Latest / The Biorevolution Podcast / The good, the bad and the overhyped - 4 years of biomedical AI on the BioRevolution podcast

Transcript

The good, the bad and the overhyped - 4 years of biomedical AI on the BioRevolution podcast · The Biorevolution Podcast

• 0:04
  Speaker 1: The Bio Revolution podcast.
• 0:06
  Speaker 0: Your hosts,
• 0:07
  Speaker 1: Luise von Stecho
• 0:08
  Speaker 0: and Andreas Roichler.
• 0:11
  Speaker 0: Welcome to the Bio Revolution Podcast.
• 0:13
  Speaker 0: This is not an ordinary episode But we're approaching four years in the field of the Bio-Revolution podcast And our fiftieth Episode.
• 0:23
  Speaker 0: Congratulations Izzy!
• 0:25
  Speaker 1: Congratulations To you I think back in the day.
• 0:28
  Speaker 1: So we started sitting on your kitchen table discussing for hours about biorevolution topics, and then we said like maybe make a podcast?
• 0:39
  Speaker 1: And I think both of us thought it would run for five episodes!
• 0:50
  Speaker 1: I didn't have these high expectations...I just thought let's try that could be cool…and now were at fifty episodes which is actually pretty nice.
• 1:00
  Speaker 0: It's just amazing and my learning curve has been really steep.
• 1:05
  Speaker 0: And I am so thankful for really exploring topics that, for me as a journalist and podcaster are close shop in real life?
• 1:17
  Speaker 1: For me it is still kind of the same because there're lot of topics we discuss also with our guests.
• 1:22
  Speaker 1: I knew about them superficially, but then really diving into it.
• 1:28
  Speaker 1: Yeah was a great experience and learning so much pretty nice
• 1:31
  Speaker 0: And we should give huge thank you out to our listeners of course because You've made this possible Because We saw that in our statistics That the growing number Of people who were interested In what we had To offer.
• 1:48
  Speaker 0: That of course Made us thrive and go on basically right?
• 1:53
  Speaker 1: Yeah, absolutely.
• 1:54
  Speaker 1: I think if no one would listen we'd probably stop doing it.
• 1:57
  Speaker 1: It's like they're doing art for arts sake but also nice.
• 2:00
  Speaker 1: If someone wants to look at or listen
• 2:03
  Speaker 0: Right now i'm pretty sure that will make it to episode hundred.
• 2:07
  Speaker 0: What do you think?
• 2:08
  Speaker 1: Absolutely You just have to open the biotech news and like crazy stuff is happening every day.
• 2:15
  Speaker 1: There are so many other things that I still want to talk about from astrobiology, too.
• 2:22
  Speaker 1: Computing based on brain cells through all these kind of crazies thinks that our happening and could be happening And we will have easily topics to fill a hundred or many more episodes.
• 2:53
  Speaker 1: What we want to dive in a little bit today is artificial intelligence.
• 2:57
  Speaker 1: And I think this is especially an area where the movement is kind of overtaking, the speed of development.
• 3:03
  Speaker 1: it's really hard to follow because when you started this podcast and our first AI episode was pre-chat GPT and then language models transform architectures they changed... They didn't only change biotech related topics but actually I would say society in a sense, and i think it's something that is noticeable across so many areas.
• 3:29
  Speaker 1: Yeah we're kind of like witnessing the revolution in real time ,and were
• 3:36
  Speaker 0: talking about it .
• 3:36
  Speaker 0: And its kinda cool!
• 3:39
  Speaker 0: have under the microscope a topic that is developing as we speak and things are changing.
• 3:46
  Speaker 0: We actually don't know what will be AI truth in a month or two weeks probably because new thing keep popping up, so it's interesting.
• 3:59
  Speaker 0: Put the title on today's episode.
• 4:01
  Speaker 0: that will be out of the ordinary and not the usual thing because we have a lot of guests.
• 4:07
  Speaker 0: We will recapture what they said.
• 4:10
  Speaker 0: you titled good, bad in the overhyped four years off bio medical I on the by revolution podcast and infact we had more than a dozen guests who were speaking to topic which is kind amazing speaks for itself
• 4:27
  Speaker 1: Absolutely.
• 4:28
  Speaker 1: And we had like, I mean more this broader over all discussions.
• 4:32
  Speaker 1: but We have also a number of examples Of people just showing really cool applications of AI which i think is Also very nice in This field where often and i Think this Is something that we noticed In the beginning there's A lot of hype and There are so many People who Have Assemblance of knowing what they're talking about.
• 4:52
  Speaker 1: But if you dig a little bit deeper You can tell That theres buzzwords around it and there's very little substance i think especially in the biotech field.
• 5:05
  Speaker 1: In the beginning felt a lot like that.
• 5:06
  Speaker 1: so you had this really overhyped situation where everything was going to revolutionize drug development changing drastically getting faster with higher probability of success, And this still has not happened.
• 5:24
  Speaker 1: That said, I think there is a lot of cool stuff happening and we'll dive into that in a second but what we discussed on episode thirty-five was the quote from a biopharma leader who actually put it to question – What are you going do with it?
• 5:39
  Speaker 1: This really sums up very nicely!
• 5:41
  Speaker 1: The general question is if your shoes have biotech or pharma leaders.
• 5:50
  Speaker 1: there's so much hype around AI and everyone is talking about it.
• 5:53
  Speaker 1: And everyone has use cases, anecdotal evidence and McKinsey forecasts of hundreds of billions of dollars per year that you somehow have to set free with AI?
• 6:05
  Speaker 1: Then at the same time You're looking at it and you might be wondering what am I going to do with
• 6:11
  Speaker 1: it?".
• 6:12
  Speaker 1: One challenge here computer science, engineering-based discipline.
• 6:23
  Speaker 1: But in biotech and in biomedicine we deal with biological systems And they're extremely complex.
• 6:31
  Speaker 1: The data behind it is extremely complex.
• 6:33
  Speaker 1: We are trying to break down into something that AI can compute.
• 6:37
  Speaker 1: This isn't always possible.
• 6:39
  Speaker 1: We had a really nice quote from one of our guests on episode forty six From Ingmar Schuster.
• 6:44
  Speaker 1: He's the founder of Provolute protein design startup, and he summed up this situation of biological complexity.
• 6:53
  Speaker 1: And what AI can do about it really nicely?
• 6:55
  Speaker 1: We have a biotech founder and CEO of Provolute today with us Ingmar Schuster who also brought as a quote about maybe the AI hype in protein design at Biotech per se.
• 7:07
  Speaker 6: So yeah currently I think most people perceive AI for biology.
• 7:17
  Speaker 6: It's like a windshield wiper when it's raining heavily.
• 7:20
  Speaker 6: So biology is super complex, AI helps a bit.
• 7:24
  Speaker 6: but then the next problem comes around and I think everybody understands what the potential is?
• 7:31
  Speaker 6: But its not as fast to realize the potential compared with areas where there are vast amounts of data like text images in video.
• 7:42
  Speaker 0: The picture of a windshield wipers puts into perspective The overall question and I think you found one snippet of episode thirty-one, where it sums up the topic of creativity versus productivity.
• 7:59
  Speaker 1: Here we just really asked a question what is AI so far doing for pharma?
• 8:06
  Speaker 1: And this was about a year ago that we discussed that in exactly the same situation back then.
• 8:14
  Speaker 1: Are they really bringing molecules into the clinic targeting targets that will be in the long run, more successful?
• 8:21
  Speaker 1: This we simply don't know yet.
• 8:23
  Speaker 1: And I have to say that the targets these drugs are targeting.
• 8:27
  Speaker 1: so what were really biologically addressing is currently not different from drug targets that are discovered without AI, which makes complete sense.
• 8:37
  Speaker 1: And I think we talked about that before.
• 8:39
  Speaker 1: it's not so good to change too many factors in a system if you want to show that works.
• 9:01
  Speaker 0: time and again We tried to explore Good use cases of AI in biomedicine Which is quite distinction.
• 9:10
  Speaker 0: two the daily use of chat GBT for instance and things are very, very different.
• 9:16
  Speaker 0: And we have examples with that too in our episodes right?
• 9:19
  Speaker 1: Yeah I mean we had so many cool guests and We cannot feature them all because otherwise The episode will get to long.
• 9:26
  Speaker 1: but we picked a couple of examples Of really cool AI use cases and first one again is from Ingmar who just heard designing proteins that don't exist in nature with AI.
• 9:38
  Speaker 1: And we talked about many times on this podcast, large language models but also other types of AI.
• 9:45
  Speaker 1: reading molecular languages and putting it into use to design new proteins has massive potential for biotech and pharma.
• 9:53
  Speaker 6: As your question doesn't have enough protein?
• 9:55
  Speaker 6: Proteins are not evolved to do the things you want them.
• 9:58
  Speaker 6: One very simple example of this is pharma production, a field that Provolute has worked in as a consulting company partnering with Big Pharma.
• 10:10
  Speaker 6: So big pharma wants to produce their compounds the therapeutic compounds and nature typically catalyzes reactions.
• 10:21
  Speaker 6: so proteins that catalyze reactions enzymes they catalyzed reactions from one tiny molecule to another tiny molecule and quoting a scientist from Zanofi in pharma production it's bulky substrate, bulky substrate.
• 10:37
  Speaker 6: And that never happens in nature.
• 10:39
  Speaker 6: so you are dealing with big molecules that your trying to assemble basically through bio catalysis.
• 10:47
  Speaker 6: It is just not something the nature ever wanted to do its not interested.
• 10:53
  Speaker 0: So those are the new proteins that don't exist in nature.
• 10:58
  Speaker 0: We also refer to something I didn't have a clue at all of which is called The Dark Genome, i think we can paraphrase it otherwise as well Which shows how limited our knowledge is right?
• 11:14
  Speaker 1: Yeah and I have to say this was even when I was studying biology, we still thought that there is really a very limited number of protein coding genes and the rest of the genome.
• 11:24
  Speaker 1: Is junk?
• 11:25
  Speaker 1: There are always in biology We always have this strange notions like we're only using ten percent off our brain.
• 11:31
  Speaker 1: Only two percent of our DNA is meaningful.
• 11:34
  Speaker 1: Usually it's just a question of okay that we consider non-functional is doing.
• 11:41
  Speaker 1: And that's the same for the genome, and we had a fascinating conversation with Possing, he's the CTO and co-founder of BerlinBest startup called Lucid Genomics.
• 11:51
  Speaker 1: they are using AI to decipher the dark genomes who actually put meaning through the ninety eight percent off our genome that does not encode proteins and still has a function?
• 12:03
  Speaker 1: This was really fascinating use case for AI.
• 12:07
  Speaker 4: maybe this where the AI comes to the game.
• 12:11
  Speaker 4: AI helps us from two parts, one is it finds patterns very well on big data.
• 12:17
  Speaker 4: of course in biology we want to know mechanisms but AI helped us share light and patterns then form other side.
• 12:26
  Speaker 4: We are able to denoise data with AI.
• 12:31
  Speaker 4: so In this dark DNA game AI can help Part, first detecting the mutations in a higher accuracy.
• 12:42
  Speaker 4: And when I talk to mutations we talked about variations.
• 12:45
  Speaker 4: We have small variations...we have larger variation structure variants..We have tandem repeats and that's one part that AI can help us from.
• 12:55
  Speaker 4: The other part is like now i know the variations how i Can make sense of these variations?
• 13:02
  Speaker 4: In different tissues individuals populations.
• 13:05
  Speaker 4: This is where AI can help us a lot with pattern finding.
• 13:10
  Speaker 0: Can't wait to see results here!
• 13:13
  Speaker 1: I have to say there are results, they're a number and we had a follow up episode with Sotakaran where you talked about the dark proteome.
• 13:19
  Speaker 1: And there were numbers of companies who actually now exploring even protein coding sequences within this dark space in the genome putting AI into it as massive potential for having new drug targets.
• 13:33
  Speaker 1: because if imagine that your missing out on almost hundred percent to target for potential applications.
• 13:41
  Speaker 1: Of course, not all of it will be targetable but I mean kind of a no-brainer that there is something
• 14:08
  Speaker 0: can find real solutions, right?
• 14:11
  Speaker 1: I would say so.
• 14:12
  Speaker 1: Yeah, i think this was also really fascinating conversation with Adela and Bannum.
• 14:17
  Speaker 1: from Gestaltmetra and Banna from Bone to Gene they are trying to use image based AI systems To actually point towards rare disease diagnosis either on the face And other morphological features or an x-ray base bone features?
• 14:33
  Speaker 1: Of course you can imagine there is a great number in the thousands of rare diseases but There's only very few people who have it, so physicians will probably not recognize this phenotype.
• 14:46
  Speaker 1: this type of app that these companies develop would point a general physician towards the potential diagnosis and then help people finding out if they have a rare disease, what kind of rare disease might be.
• 14:59
  Speaker 1: And many diseases developed in childhood age.
• 15:03
  Speaker 1: you can imagine it means for parents to have child who has challenges which simply cannot understand whats going on I think is something where we really see make patients' lives better or people's life is better.
• 15:18
  Speaker 1: It's a really great tool and it also one of these real use cases where the AI has already making a difference.
• 15:26
  Speaker 1: Gestaltmetre
• 15:27
  Speaker 3: is an AI tool that uses advanced computer vision methods, our next generation phenotyping to recognize rare diseases in simple portrait picture.
• 15:40
  Speaker 3: so the idea The pediatrician or a different doctor would just be able to take a picture with their phone, have an app the Gestaltmetcher App on their phone and within a few seconds get a list.
• 15:58
  Speaker 3: A ranked list of potential differential diagnoses.
• 16:01
  Speaker 1: Why do you think is AI actually so good at that better than us?
• 16:06
  Speaker 6: The main thing is that we increase in amount data that we have.
• 16:12
  Speaker 6: And then the improvement of algorithms,
• 16:15
  Speaker 1: it's easy
• 16:16
  Speaker 6: when you have data to train algorithms that can do this search and matching
• 16:22
  Speaker 1: in
• 16:22
  Speaker 6: a few
• 16:23
  Speaker 1: seconds.
• 16:24
  Speaker 6: so I'm not going say AI is definitely better than trained humans or maybe an experienced physician who has been working with these kind patients for forty years.
• 16:34
  Speaker 6: I
• 16:37
  Speaker 3: think for Gestaltmetscher, a visionary plan is also to combine it with the smart glasses so that pediatricians have their smart glass and don't need to take out their phone maybe even their personal phone.
• 16:53
  Speaker 3: Maybe they are concerned about that.
• 16:55
  Speaker 3: Also children...they're young.
• 16:57
  Speaker 3: They aren't like sitting still on an image.
• 17:01
  Speaker 3: So if you've got your smart glass make it work with a video and I think that would be also easier for the clinicians.
• 17:25
  Speaker 0: You were referring to artificial proteins, lab proteins?
• 17:31
  Speaker 0: And now we are speaking about artificial cells.
• 17:37
  Speaker 0: so... created cells and we had one of, I guess the most distinguished guests in that episode forty-one which is Fabian Thais.
• 17:48
  Speaker 0: He was speaking about virtual cells.
• 17:58
  Speaker 1: For virtual cells, I think at the moment it's not so much about reproducing the cells.
• 18:02
  Speaker 1: That would be more on a synthetic biology point.
• 18:05
  Speaker 1: but what they are trying to do in Fabian's group is also something that across the biotech field is gaining a lot of attention as this idea of re-creating an inner life using artificial intelligence And I mean that used to be done with rule based models and then you had like pages and pages of differential equations.
• 18:28
  Speaker 1: To say, if this happens than this happens?
• 18:31
  Speaker 1: Of course it doesn't get anywhere close to the complexity or biology.
• 18:35
  Speaker 1: It's like a windshield wiper helps little bit AI can do in more high dimensional space so option space is just much bigger.
• 18:49
  Speaker 1: That, of course this leaves a lot potential for virtual cells.
• 18:53
  Speaker 1: And yeah let's hear what Fabian has to say about it because I think that is one the most fascinating applications in the field.
• 18:59
  Speaker 1: Why we actually need Virtual Cells?
• 19:02
  Speaker 0: Yeah thanks Luis
• 19:03
  Speaker 2: first for having me and thank you guys so much for your kind introduction Quite excited to talk.
• 19:09
  Speaker 2: maybe some bigger questions We have on the field.
• 19:12
  Speaker 2: It was good time talking.
• 19:14
  Speaker 2: at the moment A bunch of revolutions coming together.
• 19:18
  Speaker 2: And that's why this excitement, sort of this reformulation off I guess an age old question in the field namely can we you know?
• 19:26
  Speaker 2: Sort of like we build cars where we have a engineering plan write up things and then not just randomly hack a few existing cars into bits and pieces to produce new ones.
• 19:37
  Speaker 2: Can be built such a blueprint for sales?
• 19:41
  Speaker 2: one ideally understand that when a particular modification happens, these things and have this in their impact.
• 19:48
  Speaker 2: Why is it relevant?
• 19:49
  Speaker 2: I'm sure we're going to discuss more later but if you want to treat diseases here... You won't understand where to sort of tune and do things.
• 19:57
  Speaker 2: so This idea of a virtual cell-I am quite excited.
• 20:00
  Speaker 2: the definition And there's very different understanding But i think the highest level picture would be decided dimmers as an example would be to build a system that can simulate cellular behavior under normal growth differentiation, as this beautiful Wilchhoff citation was just formulating.
• 20:21
  Speaker 2: But also of course on the perturbation such as external influences microenvironment but then off-course also particular disease.
• 20:33
  Speaker 2: you mentioned Aviv Regev.
• 20:35
  Speaker 2: she's a good friend one of real lighthouse figures in the field and fantastically integrative person.
• 20:41
  Speaker 2: She, together with Sarah Teichmann and many of us that the two of them are really pushing it has led this international consortium called The Human Cell Atlas.
• 20:48
  Speaker 2: That you might have heard about in some parts spiritual successor to the human genome project.
• 20:54
  Speaker 2: Eric Landner was also pushing for that.
• 20:55
  Speaker 2: so on Really nice community worldwide very integrative And that was in a sense the first type of data gathering.
• 21:02
  Speaker 2: that sort of helped even just spin off This idea.
• 21:04
  Speaker 2: over virtual cell model
• 21:06
  Speaker 1: Where would you say?
• 21:07
  Speaker 1: Are the biggest payoffs in the nearer future for these virtual cell models.
• 21:13
  Speaker 1: So how will they become useful?
• 21:15
  Speaker 1: I think there is a lot of academic groups, philanthropic groups, biotechs betting on the concept.
• 21:21
  Speaker 1: so where would you say... Would you bet your money for the payoff?
• 21:26
  Speaker 2: I'm first now speaking as a pure scientist!
• 21:29
  Speaker 2: Of course i want to understand Indian biology right.
• 21:31
  Speaker 2: some make all this fun methods that are my core expertise.
• 21:34
  Speaker 2: but then we really want to find and understand particular biological problems.
• 21:43
  Speaker 2: There was one aspect, I think the other one that I bet on and there's a whole bunch of startups as well.
• 21:48
  Speaker 2: As bigger pharma going into.
• 21:49
  Speaker 2: that is a problem that my lab sort of initially formulated.
• 21:54
  Speaker 2: something like six seven years ago.
• 21:55
  Speaker 2: we called it SCGEN.
• 21:56
  Speaker 2: so this generative modeling of a perturbation effects particular with drugs And i think This idea Of Like just screening a big library of drugs for some effect It's coming up to an end you know?
• 22:12
  Speaker 2: Some of the low-hand food And this is, of course not really low hangings.
• 22:14
  Speaker 2: But you know some these sort of easier ones have been maybe found.
• 22:17
  Speaker 2: and now the one that are there... You may be want to combine or go in your new direction.
• 22:21
  Speaker 2: so having a system it gives even just a little bit advice where to bet on?
• 22:26
  Speaker 2: Not make a full screen but let me say ten percent of that.
• 22:29
  Speaker 2: then users will have higher outcomes Or understand how we can do combinatorics which such as big experimental space that never screened.
• 22:38
  Speaker 2: This I think has major impact.
• 22:58
  Speaker 0: So again and again, we were also challenging.
• 23:02
  Speaker 0: And we may have to remind listeners that Chetche PT was in the field since twenty three when We had some episodes out already of The Birolution podcast.
• 23:16
  Speaker 0: so it tells us a story about how timely we are with this podcast and therefore, of course the challenge is the hallucination as we found out pretty soon in The Question Behind It.
• 23:32
  Speaker 0: How do we access good data that are promising for AI application on the field of the biorevolution?
• 23:41
  Speaker 0: And pharma development and drug development?
• 23:43
  Speaker 1: Yeah!
• 23:44
  Speaker 1: This hallucination problem... We have a quote from Sema Altman one the guy behind open AI.
• 23:52
  Speaker 1: The guy, you can also ask Elon Musk about this.
• 23:55
  Speaker 1: there's a really nice article about that in New York Magazine.
• 23:58
  Speaker 1: I think about his history of two of them and legal battle but he very early on pointed out that these hallucinations are a feature and not a bug.
• 24:11
  Speaker 1: And we in the number of episodes, We talked about general challenges of AI kind of delay off the land with Paul von Buneau.
• 24:17
  Speaker 1: he's a mathematician founder of an AI consultancy lab.
• 24:23
  Speaker 1: He is really good I think at putting this things into context and explained why actually have these hallucinations.
• 24:32
  Speaker 1: The model will confidently state things as if they were facts that are entirely made up.
• 24:39
  Speaker 5: It's not so easy, because the whole hallucination issue is something that
• 24:46
  Speaker 1: you...
• 24:47
  Speaker 5: That it's like inherent in technology which is a technology learning patterns from past and then generating patterns but there'll always be some thing with real pattern than data doesn't make sense.
• 25:05
  Speaker 5: I was just wondering in which area probably wouldn't we would be maybe not have an hallucination issue at all.
• 25:11
  Speaker 5: and Maybe that's music because it's art.
• 25:13
  Speaker 5: And what is wrong?
• 25:14
  Speaker 5: In music,
• 25:15
  Speaker 1: i mean That brings us also back.
• 25:17
  Speaker 1: to which doctor do you benchmark it?
• 25:19
  Speaker 1: What is wrong?
• 25:20
  Speaker 1: Because i mean humans Also make mistakes and we say wrong things.
• 25:24
  Speaker 1: Is that a hallucination or is it Just?
• 25:26
  Speaker 1: i mean our brain also auto complete stuff and A lot of the things that We come up with.
• 25:30
  Speaker 1: i mean sometimes You're also just Use the wrong word, but also sometimes you just produce information.
• 25:35
  Speaker 1: That's simply not true because your memorized it wrong or because you have strange ideas about the world Or whatever.
• 25:41
  Speaker 1: and then I mean we don't call that hallucination.
• 25:44
  Speaker 0: at The crossroads of philosophy and AI like that a lot
• 25:48
  Speaker 1: Exactly.
• 25:49
  Speaker 1: And i mean something that we asked him where does he see this field going?
• 25:56
  Speaker 1: at the moment, deep learning and especially large language models is what people consider AI.
• 26:02
  Speaker 1: So if you talk I would say to a late person on the street for them, AI as a large-language model it's a little bit like a tempo situation.
• 26:10
  Speaker 1: i mean It's not that used to be Chachi PT now its'a more diversified That People Would Also Consider Gemini or Claude Or Something.
• 26:17
  Speaker 1: But For Many People What They Consider AI Is A Large Language Model And There Of Course much more to AI.
• 26:24
  Speaker 1: There are rule-based AI models, there other deep learning applications that have different architectures and the question is should we actually bring back world knowledge?
• 26:33
  Speaker 1: So shall we teach these AI models more about the world?
• 26:37
  Speaker 1: And this will be something that remains to be seen.
• 26:40
  Speaker 1: but we had Paul's take on that which I think was really interesting.
• 26:43
  Speaker 5: That's a general issue with all machine learning systems that essentially just learn from data Because we always talk about these models as models and they are mathematical models.
• 26:53
  Speaker 5: But a philosopher would probably say, They're not model because there is no causal mechanics that we infer etc.
• 27:00
  Speaker 5: These things often black boxes in terms of epistemology or whatever.
• 27:05
  Speaker 5: So the big next frontier is combining symbolic AI which means bringing real knowledge around and combining that with something we learned from data, making these two things work together.
• 27:20
  Speaker 5: But there we are at the very beginning but that's super exciting!
• 27:24
  Speaker 1: One of key bottlenecks for AI in biomedicine is the data because it isn't made up in the way that is perfect for training AI models.
• 27:37
  Speaker 1: And we talked about that and different episodes, what Mark train says?
• 27:40
  Speaker 1: his data is like garbage.
• 27:42
  Speaker 1: you better know what your doing?
• 27:45
  Speaker 1: What are going to do with it before You collected an.
• 27:48
  Speaker 1: I think That's really interesting because when We're talking About AI and Data there Is this adage thats being used In almost every lecture.
• 27:55
  Speaker 1: Every talk about It which is garbage in Garbage out Which of course Says Like if you don't have good data, whatever your AI engine will calculate.
• 28:05
  Speaker 1: We'll also be garbage!
• 28:24
  Speaker 1: big challenge, not only when it comes to patient data in terms of their medical history but also when it come to scientific data about genetics and any kind of drug development.
• 28:35
  Speaker 1: We have a lot of missing data points that the data is not meant to be analyzed by an AI engine or put into use as we now want because I mean... It had completely different purpose than mine.
• 28:52
  Speaker 1: It's not good data, basically.
• 29:04
  Speaker 1: There's this bias in academia, which also makes sense from a strategic point of view to never publish negative data or very rarely published.
• 29:13
  Speaker 1: Negative Data because the journals are not interesting and knowing.
• 29:16
  Speaker 1: I worked five years on something And i found nothing but The algorithm would be Very interested In this information Because we basically have A strong bias towards positive outcomes?
• 29:27
  Speaker 1: We can infer Non-information as potential negatives But we cannot Be sure.
• 29:32
  Speaker 1: maybe I mean, there's the possibility.
• 29:34
  Speaker 1: no one ever explored this or someone explored that for a long time.
• 29:39
  Speaker 1: But it didn't work.
• 29:40
  Speaker 1: maybe it also did work but i don't know.
• 29:42
  Speaker 1: The PhD student decided to become a yoga teacher instead and left the project And No One Picked It Up Again!
• 30:01
  Speaker 1: Have a good model and since they usually have the best models, maybe he is proven to be right.
• 30:07
  Speaker 1: There are people who say that Dimitris Abbas has hidden the AGI so there released smart one in his basement and he just waits to release it on humanity.
• 30:16
  Speaker 1: So let's see where this going.
• 30:17
  Speaker 1: make your algorithms better.
• 30:19
  Speaker 1: you're models better.
• 30:20
  Speaker 1: You do have enough data if you were innovative on your algorithm side.
• 30:25
  Speaker 1: I found that a little bit funny, quite provocative to say name names.
• 30:31
  Speaker 1: so this is really like going against the trend of data generation we're gonna get into in second.
• 30:37
  Speaker 1: So i think it captures whole scale.
• 31:00
  Speaker 0: We also had A subtitle in this podcast, the very beginning.
• 31:06
  Speaker 0: The first like fifteen episodes or so it was not only by revolution podcasts but also we're doomed were saved so week with this because its for search engines to complicated and stuff like that and stuck too.
• 31:19
  Speaker 0: By revolution podcast still where addressing pros and cons the doomsday scenarios as well is the saving part when it comes to rare diseases, for instance and what AI can do here.
• 31:36
  Speaker 0: But the general question of course we addressed that as well how might the AI driven pharma medical world look like by a revolution where it looked like?
• 31:48
  Speaker 0: And also What could possibly go wrong For The Dooms part?
• 31:53
  Speaker 1: Yeah I mean there have been so many Dooms scenarios right AI, even before AI was really capable of doing a lot.
• 32:01
  Speaker 1: And I think it's really smart to think about that.
• 32:04
  Speaker 1: For this dual-purpose technologies like for nuclear energy or biological technology such as gene editing and artificial intelligence can be used for the good and bad depending on the actor.
• 32:17
  Speaker 1: really important to think about the question, what could go wrong before it actually does go wrong.
• 32:23
  Speaker 1: And in some of these situations both for biological applications but also for AI there is a question.
• 32:29
  Speaker 1: It's not even only bad actors But its about What can happen if it goes out-of hand and develops life on its own.
• 32:38
  Speaker 1: Lets say starts make decisions.
• 32:40
  Speaker 1: There are very recent scary examples of chat gpt hacking website by itself.
• 32:46
  Speaker 1: These are the situations there, and we talked about it in the past a lot.
• 32:50
  Speaker 1: There are glimpses of AI doing scary things but I think this is still not what should worry us most – more on people's questions what kind of bad things people can do with AI.
• 33:05
  Speaker 1: One thing that we discussed also, with Adela and Bannum is this question or facial recognition because there's a lot say very potential in their if this isn't the wrong hands to diagnose features just based on your face which you use for biometric scans are used for different things right?
• 33:23
  Speaker 1: So using these in a sense open your phone whatever but then an AI reading out some kind of medical information and informing your insurer, informing your boss.
• 33:35
  Speaker 1: Informing maybe the country that you're trying to enter about these conditions for example that you might have a mental condition or something like that is pretty scary.
• 33:44
  Speaker 1: so I mean at the moment we are talking about rare disease diagnosis but there are applications for facial recognition.
• 33:54
  Speaker 1: That could be really scary.
• 33:55
  Speaker 1: And then there is the reverse and we discussed that also, which of course it's a question what can you actually construct facial features based on genetic profiles?
• 34:05
  Speaker 1: What would do with that in terms or finding criminals?
• 34:08
  Speaker 1: so its' really the questions about which kind of biological surveillance you could imagine.
• 34:14
  Speaker 1: Let
• 34:14
  Speaker 0: us talk ethics little bit because maybe has a smell even for those who are not into science.
• 34:22
  Speaker 3: Yeah, for Gestaltmütze we also thought about this of course.
• 34:26
  Speaker 0: Of course
• 34:27
  Speaker 1: you need to!
• 34:27
  Speaker 3: Which is why we implemented this registration process and it's not open to everyone.
• 34:34
  Speaker 3: with the physicians It's quite clear that its for a medical purpose but if you give it in the hands of Everyone The different thing
• 34:50
  Speaker 1: Are there any worries?
• 34:52
  Speaker 1: I mean, privacy but also this question of someone takes a picture off my face and diagnoses me with the disease without me wanting this kind of situation.
• 35:02
  Speaker 1: Is that something you have come across as a worry or is it because its really physician limited?
• 35:09
  Speaker 1: No challenge!
• 35:15
  Speaker 1: I think facial features are something very explainable to the public.
• 35:21
  Speaker 1: It's accessible for all of us.
• 35:24
  Speaker 1: looking at an x-ray or radiograph picture, a retina picture.
• 35:27
  Speaker 1: No one would be able to make any sense of right?
• 35:29
  Speaker 1: Who is not?
• 35:30
  Speaker 1: I mean if i see a picture... If the bone has clearly broken..I probably wouldn't know what im looking at but face something that we see all time and also use for example open our phones based on image recognition.
• 35:44
  Speaker 1: so these kind things play into potential misconceptions around this technology.
• 35:57
  Speaker 0: What do you think about posing a visionary question?
• 36:00
  Speaker 0: A
• 36:00
  Speaker 1: visionary question is something good, but I want to discuss very briefly which might be visionary.
• 36:07
  Speaker 1: But also it's slightly scary this Chinese study from the Chinese Academy of Science where they did the reverse so that used a genetic sample to reassemble facial features like from images shown in paper.
• 36:24
  Speaker 1: I would not be sure that i recognize this person on the street, but it's similar maybe to what you call in English.
• 36:31
  Speaker 1: This phantom picture is like the police draws if... Like almost
• 36:35
  Speaker 0: forensic work right?
• 36:36
  Speaker 0: Exactly!
• 36:36
  Speaker 1: Yeah well.. It could be forensic of
• 36:39
  Speaker 0: course so
• 36:41
  Speaker 1: this has nothing really with Gestaltmetcher But goes in another way
• 36:45
  Speaker 1: Right?!
• 36:46
  Speaker 1: The question that If You can assume a genetic component based the morphology of a person, you can off course also go the other way around and AI seems to be.
• 36:58
  Speaker 1: this was a study from this year where which I didn't introduce properly but were scientists apparently we're able.
• 37:05
  Speaker 1: To use genetic profile to reconstruct faces Which Of Course leaves us with massive amount of ethical questions legal questions And just shows that I mean AI and the link between genetics and morphology is really moving in a new direction.
• 37:26
  Speaker 1: If this would be something where data protection for me was very important, then i'd like those tools not to be implemented because that could be quite scary.
• 37:38
  Speaker 1: to just collect DNA somewhere and say like, this is the person who left a chewing gum at the corner.
• 37:44
  Speaker 3: Now combine these with genomic newborn screening?
• 37:47
  Speaker 1: Exactly!
• 37:48
  Speaker 1: Oh my
• 37:48
  Speaker 0: god...
• 37:49
  Speaker 1: Yeah we have perfect big brother society
• 37:53
  Speaker 0: right now.
• 37:55
  Speaker 0: Yeah, indeed scary stuff promising stuff at the same time and therefore we're both I think so fascinated of what is going on really as We speak.
• 38:06
  Speaker 0: And we started with that That it's.
• 38:08
  Speaker 0: It's so fascinating to have a topic that develops not in history books but In the present tense.
• 38:15
  Speaker 1: yeah, and i think The last thing about the AI driven future?
• 38:24
  Speaker 1: Our minds with AI, will we allow the AI into our brains?
• 38:29
  Speaker 1: And I think this is for me.
• 38:30
  Speaker 1: The big question of our future which i find equal parts fascinating and scary.
• 38:35
  Speaker 1: it's really difficult to predict the future off AI in a sense that the contribution of these models Is just changing so fast.
• 38:49
  Speaker 1: I mean, i don't think that we've reached the end.
• 38:52
  Speaker 1: So at the moment We have this deep learning based mainly convolutional networks and at The moment transformer-based networks That are very efficient At doing certain tasks but they're also not so good at Doing other tasks?
• 39:05
  Speaker 1: It might be that...I'm pretty sure a lot of AI companies Are working on newer models than I at least Have never heard about But it may have similar effect as ChatchiPT in three years.
• 39:19
  Speaker 1: So that is one thing.
• 39:20
  Speaker 1: And I think already, for these large language models... ...I listened to an audio book where they discussed the potential of Chatchapiti in medicine and said it randomized the future.
• 39:33
  Speaker 1: It's a little bit true.
• 39:35
  Speaker 1: There was big hype around Chachapiti but not all is set-to be But also has say In daily life its first time.
• 39:46
  Speaker 1: everyone really feels the contribution of AI.
• 39:50
  Speaker 1: I think it's one of the fastest adapted user technologies that are out there, and this is something that has felt so before.
• 39:58
  Speaker 1: I used to be something that maybe you talk about or as you say listen to Elon Musk saying It will lead to humanities downfall And other people saying?
• 40:06
  Speaker 1: I think chat GPT is the first example where really most people or many people are using it and exploring also the interaction with an AI.
• 40:24
  Speaker 1: And i think this part of future, irrespective now of direct development that the interaction between human and computer will get a lot stronger especially if Anything that's integrated into the smartphones and potentially mainly in the next generations, also neuro prosthetics will just link us much closer to AI.
• 40:45
  Speaker 0: Absolutely!
• 40:45
  Speaker 0: And I mean we've been discussing that as well In our longevity episodes.
• 40:50
  Speaker 0: for instance you know The digital afterlife and stuff like that and the chip That helps us think or replaces Our thinking.
• 40:58
  Speaker 0: So that kind of wraps it up, fifty the episode.
• 41:06
  Speaker 0: We're ready for the sparkling wine now because we didn't expect as we mentioned in the beginning That this would last more than five episodes.
• 41:15
  Speaker 0: Thank you again For listening to us More information and also in text
• 41:20
  Speaker 1: at ScienceTales.com.
• 41:23
  Speaker 0: We already know what will do next time right?
• 41:26
  Speaker 1: We talk about AI
• 41:28
  Speaker 0: Of course
• 41:29
  Speaker 1: AI scientists to be precise and lab-and-loop models.
• 41:32
  Speaker 1: so that will be quite interesting I think.
• 41:34
  Speaker 0: That wraps it up, thank you Izzy!
• 41:37
  Speaker 1: Thank YOU!
• 41:38
  Speaker 1: And THANK YOU for fifty episodes!
• 41:40
  Speaker 0: Likewise.