# Mar 29, 2025

## Self-Supervised Learning from Images with JEPA

- The **Image-based Joint-Embedding Predictive Architecture (I-JEPA)** enables the learning of **semantic image representations** without hand-crafted data-augmentations, focusing on predicting target blocks from a single context block.

## We hacked Gemini's Python sandbox and leaked its source code (at least some)

- **Google's A.I. Gemini source code was partially leaked** during a bug bounty event, revealing internal structures and proprietary code, including sensitive **Google3** directories and **internal proto files** that should not have been exposed.

## Bolt Graphics Zeus a New GPU Architecture with Up to 2.25TB of Memory and 800GbE

- **Bolt Graphics' Zeus architecture** features up to **2.25TB of memory** and **800GbE connectivity**, utilizing a **RISC-V core** for enhanced performance and flexibility in GPU design.

## Optimizing Matrix Multiplication on RDNA3

- The optimized **FP32 matrix multiplication** on AMD RDNA3 GPU achieves a **60% performance increase** over rocBLAS, demonstrating significant improvements through iterative kernel enhancements and architectural insights.

## [R] Anthropic: On the Biology of a Large Language Model

- **Claude 3.5 Haiku** demonstrates advanced capabilities, such as **multi-step reasoning** and **poetic planning**, revealing its internal processes through attribution graphs, which enhance our understanding of model behavior.

## Low responsiveness of ML models to critical or deteriorating health conditions

- **Machine learning models exhibit significant deficiencies** in recognizing critical health conditions, failing to identify 66% of severe injuries in mortality predictions, highlighting the urgent need for improved model responsiveness in clinical settings.

## Plasmonic Modulators Can Break the Wireless Terahertz Barrier

- **Plasmonic modulators** enable the transfer of signal information from electrical to optical waves at unprecedented speeds, potentially revolutionizing **6G networks** and **AI data centers**.

## [R] DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products

- **DeltaProduct** enhances **state-tracking** in linear RNNs by utilizing **multiple steps** of online gradient descent per token, leading to improved expressivity without sacrificing efficiency.

## [R] Enhancing GUI Agent Reasoning Through Rule-Based Reinforcement Learning

- **UI-R1** innovatively combines **rule-based reinforcement learning** with **large language models** to enhance GUI agents, enabling them to learn from mistakes and adapt to new interfaces effectively.

## [R] Evaluating Multi-Step Spatial Reasoning in MLLMs Through LEGO-Based Visual Tasks

- The **LEGO-Puzzles benchmark** evaluates multimodal language models (MLLMs) on **spatial reasoning tasks**, revealing a significant performance gap between human (85.8%) and AI (59.8%) capabilities, particularly in complex scenarios.