

00:21:53: biological AI so taking indeed the set of neurons for example as basis.

00:21:59: artificial intelligence system because you can also do coding via biological computers.

00:22:06: I don't think that necessarily the AGI has to resemble human intelligence, because we're so focused on ourselves very limited in a way where you see intelligence.

00:22:24: the helpful idiot persona of JGPT.

00:22:27: Yeah, but I think this is a reflection purely on human desire for what intelligence should look like rather than anything that AI does.

00:22:36: The AI doesn't care and it would probably if you leave it alone speak in different language pretty fast or in abbreviations?

00:22:44: Or make up funny chatbots who were left alone quickly started to misuse English?

00:22:53: Well, the AI anyways doesn't want anything.

00:22:55: I'm anthropomorphizing it now as one does because It's impossible not to...I think that what we will see is or What i consider possible?

00:23:06: Is that AGI Will just look very different from human intelligence but That it can still surpass us and We Just don't know what it would Look like yet.

00:23:14: And I Think if we..it's The same question of looking for life in space I definitely think its possible that Its there But it will Not look probably will not have a humanoid shape and

00:23:26: be

00:23:27: humans in green.

00:23:28: And I think there are so many different ways that life, intelligence can look like... We know from the animal kingdom that intelligence has many different way of coming to pass.

00:23:38: it's not what we would maybe in previous definition call intelligence but also animals which aren't primates or whales can be intelligent and can be intelligence in groups like ants, like bees for example.

00:23:53: And you could consider that even a group of bacteria like our microbiome displace some kind of group intelligence the way they interact or shape their environment to respond to challenges by antibiotics.

00:24:07: So I can imagine that maybe the intelligence of artificial intelligence, if left alone would make me look more like bacterial intelligence.

00:24:18: If we can call it that way.

00:24:19: so now for a lot of things that i'm saying and think is very clear to our audience as pure speculation.

00:24:26: but this really interesting to speculate in keeping an open mind how AGI could be quite important because know there are also risks.

00:24:36: I

00:24:38: was just about to say it's extremely fascinating, of course.

00:24:41: It is quite scary at the same time that much about speculation...I think exactly the right times to speculate because we're very close to those developments taking place really.

00:24:53: and look at LLMs have been in place like what three years

00:24:58: now?

00:24:58: That's true!

00:24:59: But yeah.

00:25:01: but i don't think if AGI comes it will not come from this transformer based networks pretty convinced of, I think it will come from something different.

00:25:10: That said we were very surprised by chat GPT emerging right.

00:25:13: so whatever would come next?

00:25:15: We'll also be totally surprised and then hopefully there's a positive surprise.

00:25:39: conundrum we're in is obviously that.

00:25:41: We are telling ourselves one story over and over again, you explained why this is really fascinating?

00:25:48: That we are bound in a sense and were not free radicals if your will right.

00:25:55: And there's one path thats for some scientist appears even promising the use of psychedelics going beyond yourself.

00:26:05: And that might prove helpful in the current developments we're talking about right here?

00:26:09: Yeah, absolutely!

00:26:10: We talked about this in his previous episode with Galina.

00:26:13: it's a very nice really recommend listening to if you haven't done so because I think she explains it very well.

00:26:19: just taking a step back to look at your own consciousness basically... That psychedelics seem to allow to unwire maybe the storytelling of your brain a little bit, like this constant making sense that you're brain does.

00:26:36: That it cannot.

00:26:38: as you say You can not step out off.

00:26:40: It may be psychedelics give you The opportunity To take This step back and just A Little bit metaphorical.

00:26:47: what I'm saying is Not actually Like But Take Step Back And Look at Your Own Mind And I think this is something that can be really helpful.

00:26:57: This why it's being tested for many psychiatric disorders, but its score also a super interesting instrument of studying the brain.

00:27:05: and here comes back an intersection to artificial intelligence because artificial intelligence helps us even though they don't resemble our brains very well.

00:27:15: we make use high-dimensional, option space scanning analysis capacities of AI to actually feed in all this different data that we have about our brain.

00:27:26: and what experience from calcium imaging videos vocalizations of rats.

00:27:34: All these different types layers... To understand how does thinking end?

00:27:42: processing work in the brain, how does that translate into what we would consider intelligence consciousness?

00:27:47: Consciousness is a really hard problem.

00:27:50: Even harder than intelligence I guess!

00:27:52: Very very difficult to make sense of because where does it come from?

00:27:57: We do not know.

00:27:58: and when asked Descartes then he will have an answer but its tricky one i have to say consciousness.

00:28:04: But at least this combination of psychedelics and AI looking at brain states, looking at outputs like trip experiences people writing down what they experience and then integrating all of that over AI can really be helpful I think to get an understanding of the basic processes in our brain and also understand pathological conditions like for example, depression ruminating disorders anxiety all these kinds of things.

00:28:30: And I think there are two quite interesting routes i think also where AI and psychedelics intersect.

00:28:35: again one is just making a high exactly.

00:28:40: There was recent paper that tried to trip AI so they use different language models did psychedelic prompting.

00:28:48: So no person was high in the study, just the LLM which already hallucinates enough.

00:28:53: but yeah.

00:28:53: so then they told the Llm to act like it was on psychedelics and it was actually quite resembling in different capacities two human trip experience including this factor of mystical experiences feeling off on us with a world?

00:29:08: I mean that might seem a little bit ludicrous to do.

00:29:12: why would you want AI to trip?

00:29:14: But I mean, it can help you of course to simulate like these experiences and doing psychedelic research in people is offcourse a tricky thing because there are risks.

00:29:24: And do also

00:29:25: better.

00:29:26: I mean that's a lot of effort and you i mean if your tries new setting and you compound on your dose whatever me can be quite difficult.

00:29:33: so having tool that can already give you some hints of how it might work and, of course be very interesting.

00:29:40: Actually the same was I think also done with simulating brain states like visual hallucinations or... The other thing in that i find very very interesting is using AI related virtual augmented reality to simulate trip experiences for people.

00:29:56: so augment non hallucinogenic compounds into something experience or at least something that resembles the experience enough to get a positive effect, but maybe with fewer risks.

00:30:09: And that could of course make it also really interesting from medical and from a commercial perspective because I think there is a lot of promise in psychedelics for many psychiatric disorders But its very tricky too handle.

00:30:23: give someone LSD or a silo cyber and say, go home have fun deal with it.

00:30:29: It's not possible.

00:30:30: so you need like a good supervision You need Like A Good integration And if part of that could be done in the VR setting It could be super interesting.

00:30:40: Absolutely!

00:30:42: And would have a very low threshold of therapeutic modality for someone who is, for example dealing with acute anxiety and can get the positive effects.

00:30:52: that it's really, really interesting.

00:30:54: I mean this was at an early stage what i'm saying.

00:30:56: but... That something where you see a lot of problems.

00:30:58: Are there actually companies?

00:31:00: that are working on these kind scenarios

00:31:03: ?

00:31:03: I know they're academic groups and think their also companies But not able to name one the top of my head now, but I know that there are some academic papers just came out and i think they must be companies working on it.

00:31:30: What do you think?

00:31:31: Where will we stand as always in future questions like one or two years maybe in a stricter sense by revolution where were going to stand when it comes general AI for instance?

00:31:44: what will machines telling us bullshit.

00:31:49: That's quite honest, right?

00:31:52: No I don't believe that you really mean it.

00:31:54: no but partly they do.

00:31:56: i mean yeah and with thats the jagged intelligence.

00:31:58: so we need to figure out when they tell us bullshit.

00:32:00: and if its more work what is the bullshit?

00:32:03: oh your saying?

00:32:04: We have to figure as humans of course What the machine is telling even though there a likelihood that we dont understand them anymore exactly.

00:32:14: Thats gonna be tough

00:32:15: yes But so I do not believe that in the next two or five years, we will have a GI and if I'm proven wrong then that is fine.

00:32:22: And then I will report to my machine overlords and say that I am sorry that I mis-underestimated them but no i don't think it would look like what It's a bit like this flying cars metaphor, you know.

00:32:38: Like from the old sci-fi movies.

00:32:40: that doesn't seem to go away even though it really I mean.

00:32:43: why would we have flying cars in the future?

00:32:45: Really Doesn't make sense.

00:32:46: and then there will be drones or whatever.

00:32:48: but i think This is human like AGI That Is Even Maybe A Humanoid Robot.

00:32:53: I don't Think That This Will Be Anywhere Close And I Think maybe it Will Due To This Problem Of Biological Intelligence Not Be Possible.

00:33:02: That Said where I see the future is indeed like intersection between biology and AI to have biological computing, and have brain machine interfaces that for example have speech decoding.

00:33:16: Have these kinds of features.

00:33:18: so i think there's a lot of potential.

00:33:21: were human and Machine Intelligence can merge.

00:33:24: And this makes more sense than betting on the machine just replicating human intelligence, even though it doesn't have the same parts as we do.

00:33:34: That said I think there could be a way that it just develops very different type of intelligence than we can not imagine what would look like.

00:33:43: and then you need to deal with whatever will look like.

00:33:53: AI The Brain and the biology of Intelligence.

00:33:57: this was the by revolution podcast for today.

00:34:00: many things easy

00:34:01: thankyou

00:34:02: please.

00:34:02: Comment on ScienceTales.com, give us a like!

00:34:07: Encourage to cover one topic or the other in the field of the Biorevolution.

00:34:12: we gladly take a look at it.

00:34:15: See you around next time Thank You.