just achieved a score of 53% on the @arcprize - what a feeling! @MindsAI_Jack @MohamedOsmanML lets gooo!
Michael Hodel
99 posts
writer (of programs)
- very excited to win guys, it's been such a blast! let's goo@arcprize 2024 with more than 16k entrants just ended after 5 months, and we rank #1 (@bayesilicon @MohamedOsmanML)! We just scored 58% with a submission that finished after the deadline! We're just getting started. We hope to have an announcement about @tufalabs soon.
- While the code does not generate new tasks, it allows for an "unlimited" number of examples for each of the training tasks. Curious to hear which experiments (enabled by this) people think should be done. Hopefully it is useful to some! repo:[Paper] Current high-scoring team member @bayesilicon shares an ARC-AGI training task generator. More examples "...should enable a wide range of experiments that may be important stepping stones towards making leaps on the benchmark ." arxiv.org/abs/2404.07353
- Replying to @nanulledtalk is cheap. you wouldn't be the first person I came across in the context of ARC making big claims without ever delivering anything. make a submission and let us know what score you got. will be impressed if you get >5%. waiting
- let's go! 🔥 46% @arcprizeWoke up to find only a casual 3% improvement on the SOTA @arcprize today!! Huge kudos to our dream team @MindsAI_Jack @bayesilicon We're just getting started @GregKamradt @fchollet @mikeknoop
- "How sample-efficient can you solve ARC tasks with ML?" is a question I believe is worth studying and one of the motivations behind creating RE-ARC, which should enable this: github.com/michaelhodel/r… @arcprize @fcholletSample efficiency is the most important metric for the future of ML. The ability to improve behavior based on minimal feedback is fundamental to intelligence.
- when ARC merchLove working with this guy. @GregKamradt did an incredible job booking the @arcprize 2024 university tour. 6 down. 11 to go.
- procedurally generating examples for the 400 training tasks of @fchollet's Abstraction and Reasoning Corpus: github.com/michaelhodel/r…
- I told you Michael Hodel was cooking something hot, but this is pure 🔥. Great work Michael. 🏆 Shall we go for 60?
- Replying to @0xSMW and @arcprizeyou should clarify right off the bat that those results are not on the hidden test set, but on some subset of the public data where getting an even much better score is trivial. "reached above 50% on ARC-AGI" is misleading (not saying what that what you're doing isn't impressive)
- Replying to @PlinzBurning Man is my favourite synthetic intelligence conference!










