🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Could a major opportunity to improve representation in deep learning be hiding in plain sight? Check out our new position paper: Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis. The idea stems from a little-known

With gratitude to my phenomenal coauthors: Akarsh Kumar, Jeff Clune, and Joel Lehman Paper: arxiv.org/abs/2505.11581 Github: github.com/akarshkumar010…

Is there a cancer at the heart of modern AI, lurking just beneath the surface of its dazzling performance? Our research suggests maybe, but also shows elegant solutions are possible (though how to get them at scale remains a mystery). Much more research needs to be done to

Akarsh Kumar Great work 👏. “There are reasons to believe that current models are still suffering from FER, especially at the frontiers of knowledge where there is less data.” Even if the 10T parameters handle the issue, where do we go from there sounds like a solid foundation to invest

Kenneth Stanley Akarsh Kumar Jeff Clune Joel Lehman Congrats guys! Another excellent work ✨

Exciting new work from Kenneth et al! Check it out!

Kenneth Stanley And they say academia is dead! Nice work!

A nice example and useful discussion here! Here are a few reactions:

Really thought-provoking work! In determining what makes a good representation, it might be the journey that matters not the destination.

"Why can't my trained model generalize?" Maybe b/c its internal representation of the task is whack. Beautiful new paper from Jeff Clune's team. ~thread~ arxiv.org/abs/2505.11581

John Bohannon Akarsh Kumar Are you calling my brain FERry? I resemble that remark! re: 1(trillion) dollar question: we're working on it! We don't know the answer, but I strongly believe it will involve lots of principles from open-endedness (some known, some yet to be discovered). 😊🌱🔬🧪

attentionmech People underestimate evolutionary algorithms. Great things happen when evolution meta-optimizes us to be better at evolving. But a naive one-to-one gene-to-function mapping like is usually done in many implementations is not sufficient for this to arise

super interesting paper....

Image-text alignment is hard — especially as multimodal data gets more detailed. Most methods rely on human labels or proprietary feedback (e.g., GPT-4V). We introduce: 1. CycleReward: a new alignment metric focused on detailed captions, trained without human supervision. 2.

Revisiting Louis Kirsch et al.’s general-purpose ICL by meta-learning paper and forgot how great it is. It's rare to be taken along on the authors' journey to understand the phenomenon they document like this. More toy dataset papers should follow this structure.

We know Attention and its linear-time variants, such as linear attention and State Space Models. But what lies in between? Introducing Log-Linear Attention with: - Log-linear time training - Log-time inference (in both time and memory) - Hardware-efficient Triton kernels

oimo.io/works/life/ Incredible website by saharan / さはら visualizing Conway's Game of Life inside of Game of Life inside Game of Life ... and so on ... forever... Reminds me of the hierarchy of emergent structures in our world from physics to chemistry to biology. How many levels