⚡️ Prism: OpenAI's LaTeX "Cursor for Scientists" — Kevin Weil & Victor Powell, OpenAI for Science

0:05Okay, we're here at OpenAI with some exciting news from the AI for Science team. with us is Kevin Weil from, I guess, your VP of AI for Science. VP of Open AI for Science, yeah. Open AI for Science. And Victor Powell, who is the product lead on the new product that we're talking about today. And with me is our new AI for Science host, RJ. Welcome. So thanks for having us. Thanks for having us. Yeah, it's very good to be here. Yeah, thanks for hosting us as well. It's always nice to come over to the office. What are we announcing today? So we're launching Prism, which is a free AI native LaTeX editor.

0:40What does all that mean? Because probably a lot of people on the pod haven't worked with LaTeX in the past. LaTeX is a language, effectively, for typesetting mathematics, physics, and, you know, science in general. So if you're a scientist writing a paper, you're probably not using Google Docs because you need to, you have diagrams, you have equations, etc. But it's, and it's been the standard for decades. but the the tools that that people use to to actually write latex write their papers haven't changed in a long time and uh in particular ai can help with a lot of the tasks right because you you spend your time doing the science you need to write it up that's an important part of communicating your work but you want that to be fast and you you want that to be accelerated and ai can help in a ton of ways and we'll talk about some of those but if you if you step back right is OpenAI for Science.

1:32Our goal is to accelerate science. And the surface area of science is very large. So we're trying to build tools and products that help every scientist move faster with AI. Some of that is obviously the work that we can do with the model, making the model able to solve really hard scientific frontier, you know, frontier kind of problems, allowing it to think for a long time. But it's not only that, right? If there was a lesson from what happened over the last year with software engineering, it's that part of the acceleration in software engineering came from better models. But part of it also came from the fact that you now have AI embedded into the workflows, into the products that you use as a software engineer, right?

2:21It'd be one thing if If we were going back and forth, copying and pasting code between, you know, ChatGPT and your IDE, that would be okay. That would be an acceleration. But the real acceleration came when you embedded AI into the actual workflow. And so that's what we're doing here. So OpenAI for Science, it's both building great models for scientists and also speeding them up by bringing AI into the workflow. That's what we're doing with Prism. Yeah. I often say, like, every million copy and paste done in ChatGPT, there's probably some product to be built. Right. Exactly. That's a good analogy.

2:53Yeah. That's a good way to look at it. Especially with LaTeX, having written a lot of LaTeX papers. Yes. Yeah, me too. The number of hours as a grad student I spent, like, trying to get some diagram to line up. Exactly. Exactly. Oh, man. Yeah. Cool. And, Victor, this is your sort of baby. Yeah, I guess it started off as just a project. I left Meta about three years ago, trying to look for various different projects to start. Uh, and, uh, this was, this was one that like, when I sort of presented it to people, they're like, Oh, I get it. That's, I see what you're doing. And so I've just been focused on that building it for about a, about a year and a half.

3:34And, um, you know, it, it has now has become part of opening up and that's been very exciting. Congrats. Thank you. Yeah. So it's kind of a, kind of a fun story, right? I mean, we, as we were thinking, we had this thesis around, it's, it's not just models. It's also building models into the workflow and accelerating scientists in that way. And this is, there are obviously a lot of different ways that you can do that. But the scientific collaboration and publishing thing is definitely one of them. And I was looking around, like, what is there in this space? And there hadn't been a lot of innovation for a long time.

4:10Like, it wasn't that different from when I was, like, writing up, you know, my assignments and papers in tech in grad school. and uh and then i found on this uh reddit forum maybe it was r latech i don't remember but somewhere on this reddit forum i found this thing about a company uh called cricket and uh i was like looking around i couldn't find who the founder was it took me a little while and then i think i found you on twitter and dm'd you out of the blue and just said hey i don't know if you want to talk about this but i would love to talk about this if you're open to it and gave you my number and we talked on the phone and then like jumped on a Zoom and eventually met in San Francisco and made it happen.

4:55That's right. It's awesome to have you guys here but it's just I have a ton of respect for what you started to build. I actually never heard that full story from you until now. You gotta find that Reddit user and thank them because it might have been me. I thought you were totally in stealth because it was the hardest thing to actually figure out who the founder of this thing was. And then I was like, oh, for sure he's not going to respond to my random DM. I mean, I guess that's a part of, part of our focus has always just been entirely on product and to the point where it's almost embarrassing how little we focus on anything else.

5:29Yeah. It worked out for you. Yeah. Also full circle for a moment for you using Twitter to do your business development. Yeah, that's right. So that's kind of interesting. Twitter DMs forever. Right. Like, I actually, yeah, like, probably one of the most important social network innovations, I guess, is that stuff. And I'm sure you know a lot about that. Shall we go right into a demo or talk about it? Yeah, yeah. It's always fun to show it. Yes. Okay. I'm a fan of, like, show, don't tell, push people to the video. Yeah. All right. I'll try and arrange this so you guys can see a little bit. Yes.

6:02All right. So what you have here, so this is Prism. and what you can see on the left here this is actual LaTeX you can see why you might want AI to help you write it because it's you know a little bit it's a language it's a little bit messy and then on the right this is my colleague's paper Alex Lipsoske he's a physicist this is a paper that he wrote on black holes and so you see it over here all the all the you know you can imagine trying to write this in like Google Docs or something it'd be impossible. This is why LaTeX is super powerful. And then, and you've got kind of your files here that make up the project tech file, which is the actual main source file, bibliography files, et cetera.

6:47And, you know, you can go through and you can change it and then you compile that into the PDF itself. But here I can say, this is where at the bottom you can use the AI, you can use GPT 5.2. And I could say, you know, this introduction, maybe I want a little help writing the introduction. So help me proofread the introduction section paragraph by paragraph suggest places where, where I can simplify. There's a lot of demo and we're working on it pretty heavily, so just... Are you nervous yet? You can't be nervous. You're good. Spoken like a true founder. And so, one of the nice things is, you could do this in ChatGPT, but you'd have to go upload your files into a chat, right?

7:42You're going back and forth. Here, because the AI is built into the product, it has all of the files that are part of your project. It automatically puts them in context. It works the way you think it would work. So here, it's looking at the files. Alright. And it's given us kind of a diff here. So it's suggesting changes. You've got the part in red, which is the part that it's changing, the part in green, what it wants to change it to. And you can see the different places where it is suggesting that we change things. So, okay, we'll just keep all of them, right? YOLO. Hope Alex has one. Hope is all true.

8:20Yeah, we're changing Alex's paper. What's the big deal? So here's another thing. We were talking about diagrams in LaTeX. So I've got a, say I wanted to input a commutative diagram, right? It's really easy to draw a commutative diagram like this. It is an absolute nightmare to put these things into tech. So I will upload this photo and I'll say here, whoops. Is there a tech bench for this kind of stuff? Like a set of evals? Yeah. We totally need one. I think there's an opportunity to do that for sure. So here's a commutative diagram that I drew on the whiteboard. Can you make it into a ticks diagram and put it right after the, I don't know, right after, right before, right at the top of the introduction section.

9:24Make sure you get the details right. So I didn't want to interrupt you while you were typing, but why don't you use voice? Oh, actually, I should. I totally could. No, but isn't it interesting that we all have these voice buttons and we don't use it? Yeah. It's not second nature yet. Yeah. It's interesting. And that one I totally should have. I was going to also show something. So you have the, you know, here I am in the tech, and it's working. You also can create new parallel chats. So you can have whole sessions with ChatGPT that can be going in parallel. So here I'll ask it. There's all these equations.

10:03We're talking about symmetries of this black hole wave equation. And in particular, there's this complex symmetry here. Oh, I like how it syncs. Notice how it syncs when I highlight it. But I'll say, like, why don't you, I'll go to my chat so I can start doing this in parallel, and I'll say, please make sure, or please verify that the H plus operator in the new symmetries section is indeed a symmetry of the stationary axisymmetric black hole. Do you understand those questions?

10:48I'll say don't do it in the paper you know show it here I don't want it to actually edit the paper I just want it to prove it here right okay so I'll get that going now while we're waiting for the diagram to finish we can also get another thing going in parallel so I'll say I need to write up a set of lecture notes on general relativity. You know, say I'm a professor, right? I'm teaching a class or something. Put together a 30-minute set of lecture notes on a Ramanian curvature. Wow. That's a very different task. Put it into the file. I made this grlecture.tech. Okay? And so I've got this going.

11:44All right. Well, it came back on my earlier one, H plus symmetry. Is it really? Here you got ChatGPT doing a whole bunch of work to verify that this is indeed a symmetry of the equation. Okay. It does. It confirms it. Right. So you've got the full power of a reasoning model that can think deeply about frontier science. And now we can go back while it works on the other thing. Okay. So this was where I was making the diagram, right uh put it right below the introduction i'll compile it again so this is an auto compile uh actually you can turn that on yeah okay and look wow it nailed it um so it looks like it got it pretty much exactly just a small check the details oh yeah yeah it's pretty good but all right we can we can see if it'll get it right let's say um the C vertex should be directly to your point about voice though I do think maybe over time the code kind of might recede into the background more as you're just really interacting with the paper you're having a conversation with it when you started this product is this how you were envisioning it would be used or were there other design choices that you were considering and you didn't take that path by the way before you answer We have our general relativity lecture notes here.

13:09That was quick. 30 minutes? Six pages. Yeah. So 30-minute section. Okay, so we got curvature, covariant, derivatives. Yeah, this looks like a reasonable set of notes if you were going to go teach a class, right? It just did it for you. Or you can even think like, you know, generate the problem set for this week. Yeah, right. So it's got some examples here. We could tell it to, like, work out solutions to the examples. That's sort of a hidden feature of LaTeX, too, that it actually makes it pretty easy to generate problem sets with answer sheets and things like this. There's so many cool features of LaTeX that I think are underutilized.

13:48Yeah. So anyways, you can see we had it proofread the paper. We had it check some of the answers to verify that our calculations were correct. We generated a set of lecture notes. We added a diagram that we didn't have to actually type up ourselves, which I promise you is horrendous. And that's just, you know, we did that all basically in parallel. And, you know, you can imagine lots of other things. You can, if you have a proof that you, you know, you maybe have like the sort of the bullet points on a proof, you can just say, here are the bullet points, now flesh it out for me. You can imagine having it check all of your references before you publish, make sure all of them are real, up to date.

14:27You can imagine having it generate your references based on the topic of, you know, so like, there's just so many areas where AI can help. That's a big problem when you're trying to put together a paper is get all the references right. Well, okay. And all of this is time that used to go to, you know, typing a paper. Not science. And not science. And now it can go back to science. And that's just one of the ways that we look at accelerating scientists all over the world. Yeah. I would say definitely be careful about including references you haven't read. Right? Because, like, that's the point. Like, you can include a hundred references, but if you didn't read them, then you might as well not have them.

15:04But yeah, I think that web connection is very important. And is this stock GPT-5 or is this like a fine tune? It's GPT-5.2. Okay, yeah. And by the way, when you're looking at references, you can also ask ChatGPT to help you understand the reference. You know, read this paper, tell me the relevance. So all of the things that you might want to do to accelerate your work, you can just do from within this interface. You still have to do your work. But it should make it faster, especially like even linking to the references so you can go and verify, like, okay, this is this one. So this might also make it easier to write the paper as you do the work, right?

15:40Rather than, oh, okay, now I've got to spend two days in LaTeX land, like, trying to get my paper, right? Like a tool for thought rather than just a publishing tool. Yeah, yeah. What about collaboration? It's a great, yeah. So it's built for, I mean, you can speak to this well, it's built for collaboration, so you can bring on as many collaborators as you want. which is nice. I think most other tools in the space have hard limits and charge you money and other things. In Prism, it's as many collaborators as you want for free. Commenting. Yeah, so you've got commenting. You've got all the kind of collaboration tools that you would want.

16:17Good. And then any of the engineering choices, like what might engineers not appreciate when just looking at a tool like this? Often it would be like multi-line diff generation that you need to do because you're editing a pretty complex document. It does get pretty complicated. I mean, we're using, let me know if I'm getting too technical into the weeds, but, you know, we're relying heavily on the Monaco JavaScript framework. I'm very familiar with the lack of documentation of Monaco. That's actually, it's interesting you say that because it's very true. It's an extremely powerful library that is almost entirely undocumented.

16:54Yeah, it's just types. But you can use Codex now to generate the documentation for you. Yeah. You think Microsoft should get on that. But yeah, just stuff like that. I like to hear about the behind the scenes of building something like this. What do you struggle with? What's the model surprisingly good at? And what's the model it should be good at, but it's not? What were some of the hardest problems as you were building this in the first place? What were some of the hardest things to get right? I think initially one interesting challenge was that we really pushed on it being WebAssembly and fully just running in the browser at first, the whole entire LaTeX compilation.

17:28and that did help us in the sense that we were able to like flesh out the design and the AI capabilities early on without having to invest heavily in like the back-end infrastructure but eventually we did hit a wall with that approach and once we switched it to back-end PDF rendering like that's when we really started to hit an inflection point with like usage now fast yeah yeah yeah I think we also the AI in here benefits a lot from everything that we've learned building codecs. And as we go forward, I think we'll likely just integrate the full codecs harness into the application here. So you get all the benefits of the tools and the skills and all the things that codecs can do today.

18:08And you just sort of automatically can bring that into your environment here. Yeah. Is there a feature they're just the same app? Maybe. I think potentially it depends on, I mean, here's the reason I'm hesitating is I think the interesting thing with, um, with this and with codex is we're still mostly in a world today where people are, you have, you know, your, your main screen is your, is your document. And then you have your AI on the side, but the more that AI improves, people trust it and they're just YOLOing it, right? You're like, you're generating code and you're looking at the code as sort of secondary to instructing the AI and driving from that the UI probably changes for all of these things you don't need your document front and center because you're actually not looking at your document as much that's sort of your backup and your interaction with your AI is primary and as that happens I think you might these UIs can kind of converge over time so we'll see but I definitely would love to see a world where people needed to spend less time thinking about the actual syntax and much more about what they're trying to create.

19:21Yeah, I mean, I feel like this plus a notebook would be amazing. Yeah. Because, and something that AI can run, you know, run an analysis, generate plots, so stick that in the paper here. like oh read you know like this paper like this part of the paper like take that equation and like you know do something with it that would be a really amazing uh integration yeah like think through the different corollaries of this thing from this paper and produce some alternatives and then like yeah i completely agree yeah yeah i do think that's sort of the progression where it's like doing doing maybe work for a few seconds versus maybe we're already at a point where it's doing work for a few minutes, eventually doing work for hours, days, coming back with very complicated analysis.

20:08Yeah. I mean, that's actually maybe a good segue into some of the other questions that I had about your initiative. I mean, so stepping back to AI for science in general, can you talk a little bit, I mean, I have a million questions, but maybe start with I feel that validation of AI for science is critical to its success. You have to have some sort of real-world validation of the results that you produce with your AI. I know that there's been some publicity in the past. What are the latest and greatest hits of the things that big labs or any lab is doing with open AIs? I mean, when you step back and look at the trend, I think that's the biggest thing.

21:05Because we can debate exactly like you've probably seen in the last few weeks even. There have been a bunch of different examples of like GPT 5.2 contributing to open Airdish problems and things like that. and you get into this debate of well was it uh was it really just really good at literature search and it found an example over here an example over here when you combine the two you know that it was sort of a trivial step from there to the solution and was that novel or did it really do something new and you know that's a that it's a legitimate discussion but when you step back two years ago we were like you know this thing can pass the sat that's amazing and It progressed to like, it can do a little bit of contest math and it can start to solve harder problems.

21:54Wow. And then you keep going and it's starting to solve graduate level problems. And then you have a model that gets a gold medal at the IMO. And now we're sitting here talking about, you know, it solving open problems at the frontier of math and physics and biology and other fields. So it's just, I mean, the progression is incredible. And if you think about where we are today, then you fast forward six months, 12 months. I, I, I'm very optimistic about what the models are going to be do, be able to do to accelerate science. It's like, it's already happening. If there's one thing that I've learned from, uh, my like two-ish years at OpenAI, it's you go very quickly from this, this thing is just impossible for AI to do.

22:40Like it's too hard. AI can't do it to like, AI can just barely do it. And it like kind of doesn't work. And, you know, only early adopters are doing it because it's not particularly reliable yet, but it like sort of works to, oh my God, AI like does this thing really well. And I could never imagine not using AI for this in the future. It's like, once you start to get to, you know, five, 10 % on some particular eval you very quickly go to like 60 70 80 and we're just at the phase where ai can help in some not all but in some elements of frontier science math you know biology chemistry etc and it just means we're like right at the at the cusp and it's super exciting so i mean so it it fast forward a year or the end of the year, and we have AIs that can do a lot of this discovery process, then the bottleneck becomes the wet lab or the lab, right?

23:39Yeah. So what are you seeing in that domain? Yeah. By the way, we were talking a little bit about software engineering before and the analogies. I think 2026 for AI and science is going to look a lot like what 2025 looked like for AI and software engineering, where if you go back to the beginning of 2025, if you were using AI heavily to write your code, you were sort of an early adopter. And it kind of worked, but certainly not everybody was doing it. And then you fast forward 12 months, and at the end of 2025, if you were not using AI to write a lot of your code, you're probably falling behind.

24:19I think we're going to see that same kind of progression in AI and science. Today it's early adopters, but you're really starting to see some proof points in solving open problems and developing new kinds of proteins and things like that. But you're right, as it really starts to work, and I think this is the year that it's really going to start to work, it shifts the bottleneck. and I think we're going to be starting to talk a lot more about robotic labs and other things you know like do you need to have a grad student like pipetting things no probably not right right now you do but why why shouldn't we have why shouldn't we have uh robotic labs that where you have ai models doing what they do best reasoning over a huge amount of of different information.

25:13They have read substantially every paper in every field and can bring a lot of information to bear to help prune the search tree on a new material, for example, that you're trying to create. And then you have a robotic lab that can roll out a bunch of experiments in parallel, do them while we sleep, and then feed the results back into the AI, let it learn from them, design the next set of experiments and go. I mean, it's hard to imagine. It doesn't even have to be yellow science right to your point it's you're verifying it as you go because you have an actual lab building it in real life um but you can just do so much more in parallel you can think harder up front with ai to design the experiments uh and then again like prune the search tree so you're you're searching over a smaller number of higher value targets and then you automate the experimentation uh and and turn it around faster and again like this is acceleration like the whole if we're successful then it's you end up doing you know maybe the next 25 years of science in five years instead so in 2030 we could be doing 2050 level science and that would be an awesome outcome like the world is a better place if that happens absolutely i i guess uh so we spoke recently with heather kulik at mit and one of the things she pointed out was that there's a element of serendipity to working in a lab that you lose and so she was of the opinion that there's a class of problems especially when you have like a large search space or something like that where robotics is going to really accelerate science and there's another class of problems where even experimental science will not move forward very fast because of robotics and so then again you're at a bottleneck but i guess humans need something to do so well that what What she said sounds totally reasonable to me, right?

26:59There are probably places where the humans are adding no value because they're literally just trying to pipette a certain amount of a thing into another thing or, you know, do some, uh, the same motion repeatedly in a bunch of different ways. And then there are places where it's less well understood. You want the full flexibility that you have for the really smart human thinking about the work that they're doing. Um, by the way, the same is true in, in, uh, the more theoretical fields as well, where it's not, this isn't about, let's automate all the humans out of their jobs. This is about accelerating scientists.

27:32It's scientists plus AI together being better than scientists alone or AI alone. And I think the same is true whether you're talking something that's happening in silico proving a theoretical problem or happening in the real world with a lab. Like find the parts that you don't need a human to do and try and automate them as much as you possibly can so that the humans can spend their time on the most valuable things. Yeah. I'm very pro the in silico acceleration, because obviously you have more control over that and you can parallelize and repeat and do all those things. Yeah. I think there will be a huge amount of value in, you know, a lot of fields are heavily simulatable and they, you know, and so, you know, nuclear fusion, for example, they're running a lot of simulations before they do any particular experiment because the experiments are very time-consuming and expensive.

28:21But I'm excited to see what you can do when you have a loop between, you know, a very intelligent reasoning model that understands fusion and a simulation, and you get the model thinking about what parameters to set for the simulation and then running, you know, a bunch of simulations in parallel, feeding that back. And you have that same sort of lab loop, except it's all in silico and in an experiment in running on a giant GPU cluster yeah and then when you really have like gotten to the end of that calculation then you go run it in yeah this is bringing it back to prism this is sort of a nice aspect that you're you're getting a more sophisticated view of your result right instead of just um you know like a chat output in it i would i would hope as it develops as a way for a scientist to be able to interact with the information before you kick off your nuclear fusion experiment for, you know,$10 million or whatever.

29:19And the human can learn from more things, right? You just, you get more data that you can, that you can look at and evaluate. So, yeah. So this, by the way, this fusion discussion makes me think like, you know, if one day opening after science, you know, it gets serious enough and starts to self-accelerate, you should solve cold fusion and, you know, be your own power source. Well, I mean, this is why we're so excited about this, right? I mean, imagine our mission is to bring AGI to the world in a way that's beneficial to all of humanity. It's right there at the lobby. Yeah. It's amazing. You see it every day you walk in, you see it.

29:56Yeah, absolutely. And imagine, I mean, if we had GPT-9 inside of ChatGPT today, it would be awesome. You could do lots of things. But if you had GPT-9 and it could, which I'm using as a stand-in for AGI, right? And it could create new materials and we were, the devices we were using were all incredible and, you know, had 30 day battery lives and things like that. And we had personalized medicine and we all knew someone whose life was saved because we were developing personalized, you know, cancer treatments and things so much faster. Like that's the real benefit of AGI. That's, I think, maybe the most tangible way that we're all going to feel AGI as it starts to be real.

30:39And that's why this work is so mission-driven for us. So it brings up kind of two questions in my mind. The first one is, so then who owns the invention? And then the other half of that is, okay, so then does OpenAI become a drug company and a fusion company? Because this is how, I mean, you laugh, but it's a little bit serious that all the AI for drug discovery companies ended up being drug companies. because they couldn't sell the, so far with some exceptions now with Noetic, for example, but they end up being drug companies because they can't sell the drug. But in any event, there's a lot of precedence for basically building your own portfolio using AI.