# May 20, 2025

## MAGI-1: Autoregressive Video Generation at Scale

**MAGI-1** is a **world model** that generates videos by autoregressively predicting sequences of video chunks, achieving high **temporal consistency** and scalability through innovative algorithms and infrastructure.

## The Behavior of LLMs in Hiring Decisions: Systemic Biases in Candidate Selection

**LLMs exhibit systemic gender bias**, favoring female candidates in hiring decisions, selecting them 56.9% of the time despite identical qualifications, indicating a significant deviation from expected rational behavior.

## Introducing Veo 3 and Imagen 4, and a new tool for filmmaking called Flow

**New generative media models** like **Veo 3** and **Imagen 4** enable the creation of high-quality images, videos, and music, enhancing artistic expression and creativity through advanced AI tools.

## I got fooled by AI-for-science hype–here's what it taught me

**AI's application in plasma physics research revealed significant limitations**, as methods like PINNs often performed worse than traditional numerical techniques, challenging the narrative of AI's transformative potential in science.

## Deep Learning Is Applied Topology

**Deep learning leverages topology** to manipulate data surfaces, allowing complex datasets to be separated through transformations akin to bending and stretching, as illustrated by neural network operations.

## Gemma 3n preview: Mobile-first AI

**Gemma 3n** is a **mobile-first AI** model designed for **real-time** applications, enabling developers to create efficient, on-device experiences that respect user privacy and operate without internet connectivity.

## llm-d, Kubernetes native distributed inference

**llm-d is a Kubernetes-native framework** designed for high-performance distributed LLM inference, optimizing deployment with features like KV-cache aware routing and disaggregated serving to enhance efficiency and reduce latency.

## Robin: A multi-agent system for automating scientific discovery

**Robin** is the **first multi-agent system** that automates the entire scientific discovery process, integrating literature search and data analysis to generate and validate hypotheses autonomously.

## AI's energy footprint

**AI's energy consumption is escalating rapidly**, with projections indicating that by 2028, AI could consume as much electricity annually as **22% of all US households**, driven by the increasing integration of AI into daily applications and the construction of energy-intensive data centers.

## Thinkless: LLM Learns When to Think

**Thinkless** introduces a **learnable framework** that enables LLMs to choose between short-form and long-form reasoning, optimizing performance based on task complexity and model capability.

## [P] OpenEvolve: Open Source Implementation of DeepMind's AlphaEvolve System

**OpenEvolve** is an open-source framework that **evolves entire codebases** using LLMs, enabling the discovery and optimization of algorithms through a structured pipeline of code generation, evaluation, and selection.

## NVIDIA-Powered Supercomputer to Enable Quantum Leap for Taiwan Research

The **new AI supercomputer** at Taiwan's National Center for High-Performance Computing, built by ASUS, will deliver **over 8x more AI performance** than its predecessor, Taiwania 2, significantly enhancing research capabilities in climate science, quantum computing, and large language models.

## Questioning Representational Optimism in Deep Learning

The **Fractured Entangled Representation Hypothesis (FER)** posits that while both evolved and SGD-trained neural networks can produce similar outputs, their internal representations differ significantly, with SGD networks exhibiting a disorganized structure that may hinder generalization and creativity.

## NVIDIA Grows Quantum Computing Ecosystem With Taiwan Manufacturers and Supercomputing

**NVIDIA is enhancing its quantum computing ecosystem** by collaborating with Taiwanese manufacturers like Compal and Quanta, integrating AI supercomputing hardware to accelerate quantum research and development.

## [R] [Q] Why does RoPE need to be decoupled in DeepSeek V2/V3's MLA? I don't get why it prevents prefix key reuse

**RoPE's decoupling in DeepSeek V2/V3's MLA is essential** because it prevents the absorption of projection matrices, which is crucial for efficient key reuse during inference.