WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=XLEYtv3cMlw NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/XLEYtv3cMlw.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:00.719 --> 00:00:04.240 Hello everyone, my name is Sina Shahande. s2 00:00:04.240 --> 00:00:12.320 I quite my pleasure to present you this talk about uh running autonomous agents for scientific tasks. s3 00:00:12.320 --> 00:00:14.559 Um let's let's dive in. s4 00:00:14.559 --> 00:00:23.279 Um so here um I think everyone is quite familiar with the concept of auto researcher or auto research. s5 00:00:23.279 --> 00:00:40.719 This is the original Andre carpat is um GitHub repo where we have a ML model and we ask a coding agent to uh find def find a particular metrics and then optimize the code in order to minimize the error basically s6 00:00:40.719 --> 00:00:45.440 do a hill climb over model optimization. s7 00:00:45.440 --> 00:01:00.000 Now for many of these coding tasks um this works very well but when the problems become very much open-ended and um sometimes a long horizon like most of the scientific task s8 00:01:00.000 --> 00:01:12.560 you have this case where AI agents uh usually kind of saturate to a certain level um simply they they are very good at implementation of the of the code or changing the or running the experiments s9 00:01:12.560 --> 00:01:20.479 over lots of data and so on but the problem is they ran out of ideas uh or you know what people call them research taste. s10 00:01:20.479 --> 00:01:32.079 Now you can see that in this situations uh you know good humans keep going uh higher and the top 1% humans you know they keep even improving better and better over time. s11 00:01:32.079 --> 00:01:46.000 Um now um the the the really the difference um from here is the way that uh good ideas uh can or good hypothesis on how the model could be improved or how the the problem could be solved s12 00:01:46.000 --> 00:01:50.479 um keep um keep coming coming up humans keep coming up. s13 00:01:50.479 --> 00:02:03.200 So in the scientific task uh you you know you you have this um scientific method um where you know we observe a situation we make a questions we come up with a good hypothesis and s14 00:02:03.200 --> 00:02:15.440 come up with a hypothesis how to solve this problem and then we do experiments and implement and do the experiment um do the loop um and we as each of the iterations we learn something and we improve now s15 00:02:15.440 --> 00:02:28.160 the components of the learning component uh I think those are all questions of memory and implementation from learning the mistakes but which is one of the bottlenecks of using coding agents but I think this quite solved by just simply s16 00:02:28.160 --> 00:02:38.400 organizing patterns activity I think what is much more difficult is it coming up with a hypothesis so how can we come up with a good hypothesis good ideas for our s17 00:02:38.400 --> 00:02:50.239 uh coding agents to keep improving better uh the process and this is something core things that I would like to kind of um um focus on this talk and plus a little bonus at the s18 00:02:50.239 --> 00:02:56.879 Um so let's look at the problem here we're trying to achieve as an example so you have a good a good idea of what we're trying to do. s19 00:02:56.879 --> 00:02:58.959 So here is what we're doing at Radicate. s20 00:02:58.959 --> 00:03:10.400 We're building uh insilicopet meaning you have a CT and we want to generate um a PET PET image um a PET scan from the CT scan. s21 00:03:10.400 --> 00:03:18.480 Example of this you can see here um we have CT scans and of slices of scans of the of the of a patient. s22 00:03:18.480 --> 00:03:22.879 they might have a nodule in the the lung and the question is is this cancerous or not? s23 00:03:22.879 --> 00:03:24.319 Um is it a lung cancer? s24 00:03:24.319 --> 00:03:34.000 So they do a PET scan which is a difficult process and very timeconuming and uh to do um but here we do an ML through ML model image translation s25 00:03:34.000 --> 00:03:44.000 one can kind of change the modality learn the structure of the body and infer what would happen in a PET scan if the hyper the activity of the tissues. s26 00:03:44.000 --> 00:03:54.080 So you know certain tissues absorb more radioactive tracer and they shine up in this PET scan um and the the tumors usually that's the case. s27 00:03:54.080 --> 00:04:05.439 Now to generate this relationship uh we need um we need a model uh to um to do the translation but the problem itself has many components. s28 00:04:05.439 --> 00:04:21.040 So really uh this like any other scientific task the problem is decomposing the problem entire long-term horizon two years 10 years research process into steps and each of those steps is really fundamentally are the goal are a loop. s29 00:04:21.040 --> 00:04:26.800 So I'm going to focus on one of these particular ones right now and that is on training of a machine learning model. s30 00:04:26.800 --> 00:04:37.680 Uh so we have here you know decoder encoder type of situations and so encoding the CT and then decoding it into PET. s31 00:04:37.680 --> 00:04:41.840 Um so that's the typical GAN model which kind of generates the the image. s32 00:04:41.840 --> 00:04:53.919 Now for this we can kind of define this kind of you know the architecture and we try it capture data and do all the 80% of work to basically bringing the good data set and here s33 00:04:53.919 --> 00:05:03.840 uh we create the metrics and so on and um around the image you know fidelity of synthetic pet to to real pet and so so on. s34 00:05:03.840 --> 00:05:16.400 But the challenge is how can we improve the situation given a certain um initial point and we'll go back to our idea of uh hill climb around this optimization. s35 00:05:16.400 --> 00:05:30.639 So you can see an example of iterations coming from a real run in codeex where you know we improve the data and so on and then we the model goes around and tries to do the optimization and you can see there's a range of possibilities s36 00:05:30.639 --> 00:05:44.960 and some of them become dead end some of them not don't improve anything but um we kind of desaturate at a certain point and then you need really need a good idea a good idea has to come up so in this case we have a slices of CT s37 00:05:44.960 --> 00:05:47.600 and we feed these as channel into the model. s38 00:05:47.600 --> 00:05:52.080 So initial problem initial model that was trained was 2 and a halfd. s39 00:05:52.080 --> 00:05:58.160 So treating each CT slice as a be the 2D convolutions but stack over channel. s40 00:05:58.160 --> 00:06:12.800 Now if you give this to a typical ML model they would not think about it as you know go through hyperparameters or you know some sort of you know playing around with problems that it knows but it wouldn't do a very radical change. s41 00:06:12.800 --> 00:06:20.560 uh for example to come up with a 3D idea convolutions or change the whole problem upside down. s42 00:06:20.560 --> 00:06:38.880 So to create those ideas um for for for the for the model to try um I had to kind of in the midst of the codeex loop say oh what about what about this idea what about that idea or go read papers out there and see what the papers are are are saying what's other people are trying s43 00:06:38.880 --> 00:06:54.080 um so to induce uh that hypothesis generation we need to do something about about about our our ML model LLM models so this is a trick um that I found that working very efficiently s44 00:06:54.080 --> 00:06:58.000 and it's very similar to that chain of thoughts step-by-step problem. s45 00:06:58.000 --> 00:07:05.919 First is to decompose the problem into its subcomponents but it's an explicit act um um action. s46 00:07:05.919 --> 00:07:18.960 So you know you could ask actually go through your problem here in this case you know create a translated polymer nodule CT patches into equivalent pet um that's our top problem that I just explained to you s47 00:07:18.960 --> 00:07:31.039 and then it has components into it so different domain in this case you can see um you know the the the data component the actual core the learning the architecture s48 00:07:31.039 --> 00:07:45.599 the training loss the you know the operational part of the the ML modeling the metrics and evidence and um you know the peripheral scripts that kind of run the model and data preparation itself is very important pieces. s49 00:07:45.599 --> 00:07:52.080 Now what we have here is this hierarchy of components of this model that's induced. s50 00:07:52.080 --> 00:07:56.160 So this itself can be induced very easily using a prompt. s51 00:07:56.160 --> 00:08:10.639 Uh so basically you know coding agents can itself go in with this prompt of go in through this codebase and create this series of hyper um documents that are linked to each other. s52 00:08:10.639 --> 00:08:25.360 Now what we're trying to do is is to give our um our coding agent ability to look at this problem in this component uh where it might not do so and then induce s53 00:08:25.360 --> 00:08:28.479 a a a change within those plateau. s54 00:08:28.479 --> 00:08:31.680 That's exactly what uh human researchers would do. s55 00:08:31.680 --> 00:08:34.560 Uh the question is look at what's important what's not important. s56 00:08:34.560 --> 00:08:38.080 So let me show you let me show you this graph here in obsidian. s57 00:08:38.080 --> 00:08:43.120 So the documents that's generated out of this hierarchy generation would looks like something like this, right? s58 00:08:43.120 --> 00:08:55.440 So there's there's an initial readme says here's a problem statement the level one and here is our you know the documentation and then when we go to a next level s59 00:08:55.440 --> 00:09:06.320 uh of the of the problem you see here um the level two level two itself goes in and have other level three components going on. s60 00:09:06.320 --> 00:09:09.519 So one of these pieces for example is the model architecture. s61 00:09:09.519 --> 00:09:12.640 There's all those kind of components I showed you in the graph that you can see it here. s62 00:09:12.640 --> 00:09:14.080 And each of these documents are very simple. s63 00:09:14.080 --> 00:09:22.800 We just ask them the model to to document different levels of um this hierarchy as it goes in um through the process. s64 00:09:22.800 --> 00:09:33.440 So here you can see the generator now it says it's a 2 and a halfd generator and so on and we go shows the details and at the end of these hierarchy components is the actual code s65 00:09:33.440 --> 00:09:36.399 um that is that is in our system. s66 00:09:36.399 --> 00:09:50.560 Now what we do during before we actually start the loop the question becomes okay now that you have this hierarchy and now the our model can start thinking uh step by step through the hierarchy s67 00:09:50.560 --> 00:09:55.839 and going forward uh for the actual uh implementation. s68 00:09:55.839 --> 00:10:07.760 So we ask um then the task is given this hierarchy that you have come up with ideas that can improve improve this problem and you can see here before s69 00:10:07.839 --> 00:10:25.680 if I just say say here's our code base and here's my objective Google optimized this process similar to what originally Carpathy's readme file in this program MD it would not it would not generate it would saturate after a while but now that we have this hypothesis s70 00:10:25.680 --> 00:10:30.000 generation process that can work in an adversarial and collaborative way. s71 00:10:30.000 --> 00:10:47.200 Um now you have a process where you can create say you know um you you basically create 100 um you can create um 100 different solutions. s72 00:10:47.200 --> 00:10:55.519 So this goes in and because we have all the components we actually have a have a process in which going and modify every component of it. s73 00:10:55.519 --> 00:11:08.240 So it becomes a very much more comprehensive search because you have a scaffold that are reasoning LLM can go in and make a decision around improving each of these and s74 00:11:08.240 --> 00:11:16.800 very much you can see here that you can come up with the idea around uh you know creating a three-dimensional um uh structure rather than 2D. s75 00:11:16.800 --> 00:11:25.519 So basically creates a much more radical change that it's otherwise would have not come up or or would have missed. s76 00:11:25.519 --> 00:11:40.959 Now you can also review this of course um this plan uh with a with a adversarially and um or collaboratively with another agent or another model to basically loop in and uh and improve this s77 00:11:40.959 --> 00:11:41.440 for. s78 00:11:41.440 --> 00:11:49.600 Ultimately the idea would be um to go through this in in in in a loop like um similar to like this. s79 00:11:49.600 --> 00:12:01.279 So we have basically um you know a codebase a metrics um goal achieved no then you hypothesize a required change and then kind of go in and implement and this loop can go s80 00:12:01.279 --> 00:12:06.480 much faster and much better because here we create a better hypothesis. s81 00:12:06.480 --> 00:12:22.639 Um now let me give you another example uh where um we also in the parts that data preparation another way to kind of improve this scientific loop um is around the concepts of um um s82 00:12:22.639 --> 00:12:24.959 so you can see an example of it as a registration. s83 00:12:24.959 --> 00:12:34.880 I want to show you how you can use this additional skills additional kind of collaborative and addial process within your loop. s84 00:12:34.880 --> 00:12:40.880 problem is that we have this case where we have to do um image registration. s85 00:12:40.880 --> 00:12:49.760 It's another type of optimization loop in a sense that you you the idea here is so you can see two process two two two scans are overlapped here. s86 00:12:49.760 --> 00:12:56.480 So the CT scan the gray and then the PET scan is the color one that shows the metabolic activity of the of the tissue. s87 00:12:56.480 --> 00:13:04.079 So you see liver is high and the lung is is black is not absorbing that radio trace there. s88 00:13:04.079 --> 00:13:10.320 Um now at the beginning these two scans are not aligned because they they're scanned at different times. s89 00:13:10.320 --> 00:13:12.800 Uh at the beginning they're look like this. s90 00:13:12.800 --> 00:13:18.639 Um it's off and through that process of optimization which we call image registration. s91 00:13:18.639 --> 00:13:21.279 We need to align these images together. s92 00:13:21.279 --> 00:13:24.800 This itself is has a goal to make a good registration. s93 00:13:24.800 --> 00:13:30.639 But there are many different components to it many metrics and sometimes the there's a very long process. s94 00:13:30.639 --> 00:13:39.839 it it would can fail in different aspects of this as this you know the bodies may not be the same well the the body's not the same because the because respiration s95 00:13:39.839 --> 00:13:52.639 because changes in the system they they're not always aligned and the patients moving and coming in the next next day or next week do the same scan uh with a different scanner would be very different um s96 00:13:52.639 --> 00:13:53.680 positioning. s97 00:13:53.680 --> 00:14:05.760 So to do so um we have to kind of align these and we go through the same loop here but we have many different um type of metrics here. s98 00:14:05.760 --> 00:14:17.920 Some of them not necessarily um a quantitative but qualitative and also hard to know which bug or which issue has has come across. s99 00:14:17.920 --> 00:14:27.920 So as part of that loop process um one of the idea is that as a as a scientist I would be come here and look at this images one by one myself. s100 00:14:27.920 --> 00:14:34.160 I could have tried to say oh is this kind of quality control metric type story where you look at this image say does this make sense? s101 00:14:34.160 --> 00:14:40.959 So here you can see a long mask is the mask applied correctly on the lungs or not. s102 00:14:40.959 --> 00:14:55.519 If this process next steps is going is it scans properly truncated or cropped but each of these processes there's a loop in itself and could be this con adversarial could also be collaborative. s103 00:14:55.519 --> 00:15:10.000 So as part of the metrics uh you can kind of call in another model that has a better multimodal capability uh to review the image uh and what's the opinion about this image does it scans s104 00:15:10.000 --> 00:15:23.839 correctly aligned um and then as part of the hypothesis generations or critique of the changes that has happened you can invoke the another model such as that is much more better at reasoning s105 00:15:23.839 --> 00:15:37.760 545 uh to generate the hypothesis because they have much bigger um test time compute than and in first time compute to do so you need to equip your loop with additional skills s106 00:15:37.760 --> 00:15:48.000 to do these these processes by themselves and it's very easy now in the coding agent of course so here I build a skill for review of that image generated s107 00:15:48.000 --> 00:15:54.399 with say with Gemini model and so on so as part of that loop you can kind of implement these these loops there. s108 00:15:54.399 --> 00:16:06.399 Uh, and of course to use GP5 and pro that's something we really like is Peter Spinberger's Oracle CLI which packages the code and package the data and send them to s109 00:16:06.399 --> 00:16:19.839 API uh of the the GP 5.5 Pro and then we'll be able to kind of close the loop um on the on the outcome and you I've seen a much better improvements s110 00:16:19.839 --> 00:16:33.279 during the especially the hypothesis improve generation uh but also So the critique of is this implementation was really effective because every loop that goes in the loop basically asks say this is what we implemented this is the outcome of it s111 00:16:33.279 --> 00:16:42.800 does this make sense what would you do next so the process of the loop is very important uh to kind of keep going and improving the model out of this story s112 00:16:42.800 --> 00:16:56.959 I think what you see as this um back to the concept of autonomous researcher is that you have the ability to go through this loop of scientific discovery with much more rigor in terms of hypothesis s113 00:16:56.959 --> 00:17:08.799 and and I think implementation is already quite a bit solved especially if your problem is um you know b you know in in a basic can be done in a basic simulated environment s114 00:17:08.799 --> 00:17:23.600 uh in in a box of course the big thing in science is um better observations the multimodal model currently they lack very much understanding of these components so for example here most s115 00:17:23.600 --> 00:17:34.880 um understanding of this tin tiny changes for example the existence of this nodule no LLM today is able to identify these very well because they're just not simply trained s116 00:17:34.880 --> 00:17:45.919 on scientific images and scientific data so there's a big space for future work for people who build these models to be able to fine-tune them or be able to adjust them so they can s117 00:17:45.919 --> 00:17:58.720 the model themselves can close the loop by observing the same way as a trained scientist would be able to observe that particular uh issue and then um so that's I think one of the biggest bottlenecks for uh for s118 00:17:58.720 --> 00:18:11.440 not having a full you know scientist in a in a in a data center and then the next thing would be this strength hypothesis but I feel the same way as the chain of thoughts s119 00:18:11.440 --> 00:18:25.280 came on top of the a basic um [clears throat] base model like you know 40 GP 4.0 know and allowed it to do much more better reasoning. s120 00:18:25.280 --> 00:18:38.080 This type of trick with this type of um hierarchal structures is the same trick basically applied on the current model and I think going forward with the newer models that are much better post-trained s121 00:18:38.080 --> 00:18:41.520 to compartmentalize the problems and break down the problems. s122 00:18:41.520 --> 00:18:46.080 Uh we probably get need less and less of these tricks down the road. s123 00:18:46.080 --> 00:18:58.400 But in the meantime and in the absence of a very good models I think you you can do something like this or manually well in a different explicit step steps decompose a problem into hierarchies s124 00:18:58.400 --> 00:19:03.679 and then going through those hierarchies for for improvements in each component of it. s125 00:19:03.679 --> 00:19:16.320 And basically that allows a very uh structured way to scale the test time compute to to generate more and more tokens on this problem and uh close and continue the loop. s126 00:19:16.320 --> 00:19:18.080 Well, thank you very much for this talk. s127 00:19:18.080 --> 00:19:21.840 Uh hope uh to see you around.