How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI

Talk transcript

1. 0:00

   [on-hold jingle] This April, OpenAI ran a hiring challenge, a competition called Parameter Golf.

2. 0:19

   The top contributor was one candidate that they couldn't hire. It wasn't a person. It's an agent we built called Aiden.

3. 0:31

   In Parameter Golf, the goal is to train the best language model you can under size and the computation constraints. About one thousand machine learning engineers, researchers participate.

4. 0:47

   They filed two thousand submissions. Only forty-seven passed OpenAI's review and made into the leaderboard. Seven of those are actually Aidens, more than twice what ev-- any human contributed.

5. 1:07

   You've seen a lot of auto research today. Agents are hill climbing benchmarks. Those are really impressive results. The question I want to ask is a bit different here. Can the auto research agent produce work that a human community actually recognize?

6. 1:26

   Beyond a good score, agent is optimizing for something that other engineers can merge, fork, and build on.

7. 1:36

   So instead of having an agent just hill climbing locally, we build one that publishes its own work, and that's Aiden.

8. 1:47

   Quick context on us. Weco is a auto research company that founded about two and a half years ago. Uh, I'm co-founder and the CEO, Zhengyao. Um, got my PhD at UCL on reinforcement learning.

9. 2:01

   About two years ago, we built AIDE, the top auto research agent independently evaluated by OpenAI in their MLE-bench paper.

10. 2:13

    Uh, even though back then there was, um, n-no such name called auto research. People call it machine learning engineering agent.

11. 2:22

    Aiden is the next step in a, a experimental prototype. 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.

12. 2:46

    We send Aiden to Parameter Golf competition, and it ran for about twenty-two days. By the end, Aiden has set seven leaderboard records. Each one is a new best for the competition, stamped by OpenAI, and the best human only made three.

13. 3:07

    Passing the host review is a one signal for the quality. A second, maybe more important one, is whether other participants would build on your work. And it turns out Aiden's work had the highest impact within the whole community.

14. 3:27

    Here we are using a inference measure that used widely in academia. It's called a h-index. Roughly, if you have X papers get cited X times, then your h-index is X.

15. 3:44

    Computed over PRs, Aiden was ten, and the next human was seven. The whole community was building on a AI system's work, including many of other leaderboard entries.

16. 4:01

    To break it down a little bit, why can a autonomous AI system be so powerful? One obvious reason is that, uh, it's an AI, it can run tirelessly. Over twenty-two days, it ran about one thousand and three hundred experiments on a single H100 node.

17. 4:25

    But throughput isn't the whole picture. A well-tuned AI system can also keep its output quality high.

18. 4:34

    On the compute side, it uses at most four percent of competition's total compute.

19. 4:44

    And it made about fifteen percent of the records.

20. 4:49

    Also, twenty-eight percent of its submissions made the leaderboard, roughly six times higher hit rate than the community average. So Aiden actually lifted the signal-to-noise ratio within the whole community's public communication channel, which is, uh, PRs.

21. 5:10

    It didn't win through massive parallelization, even though auto research have tons of potential of parallelization.

22. 5:21

    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. Humans and AI are actually contribute in very different ways.

23. 5:41

    When we trace the ideas, Aiden, Aiden's record PRs,

24. 5:47

    almost all of them come from human research papers, other participants in Parameter Golf or in similar communities like NanoGPT. Those ideas are not necessarily a merged PR.

25. 6:03

    Sometimes it's a, a note, um, a human researcher said- Oh, I give up this idea because of some implementation, uh, implementation difficulty, and the agent is good at, uh, finding them and actually implement them.

26. 6:19

    There are also a very small fraction of original ideas that Aiden came up by itself, which, uh, emerged from its efforts to navigate the file size constraints.

27. 6:32

    Here's a concrete example that, uh, traces the patterns I just talked about.

28. 6:39

    So Aiden picked up an idea from Quen paper called the Gated Attention, and it worked. But, uh, uh, it introduced more parameters, and it broke the sixteen megabytes file size limit.

29. 6:56

    So it figure out a quantization mechanism to bring the file size down. But with those two primitives combined, the score barely moved.

30. 7:09

    Then another contributor posted a tokenizer improvement. Aiden recognized the idea, combined it with the architectural work. It just, uh, worked for five days or so.

31. 7:23

    And after this combination, the three ta-- the three ideas turns out to, to have a huge synergy that lead to a big jump in performance and it become one of the Aiden's leaderboard records.

32. 7:41

    So to sum up how I interpret Aiden and in general autoresearch systems' effectiveness, it's very strong at finding and implementing ideas. 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, uh, promising ingredients out of, uh, the

33. 8:05

    parameter golf community, even though the public channel is actually very noisy information-wise.

34. 8:14

    It can also came up logically straightforward ideas. For example, in this case, once you add the parameters and it, uh, breaks the file size limit, one obvious next move is just a quantization.

35. 8:30

    And it's really fast and, uh, really efficient at finding right combinations across a huge search space.

36. 8:41

    Okay, maybe none of those sounds very sexy. Most of them are just a good execution. But in reality, execution is, uh, mostly the bottleneck.

37. 8:53

    What moves the frontier is usually exactly some belief on existing ideas and tons of good executions.

38. 9:06

    Okay, to step back, the state of a human-AI collaboration is a human collectively provide a lot of creative ideas and the agent do the execution

39. 9:19

    to solve a concrete challenge. What we are looking at is a, a large group of human and one AI system. Does it mean a single human engineer's contribution marginally gets smaller?

40. 9:35

    I'd say even for that, uh, not really. In parameter golf competition, it's easy to only focus on engineers that's actually doing hill climbing. But the design behind the competition itself is tremendously important.

41. 9:52

    A bad design can make the whole community effort useless. And their eval design work will have a fe-- huge leverage in the autoresearch era.

42. 10:04

    I really like one tweet from Andrej Karpathy about ten years ago, where he said, "Gradient descent can write code better than you. I'm sorry."

43. 10:17

    For the context, about ten years ago, deep learning was starting to eat up a lot of, uh, software engineering, like conventional coding work. And his tweet was arguing against those people who thought they can hand-write better code than a trained model.

44. 10:37

    Okay, now obviously no one is seriously trying to hand-write code to beat a model. However, software engineering, I mean, as a job still exist. And so many people's job are just, uh, training those models, and those are one of the most well-paid job today.

45. 10:58

    I think how gradient descent change coding is a great metaphor for how autoresearch will change research in the ML engineering.

46. 11:08

    It commoditize certain execution skills. At the same time, it makes some higher-level skills far more valuable.

47. 11:19

    So actually doing autoresearch is, is a lot like training a model. Your codebase abstraction is essentially the architecture. It sets the constraint and the priorities, um, for what the agent can explore.

48. 11:35

    Your eval is the loss function and the data. It sets what agent optimizes for.

49. 11:43

    Take the eval first. The eval is the signal you use to train a model. In this case, it's training your code.

50. 11:53

    It plays the same role that, uh, like data and the loss function, uh, in model training, or in a reinforcement learning setting, it's like environment that the agent is training.

51. 12:07

    Nowadays, no one would argue data or environments, uh, don't matter.

52. 12:14

    And, uh, this is where a vertical moat can also be built. 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.

53. 12:30

    And a good evaluation would be amplified more and more as autoresearch are getting stronger.

54. 12:40

    The other one I think is really underrated is codebase abstraction.

55. 12:46

    The abstraction provides the framework that autoresearch can iterate on.

56. 12:53

    And, uh, that's also... That starting point hugely bias the whole search direction. This is a lot like architecture design in neural networks. Different architecture in theory can represent the same function, but the architecture systematically makes some of the functions easier to be learned.

57. 13:19

    And a good architecture biases the optimization towards solutions that generalize better, perform better, even when the training loss might looks the same.

58. 13:32

    That's exactly the same for autoresearch. Here's an example. We run autoresearch for a, um, fraud detection pipeline, um, and we're trying to optimize the data preprocessing.

59. 13:48

    And first we goo- give it a loose API where the same function process both the training and the testing data,

60. 13:59

    and the score looks great. But the solution was polluted because, uh, there's a certain test set information got leaked to the training information.

61. 14:14

    We then tightened the abstraction to a more strict API where the test data couldn't reach the training, and the data leakage rate just dropped to zero. In this case, a good abstraction leads to better solutions even though if the agent really want, they can still reward hack.

62. 14:38

    So my point is, uh, using autoresearch is a new craft. It's about designing a hill for an agent to climb, and we are still very early on it. I think, uh, that makes this an extremely exciting time to be an AI engineer.

63. 14:57

    Autoresearch will change what skills matter most. Creativity, the judgment to design a good eval or an abstraction, those will soon get exponentially more important.

64. 15:12

    Driving those system itself is where will be a new skill, and, uh, that one is, like, uh, barely existed one or two years ago. So the search is automated.

65. 15:25

    The human would just move up the stack, not out of it.

66. 15:32

    Again, um, WeCo is a autoresearch, um, product research lab. We, we keep sharing what we are learning as we build, uh, on our blog, and I will also post some of my thinking too on X.

67. 15:48

    Uh, if you think some of this are useful to you, feel free to follow me on X. Thank you. [audience applauding] [upbeat music]