AI and the Future of Scientific Publishing

Hello everyone. Thanks for joining us. Here at another live stream at the Academic Innovation Channel at Wolfram Research, we are joined by three very distinguished guests today. We have Victor Galitsky, who is a professor of Quantum physics at the Joint Quantum Institute, University of Maryland, also editor for the Annals of Physics and CTO of ScienceCast, which builds AI tools to process research data for archives, publishers and digital repositories. We're also joined by Richard Sever, Assistant Director of Cold Spring Harbor Laboratory Press, co founder of the BioRxiv and MedRxiv services. And we have Stein Sigurdsson, who is the Director, Scientific Director of Arxiv and professor in the Department of Astronomy and Astrophysics at Penn State University. We're also joined by Michael Trott here from Wolfram Research, Executive Chief Scientist for Wolfram Alpha. And today's topic we're going to be discussing the impact of AI on the future of scientific publishing. Now this is a very broad ranging topic, is likely to have a huge import and impact on the way that research is carried out in the future. So the format, in case this is your first time joining us for a broadcast like this, is that we're going to spend some time where our guests are going to introduce what they've been up to both in the, you know, introducing the concepts of what are archives for those of you who may not be familiar. And then we're also going to get into what they've been up to recently with projects like ScienceCast and Archive Labs. And then we'll have also a bit more freeform discussion where we discuss some specific questions within this topic. And if you're here in the audience, as always, feel free to put questions in chat. And after we go through some of the introductory materials, then we'll be more than happy to discuss audience questions that make their way into that chat. So thank you all very much for being here. It is a real pleasure to be speaking with all of you. So why don't we start off with just a quick introduction? What is Arxiv and why is Arxiv something that came to be in the first place? And so we can start with Stein discussing this topic.

Thank you. So Arxiv actually predates the web. It was founded in August 1991 over a casual conversation at the Aspen center of Physics by Paul Ginspark, who is currently a professor at Cornell. It is a curated research sharing platform. It is intended for PhD level research and the basic idea is when your paper is ready to be submitted or when you're happy with it. You can share it quickly and for free with your with other researchers and for other people to see. It is set up to be by scientists and for scientists, although obviously anybody can read it. And we also encourage interested people to read about research, as in the it's not restricted to scientists, it's just our driving audience is scientific researchers who are interested in finding out about research. And it is free to read and it's free to submit. We don't charge anything to our authors or our readers. ARXIV is hosted at Cornell University and has been for the last 20 years. It covers specific subject areas, specifically physics, where it started, mathematics, computer science, statistics, quantitative finance and quantitative biology, economics and electrical engineering system science. And we are hoping to expand further as and when we get capacity. Next slide, please. So scope of Archive it's worldwide. In fact, our biggest impact is outside US and Europe now. We have billions of downloads. Our current hit rate is a few hundred thousand per hour. And we have 5 million roughly monthly active users, distinct active users. We host 2.4 million research articles in the subject areas we cover. And our current growth rate is about 18,000 papers a month, roughly 1,000 papers per working day. We are supported by member institutions, mostly universities and libraries and research centers, over 200 of those, and by donations from individuals and foundations, in particular the Simons Foundation. It is a curated platform. It is not the Internet. It is not just a website where you can upload anything. Content is categorized and moderated both through automated tools, including some AI tools, and by over 200 volunteer moderators. And the typical volunteer moderator is a tenured faculty member. In the subject area, they're moderating. Works in the subject area, they're moderating. We currently have 153 categories and we have a governing board, actually three, three governing boards covering different aspects of arXiv. And we have, we have staff, we have 22 part time and full time paid staff. Most of them are Cornell employees. Next please. Feature of ARXIV that we started a little bit before COVID Fortunately, before COVID is Archive Labs. Archive Labs is basically a space where we can experiment to stuff, in particular with outside partners, because we recognize that we don't have the capacity in house to try out a lot of stuff that is interesting and of potential use to Archive users. And so we created a space where people can formally collaborate or partner with us to do stuff. And we currently have 15 partners covering things like recommendation engines, linking code to source, summaries of research content, and lots of fun tools. The Archive Labs by The way are found at the bottom of the abstract page. Nobody knows they're there. But if you go to the bottom of the page and click around, you'll see stuff. And the idea is to have a framework which allows collaborations and interactions and enriched content, some of which may be brought in to be in line with ar. Particular examples R5, which generated HTML papers, versions of archive content, and which we brought in house and is now part of arXiv Live service, where every paper, we make an attempt to convert the text source to HTML next. And that hands over to Richard for his time.

Thanks.

Excuse me, sorry. So, yeah, Richard, if you could please now also set up for audience. So our archive was the first one on the block. But now BioRxiv and MedRxiv are also very important within their respective fields. So, please, if you could motivate these for us.

Yeah, well, thanks very much. So BioRxiv and MedArchive are very much inspired by archive. And actually, Paul Ginsburg, who Stein mentioned, was on the launch advisory board for bioarchive. And really what we wanted to do with Bio Archive and MedRxiv was bring the culture of Arxiv to biologists and clinicians, respectively. You know, I often say, you know, the inspiration was that 2 million physicists can't be wrong. But there was a recognition that there was, you know, there are cultural differences and there are some concerns around medical information and biological information. So hence a separate server with slightly different procedures. They're both nonprofit. Bioarchive is based at Cold Spring Harbor Lab. So Cold Spring harbor is playing a role analogous to Cornell here. And medarchive is a partnership with BMJ and Yale. They're both nonprofit, just like archive. Submission and access are entirely free. Anyone can submit. The content is not peer reviewed, but it is green. So much like arXiv, there are essentially senior investigators who look at papers and say they're okay. They're not reviewing them, but they're just. They're saying they're not. Nonsense. We now have a quarter of a million papers across the two servers. We get something like 5,000 papers a month, you know, so a few hundred a day and around 10 million people looking at these papers every month. Next slide, please. And this was really our motivation for setting up Bioarchive and MedRxiv, which was this recognition that in the traditional publishing process, in which scientists will submit to a journal and then basically have to surmount the hurdles of the editor, the referees, multiple rounds of revision, and on average, this takes a year that this could significantly delay the progress of science. And actually, you know, academics are routinely seeing material at conferences and unpeer reviewed material, so they're perfectly equipped to judge it themselves. So if it can be distributed in a matter of hours or days rather than months or years, that could have a significant impact. So, next slide, please. And you can do sort of back of envelope calculations around this as to how much of the, how much of a problem this delay is. And if you were to eliminate the delay by this rapid sharing of information like arXiv, bioarchive, med, arXiv do, you could have a significant effect on the progress of science and really speed science up. And Steve Quake at the biohub in San Francisco did a few back of envelope calculations in which he concluded that within 10 years, you really could speed up science fivefold. Next slide, please. And so, you know, I mean, that's, that's a kind of backup envelope calculation. So rather hypothetical. But of course, when the pandemic arrived, we, we began to see this in real time, see concrete examples of the benefits of speeding up science in this way across BioMed, BioArchive and MetaArxive. We had 29,000 papers on Covid in those four years. And these, this started almost immediately. We got our first paper in January 18th of 2020, and we started seeing waves of submissions from China about the epidemiology of the spread of COVID the structure of the SPIKE protein, et cetera, the fundamental biochemistry of how the virus was working. And then, you know, amazingly, the clinical trial showing that dexamethasone, a steroid, was an effective treatment for Covid and the hydroxychloroquine wasn't. Then we saw details, the viral variants, the performance of the vaccines to the extent that most biologists, and certainly not the chief medical advisor to the US President, hadn't thought of preprints before. But by the time Covid came, you had Anthony Fauci advocating for this. You had Samuel Swimminathan, the chief scientist at the who, advocating for it. And, you know, and a Med Archive author told me recently that, you know, it's very clear there are people who are alive today because of this rapid sharing of information who wouldn't have been. And that is thanks to this preprint to this early sharing of information. Next slide, please. So, I mean, our philosophy has been about decoupling the dissemination of information from its subsequent evaluation and certification to kind of speed up the scientific process. But much like Arxiv labs, we have an analogous thing called the bioarchive Dashboard, where we feel that this decoupling allows experiments in peer review, experiments in evaluation, experiments in community building. So if you go to a bioarchive paper, you can launch this dashboard which has third party services in this case. What I'm displaying here is a peer review from a journal that's entirely separate from BioRxiv, but it's peer reviewing this content on BioRxiv and there is communities here, as I said, and social media mentions etc, and automated tools, one of which is science cost. And you'll hear a bit more from Victor about that later. That's me over.

Thanks so much, Richard. That really, you know, having a very concrete example of where advancing scientific process benefits people is kind of unique, I think, in some ways with recent experiences that everyone has had. So. All right, so Victor, I know that there's been a lot of great work going on in terms of how AI tools can help with a lot of these dissemination of information and how the scientific publishing process should happen. So if you could please give us a quick brief overview into the sorts of things that have been ongoing in that domain before we get into the more freeform discussion.

Of course, yes. Thank you, John, first of all for organizing this event and for inviting us to participate. So yeah, I'll tell you a little bit about ScienceCast, which you can view as a sort of AI substrate for various archives and open size repositories in general. So we have Moshe Levy here on the call who is doing most of our original AI work, and a wonderful group of people working with us. You can see them here. So I will start kind of with a brief history. So we're discussing scientific publishing. So what is scientific publishing? So it's an ancient industry, so you can go way back to the 15th century when the printing press was invented by Gutenberg and the first journal was founded back in the 17th century. And actually this journal, Philosophical Transactions of Royal Society, still exists up to this day. So the first peer review appeared in 1731. And around the beginning of the 20th century an external peer review system was introduced kind of en masse. The publishing system as we use it today. In particular, the for profit publishing industry was invented, quote unquote around 1948 by Robert Maxwell, who was rather controversial figure. And I would argue that, well, one of the best things to scientific publishing that happened recently was what Stein mentioned, namely the invention of arXiv, which is open access repository in 91 by Paul Ginsburg, who launched it in Los Alamos and arguably we are now going through another disruption which is driven by AI. And this is sort of the topic of today's discussion. But all in all, we're still using the kind of ancient publishing model. Now I would like to start with the emphasizing sort of pain points of the traditional scientific publishing, which AI may resolve, or on the other hand make them worse. And it remains to be seen what will happen. So in my opinion, a key problem facing science in general and scientific publishing in particular, is actually overflow, exponential overflow of data. So for example here what I'm showing is how many papers were published in quantum physics by Einstein, Feynman, Schwinger, Dirac in the first 50 years of the 20th century. It's about 15, 1600. In this century we have about a million papers published in quantum physics, not by Einstein, Dirac and final. And it would take a human being about 200 years to read them. If you don't do anything else, if you just read, you're not allowed to think, check references or eat. It will take you about 200 years. Now, another example from Richard's territory. So Covid. So obviously it's only we're talking, it's COVID 19, so it's by definition four or five years. So we already have about half million papers, almost exactly I checked yesterday, which will take a human about 100 years to read. So here's a problem for you right there. So that we need some new tools to process information. And the word exponential here is not just a figure of speech. So here what's plotted is actually submission rate, monthly submission rate to archive.org which arguably is actually exponential. So here is actually Covid plateau because there was no experiment. And this, I'm afraid, are LLM generated papers contributing to further increase. But all in all, we're kind of no longer capable of keeping up with the latest research, even in highly specialized fields. And the reason we started this project originally, even unrelated to AI, was try to resolve a little bit this issue. Now, kind of a big question you can ask. So if Einstein were to write his miracle year papers now in 2024, would we even notice them in this firehose of publications and data? So I think one of the key issues facing science is how to extract gems from pile of firehose of data. Now let me just say a few words about the current publishing system and then I will actually move to AI. So, so the current publishing system is a big industry. It's an old industry. And the business model of this industry is the following. So we Write papers for free. We refer you papers for free. We oftentimes work as editors for free or small pay, and then we pay to read our own papers without noticing. So this is actually a budget of a typical university. $33 million total budget, about 12 million goes to subscriptions. So it's about 10 million per university per year and $25 billion in revenues for science publishers. Actually it's profit margin more than Google and Coca Cola. So this is the type of industry we're dealing with and it has its positive and negative sides and I'll talk about that. But sort of this is what we're dealing with now. Why are we publishing in journals? Why do we actually do it? So there are a lot of pressures to publish well related to funding, promotion and tenure, et cetera, but also at some high level. So the value that journals actually provide, they're an attention focusing, basically vehicle. So if you get a paper published in a high impact journal versus the same paper published elsewhere, so the former will have a much higher impact. And so journals obviously select papers based on peer review. But many people, myself included, believe that the traditional peer review system is broken primarily because of the information overload, because people just don't have time to review, especially for free. So, and this leads to many problems and some of them I will mention later. Now, how does AI enter this picture of science and scientific publishing in general?

So, and so here in this column I will mention briefly and then elaborate on that some possibilities provided by artificial intelligence in particular. These things we actually have implemented with Archive Biorxiv and within this science case project. So an obvious1 where LLM is actually good is summarization. So kind of elevator pitch. So get to the point essentially. So if people don't have time to.

Read and process all this data, so.

AI is pretty good at kind of summarizing things. And that's one thing where we can use AI. Another I'll show you is sort of discovery, for example conversational search, sort of specialized search engine. So another thing which is actually a very big issue, in fact, I know Archive has special sort of task force to contribute to it. So it's accessibility, because open access doesn't automatically imply accessibility. So there are people with disabilities, there are people who have language barriers, and AI can certainly help with accessibility. For example, what we do is we transform the summaries into audio format and there's also translation, which is pretty easy to do with AI. So this is not a good thing. So with due to this kind of avalanche of data So a big problem. And Stein can correct me if I'm wrong, but I think it is indeed a big problem. I'm a moderator at Arxiv, so there is a kind of shortage of manpower.

For.

Moderating such a huge volume of submissions. And we need a reasonable automated initial, not final, but initial screening of the content. So I'll talk a little bit maybe about smart referee selection and some AI research assistant, but also I'll mention that on the problem side, where AI can basically make these problems worse, is that if people use large language models to write poor quality papers, and unfortunately that's happening, in fact, as a journal editor, not only I see LLM written papers, I also see LLM written reviews and responses to these reviews kind of closing the loop, which is obviously ridiculous. So that's something we don't want. So obviously misinformation. I'll show you pretty striking examples of that. So misinformation by hallucinations of AI. So I mentioned already I generated quote unquote peer reviews. And while a big issue on which I don't have an opinion, it's a copyright issue when training. So yeah, now just give you a few examples of things we actually have been doing with. For example, this is an example of a feature which we introduced embedded into bioRxiv. Well, as it turns out, most people accessing archive and bioarchive content actually not necessarily professional scientists. So Stein mentioned that they have about 300,000 visitors per hour to archive, which.

Means it's millions every day.

So there are not that many scientists in the world. So most people who read these papers actually are not familiar with the jargon and terminology. And there's some value in having sort of summaries which are tunable summaries with tunable level of expertise and we can do it and we have embedded it in bioarchive, which provides the summaries kind of tailored to a specific audience and.

You can change it.

So another thing, and this is done in collaboration with eleven Labs, which is a text to speech company and also that's how we started with Stein and Arxiv is that we provide audio summaries for papers.

So if you.

Let me just show you. So if you go down here, so.

You will see quantum entanglement is a fundamental aspect of quantum mechanics.

Yeah, so basically what happens here is that AI through API, well we get the paper, we process it into summaries, in this case at the non expert level of a reader and then transform it via text to speech into an audio abstract which helps accessibility. So another thing I mentioned is conversational search engine where you can just basically ask a question.

Primary entry point into ScienceCast's suite of tools is the conversational search engine inspired by Perplexity AI. The sources used by the conversational AI are shown here placed along the X axis in order of publication date and along the Y axis according to their semantic relevance. Each source can be used as a starting point for another search that shows similar papers. Once you find a paper you'd like to explore, you can go to the corresponding cast, which is an audio summary of the paper.

The paper proposes a new neural network architecture.

So basically here what happened is that we take papers and we embed them, which means we create essentially a vector in some multi dimensional space which is like a fingerprint as Richard put it, of a paper. And this vector embedding allows us to search for similarity between papers and or queries that people provide. And this is a useful search engine. So just I will mention so sort of another example which I personally love because I don't like preparing talks is, you know, you can just Type in Archive ID and generate for example a PowerPoint presentation on the fly. And I love it because I hate preparing PowerPoint presentations. So and it can save many hours of work. So just kind of wrapping up. So another things we're working on is screening for contradiction and similarity between papers. I mentioned that one can embed papers into vectors in some multidimensional space and then look at similarity if vectors are kind of almost aligned or in some subspace if they're 180 degrees, which means they contradict each other. This is kind of red flagging. So predicting potential impact of papers through new metrics unrelated to citations because citations often lag and they're oftentimes ceremonial citations. It has actually interesting also applications to other spaces like government. US government is very interested in potential impacts impact of various events, not necessarily papers. And some algorithms that we develop may be of interest to that too. So flagging low quality content and fraudulent data, tracing ideas and facts to the original sources and potentially help in peer review. But not, not peer review, but for example recommendation engine is actually kind of dual to the problem of referee selection. So on the negative side of things, so PI is creating and that's not what we do, but that's just happening organically. So there are papers in this case they actually peer reviewed papers which appeared in the journal. So and you can see here that it went through peer review with flying colors and it says please note, as an AI language model. I'm unable to generate tables. And perhaps the most well known is also another paper which I think is one of the most downloaded recent papers in Richard's territory is this paper where the image was generated using artificial intelligence. And I would, you know, if Richard wants to comment on that, please do. I'm not an expert on this particular issue, but both these papers were retracted. But this is just the beginning of this negative, I guess, trend where there are papers being generated by LLMs with very low quality. I think I am out of time and so I'll just kill my screen share and I'm done.

So thank you, thanks all so much for that sort of introduction to some of the, both motivations and some of the recent work that's been going on. So there's a lot to unpack here. We also have a couple of audience questions already. So building off one of the audience questions that we have, let's back up a little bit from the topic of AI and just talk about how one thinks about preprints in the context of disseminating scientific knowledge. The audience question is that, you know, the, the term academic paper and therefore the form in which we disseminate knowledge carries a lot of baggage from when the printing press was the best that we could do. Right. And so the, the question is.

How.

Does one think about the fact that now we have the Internet and electronic means of disseminating this and how does that bear on the relationship between the value of preprints and the value of, you know, peer reviewed published publications? I wonder if we could briefly just touch on that topic. So let's maybe go in sort of the same order that we had gone in before. So we'll start with Stein. So how does one think about the relationship between preprints to peer reviewed things given that in the past all one could do is have a printing press, but now we have these, these new means of disseminating information.

So obviously this is something we've considered quite heavily. My, my take on this is that when you're, when you're ready to share your research, you need to accept the fact that you may have made errors. There's no, it's not, not a sin to be wrong. And the whole point is to get feedback on your research and get correction and input for improvement. In practice, the majority of preprints have relatively small changes between coming out on archive in particular and being published that's actually been studied several times and consistent results. Although as I noticed, the introduction of the word not is a Very small change to a paper, but makes a very big difference to the conclusion of the paper in principle. So the size of the changes is not necessarily indicative of the impact of the changes. This depends a little bit on what research area you're in. There are different norms in different fields about what is important. Mathematicians, for example, are very concerned about proofs and formally show that things are correct in some strict sense. Others are more. You want to get the data out there. In astronomy, for example, there's high value in unique one time data being shared because it can be used over time and in conjunction with other data. The other way I think about it.

Is.

It'S not just a matter of correctness, it's also a matter of whether things are interesting. And that's really where you want to get the speed of things going, where you want to get the conversation going quickly and shake out what's interesting and what isn't. There's stuff that's incorrect but interesting in the sense that it highlights issues that should enter the scientific conversation. And there's stuff that's absolutely correct and of very little interest. And sadly that is a fair amount of research, because you don't know often what's going to be interesting until you try it. And famously there's a need to produce negative results. I tried this, it didn't work. Everybody don't bother trying this.

Yes, those are all excellent points. So one of the things that you brought up is this difference in norms between various fields. So perhaps, Richard, you'd be in a great position to comment on. So in the context you already alluded to in medical research, perhaps there are some different processes that are necessary for thinking about the relationship between preprint and publication in those fields. So could you elaborate on that a little bit for us, please?

Sure. I mean, I would, I would certainly echo what Stein said and sort of, before I say anything, underscore the importance of the ability to revise in response to feedback and the importance of kind of peer criticism and that whole kind of like sort of dialogue that goes on. But I think one thing that's really important in the context of that original question to understand is, is the history of scientific communication in papers. A lot of the time there's this sort of somewhat false narrative that, you know, 300 years ago journals were created much as they are today, and that everybody's been doing this for 300 years and it's never changed. The reality is actually that peer review, as we currently perform it, is only about 40 years old. So, you know, The Lancet, the premier medical journal, and Nature Science, all those journals only started doing what we now call peer review after World War II. Actually, what happened prior to that was the papers were refereed when they came to journals. And what they. That was mainly by societies wanting to make sure they weren't putting stuff out. That was really crazy and an embarrassment to the society. So, funnily enough, that actually looks like the kind of screening that Bioarchive and Med Archive perform. But then what happened was because of the explosion of research after World War II, there was a real drive for papers needed filtering. There was concern about this concept of peer review because the politicians might co opt messages so peers had to evaluate them. And you've got the conflation of a whole bunch of different things. So it's really important to say, you know, the thing where that, you know, in Victor's timeline, the thing we're changing is not this thing that's been happening for 350 years, it's this thing that's been happening for 40 years and has become a. You know, I think one of the dangers is a distortion to the academic process because we end up having this conflation where we equate the importance of the work, the quality of the work and the impact of the work. And, you know, and the worry is that people actually say, oh, you know, there's a phrase we use in biology quite a lot, which the journal that you publish in is more important than what you publish. And it sounds facetious, but actually for people's careers that's probably true. So I think that you have to set against that background. And then there's the other issue, which is when you get kind of clinically actionable information and you have worries about misinformation. So one of the things we're very keen on at medrxiv is we have a kind of category of paper which we say is better after peer review. There's some things that could immediately change public behavior that you really want to be a bit more sure of. And so, you know, there's something. So I'm all for rapid dissemination, but there's a class of paper that shouldn't be rapidly disseminated. But I always point out that's probably like 0.01% of work. And we need to have a conversation about educating the public about the nature of scientific evidence, at what point something becomes actionable. You know, the FDA don't approve a drug just because it's appeared in a, in the New England Journal of Medicine. So that's a more complicated process. So that was a bit of a long winded answer to your question.

No, thank you. So, Victor, I saw that you wanted to interject about one of these points. And then also as a follow up, you also brought up this, this couple of different things that get conflated in the way that things are done currently in scientific publication where there's this question of first is it nonsense and then is it interesting? So what did you want to add to Richard's comments? And then also if we could comment on how do you think about deconflating these various things?

Yeah, so I do want to comment a little bit on the current state of peer review, but more as an author and referee and also an editor who deals with this. So I recently asked my colleagues, and my expertise is quantum physics, have you ever received a useful review ever? So I've never received actually useful peer review and most people said not really. So it's more like, at least in our field, it's more like exercise to get things published than to seek the truth. And I think, I don't know why it's the case, but I think the reason for this, I mean, I think it may be different in biology, it certainly is different in mathematics. It is very much field specific. But I think the real reason for that is because, well, let's say we spent a year writing a paper, we submitted it to a journal and then editor sends it to referees. Referees don't have time, they don't have a year to check our results. So they have five minutes. They look at it superficially, they may have their own biases. And I am also in these shoes. So people don't have time to do work for free, which is actually difficult work, which is due diligence basically on very complicated research. So I think one big question to answer in general is whether the current peer review system is working at all. And I think it's very much field specific. And mathematics, which is a smaller community where there are well defined problems that mathematicians consider important, I think they actually do a pretty good job in verifying their proof. So I have a question to mathematicians though. How do they select which problems are important and nobody can answer? They say, oh, Hilbert told us 200 years ago, 100 years ago. But if you start asking why, nobody can actually answer. So it's a bit of a club. So in this situation, I think peer review works. But if you have a fast developing field, very large field with a lot of competing interests, et cetera. So peer review needs to be rethought in my opinion. And I also certainly agree with Stein that interesting and correct are not necessarily the same thing. So there have been a lot of high impact papers which were originally wrong. For example, the first paper by Abrekosa on vortices was wrong. It predicted. He got Nobel Prize for this. He predicted a square lattice. It was supposed to be triangle lattice. So there was a paper, there were many examples of papers that got eventually basically Nobel Prize that were actually wrong at some level. But there are also a lot of papers which are uninteresting and wrong. So those we don't want for sure. So but I don't have a solution actually to this very difficult question. But I think it's important to recognize that this problem exists.

All excellent points. We have some more audience questions. Before getting to some of these audience questions I want to also ask Michael so you spend a lot of time learning what there is in the very rapidly evolving field of large language models and how those are relevant for building AI tools and part of the dissemination of information through preprint seems to me to be critical to this process. So I was hoping you could comment on that. Quickly moving aspects of why it's important to have a dissemination mechanism that can move at the speed at which people are actually having the ideas and communicating about them.

I think the mentioned five fold speed up is probably an underestimate. If you think on LLMs, what happened over the last I would say for most people it became visible maybe in November 22nd with some of the Galactica papers or Minerva. Some people outside of the direct CS field noticed this and then what happened the last 14, 15 months is pretty amazing and I find it nearly daunting by now to look on Monday evening what has been published in the arxiv CS. But I think last Monday it was 910 papers. So just reading the title and have a look at the abstract, what's interesting takes me one and a half hours to go through on the other side it really is the only way I can see that one decimates the knowledge quick enough and sees what people do and the breadth of what is done in the especially LLM world. On the one side on the actual implementation of LLMs and techniques, but also on the application side and the whole prompting technique. Something that I think most people would have never thought would become important that you spend a fair amount of your time thinking how do I ask a machine? Something like an old science fiction book would have probably always thought, you just ask the machine and the machine will answer it. But now it's so that nearly every day there's at least one, not every single one is sensible, but one useful paper about prompting techniques. Like there was this one on Monday that really on GDP 4 proved that with a little bit more than 1%, with sufficient statistics, you get a better result if you have enough new line characters in your prompt so that your prompt is sufficiently structured. And I think it's very important that all these little things get distributed quickly in the world so that we can stand on the shoulder of giants and quickly, within timescale couple months, come to really new levels of research and usage of the things.

Richard, it looks like you want to make a comment on this topic as well.

Yeah, I mean, I just thought I'd give an example from biology of where that kind of where things are moving too fast for the current system. And one, you know, many years ago when we had the first SARS outbreak, somebody pointed out that 99% of the papers about the SARS virus were published after the epidemic had ended. And of course this was actually a very, this was a call for, you know, something like arxivin in biology. And of course, come Covid, we had exactly this experience. And there was, there was examples where, you know, people pointed out that if you were writing a paper on the beta variant of the SARS CoV2 variant of the SARS CoV2 virus, by the time that was published, the beta variant no longer existed and everybody was infected with delta. Whereas. So if you can disseminate your material within 48 hours, then you actually have a time to be talking about something on time concurrently.

There are two audience questions which are very related to points that all of you have alluded to and Victor explicitly stated. So one question is more of a comment actually that's one of the great things, is that anyone can look at these preprints and get direct access to the most recent things, right? Or perhaps actually it was Stein who had mentioned that the volume of people who are looking at these services are greater than the number of scientists that are out there. And so, you know, just curious, people who are out there are able to actually gain benefit from this. And then a related point is answering this what's interesting question. So we have a question about looking at citation networks and what might be some ways that as we start to segue a little bit into talking more about AI tools and how this is perhaps going to shape things in the future. What you mentioned new metrics. So let's start with Victor. What Sorts of things might there be out there that are either being actively looked into or are sort of in the early phases of thinking about what would be other ways of trying to solve this problem that you posed of? Let's say that the equivalent of the Einstein miracle papers were written amongst the sea of other things. How do we make sure that those are discovered in a way that makes them useful to people?

Yeah, so I think actually there is a very high probability that there are in fact Einstein papers being written that people have missed. We just don't know about it.

So I find that plausible.

Yeah.

So how to solve it is a billion dollar question. So but I think some simple ideas that I'd like to comment on. So first of all let's look at arxiv. So I think Michael Trodd just mentioned that there were I think 910 papers if I'm not mistaken in LLM just in one day. So it takes Michael hour and a half to read just the title. So obviously it's a very time consuming exercise. So it's similar for all fields. Right. So and at the moment ARXIV papers are just time ordered so there is a cutoff for submission and it's time ordered and there is this pretty curious kind of data and I think Stein will correct me, please correct me if I'm wrong. That paper that appears first is cited 10 times as much than paper that appears at a number 55 independently of the content. I don't know Stein, can you comment on the actual numbers?

The ordering effect is well known and studied. The it's not as simple as you think because there's also a factor where people who are motivated in certain ways work at this they game it and that correlates with in principle with the impact of the result. There's basically people who are fired up and think they have a high impact result or from or from a large well structured research group organized to take advantage of this and that co correlates with the impact of the research anyway so it's not a trivial problem to de factor we are aware of it and all the solutions we thought of are worse than the problem.

Yeah, but arguably of papers is not necessarily the best indicator of disruption and an obvious, I think a solution that AI could help with is at least some sort of AI assistant ordering of.

Papers which is specific to a user.

So if I'm interested, listen quantum physics and not interested. I don't know first principle density functional theory. So then I won't see the latter papers, I will see them in the.

Bottom and I will see what interests me on top.

So I would say that attention should be specific to the user and.

That.

Clearly is something that needs to be automated and it's not that difficult actually. So in some sense every time you go to Netflix or something, so it presumably shows you movies that you're interested in as opposed to those that you're not interested in. So I think this is an obvious thing that needs to be done. In fact, we are actually working on that.

Archive has not implemented, that is there's a computational cost to it. We actually struggle to get our current simply ordered output out to the users every day. And the other reason is we, per our advisory committee in particular, we worry intensely about siloing, that we narrow people down too much and they become disconnected from stuff they should be aware of. That is a broader issue in recommendation systems. It's not just specific science and we would are very interested in solutions to that.

Yeah, so but on your question, John, also I think you asked what are.

The predictions of impact? So first you have to define what impact actually means. What does impact? So I mean, I guess person not getting a Nobel Prize is impact, cure.

For cancer is impact.

But I mean that I don't think is quite possible to predict on the.

First day when the paper appears.

But if you define impact as how it affects scientific community. So I think Stein, please correct me if I'm wrong, but according to Paul Ginsberg, actually initial download rate of papers correlates with their long, long time impact as defined by citations and or long term downloads. Is this correct?

With the exceptions of some extremely famous outliers that are well known, Weinberg's initial paper had very little reaction for a couple of years. There are a number of things that are sleepers and then suddenly take off. So it's again field specific and also level of impact. Papers that are steady, sustained citations over multiple years in aggregate can be more impactful than things that have very high spikes of citation and then taper off because they get obsoleted.

So we have a follow up audience question which is related to this point. So one of the things that can help with preventing siloing sometimes is simply people in community talking about various topics. And so we have an audience member who's wondering about this idea of having what they're calling a micro forum, but. But basically a way for people to have discussions around particular preprints. Is this something that either is part of Archive Labs or Sciencecast is something that people can interact with right now, today? And if so maybe we could help direct people to where they can find such things.

Yeah, so we have this capability. I mean, to be honest, it's not very widely used at the moment and people do interact with authors and that's actually not only community, you know, just.

Readers among themselves, but most, most importantly.

With authors, direct access to the authors. But I think there is also a problem with inertia. Like if people spend, I don't know, 90% of their time on Twitter or Facebook or wherever, so they just keep doing that. And unfortunately, at least in my opinion, I mean, I just don't understand Twitter. I think it's a. In my opinion, it's just spoiled with a lot of low quality information. It's not necessarily the best forum to discuss serious research. So we, I think, badly need. I agree with whoever asked this question that we absolutely need a forum which is free of white noise where we can discuss research. Actually, mathematicians do have such a forum, it's called Math Overflow, which I've used and I think it's excellent. But in their case, and you can see people like Terence Tao and other fields medalists chiming in on questions. So this is, I think, a model to be replicated in other fields which, yeah, we're trying to build and connect it to Arxiv and Stein. I don't want to. I think you should comment on Archive Labs and how you see integration.

So, so part of, part of Arxiv's motive for setting up the Labs framework was to have spaces where you could be tied into archive and provide services that either we can't provide with an archive because of lack of resources, we have a small number of people and a small amount of money, so we focus on our core competence and keeping everything alive. But also there are some things that are better done outside the core. In particular, discussion forums are notoriously difficult to manage. First you have to get into critical mass, which is an interesting problem, and anyone who solves that will become very rich. And the other one is how to stop them from degenerating into a toxic swamp. And to be honest, we just don't feel the capacity to police that or take the responsibility for it. So we, we can't.

Yeah, I mean, we kind of. We've had some similar experience. I mean, I think my feeling is that, you know, I mean, as was said before, it's a kind of a billion dollar question about how to, how to make this happen in one place. Our feeling is that we couldn't do this. So we have like a combination of you know, recognizing that the discussions will happen in different places, but being able to point readers to where those discussions are. I mean, I think, you know, I mean, what Vector says about math is definitely the case. Twitter has been a strange place for genomics and actually very early on in genomics there's a lot of people on Twitter. So you actually did. I mean, things have degenerated in the past few years, but early on there was a lot of conversations in genomics where people saw that very well respected people in the field were on Twitter and having serious discussions. But that has not been replicated in immunology. And the minute you start discussing vaccines, then that kind of deterioration happens almost immediately. For example, we, one of the things we do in our dashboard is we do have a comment section on bioarchive and on site when there are some discussions there. It's not heavily used, maybe about 5% of papers. But you can go to conversations on Twitter. We want to be able to go and link out to the conversations on Bluesky, on Mastodon and we have some of these. There are a small number of self organizing communities. A couple of things sprung up during COVID Like one was Rapid Response COVID 19, which is an initiative from MIT. There was an initiative from Mount Sinai here in New York and Oxford where they were like, there's these Covid papers coming out. There needs to be some expert interpretation for all these lay readers that we've all observed are coming in. So there are a small number of these forums. But a real challenge is, I always say the most precious commodity for an academic is time. And so, you know, it's what is what, what is in it for them. And so people will engage if they can really benefit. But trying thinking that we can sort of recreate the peer review process, that may be tricky because you know, there's time is of the essence for all these individuals.

Victor, go ahead.

Yes, I just wanted to provide an example, John, of recent paper which was peer reviewed, quote unquote in the social media space. And actually it was launched on sciencecale. So I'm moderator, archive moderator in superconductivity. And I believe it was in July 2023, Stein texted me saying, oh, there is this paper, LTA 99 which claims to have discovered room temperature superconductivity. And the authors mentioned the video, which was not actually in the original submission. And I contacted the authors and ask them for the video and I put it on the website and linked it to archive. And then in the middle of so and I Forgot about it sort of. And then in the middle of the night, so I woke up because I got a notification on my phone saying that I owe, I put it on my server, I owe, you know, thousand dollars because there were a million people who looked at it. So. And as you may have heard, so this LK99 became a sensation. So we had a lot of traffic on the site and then it kind of migrated to Twitter and where there were a mixture of high quality quote unquote reviews and lower quality comments. But all in all, I think very quickly people, you know, concluded that this was not actually a true result well before it went through peer review. So I think for such potentially high impact papers, whether it's Covid or high room temperature superconductivity or something like that, so we definitely need kind of fast response peer review system which journals just don't provide. And yeah, I guess we're trying to build that within Archive labs. Yeah.

One of the tools that we have an audience comment that they're particularly excited about is one of the things that you showed of making the content on a slider where you can say for experts or less for experts in terms of what the summary looks like. And Richard also just brought up the point of, you know, suddenly a few years ago, many more people who are not experts in immunology had an interest in immunology. Right. And so I wonder if maybe we'll start with Richard and then the table can comment on what you see. The relationship between these AI tools and that very important topic of taking something which perhaps there's a topologist who's interested in large language models or the reverse is true. How can we use AI tools to make these things accessible to people who are not currently an expert in whatever field it happens to be.

Well, I mean, I guess the, you know, the obvious thing and a point that Victor frequently makes that, you know, I think having seen some of our articles on bioRxiv, is that the propensity to use jargon and to use acronyms and abbreviations that make something impenetrable, it's so common in biology because it's a shorthand, we know why people do it. So one of the things that, you know, it was noticeable when we started looking at some of the lay generated summaries from Science Cast, how they were so much more readable than the abstracts and in part was because it wasn't stuffed with, you know, acronyms like PDK1 that you wouldn't know anything about. I mean, I guess the, I Mean, the one thing that I would say is, you know, I always very keen to point out in bioarchive in particular, the vast majority of these papers are incredibly arcane, of interest to a small number of experts. So we shouldn't be under the illusion that we can somehow give somebody a kind of like a, you know, a rapid BSC and four year PhD to allow them just to understand a paper. But I think on the medrxive front that will, that will be interesting. But I think that goes to this broader question of providing context and how we do it. And I think there's a lot that AI contribute potentially there beyond just summarization.

Stein and Victor, do you have any additional comments you want to make on that topic of making things more accessible?

Yeah. So in addition to kind of providing information about things to general public, one of the things that ARCA we're acutely aware of and are hoping that these tools will expedite is for example, there are known unknowns for people. And my canonical example is statistics does a lot of research on obscure statistical techniques and then they write R code, dump it into the R library, it's validated and it floats there. I know and I can't think of a specific example, but I know of examples where somebody has a data set that is really. They need to know the statistical technique. We can help you find it. It'll enable you to get a actual robust result or solve a problem that is intractable for you. And I'm actually acutely aware of it because when I was a young postdoc I really needed to a statistical test on some multidimensional data and we had to figure out the hack up one by ourselves because I couldn't find one in the literature that did it. It turns out years later I found it and we sort of did the right thing. But even more intriguingly, and that's where I think some of the stuff Victor's working on are the unknown unknowns, the research direction. We're blind to the gaping hole in our understanding. Just nobody's looking in that direction. And sort of my dream is that we will take the set of knowledge we have and where it's moving and see where we should change direction or we should stop looking because we had to hit a wall and really find ways to expedite discovery even faster and better by finding the stuff we really hadn't thought to think of. But that's getting science fictional. But it's fun.

That's a great point that you make Stein, about how the the accessibility question is not just for lay people. It's also for experts across domains. It's a really subtle and interesting point to bring up there. And Richard, please.

Yeah, I mean, I think that's an amazing. The idea that people will make connections across disciplines and sub disciplines that they wouldn't make because you've got this. I mean, if you're trying to, you know, if you're working on quantum physics or you're working on pancreatic cancer, you can't read everything in your own space, let alone things that could be relevant from elsewhere. And I'm always, always struck by. Several years ago, I remember going to a Data in Medicine conference where there was a radiologist talking about identifying melanoma. And if you've ever looked at, like, melanoma versus benign blemishes on skin, if you're not a radiologist, it is almost impossible to tell. There are things that look horrific, that are benign, and there are things that look sort of pretty benign and are melanoma. And you have to have a lot of training to do this. And they had trained an AI to do melanoma detection and got to the point where this was pretty much almost as good as a trained radiologist. But what was very interesting was at one point, the radiologist who developed the tool recounted the story of having a patient who he'd seen, and he said, don't worry, there's nothing to worry about, this blemish that you have. And then he was like, it kind of looks a bit weird. I haven't seen something like that before. I'm just going to run it through the eye, just aic, just to see what it says. And he ran it through the AI, and the AI came back and said, it's melanoma. So they biopsied it and it was melanoma. So the AI was seeing some dimensionality to this that the expert radiologist, despite years of draining, couldn't see. So the idea that, you know, you can imagine at scale the kind of things Stein's talking about, with connections being made and avenues that the human being, through lack of time or simply not being able to make those connections seems. I mean, maybe it's science fiction, but it feels like that should be something that we ought to be able to at least try and pursue.

I think there's also another dimension that I have some hope over the next years that I will hope a lot. So there are a lot of things that interest people, but they just don't have time to follow up on it because they have a daytime job in a totally different area like myself, interest, say in quantum friction, quantum embezzlement or some other a little bit exotic things. And if I have time, I read one or another preprint from this. But would it not be great, Kenda, if really I could say, take the last two years of archive papers on this subject, put them all coherently together and make me, in the spirit of what Victor showed at the beginning, some kind of PowerPoint slide, what has been found out over the last two years. I think that would really speed up potentially cross connection to other sciences and spoil some ideas of what you can do in your own field.

Yes.

And Victor, it looks like you wanted to make a point about this topic as well.

Yeah, so I just wanted maybe if you let me share a screen for a second, I want to show how it actually works in practice. So this is, for example, a paper by Dennis Urs, who is Vice Provost for Research at Johns Hopkins.

So here is the paper.

And so he. Well, he obviously knows what he's talking about. So here is how it works, basically so that we take a paper we actually access by archive every 15 minutes. We take all new papers when we process them into summaries at two levels. We could do it at 22 levels if need be. So but this station is two levels. And so this is kind of simplistic and this is expert. I think this has a lot of value to people, especially those who are not familiar with a particular topic. And we would like actually to expand it to other archives and repositories. But also I'd like to mention so on your question, John, about making connections between different fields. So I don't know how to do it, but I'm thinking that potentially these embeddings that we work with could look for hidden links between seemingly unrelated topics, maybe. Obviously embeddings, you know, they project semantic data into just numerical vectors basically. But presumably there are some subspaces there where they can intersect and one can look for maybe something that is invisible to humans, but something that can potentially be of relevance to people from different fields and say, hey, look at this paper. Even though it sounds irrelevant, there is maybe a connection. So to develop something like this could be very powerful. I don't know, Michael, maybe just had.

A discussion with somebody from the U of I about this topic just last week. Just some random idea that we discussed was maybe the information isn't in a single embedding vector, but maybe the information is in, in the sense of topolog analysis. Between a bunch of preprints about one field and a bunch of preprints from the north side. And then look how they topologically kind of form some simplex in some Howard space that has some topological similarities.

Yes, and find the holes.

Yeah, exactly. Higher dimension at once.

Yep.

Richard, please.

Sorry, Just very quick. I mean, one way in terms of accessibility, I think another point to make here is a much more simple one. I mean, I'm conscious on this call that I think I'm in a minority of people who are speaking English as a first language. As a kid I lived in France. When I was a teenager I lived in South America. But I'm aware of the challenge that I was never able to rise to in the way that three of the people on this call have done. One of the things that often comes up a lot in scientific discussions is the fact that just like the web, everybody's using English. It's the universal language. And this is a challenge for an awful lot of people. And when you try and solve it by saying, oh, we should translate things into French and Spanish, then people say, well, what about Hindi and Mandarin? And then you just keep going until you know you're not going to solve that problem. But I really has the possibility of solving that problem, both from the author perspective. Authors are often at a disadvantage when writing. They know that people speaking of the first language often have an advantage here. And you know, I know colleagues who, when they have postdocs who come, whose English isn't so good, they use AI to make their papers much, much better. So from an authoring and kind of consumption point of view, I think there's a huge amount that I can do right now. And that's, you know, that's probably the lowest of the low hanging fruit.

Yeah, we're seeing that too. We're seeing people clean up their particular introductions and method sections using AI tools to make the English less tilted and flow better.

It also leads to hallucinations such as was shown.

You have to be careful. There's a difference between generative AI where you're making stuff up and AI as a sophisticated clippy, cleaning up your grammar and getting your prepositions and whatnot.

Yes, this is a great time to bring up this topic, which I'm actually a little surprised by the lack of audience questions on, but I think is an important topic, which is that we've been describing humans using AI in order to help them accomplish things. That's the good outcome that people at Archive Labs and sciencecast are working on. Could we Maybe help draw a clear line between. And Victor showed some very striking examples of where we've seen the bad outcome of AI getting involved in this whole scientific publishing endeavor. So can we maybe delineate between what, what are some of the things that we don't mean when we say AI and scientific publishing and have a positive connotation with that versus the things that really we do mean that we've been discussing? So Victor, you showed some quite striking examples. Maybe we'll start with you. So what do we, what do we not mean when we say AI in scientific publishing?

Yeah, so I think essentially at the moment, I guess the main problem that we are facing with using large language modeling in particular is hallucinations.

Right.

So. And the probability of those is never zero. And I guess if you're talking about quantum field theory and you know your AI hallucinates a little bit, well, it may be embarrassing, but you're not going to kill anybody. So if you're talking about a vaccine for. There was a recent paper, I believe, related to that. So. And there are some hallucinations about medical matters. Well, that's a different territory. Right. And in general we want to minimize hallucinations. So I don't obviously I'm not an expert on training large language models. It's obviously question everybody's aware of. But one simple solution we came up with with Richard and his co founder John, specifically for Biorxiv, is that we let authors edit the summaries. Actually human input. I love AI, but I actually kind of like humans too, even more so. And if there is some human input, it's good actually. So I would rather have authors go over this and correct it if there's something wrong. I think it's been actually quite successful experiment. So Richard, please correct me if I'm wrong, but I believe people have been using it. Right, so.

Absolutely.

So, yeah. So this is kind of, maybe it's a band aid kind of solution, but you know, I think it's a hybrid of AI generated content with some sort of sanity checks from humans, actual authors of the paper who own the content, by the way. So that would be successful in this, in this direction at least. So. But my personal opinion is that I think it's not a good idea to use, it's again, my personal opinion, large language models and AI to write papers. Because even without large language models, we're bombarded with papers. We don't need more, we need less. And I think that's where the strength of AI is focusing your attention, not making More stuff up. But maybe I'm wrong. So I don't know.

And Michael, you said quite a few times you are a proponent of using language models, the things they're good at. And when it's something that the language model maybe isn't the right tool for the job using other tools. I wonder if you could quickly comment on that point.

Yes, like for some of the project that we are using large language models in house, I always say do not be lazy and say here large language model do everything carefully. Really do it only there where you can't do it with programmatic way or large scale manually thinking really this simple one. And I was wondering like one should maybe even use the large language models if they would be a little bit more up to date. Now they're at least half a year, maybe a year old. And even my experience is if I asked for a specific archive blueprint, just sometimes heard about the paper title, but typically it cannot tell me what's in section 4. It has no idea about it. If it would be really more knowledgeable about the whole ARXIV corpus, which with more than 2 million is really substantial. Giving it a paper and say hey, tell me what is really new in this one and tell me all the things that I can already find somewhere else that might be a possibility to reduce the overflow of new material.

Stein, please. Yeah, you want to make a comment on this point?

Yeah, so a couple of comments. One is I'm really looking forward to a human reinforcement. Human learning by AI reinforcement kind of close the loop on us. But on the more general kind of abuse of generative AI and language models, we noticed that the translation capacity was very early abused. People would translate paper into another language and then back translate it and then resubmit it as a new paper because the transliteration tended to avoid plagiarism detectors. But it. But it triggered. I can't remember the technical phrases but. But basically synonyms for technical phrases became a anomaly signature tortured phrases. Is that tortured phrases?

Thank you.

Yes. Like artificial intelligence became fake brains or something like that. But then we. There was actually was a prominent social media. There was a paper on arXiv which was a copy of another paper paper where every word essentially had been transliterated and the author of the paper spotted it because they are in fact the world expert on such techniques and AI. And this unveiled not just a minor industry doing this, which we are fighting, but it also fed into citation circle fraud. So and in fact fake science identities being set up on science communities to generate plausible Backgrounds and stuff. And this is a systemic problem for us, not just arxiv, but for everyone in science publishing. And it is driven not just by the technology being available, but by the fact that the whole system provides perverse incentives or rewards. And that is one of the near term, possibly catastrophic impacts of AI is basically human abuse of pretty basic AI technologies to invoke Goodhart's Law with vengeance.

I think it's certainly highlighting a few things that, you know, people have kind of known for a while but been unwilling to deal with. You know, we're seeing this paper mill problem that Ein referred to, where people are generating entirely fake papers because they need them as career currency to get promotion. Similarly, citations. There's now this phenomenon, citation mills, they're called, where fake papers are created purely for the purpose of boosting other citations. And of course, what this is telling us is that, you know, we have our kind of incentive structured wrong. And people should not be acting as bean counters in academia when they're deciding who to promote and who to hire. So that's probably a good thing. It doesn't really solve the problem, but it says you've got to really, you've got to really think about that. I mean, similarly with misinformation. I mean, there was this sort of farcical example that Victor showed of the rat image that, you know, anybody could spot. But, you know, there's probably a lot, a lot of other stuff out there and embarrassed people have. You know, there's a lot of. There's a lot of things that, you know, humans are already doing really badly. And the question is whether or not AI is going to be just as bad bad. I mean, humans are incredibly good at creating misinformation. We don't need AI for it. So it really just focuses on some of these, these issues. So maybe that's just one thing AI is doing, is saying, you know, we've got to think hard about these things.

I like to remind people that people also hallucinate and, and AI has just sped the process up.

In some ways, they learned it from us, most certainly. So another thing that the audience has brought up that I want to make sure that we touch on a little bit before we come to the end of the broadcast is that the process of scientists thinking about their research is somewhat different from what comes out in just the preprint and then later publication steps. In other words, there's a lot of collaboration that happens before even getting to that. Are there any things that either science cast archive labs or just in General would be interesting to say about how could some combination of AI tools and the great dissemination methods that we have in terms of the various archives also maybe show some of the thinking process that goes into even creating the preprint or is that sort of before when the problem sort of gets picked up and run as far as these organizations are concerned? Yeah, Richard?

Well, I mean I'm going to sort of answer a bit of a broader question, but one of the things I always want to say, particularly going back to that point about the bean counting, is to make clear that what goes on arXiv and what goes on bioarchive is a narrative.

Right.

It's not the experiment, the astronomy or the, you know, the, the gene sequencing. All the data is some somewhere else. And I think maybe, you know, thinking about that, the kind of network that Victor described, the network is not just papers. Each of those nodes is collected to a different type of node which is loads of data, you know, and in clinical, in clinical medicine it may be connected to a trial plan that is registered two years before. So the possibility of creating a much bigger network that can be explored which links not just the narrative but the data, the provenance and research plans and the kinds of things that you're thinking about. I think that could be very productive.

Yeah. So maybe I can comment too, John, on your question. If I understood it correctly, you're asking whether some AI tools could be helpful not only in disseminating information but in actual research.

That's part of. Yeah, I think that's implicit in the audience member question. Yes.

So I mean at this stage, I mean in my field again I'm only, I'm an expert hopefully in quantum physics and ganess matter physics. So I don't see it too much, at least on the theory side and experimental side, too much activity in sort of AI driven research. So nobody types in ChatGPT, please tell me what problem to solve next. I mean I think it would be kind of a dystopian scenario, frankly. I hope never happens. But, but there are a lot of tools which facilitate things, things that I don't want to do, basically like organizing data, processing data. Well, I've used quite a few times Wolfram's plugin chatgpt so just to run some simple senti checks. So on some calculations it's extremely valuable. Obviously I use Mathematica all the time. So these types of tools, so they're just an extension of already existing computational tools and other tools which may be sort of an assistant to a researcher. So I don't know when artificial intelligence will become competitive with the researchers. So. But I think the more abstract the field is, the more chess like it is, basically the more formalized it is, the quicker it will. AI will basically beat humans. So fortunately or unfortunately, I think for me it's fortunate. So in our field it's also partly a social activity and what's important, what's not important, it's a bit of a community consensus and also it's driven by, I don't know, commercial interests, et cetera. And I'm not an expert in biology or medicine, but I assume it's similar. So it's less formalized and it's less of a proof driven. Like for instance, in physics, very rarely we have exact proofs of anything actually. Like even Feynman path integral, which is a key ingredient of quantum mechanics. Mathematicians say, well, it's just wrong, it makes no sense. And I use it all the time and I'm just fine. So basically my point I guess is so I think the fields which are very formalized, pure math, I would be worried about pure math, but about less formalized fields, maybe less so in terms of humans losing to AI. Yeah, I spent a fair amount of.

Sorry, sorry, go ahead.

No, you go ahead.

I spent a fair amount of time over the last year, kind of mostly with GP4 testing out of limits, kind of what it can do, what it cannot do. And I give this victor for the next year, probably we don't have to be worried, it will take over and beat us in some ideas what we should do in quantum mechanics. I think it's far, far, far away from this. But one thing that I noticed, and in general, if you ask it, give me something interesting about this, it just does the opposite. It falls back to the most simple thing it has somewhere heard really boring and probably because it has read it 120,000 times on the Internet. But what I found already right now interesting is if you give it a couple of disconnected things or for the something that still I'll say always in science, quantum physics, plus minus and say, I have heard about this, this, this and this. There must be some relation about them, say weak measurements and relation to quantum energy teleportation. And I encourage it actually to have hallucinations and come up with something. Sometimes what it comes up with, it's pure hallucination. It's still not uninteresting and has some interesting ideas inside.

So a long time ago I beta tested a new thing called Mathematica and it had errors in it back then, sorry, but it was fun. And at the time when I wrote a paper, you had to say I used this Mathematica tool because otherwise people wouldn't know where you got your calculation from. We don't do that anymore. If I, if I'm doing a tensor calculation, of course I'm using Mathematica to do the connections or whatever. It's become a productivity tool that is ingrained into the process. So the summer of 22, a couple of my colleagues waved me over at lunch, Paul Ginsberg and Michael Douglas and they were playing with a hugging face sandbox and I sat next to them and said, okay, ask it what the extension of the standard model is. I saw the answer and I was like, okay, this is going to be interesting. And then that November that year I actually alerted the archive staff that we now had with a problem and an opportunity because I saw what was about to happen. AI is going to be certainly and will be a part of our toolbox unless we about jihad and eradicated, which I doubt it is going to lever our brains. I mean it already is. I can do stuff faster and more productively and move through things, but I think it is going to go beyond doing the routine stuff for us and make qualitative differences. I just don't know where. If I knew where I'd be doing it, not speculating about it, but I can see it coming and I'd be very surprised if we hit a wallet right now. I think we have some more breakthroughs coming fast.

So I see that there's a lot of thanks in the audience for all of your participation in this. And so as we're coming up at near the end of when our scheduled time is, I want to make sure that everyone has a chance to give any additional thoughts could be about any topic related to AI and scientific endeavor that you know, you feel is important and you want people to think about as something that we're bringing up here.

So let me just give a quick shout out that there are right now scientific AI collaborations being set up and the NSF has a call for scientific AI institutes and so for example polymathic at the Simons foundation. These are formal collaborations to work on this and, and they're open, open source codes mostly. Actually any other government funded ones are open source and they are actually there's index of them somewhere on the web which I don't have the link to at hand. So things are happening.

I'll step back so we don't have a hard stop coming up here. But I just want to make sure that since we are nearing the end of when people's schedules might be pulling them to other things, that we get a chance to sort of go around the table again and, and mention any such important thing. So let's maybe start with Michael. Are there any thoughts that you want to leave the audience with as we're getting ready to wind things down here?

I think there are exciting times ahead and many things will change over the next two years that we five years ago thought would never change in our lifetimes. And some of the things will be amazing, but some of them will probably also be a little bit scary and frustrating.

And so Stein, you mentioned that there are already scientific cloud, excuse me, scientific collaborations and calls for proposals out there in this space. So aside from this, what are the sorts of things you want to leave the audience thinking about?

Just to remind people that there's also a well funded, organized initiatives who, who are setting up AIs and posting the question how do I best kill my enemies? It's a two edged sword, but it's.

A sobering point to make for sure. So Richard, there's I'm sure a lot to be said here in terms of applications, medical sciences and to biology. So what are some parting thoughts that you'd like to leave the audience with as we get ready to wrap things up here?

I mean, I think I would just underscore again this, this combination of the human and the AI. I think that that's critical that you know, this, this, you know, the, the caution every step of the way not to, just because of the late hallucinations, because it may be something you don't know about. You can't trust it. You know, you need, you need some human verification. You know, going back to that talk by the radiologist I mentioned, he, he said, you know, AI will not replace radiologists, but radiologists who use AI will replace radiologists who don't. So I think every, every scientist should be thinking like that.

Collaboration between human and tool as we've done for a very long time. It's just a different set of tools we're getting used to now. And Victor, what, what would you like to leave the audience with as we get ready to wrap things up?

Yeah, so first of all, thank you so much for organizing it, John. Yeah, thanks the audience for their time. And I'll just say that Stein, Richard and science guests were working on developing these tools, but if people have ideas about what they want to see, they should feel free to reach out. And we'll be happy to discuss and perhaps implement these tools with Archive and Bioarchive and other repositories. So just thank you very much for this event.

Very, very positive note of asking people to contribute to Building the Future in this area because it really is fast moving and exciting and both from the bottom of my heart, thank you all for the work that you do in keeping these archives up and running because I know personally I've gotten a lot of good use out of them and Michael can add his own, but I know he agrees with me completely on this point. I think both of our favorite websites, so I expect that many people in the audience are in similar boats. So thank you all so much for really deeply thinking about how an already good thing can be made both better and what we want to avoid in terms of making it worse in the future as our tools evolve. So thank you all so much for joining us. And if you're in the audience, feel free to leave more comments as we see the video after the fact and also reach out to sciencecast and collaborate with Archive Labs. So thanks all very much.