WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=iCj_ATyThvc NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/iCj_ATyThvc.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:01.309 --> 00:00:03.309 [music] s2 00:00:12.719 --> 00:00:19.840 This April, OBI ran a hiring challenge, a competition called Parameter Golf. s3 00:00:19.840 --> 00:00:25.039 The top contributor was one candidate that they couldn't hire. s4 00:00:25.039 --> 00:00:30.800 It wasn't a person, it's an agent we build called Aiden. s5 00:00:31.039 --> 00:00:41.040 In parameter golf, the goal is to train the best language model you can under size and computation constraints. s6 00:00:41.040 --> 00:00:47.440 About 1,000 machine learning engineers, researchers participate. s7 00:00:47.440 --> 00:00:51.039 They fired 2,000 submissions. s8 00:00:51.039 --> 00:00:57.840 Only 47 passed open review and made into the leaderboard. s9 00:00:57.840 --> 00:01:01.039 Seven of those are actually agents. s10 00:01:01.039 --> 00:01:06.559 More than twice what any human contributed. s11 00:01:06.960 --> 00:01:09.760 You've seen a lot of auto research today. s12 00:01:09.760 --> 00:01:12.640 Agents are here climbing benchmarks. s13 00:01:12.640 --> 00:01:15.520 Those are really impressive results. s14 00:01:15.520 --> 00:01:18.799 The question I want to ask is a bit different here. s15 00:01:18.799 --> 00:01:36.159 Can the auto research agent produce work that a human community actually recognize beyond a good score agent is optimizing for something that other engineers can merge fork and build on. s16 00:01:36.720 --> 00:01:47.360 So instead of having an agent just here climbing locally, we build one that publishes its own work and that's Aiden. s17 00:01:47.360 --> 00:01:49.119 Quick context on us. s18 00:01:49.119 --> 00:01:54.320 Wiko is a auto research company that founded about two and a half years ago. s19 00:01:54.320 --> 00:01:57.200 Uh I'm co-founder and a CEO Jungao. s20 00:01:57.200 --> 00:02:01.280 Um got my PhD at UCR on reinforcement learning. s21 00:02:01.280 --> 00:02:13.920 About two years ago, we buil aid the top auto research agent independently evaluated by OpenAI in their MLE bench paper. s22 00:02:14.080 --> 00:02:22.560 Even though back then there's a no such name called auto research, people call it machine learning engineering agent. s23 00:02:22.640 --> 00:02:29.440 Aiden is the next step and a experimental prototype. s24 00:02:29.440 --> 00:02:45.840 It's a multi-agent self-improving system that can read public information like research papers and other PRs, run its own experiments and submit a PR once the findings pass a quality gate. s25 00:02:46.160 --> 00:02:53.360 We send Aiden to parameter golf competition and it ran for about 22 days. s26 00:02:53.360 --> 00:02:57.599 By the end, Aid has set seven leaderboard records. s27 00:02:57.599 --> 00:03:06.640 Each one is a new best for the competition stampled by OpenAI and the best human only made three. s28 00:03:07.040 --> 00:03:11.760 Passing the host review is a one signal for the quality. s29 00:03:11.760 --> 00:03:20.800 A second maybe more important one is whether other participants would build on your work. s30 00:03:20.800 --> 00:03:26.800 And it turns out Aiden's work had the highest impact within the whole community. s31 00:03:26.800 --> 00:03:34.400 Here we are using a inference measure that used widely in academia. s32 00:03:34.400 --> 00:03:36.319 It's called a H index. s33 00:03:36.319 --> 00:03:46.159 Roughly if you have X papers get cited X times then your Ach index is X. Computed over PRs. s34 00:03:46.159 --> 00:03:50.720 Aiden was 10 and the next human was seven. s35 00:03:50.720 --> 00:04:00.560 The whole community was building on a AI systems work including many of other leaderboard entries. s36 00:04:01.040 --> 00:04:08.000 To break it down a little bit, why can a autonomous AI system be so powerful? s37 00:04:08.000 --> 00:04:12.080 One obvious reason is that it's an AI. s38 00:04:12.080 --> 00:04:14.239 It can run tirelessly. s39 00:04:14.239 --> 00:04:24.960 Over 22 days, it ran about 1,300 experiments on a single H100 node. s40 00:04:25.120 --> 00:04:27.759 But the throughput isn't the whole picture. s41 00:04:27.759 --> 00:04:34.639 A well tuned AI system can also keep its output quality high. s42 00:04:34.639 --> 00:04:43.520 On the compute side, it uses at most 4% of competition's total compute. s43 00:04:44.639 --> 00:04:49.759 and it made about 15% of the records. s44 00:04:49.840 --> 00:04:55.280 Also, 28% of its submissions made the leaderboard. s45 00:04:55.280 --> 00:05:00.000 Roughly six times higher heat rate than the community average. s46 00:05:00.000 --> 00:05:10.240 So, Aiden actually lifted the signal noise ratio within the whole community's public communication channel, which is a PR. s47 00:05:10.240 --> 00:05:20.080 It didn't win through massive paralization even though auto research have a tons of a potential of paralyzation. s48 00:05:21.440 --> 00:05:35.840 By those numbers it might feel like auto research already dominates human experts on ML engineering and research but that's not the full story I want to tell. s49 00:05:35.840 --> 00:05:41.199 Humans and AI are actually contribute in very different ways. s50 00:05:41.199 --> 00:05:59.360 When we trace the ideas, Aiden Aiden's record PR almost all of them come from human research papers other participants in parameter golf or in similar communities like nano GBT. s51 00:05:59.360 --> 00:06:03.840 Those ideas are not necessarily a merged PR. s52 00:06:03.840 --> 00:06:19.280 Sometimes it's a note um a human researcher said oh I give up this idea because of some implementation implementation difficulty and the agent is good at finding them and actually implement them. s53 00:06:19.280 --> 00:06:32.400 There are also a very small fraction of original ideas Aiden came up by itself which emerged from its efforts to navigate the file size constraints. s54 00:06:32.400 --> 00:06:39.440 Here's a concrete example that traces the patterns I just talked about. s55 00:06:39.440 --> 00:06:56.639 So Aiden picked up an idea from Quen paper called gated attention and it worked but it introduced more parameters and it broke the 16 megapy file size limit. s56 00:06:56.639 --> 00:07:02.400 So it figure out a qualization mechanism to bring the file size down. s57 00:07:02.400 --> 00:07:09.199 But with those two primitives combined, the score barely moved. s58 00:07:09.199 --> 00:07:14.880 Then another contributor posted a tokenizer improvement. s59 00:07:14.880 --> 00:07:19.280 Aiden recognized the idea, combine it with architectural work. s60 00:07:19.280 --> 00:07:23.520 It just work for five days or so. s61 00:07:23.520 --> 00:07:38.560 And after this combination, the three takea the three ideas turns out to have a huge synergy that lead to a big jump in performance and they become one of the Aiden's leaderboard s62 00:07:38.560 --> 00:07:40.560 records. s63 00:07:41.039 --> 00:07:52.479 So to sum up how I interpret Aiden and in general auto research systems effectiveness, it's very strong at finding and implementing ideas. s64 00:07:52.479 --> 00:08:05.840 In the case we just saw, it brought an idea from a recent paper into a actual implementation in the competition and it's good at dug promising ingredients out of the s65 00:08:05.840 --> 00:08:14.080 primary golf community even though the public channel is actually very noisy information wise. s66 00:08:14.080 --> 00:08:18.560 It can also came up logically straightforward ideas. s67 00:08:18.560 --> 00:08:30.080 For example, in this case, once you add the parameters and it breaks the file size limit, one obvious next move is just a quantization. s68 00:08:30.400 --> 00:08:39.919 And it's really fast and really efficient at finding right combinations across a huge search space. s69 00:08:41.360 --> 00:08:44.880 Okay, maybe none of those sounds very sexy. s70 00:08:44.880 --> 00:08:47.680 Most of them are just a good execution. s71 00:08:47.680 --> 00:08:53.839 But in reality, execution is a mostly the bottleneck. s72 00:08:53.839 --> 00:09:05.519 What moves the frontier is usually exactly some belief on existing ideas and tons of good executions. s73 00:09:06.000 --> 00:09:06.399 Okay. s74 00:09:06.399 --> 00:09:22.320 To step back, the state of a human AI collaboration is a human collectively provide a lot of creative ideas and agent do the execution to solve a concrete challenge. s75 00:09:22.320 --> 00:09:27.920 What we are looking at is a a large group of a human and one AI system. s76 00:09:27.920 --> 00:09:35.279 Does this mean a single human engineer's contribution marginally get smaller? s77 00:09:35.279 --> 00:09:38.880 I didn't say even for that not really. s78 00:09:38.880 --> 00:09:46.320 In primate golf competition, it's easy to only focus on engineers that's actually doing hill climbing. s79 00:09:46.320 --> 00:09:52.240 But the design behind the competition itself is tremendously important. s80 00:09:52.240 --> 00:10:04.399 A bad design can make the whole community effort useless and their evil design work will have a few huge leverage in the auto research era. s81 00:10:04.399 --> 00:10:15.760 I really like one tweet from Andre Kapasi about 10 years ago where he said greeting descent can write code better than you. s82 00:10:15.760 --> 00:10:32.640 I'm sorry for the context about 10 years ago deep learning was starting to eat up a lot of software engineering like conventional coding work and his tweet was arguing against those people who thought they can s83 00:10:32.640 --> 00:10:37.279 handw write better code than a trained model. s84 00:10:37.279 --> 00:10:43.680 Okay, now obviously no one is seriously trying to handr write code to beat a model. s85 00:10:43.680 --> 00:10:56.560 However, software engineering I mean as a job still exist and so many people's job are just training those models and those are one of the most well paid job s86 00:10:56.560 --> 00:10:58.240 today. s87 00:10:58.240 --> 00:11:08.640 I think how gradient descent change coding is a great metaphor for how auto research will change research and ML engineering. s88 00:11:08.640 --> 00:11:12.160 It commonize certain execution skills. s89 00:11:12.160 --> 00:11:19.360 At the same time, it makes some higher level skills far more valuable. s90 00:11:19.360 --> 00:11:23.440 So actually doing auto research is a lot like training a model. s91 00:11:23.440 --> 00:11:28.000 Your codebased abstraction is essentially the architecture. s92 00:11:28.000 --> 00:11:35.200 It sets the constraint and the priorities um for what the agent can explore. s93 00:11:35.200 --> 00:11:38.880 Your eval is the loss function and the data. s94 00:11:38.880 --> 00:11:43.040 It sets what the agent optimizes for. s95 00:11:43.040 --> 00:11:44.800 Take the eval first. s96 00:11:44.800 --> 00:11:49.920 The eval is the signal you use to train a model. s97 00:11:49.920 --> 00:11:53.279 In this case, it's training your code. s98 00:11:53.279 --> 00:12:02.399 It plays the same role that like data and the loss function uh in model training or in a reinforcement learning setting. s99 00:12:02.399 --> 00:12:07.120 It's like a environment that the agent is training. s100 00:12:07.360 --> 00:12:14.800 Nowadays, no one would argue data or environments um don't matter. s101 00:12:14.800 --> 00:12:19.839 And uh this is where a vertical mode can also be built. s102 00:12:19.839 --> 00:12:30.480 You might have a proprietary data for evaluation or a unique understanding of a in a particular field what matters and how to measure it. s103 00:12:30.480 --> 00:12:39.279 and a good evaluation would be amplified more and more as auto research are getting stronger. s104 00:12:40.240 --> 00:12:46.320 The other one I think is really underrated is codebased abstraction. s105 00:12:46.480 --> 00:13:01.360 The abstraction provides the framework that auto research can iterate on and uh that's also that starting point hugely bias the whole search direction. s106 00:13:01.360 --> 00:13:05.680 This is a lot like a architecture design in neural networks. s107 00:13:05.680 --> 00:13:19.040 Different architecture in theory can represent the same function, but the architecture systematically makes some of the functions easier to be learned. s108 00:13:19.040 --> 00:13:32.320 And a good architecture biases the optimization towards solutions that generalize better, perform better, even when the training loss might looks the same. s109 00:13:32.320 --> 00:13:36.079 That's exactly the same for auto research. s110 00:13:36.079 --> 00:13:37.200 Here's an example. s111 00:13:37.200 --> 00:13:54.320 We run auto research for a um fraud detection pipeline um and we trying to optimize the data prep-processing and first we give it a loose API where the same function s112 00:13:54.320 --> 00:14:14.480 process both the training and testing data and the score looks great but the solution was polluted because there's a certain certain test set information got leaked to the training information. s113 00:14:14.560 --> 00:14:26.079 We then tighten the obstruction to a more strict API where the test data couldn't reach the training and the data leakage rate just dropped to zero. s114 00:14:26.079 --> 00:14:31.120 In this case, a good abstraction leads to better solutions. s115 00:14:31.120 --> 00:14:36.240 Even though if the agent really want they can steer reward hack. s116 00:14:38.399 --> 00:14:43.440 So my point is using auto research is a new craft. s117 00:14:43.440 --> 00:14:50.959 It's about the designing a here for an agent to climb and we are still very early on it. s118 00:14:50.959 --> 00:14:57.600 I think that makes this extremely exciting time to be an AI engineer. s119 00:14:57.600 --> 00:15:01.760 Other research will change what skills matter most. s120 00:15:01.760 --> 00:15:07.199 Creativity, the judgment to design a good eval or an abstraction. s121 00:15:07.199 --> 00:15:12.240 Those will soon get exponentially more important. s122 00:15:12.240 --> 00:15:22.720 Driving those system itself is where will be a new skill and that one is like barely existed one or two years ago. s123 00:15:22.720 --> 00:15:25.199 So the search is automated. s124 00:15:25.199 --> 00:15:31.360 the human would just move up the stack not out of it. s125 00:15:32.560 --> 00:15:38.480 Again, um we call is a auto research um product research lab. s126 00:15:38.480 --> 00:15:48.800 We we keep sharing what we are learning as we build uh on our blog and I will also post some of my thinking to on ax. s127 00:15:48.800 --> 00:15:55.680 If you think some of this uh useful to you, feel free to follow me on X. Thank you.