“…A lot of times the new discoveries are made on the boundary between fields where one field is well defined and another field is well defined. But at the boundary something happens, something is new and there a new field is born.”
What happens when you ask a physicist-turned-theologian-turned-AI-researcher to teach middle-schoolers about machine intelligence and build a hypothesis engine for scientists? In this episode I sit down with Dr
—creator of Innovation Lens—to map the maze between AI literacy and breakthrough discovery.We start in the classroom, where Jonah uses blindfold-maze games to demystify algorithms. His approach echoes new federal and NGO pushes for early AI education—from the U.S. Presidential order to “promote AI literacy” (2025) The White House to MIT RAISE’s open, ethics-first curriculum for middle-schoolers MIT RAISE. But we also confront the cautionary data: a Wharton-led field experiment shows students’ grades soar 48-127 % with GPT-4—yet their underlying skills atrophy once the crutch is removed SSRN.
From there we zoom out to research culture. While labs such as Sakana’s “AI Scientist-v2” can already draft peer-reviewed papers end-to-end Sakana AI, Jonah argues that human curiosity still owns the out-of-distribution frontier. His own model doesn’t write your paper; it points you to the unexplored coordinate where the next 1-k-citation breakthrough might lie—then hands back the shovel. Think Richard Feynman’s reminder that science marries “disciplined thinking” with imagination “different from that of the artist” Lib Quotes.
Along the way we debate guardrails, risk, and why letting learners “run into walls” may be healthier than bubble-wrapping knowledge. If you’re an educator, researcher, or just AI-curious, this conversation is your map—and maybe the pickaxe—for digging into tomorrow’s gold.
Chapters
00:00 Understanding AI in Education
05:47 The Role of Hard Work in Learning
06:11 Exploring the Nature of Truth
12:48 Innovating Research with AI
16:34 Comparing AI Tools for Research
20:41 Generating Hypotheses for Scientific Research
25:32 Training Data and Its Implications
29:51 Ethics in AI Research
36:13 Exploring Latent Space in AI
42:07 The Eureka Moment in Development
44:10 Collaboration and Inspiration in Science
46:50 Misconceptions About AI and Science
Full, Unedited Transcript
(Apologies, it may contain some errors)
Jonah Lynch (00:03)
One project I'm working on now is about teaching AI to elementary and middle school students. And the premise that I'm starting with is that it's important to get inside the black box and try to understand as much as possible without the extra complications of this magical machine that looks like an oracle doing things that you can't understand. Actually, a lot of the operations that we call AI are pretty simple. They have to do with rules. They have to do with linear regression, looking for patterns and...
predicting the future based on past patterns. And a lot of that can be taught at a fairly easy level without using computers, just using different experiences. Like for example, helping somebody blindfolded pass through a maze by hand is easy enough. But then if you try to abstract it and say, okay, give a set of rules so that a person who's blindfolded can get through the maze, then you start to think at a meta level, which is a lot like what you're doing in the first iteration of a self-driving car.
This is a project that I'm presenting in a United States school district in a couple of weeks. Then I'll be talking about it at a university in Slovenia that invited me to talk about this. so there's some teaching I'm doing there too, but more on a kind of spot basis, trying to think about how to help students today navigate the coming AI world to the degree that it will be AI. I think there are a lot of things that aren't going to change too.
Alberto (01:29)
Yeah, very interesting. I want to make a quick digression there because I'm both excited and concerned about the role of AI in education. So I just wonder what's your take, what's good and what we should be cautious about using AI in the context of education.
Jonah Lynch (01:46)
Yeah.
I guess I think about it a lot in the terms of like a muscle that if you don't use muscles, the atrophy. Now, if you use force multiplying machines, you can do a lot of things you can't do with muscles. And I don't think that as humans, would, I don't think we're nostalgic for the time when enormous groups of slaves would dig ditches and build foundations for buildings. I think we're glad that we have bulldozers that can do those things. But.
There's also something to be said for maintaining the full strength of the human and choosing our priorities carefully. So I think that education has a lot to do with those priorities. It has to do with which kinds of things do we want to simplify and which kinds of things are better kept local, know, are better kept on our local machine instead of outsourced to somewhere else. Writing is a great example of that, right? So, you know, before writing, humans had incredible memories and oral cultures.
are able to memorize enormous amounts of text and myth and story and history. We have the same architecture and we've chosen to outsource a lot of that computational power to the written word. So that has given us certain advantages which are undoubtedly unquestionable, but it's also given us certain disadvantages. And we've made that choice. We've said, okay, this is a trade-off that we're willing to make. With AI, I guess we're...
probably talking about large language models today in 2025 when they say AI, but there are other things we could be meaning as well. But with automated decision-making systems or with large collections of text in a large language model, that trade-off is not yet quite as clear. So we're currently navigating what is important for the future and what are we choosing to make not important. And we're probably not getting that entirely right yet. So I think there's a lot of...
reflection that needs to happen in this stage. I'm also pretty hopeful. know, 10 years ago I wrote a book about technology and human relationships in which I was pretty concerned. I was pretty concerned about called the scent of lemons and I was pretty concerned about the effect that social media would have on our ability to connect with each other. But I have to say 10 years on, I think we have learned to a large degree how to do it well and we're at least conscious of the trade-off.
So maybe you get sad and FOMO because you look at Instagram too much. But we know that that's happening. And that sort of meta-consciousness is helpful as we navigate it further and decide what persons do we want to be. So I think education has a really important role to play there. But also part of it is just running into walls and finding out that it hurts.
Alberto (04:43)
Yeah, that's right. speaking of hurting and fatigue, there's also pedagogical value of hardship and doing the hard work rather than taking shortcuts. I think it's another good value to educate.
Jonah Lynch (05:02)
Yeah, no, I agree. agree. But I have to laugh a little bit because I totally remember when I discovered that with a calculator, I think it was in the fourth grade, that with a calculator I could do my math homework much more quickly. And by doing it more quickly with a calculator, I became dumber. And there's new research came out of the Wharton School last year that said basically the same thing. Students who use AI on their homework do better.
as long as they've got AI. As soon as they stop using AI, the people who didn't use it before do better. So, you know, is this, is this crutch good for you, like for what you need in your life? That's a, it's an open question, but it's an important one.
Alberto (05:37)
Yeah.
Yeah, that's right. That's Great. moving to the research work that you've done and that if I understand well, it shaped some of your thinking along all the thinking that brought you to technology or product that you're building. I would like to hear more about that. What's the...
Jonah Lynch (06:11)
Yeah, sure.
Well, okay, so the question of what is true, I guess, has always intrigued, delighted, infuriated, attracted. It's a question that I care about a lot. What is real? People have told me, stop saying what is true because that word is passe. But I don't know, I think it's the best word we've got. It's not perfect, but words are approximate references, I guess.
Anyway, so truth is something I care about. And that's brought me to study a lot of different things. When I studied physics, I really also wanted to know what came before the Big Bang and what's going to happen after, know, that kind of thing. So that broadens out into philosophy and theology and so forth. Then at another point, I felt like the whole edifice of, you know, humanistic research, especially in philosophy and theology,
was something like an inverted pyramid where there's an enormously beautiful flowering of ideas on a certain level. And then as you come down, it's all resting on smaller and smaller pillars. And then at the very point, the bottom point, what is it resting on is a question that I think I arrived at several times in my life. the last time that I arrived to that point, I said, okay, right now I'm
I don't know. And right now I'm also not particularly interested in pursuing that path because I really like something solid. guess somebody called me Cartesian the other day and I think maybe that is what happened to me that I ended up wanting some kind of trustworthy foundation. And I went to math. Math was the place where I thought I could find something more trustworthy. Now it's very poor in a sense. Math can't say much. But what it says it's very certain about.
I appreciate that. So that's part of my story. Then the next part of it, which is the research I did in my second doctorate when I was working with a group of historians and archeologists, was about, okay, so I was trying to model their discourse, model how do archeologists argue about their interpretation of history. And I tried to build software that would help us to remember the arguments.
and create a weighted graph where the arguments that each person brought to bear would be connected to source data, so basically, like not open to opinion data, sort of not open to opinion, right? And then through successive levels of abstraction and interpretation get to the level of an overall discourse where you've got an interpretation of information of all the pieces of pottery coming out of the ground.
I was also modeling the various levels of authority of the people speaking. And I was trying to build a graph structure so that you could get back to the original disagreement. Because you often have situations where schools of thought disagree on many things. But if you can get a bird's eye view of their disagreement, you realize that what they really disagree on is somewhere upstream. And one reason why I was doing this was
as a way to help new researchers who are being onboarded into our team to get up to speed quickly so that they could quickly learn where did our team take a stand, why did we take that stand, and who thought differently. So that was sort of a first version of things. And then as I did that, I realized that text vectorization could be useful. A lot of NLP things that had been developed in the previous years.
would be useful in what I was building. So I started to learn a lot about NLP and the AI techniques that were coming out. This was in 2020, 2021. Then when ChatGPT hit the big time in 2022, I was well prepared for that. And it was absolutely, it was a very exciting moment for me when I realized that the convergence of linguistics and structures of knowledge and graph theory and algebraic topology and a bunch of different ways of thinking about
the structure of information converged for me. And that was the research that I did for my doctorate, which is publicly, it's been published, so that's available. After that, I saw some ways, I was thinking, how could this be useful for more people? archaeology, first of all, the archaeologists generally didn't understand or care what I was doing very much. Because it was kind of too mathy.
Alberto (10:46)
Mm.
Jonah Lynch (11:10)
They're used to arguing things out and that's the way that they like to do it. They don't all have a Cartesian bent like me. So then I started to think, where else could this sort of process be useful? And the idea that I've developed since then is now part of what I'm calling innovation lens, which is it's a basically topic agnostic tool to help scientists think about what's the next step in their specific domain.
Alberto (11:14)
Yeah.
Good.
Jonah Lynch (11:39)
So it's based on those same structures and ideas that I was doing in my doctorate. But it's also forward-looking. It's looking... We found a... I found a... Say we because it makes it sound more broad, but I guess it was just me. Anyway, I found a pattern in the distribution of articles and the way that scientists talk about their work, which has some predictive value. it...
Alberto (11:40)
Mm.
Okay.
Jonah Lynch (12:09)
I've started to show how that can be useful in predicting what's a good next article subject or a good next problem statement for research in some given domain. Sorry, that sort of got confusing. Maybe you can ask some questions to guide it.
Alberto (12:25)
Yeah, yeah, definitely. That's super interesting. I understand you're building a tool to help a researcher to generate hypotheses or generate areas of research that have not been explored yet. Would that be an accurate description?
Jonah Lynch (12:48)
Yeah, yeah. well, guess what I can, you know, building on my previous statement that I want to believe in math, I believe in the statistical validation. So here's how I did it. One was, first of all, what am I looking for? So the target is I'm looking for articles that receive more than a thousand citations. That's a pretty good article, right? You could say that thousands of citation articles, they tend to be the articles that start something.
Alberto (13:07)
Mm.
Jonah Lynch (13:17)
They're the ones that the obligatory reference in every paper after that in some domain. How did I validate? I thought, okay, so how do young researchers decide what to write about? Generally, they choose something that's within the existing literature. So you can try to look at a field and then you look for a little gap in the literature and that's where you go. Something missing, but within a domain that's already been explored.
Now I can quantify that because I can translate, this is something that word vectorization allows us to do, can translate words into a geometric space. So I can translate a scientific domain into a geometric space and say, given a distribution, so given an existing collection of articles in this space, in the future, how many
Thousand and up citation articles are found in that same domain, right? So that's that's test you can do So I did that and I compared that against my algorithm my my thing that I discovered the pattern is covered so that's one one baseline another baseline would be Go random cover the whole space cover the whole domain. You've got basically unlimited resources study everything, you know and
Now as you study everything, of course, you're going to beat my algorithm, but you don't have unlimited resources. So I was also able to show that in that case my algorithm beats the scattershot way of proceeding up until like, you know, a high number of test points. So that validation tells me that, okay, so there is an intermediate distance. Like if you just sort of go randomly looking, use the archaeologist metaphor.
If you just go dig somewhere randomly, you might find an Egyptian tomb filled with gold. It's not very likely, but it is possible. If you go where the known pyramids are and dig, you're likely to find an empty tomb because somebody else has been there before you. So is there an intermediate distance where the information from the past is predictive?
but is also new, right? So it's not randomly just digging, but it's going the right distance away from the existing objects. So it's a metaphor, but that is kind of how the system works. It measures a distance and a location, and based on that location, it says, dig here, and you're more likely to find gold.
Alberto (16:04)
Yeah, that's fantastic. Yeah. And we hear these days a lot about tools that other companies are launching to do research. One was the AI scientist by Sakana. We have heard all the labs launching their version of deep research. Can you explain, like, I think I can guess what's the difference between the tool you're building and these kind of tools?
But can you explain better how the tool works in relationship with those others and what's the role of the researcher? I think there is a difference between the role of the researcher in one and the other approach.
Jonah Lynch (16:46)
Yeah,
absolutely, absolutely. Now that's a good comparison. You're making me think I need to update my white paper and take into account some other things. So what can a large language model do even with iterated prompting like the deep research or R1, the various ways that they do? What can that do as compared with a human researcher? And the way that I understand it is
mostly having to do with in-domain or in-distribution and out of distribution inference. So if the information is already there, the LLM is a wonderful librarian. I mean, you can retrieve information that would have been very hard to find in a very short amount of time. It's really wonderful from lots of points of view. But how is that actually going to do something that's new?
So maybe it's doing the kind of research that middle and low level PhD students are doing where they're really just filling small gaps and everybody knows it and that's fine. It is useful. Sort of like compiling tables of logarithms. It's just awful. But before you have calculators, you need them. So fine, you do it.
Now that's one kind of work, right? That's a kind of work that I think could be automated with the current systems on LLMs, and that's fine. What a creative researcher does is pull from different fields, or different times, or have intuitions that have a mathematical basis, and then they're expressed in some other field where you didn't expect them. Like quantum physics is full of this where
You have mathematical intuitions that precede by decades or hundreds of years their physical interpretation in the quantum world. It's very beautiful and very wonderful that somebody who happened to know that math was able to then solve the physical problem at a later time when it was possible to understand there was a problem and do the physical experimentation to validate it.
That's a different thing and a large language model is not able to do that right now. So what am I doing? What I'm trying to do is a little bit different from both of those paradigms. One thing I'm trying to do is, first of all, I did find a pattern, I measured it, it's visible, so that's the basis. What's the pattern good for? Well, since the data I've been working on so far is scientific articles,
It's good for helping scientists prioritize which topics to study and which ones to kind of say, ah, it's all right, we'll leave that one for now. So hypothesis generation, because then there's a next step, right? Finding, finding a, being able to the scholar where to dig is useful, but that abstract space is not where we work and live. So you have to also translate that abstract space into words.
Alberto (19:45)
Right.
Jonah Lynch (20:02)
and that's where my generative AI part is coming in. But it's again not working in the same way as things like Sakana or DeepSeek, DeepSearch, because what I'm doing is I'm literally translating the vector into words. I'm not prompting a model using words to give me more words. That's the fundamental difference.
Alberto (20:28)
And so how researchers in different fields would use this to start off and accelerate their work if you can bring it to life.
Jonah Lynch (20:41)
Yeah, so the demo is online. It's already sort of working. We're trying to improve the front end because there are lots of technical difficulties when you're handling millions of articles and trying to turn it into a responsive web app. But whatever, that's just the surface level stuff. I think that the way that it's useful is for scientists who are looking to decide what is their next study. What's their next topic?
Alberto (20:52)
Yeah.
Jonah Lynch (21:08)
to generate a list of tens or hundreds of hypotheses within their domain, right? So part of the thing is, like take computer science, there are literally 100,000 computer science articles published every year just on archive. So you can't read it. It's not possible. Now you can sort of, you can say, okay, well, I know about large language models so I can narrow it down and then I'm really mostly about...
whatever transformer architectures and so that's my specialty. That's one way of thinking about things. But a lot of times the new discoveries are made on the boundary between fields where one field is well defined and another field is well defined. But at the boundary something happens, something is new and there a new field is born. So one thing that my tool can do is help you to pay attention to those boundary layers in an extremely updated way. So you don't need to read
100,000 articles and remember them, which is also impossible. But you can use a tool that every week is updated with all the recent archive articles and graphically see where is the boundary and then zoom in on the boundary that's most interesting to you and then automatically generate hypotheses there. now if you throw away 95 % of the hypotheses, I would say that's fine. I mean, I don't expect all of them or even many of them to be useful.
But if you find two or three that are good, that spark an idea that you didn't have before, that I think is the power of the latent space inside the large language models, where concepts from the past, right, so it's in distribution, have never been juxtaposed in this way, this particular way. And that juxtaposition can bring the researcher to have an original thought and say, ⁓ if I go down this direction, I bet I can make something that actually makes progress. Now this is my hypothesis.
but we are in the stage of validation. So I'm inviting scholars to use it and giving away free coupons to do so. And I'd love to have feedback about that to understand which domains is it useful and which is it not, because I don't expect it to be useful everywhere. My validation was done on computer science articles. So I know that it does work there. Then for everything else, I still need to do the validation.
Alberto (23:32)
Yeah, absolutely. I'm going to share with these recordings the links to try out the tool. And my next question was exactly on the field, but you kind of already answered. You envision this tool to be used in any field, virtually.
Jonah Lynch (23:38)
Awesome.
I mean, okay, so first I did it for computer science because I thought that's where the money is, that's where the interest is today, so that was my Silicon Valley startup dream, that's where I started. Then I thought, well, why not all of archive? I once you've got it set for this kind of article, why don't you just take everything else too? And so was six subjects then, computer science, biology, okay, so biology, economics, or quantitative finance.
Alberto (23:56)
Yeah.
Jonah Lynch (24:17)
and electrical engineering, signal processing are very small fields, but I've distinguished them as three distinct fields. The big ones on archive are computer science, physics, and math. And so that's already, you know, that's interesting. But then I had somebody say, what about medical? You I'd really like to try this out for pharmaceuticals or for health innovation, know, health tech.
So I also have loaded up all of PubMed, which is a lot, 27 million articles. And why these two archives? Well, because they're free and they have good APIs, so it's easy to download. But I think that the basic underlying procedure could be applied to many other fields too. And I think from a commercialization point of view, that's actually where the money is going to be, is not so much serving individual scientists, but finding
like pharmaceutical industry or maybe it's insurance or hedge funds or something, know. I think that'll be more interesting in the long term and I would love for that to be the way for the scientist-facing tool to be free. That's kind of what I'm hoping to do.
Alberto (25:32)
Yeah, yeah, that's wonderful. using developing and using the tool yourself, have you already had some sort of Eureka or aha moment where you saw something on these boundaries that you were describing that actually is really interesting and
Jonah Lynch (25:52)
That's a good question. You know, I don't have an example I can point to right now. This is what I've started to do with specialists in a couple fields, but I've just started. So I'm going to keep that question open, and I'd like to come back in a couple months and tell you, OK, I've made these two or three discoveries that seem really interesting, and talk again.
Alberto (26:16)
I'm gonna try myself and let you know what to Can you describe a bit better the training data that you use to build this tool? I want to ask you this question on two directions. One is to understand the type of training that you've done and what are the implications on training on those data. And on the other one, with training data comes also sorts of...
Jonah Lynch (26:18)
Yeah, cool. Yeah, yeah.
Alberto (26:46)
some biases or risks of topic clumping, these kind of things, whether you consider that.
Jonah Lynch (26:54)
Sure. So the training data is the archive. Now I have to qualify that because parts of the system are trained in certain different ways, but the essential architecture is that the data is the scientific articles. In the first iteration, the data is the title and the abstract of the scientific articles.
Alberto (27:00)
Yeah, okay.
Jonah Lynch (27:23)
just a computational problem, but it can certainly be expanded to include the full text and the images and so forth. So that's the basic data of the system. Then it goes through a couple of different transformations. One is using a language model which is specialized on scientific articles. So there's a layer of modeling which is fine-tuned for the kinds of text that scientific articles contain and also their citation networks, the bibliographic, the graph of connections between
between articles and people. So that's one part that's important. Then there's another part which is how are the embeddings of the articles manipulated in latent space? That's purely mathematical, so it's not open to question of bias. And then the final stage, I guess two pieces to the final stage. One stage is translating the abstract vector into text again, so reverse engineering the vector.
which is done through an iterative process where you try what might be the text for that vector, embed that text, measure it against the original vector, and use the difference as a way to get closer. And so after some number of iterations, you arrive at a text which closely corresponds, the embedding of which closely corresponds to your original vector.
That stage is, it works pretty well. It's also probably the most open to optimizations and improvements. And then the final stage is a very basic prompting layer in two steps using a publicly available commercial LLM where I literally say, take this text and make it in complete sentences. Like that's it. I don't add anything, I take anything away, but sometimes...
Sometimes the text that comes out of the reverse engineering the vector is a text that is truncated or expresses an incomplete thought. So just to clean it up for the user. This is something like, think in version two, I'd like to offer the option. Would you like to see the raw output or would you like to see the prettified output? Because I think that as a customer facing product, you have to give
kind of pretty output, so it makes sense to say turn it into complete sentences. But I also think that the raw information is probably more useful in the real case. I don't want to push scientists away because they have ugly answers, but I also don't want to push them away because the information is too pretty and not useful. Because we are becoming, you know, we're becoming like suspicious of LLM bullshit. It's too prevalent, you know.
Alberto (30:12)
definitely.
Yes.
Jonah Lynch (30:21)
And then the final, final stage.
Alberto (30:21)
Yeah, think human intelligence
has a better detector.
Jonah Lynch (30:25)
Yeah, yeah, absolutely,
absolutely. yeah, then the final stage is again a prompt to a commercial LLM, but it's just doing ranking. It's just saying, does this article, does this text look valuable or not? know, score it between zero and 10. And I use that to try to give the best things first to the user. that's the overall architecture.
Alberto (30:45)
Yeah,
that sounds great. And speaking of, let's say, some ethical and associative angles of this product, have you thought or put in place already any safeguards to ensure the tool doesn't generate unethical research directions, harmful directions, stuff like that?
Jonah Lynch (31:09)
No, I haven't. I don't think I want to either. think that, yeah, the way I think about research is that it is basically infinite field. And there are domains where...
Yeah, it's even hard to say. Where you shouldn't go, where you don't want to go, where we have discovered it's not worth going. You know, there are different ways to describe why would you not study a subject. But I think I would rather leave that possibility open and not outsource the moral decision to a machine, but let the human decide. Because what I'm trying to do is spark thoughts.
you're going to throw away 95 % of the thoughts sparked by my tool for sure, right? For sure. Those 5 % that you keep, it's up to you, which they are. And I don't think I really want to hide ideas that might be unsavory. I think I'd rather leave it up to humans. I don't know, what do you think?
Alberto (32:28)
Yeah, I like
that. I like that. I like the ethos that I also think differentiated this approach with the completely automated deep research or Sakana AI scientist agent where you offload everything to the machine. So the scientist is no longer on the driver's seat. It's more like, do this for me.
Jonah Lynch (32:58)
They're
just peer reviewing the results.
Alberto (32:58)
So at that point, exactly, exactly.
So I think it not only becomes dehumanizing, but I already see, I already see myself the way I use this tool that become lazy. So I would become lazy to review entire papers generated completely by an agent. Whereas this approach of first redefining a hypothesis in the research question that is still the job of the human researcher.
and giving the tool just to do that better, faster, and in a non-obvious way is already very valuable. And I like this bias towards the human agency and human responsibility and free will, rather than putting constraints into the machine, then the debate becomes like, who put the constraints and why and so on.
Jonah Lynch (33:54)
Yeah, well, mean,
and it's inevitable because how else, I mean, we've been disputing this for as long as humans have been around. You know, who says what's good and right and true? And we know that it's a very important question. At the same time, we know that there isn't an easy answer to it. You can't, like, the whole idea that you can put guardrails in places begging all of the relevant questions. You know, how do you know that those are the right guardrails? And...
And I'm not saying that as a sophomoric way of just getting out of the argument, but to say that is really the relevant question. How do you know? How do you know? And I guess I don't want to make a universal judgment by saying this. I don't think that everybody has to live the way that I have lived. But I would say that I have had to run my head into the wall a few times and discover that it hurt.
I don't think I could have learned certain things in another way. in any case, I didn't. So that makes me say I would rather a world where I'm not too protected from the walls so that I can actually find where they are and not stop short of them because somebody told me, but there wasn't actually a wall there. And then, yes, I bear the wounds of that too. And I'm not proud and not.
Alberto (34:52)
Yeah.
Jonah Lynch (35:20)
not happy to bear trauma, but I don't know if there's another way to get through life than bearing a couple of scratches.
Alberto (35:28)
It reminds
me a lot that long essay that Jonathan Hyde wrote a few years ago on the Atlantic, the Cuddling of the American Mind. I don't know if you came across that.
Jonah Lynch (35:37)
Yeah. Yeah, yeah, Yeah.
Alberto (35:43)
Yeah, very interesting. And if I can, there was a question that it's a bit probably off on another tangent, but it may be relevant. I was reading the technical description of how you tested this. And so I understand that you have, know, this training data creates your, let's say mathematical space, your geometrical space.
of topics, directions, et cetera. And we were discussing what I believe the test data is. You took out some of the top-sided papers and tried to predict whether your tool predictions were close enough to the same
point in the space where those top papers reside. And now my question is if the space, if the geometry and the mathematics of that space is created by the training data, isn't that a risk that we are missing something because it's not just part of the definition of that geometry? Not sure if the question is cleared.
Jonah Lynch (37:08)
Well, specify it further if you want or I'll respond as it is.
Alberto (37:15)
Yeah, so I think that the choice of the training data influenced the space that we look because we can only move in a space that has been defined by that training data. given that the training data, they say, doesn't have tomorrow's discoveries, doesn't that limit the geometry where we can explore?
Jonah Lynch (37:28)
Yeah, yeah, yeah, yeah.
Yeah, for sure, for sure. This is, yeah, I'm smiling because this is the thing that I was trying to study most specifically in my doctorate. With the image of, know, bootstrapping where you're like literally lifting yourself out of the bog by your hair, you know, which is not possible, right? But in a sense it is possible and it's exactly the history of human thought. Like.
We started out on the steps of Africa not that long ago, 60,000 years ago, something like that. I mean, even if you want to go back to the skulls that they found in Dmanisi, Georgia, two million years ago, or whatever beginning point you want to put, it's not been that long that humans have been around. And everything, all of our cultural artifacts have been produced in...
really short length of time.
And as far as I can tell, there weren't any angels that came to tell us how to invent language or how to domesticate fire or animals or how to extract iron out of stones or how to use yeast or how to turn clay into an instrument of writing and all of these things that we have discovered. And now here we are, you're in Australia and I'm...
about as far away as you can get on the globe. And we're talking in real time across video connection. 20 years ago that was science fiction essentially and here it is. So we have expanded the latent space of the model by bootstrapping our way into new dimensions and new spaces. So first of all that's to say, somehow humans are able to
expand the latent space? And then another question is, are we expanding the latent space or are we exploring it? And so are we finding things that were already there? And this brings us back to like the Socrates, know, the idea that the ideas already exist somehow and we discover them rather than create them. Anyway, really interesting questions, but.
Where that brings me is to say, if we're talking about the latent space of a large language model that has been trained on a certain kind of data, what does that say about the conclusions that can be reached?
What I think it says is there is information in the specific contents of that space for sure. So you query the certain vector, you get out the text, the original text that trained it or the original concepts that trained it. Okay, that's clear. What else is there? Well, I think there is some geometry beyond that. There is some clustering and some distance metrics that can be used inside of that latent space.
that are interesting. So we have concepts like concept vectors, where you can add a scalar to all of the vectors in some dimension and move things toward Claude always thinking about the Golden Gate Bridge. don't know if you remember that wonderful experiment that Anthropic did last year. concept vectors, okay, that's another level. What I'm saying is, I guess this is a hypothesis, right? This is something that needs to be
further clarified and validated, I think it's vero simile. It looks like it's true that novel location in latent space does have some content. Novel meaning there isn't any vector in the training data that points to that location. It does have some content which you can deduce by
it's clustering by what it's close to. And then my tool is kind of just basically saying, okay, so just take the juxtaposition of those concepts from that cluster. And sometimes you get out interesting information.
Alberto (42:07)
Really cool. Yeah, it's very exciting. Speaking of which, what was the most exciting part to developing this tool?
Jonah Lynch (42:17)
You know, I should go find out what day it was exactly, but it was something like October 2nd, 2024. I've been working on this thing for five years and I had this idea, you know, and there are lots of steps to learning enough Python to be able to actually program it and learning enough about how to manage multi-gigabyte or terabytes of data. All that, you know, all the plumbing when you finally get all the tubes together.
actually hitting run, know, okay, so do my test, validate this, does it actually work, you know, after five years of trying to make it work. I was really scared. I thought, you know, nobody's gonna know, I'm here in my room alone, but I'm gonna know, and I've just either spent or wasted a good chunk of time. And, you know, for a couple hours I was going back and forth and I wasn't hitting run, I was waiting.
And then I thought, you know what? I want to know, is it real? Is this idea actually real? And so off we go. And I ran it and it didn't work. It was like null hypothesis. It's like, man, this is terrible. So I thought, okay, well, let's do a grid search along the parameters and see if maybe I just was really unlucky and chose the wrong parameter right away. And that is what happened. I did a simple search and I found that
Basically, all the other parameters gave some result except for the one that I used at the start. Then I started to drill in and do the comparisons against randomness, against the existing distribution, and so forth. I was jumping, jumping up and down. Yeah, that was an exciting moment. I was so happy that it worked. That was the eureka moment.
Alberto (44:10)
Yeah.
Fantastic. If the innovation lens could generate a hypothesis about any unsolved problem in science, what would you want it to tackle?
Jonah Lynch (44:23)
I mean, you know, the Riemann hypothesis or quantum gravity, you know.
You know, I don't have better answers than that. I... Yeah. What would I really like to know?
kind of really like to know a lot of things and I hope that thousands of people use it and discover interesting things and useful things and beautiful things. Yeah, I'm sorry, I don't have a good answer to that question. I should make one up, so.
Alberto (44:55)
Yeah, exactly.
No problem.
If you could collaborate with any scientist, living or dead, who would it be and why? I don't know. I'm a big fan as well. I'm a big fan as well. My undergrad was in physics and I remember watching a lot of his lessons and reading his books.
Jonah Lynch (45:07)
Feynman for sure. Richard Feynman. Yeah. No, I he's got, I love his personality. What's that?
Yeah.
Yeah, no, I love his personality. I love his sense of humor. I love his penetrating, penetrating. But also, like, there are other people who are penetrating, but he's also so flexible and so mobile, so polyhedric. He's a drummer, and he would do 12 against 13 rhythms. I I can do four against five, but 12 against 13, man.
Alberto (45:52)
Yeah,
another Renaissance man. And I think in today's world of AI, I can't help but noticing that there are many computer scientists and engineering, let's say people from engineering department or well, some linguistics as well, but most of people spend
Jonah Lynch (45:57)
Yeah,
Alberto (46:21)
most of their lives in computational, some computational science. And maybe they achieved some success and suddenly start feeling like an expert in philosophy and start writing some funny philosophical statements here and there. I've seen a lot of that and I like approaches from people who come from
very different walks of life and different disciplines. They went through the depth of different fields because that's really enriching. if you in the vast breadth of experiences that you had, what would you point out as misconceptions people have today about AI in general for the general public?
and specifically in AI with incentive.
Jonah Lynch (47:25)
That was a lot. Yeah, I guess I would...
There's a lot to be gained by going deeply and studying something difficult far enough to become competent in it. There's so much to be gained in that. And one of the things that it does is give you a sense of pride in your powers, the powers of your mind. And that's a very good thing and it's also a powerful temptation to think that
Alberto (47:37)
Yes.
Jonah Lynch (48:03)
the powers of your mind cannot resolve other problems too that perhaps you aren't equipped to solve. For me it's also been helpful to have experiences in completely different domains like art, like painting, like music, like playing the violin. Because physical competence also teaches you a certain humility. You literally can't play the violin for more than a couple hours a day.
mean, you can push it, you can do like five, but you're gonna hurt yourself after a while, unless you have absolutely perfect form, which very few people do, and unless your body is made in a certain way and you can't really change the way your body is made. the experience of physical limits is also really helpful to learn what I guess is sort of just a lightheartedness that you can't resolve all the problems.
You know, like the beauty of human intelligence is magnificent. It moves me to think of the beauty. Sometimes I move to tears by the beauty of math. The beauty of a discovery. I had a problem recently that I had to solve. How do you compute a lot of nearest neighbors in a short amount of time? And when I discovered the KD tree,
algorithm is like, this is so beautiful. Just by recasting the problem in a different way, all of a sudden it's computationally totally tractable. You can do it in a fraction of a second, it's great. Whereas if you do it sort of like a bull straight ahead measuring every distance, you're never gonna get done. that kind of thing is really beautiful. And at the same time, how to handle the fact that there are some limits too.
I think is really important. So to answer your question about.
You know, if you think that you've created a god, if you think that AI is like a personality, and it can certainly fool you to think that way because it's so good at mimicking human speech, and we are so easily fooled, we're so gullible, it's just the way that we're built, that can lead you down some stupid paths, know, paths of, you know, conclusions that aren't justified and...
and are dangerous and damaging to yourself and to others.
And then what was the part about science, the last part of your question?
Alberto (50:43)
Yeah, any misconception of the use of AI into science and scientific discovery.
Jonah Lynch (50:47)
Yeah.
Yeah, I mean we've been talking about the problem of going out of distribution, that the interesting things are the things that haven't been done before, and by definition those are the things that the model doesn't know about. So, you know, that's something that's important to remember. One of the talks that I gave last year at Oxford University, and I'm going to give again soon in Slovenia, is about human creativity and AI, and one of the things I use to try to show the difference is
the movement from billions or trillions of words or tokens down to a summary, which an LLM can do, and the human movement from a very banal musical theme, like the theme that Diabelli sent out to a bunch of composers for the Diabelli variations, and then the seemingly infinite variety of responses to that prompt.
I think it's worth thinking about and to not have the misconception that science is different because I science is totally the same, that the really interesting things in science are the things that you didn't see coming. Some of them, Kepler and Copernicus, have to do with observation, just being so precise in your observations that you see a difference that nobody had noticed before. A lot of them have to do with thinking in a way that...
Einstein thinking about the speed of light and what would happen if you were riding on the light beam, that kind of thing. It's highly creative, you know, and that doesn't come out of just doing the thing that had been done before or staying within distribution. yeah, I mean, there's so much else to say there, but anyway.
Alberto (52:40)
Beautiful, beautiful. It was amazing, think it's a fantastic concluding thought to say, a hymn to the beauty of human intelligence.
So thank you. Thank you very much. It was a wonderful conversation and I'll surely share all the links about the tool that you built about yourself, cetera, along with this podcast. It was an honor. Thank you, John.
Jonah Lynch (52:58)
Thanks for having me.
All right, well thank you, Alberto.
And thank you for speaking honestly about AI. I really appreciate your headline there.
Alberto (53:17)
Yeah.
Thank you, thank you so much. I appreciate it. Bye.
Jonah Lynch (53:21)
Bye bye.
Share this post