How Autoresearch is changing ML research — Zhengyao Jiang, Weco

AI Engineer · 16 min · 127 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:01[music]
2. 00:12This April, OBI ran a hiring challenge, a competition called Parameter Golf.
3. 00:19The top contributor was one candidate that they couldn't hire.
4. 00:25It wasn't a person, it's an agent we build called Aiden.
5. 00:31In parameter golf, the goal is to train the best language model you can under size and computation constraints.
6. 00:41About 1,000 machine learning engineers, researchers participate.
7. 00:47They fired 2,000 submissions.
8. 00:51Only 47 passed open review and made into the leaderboard.
9. 00:57Seven of those are actually agents.
10. 01:01More than twice what any human contributed.
11. 01:06You've seen a lot of auto research today.
12. 01:09Agents are here climbing benchmarks.
13. 01:12Those are really impressive results.
14. 01:15The question I want to ask is a bit different here.
15. 01:18Can 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.
16. 01:36So instead of having an agent just here climbing locally, we build one that publishes its own work and that's Aiden.
17. 01:47Quick context on us.
18. 01:49Wiko is a auto research company that founded about two and a half years ago.
19. 01:54Uh I'm co-founder and a CEO Jungao.
20. 01:57Um got my PhD at UCR on reinforcement learning.
21. 02:01About two years ago, we buil aid the top auto research agent independently evaluated by OpenAI in their MLE bench paper.
22. 02:14Even though back then there's a no such name called auto research, people call it machine learning engineering agent.
23. 02:22Aiden is the next step and a experimental prototype.
24. 02:29It'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.
25. 02:46We send Aiden to parameter golf competition and it ran for about 22 days.
26. 02:53By the end, Aid has set seven leaderboard records.
27. 02:57Each one is a new best for the competition stampled by OpenAI and the best human only made three.
28. 03:07Passing the host review is a one signal for the quality.
29. 03:11A second maybe more important one is whether other participants would build on your work.
30. 03:20And it turns out Aiden's work had the highest impact within the whole community.
31. 03:26Here we are using a inference measure that used widely in academia.
32. 03:34It's called a H index.
33. 03:36Roughly if you have X papers get cited X times then your Ach index is X. Computed over PRs.
34. 03:46Aiden was 10 and the next human was seven.
35. 03:50The whole community was building on a AI systems work including many of other leaderboard entries.
36. 04:01To break it down a little bit, why can a autonomous AI system be so powerful?
37. 04:08One obvious reason is that it's an AI.
38. 04:12It can run tirelessly.
39. 04:14Over 22 days, it ran about 1,300 experiments on a single H100 node.
40. 04:25But the throughput isn't the whole picture.
41. 04:27A well tuned AI system can also keep its output quality high.
42. 04:34On the compute side, it uses at most 4% of competition's total compute.
43. 04:44and it made about 15% of the records.
44. 04:49Also, 28% of its submissions made the leaderboard.
45. 04:55Roughly six times higher heat rate than the community average.
46. 05:00So, Aiden actually lifted the signal noise ratio within the whole community's public communication channel, which is a PR.
47. 05:10It didn't win through massive paralization even though auto research have a tons of a potential of paralyzation.
48. 05:21By 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.
49. 05:35Humans and AI are actually contribute in very different ways.
50. 05:41When 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.
51. 05:59Those ideas are not necessarily a merged PR.
52. 06:03Sometimes 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.
53. 06:19There 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.
54. 06:32Here's a concrete example that traces the patterns I just talked about.
55. 06:39So 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.
56. 06:56So it figure out a qualization mechanism to bring the file size down.
57. 07:02But with those two primitives combined, the score barely moved.
58. 07:09Then another contributor posted a tokenizer improvement.
59. 07:14Aiden recognized the idea, combine it with architectural work.
60. 07:19It just work for five days or so.
61. 07:23And 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
62. 07:38records.
63. 07:41So to sum up how I interpret Aiden and in general auto research systems effectiveness, it's very strong at finding and implementing ideas.
64. 07:52In 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
65. 08:05primary golf community even though the public channel is actually very noisy information wise.
66. 08:14It can also came up logically straightforward ideas.
67. 08:18For example, in this case, once you add the parameters and it breaks the file size limit, one obvious next move is just a quantization.
68. 08:30And it's really fast and really efficient at finding right combinations across a huge search space.
69. 08:41Okay, maybe none of those sounds very sexy.
70. 08:44Most of them are just a good execution.
71. 08:47But in reality, execution is a mostly the bottleneck.
72. 08:53What moves the frontier is usually exactly some belief on existing ideas and tons of good executions.
73. 09:06Okay.
74. 09:06To 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.
75. 09:22What we are looking at is a a large group of a human and one AI system.
76. 09:27Does this mean a single human engineer's contribution marginally get smaller?
77. 09:35I didn't say even for that not really.
78. 09:38In primate golf competition, it's easy to only focus on engineers that's actually doing hill climbing.
79. 09:46But the design behind the competition itself is tremendously important.
80. 09:52A 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.
81. 10:04I really like one tweet from Andre Kapasi about 10 years ago where he said greeting descent can write code better than you.
82. 10:15I'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
83. 10:32handw write better code than a trained model.
84. 10:37Okay, now obviously no one is seriously trying to handr write code to beat a model.
85. 10:43However, 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
86. 10:56today.
87. 10:58I think how gradient descent change coding is a great metaphor for how auto research will change research and ML engineering.
88. 11:08It commonize certain execution skills.
89. 11:12At the same time, it makes some higher level skills far more valuable.
90. 11:19So actually doing auto research is a lot like training a model.
91. 11:23Your codebased abstraction is essentially the architecture.
92. 11:28It sets the constraint and the priorities um for what the agent can explore.
93. 11:35Your eval is the loss function and the data.
94. 11:38It sets what the agent optimizes for.
95. 11:43Take the eval first.
96. 11:44The eval is the signal you use to train a model.
97. 11:49In this case, it's training your code.
98. 11:53It plays the same role that like data and the loss function uh in model training or in a reinforcement learning setting.
99. 12:02It's like a environment that the agent is training.
100. 12:07Nowadays, no one would argue data or environments um don't matter.
101. 12:14And uh this is where a vertical mode can also be built.
102. 12:19You might have a proprietary data for evaluation or a unique understanding of a in a particular field what matters and how to measure it.
103. 12:30and a good evaluation would be amplified more and more as auto research are getting stronger.
104. 12:40The other one I think is really underrated is codebased abstraction.
105. 12:46The abstraction provides the framework that auto research can iterate on and uh that's also that starting point hugely bias the whole search direction.
106. 13:01This is a lot like a architecture design in neural networks.
107. 13:05Different architecture in theory can represent the same function, but the architecture systematically makes some of the functions easier to be learned.
108. 13:19And a good architecture biases the optimization towards solutions that generalize better, perform better, even when the training loss might looks the same.
109. 13:32That's exactly the same for auto research.
110. 13:36Here's an example.
111. 13:37We 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
112. 13:54process 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.
113. 14:14We 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.
114. 14:26In this case, a good abstraction leads to better solutions.
115. 14:31Even though if the agent really want they can steer reward hack.
116. 14:38So my point is using auto research is a new craft.
117. 14:43It's about the designing a here for an agent to climb and we are still very early on it.
118. 14:50I think that makes this extremely exciting time to be an AI engineer.
119. 14:57Other research will change what skills matter most.
120. 15:01Creativity, the judgment to design a good eval or an abstraction.
121. 15:07Those will soon get exponentially more important.
122. 15:12Driving those system itself is where will be a new skill and that one is like barely existed one or two years ago.
123. 15:22So the search is automated.
124. 15:25the human would just move up the stack not out of it.
125. 15:32Again, um we call is a auto research um product research lab.
126. 15:38We 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.
127. 15:48If you think some of this uh useful to you, feel free to follow me on X. Thank you.