# Jun 28, 2025

## AlphaGenome: AI for Better Understanding the Genome

- **AlphaGenome** is a groundbreaking AI tool that predicts the effects of **genetic variants** on biological processes, utilizing long DNA sequences of up to **1 million base pairs** for high-resolution insights into gene regulation.

## No One Is in Charge at the US Copyright Office

- The **US Copyright Office** is currently leaderless, following the abrupt firing of **Shira Perlmutter**, which has raised concerns about the validity of copyright registrations being issued without her signature.

## Lossless LLM 3x Throughput Increase by LMCache

- **LMCache** is an **LLM serving engine** that optimizes performance by storing KV caches across multiple locations, achieving **3-10x delay savings** when integrated with vLLM for various use cases like multi-round QA and RAG.

## Qwen VLo: From "Understanding" the World to "Depicting" It

- **Qwen VLo** represents a significant leap in **multimodal AI**, enabling users to not only understand but also generate high-quality images based on natural language prompts, enhancing creative flexibility and control.

## SymbolicAI: A neuro-symbolic perspective on LLMs

- **SymbolicAI** merges **classical Python programming** with the **differentiable nature of LLMs**, allowing for a seamless integration of syntactic and semantic operations through its modular design, enhancing user customization and extensibility.

## Reinforcement learning, explained with a minimum of math and jargon

- **Reinforcement learning (RL) has revolutionized AI by enabling models to learn through trial and error, overcoming the limitations of imitation learning, which often leads to compounding errors.** This shift has allowed for the development of more reliable agentic AI systems capable of complex, multi-step reasoning.

## Normalizing Flows Are Capable Generative Models

- **Normalizing Flows (NFs)** have been shown to be more powerful than previously recognized, with the introduction of **TarFlow**, a scalable architecture that enhances performance in generative modeling and density estimation tasks.

## [R] OpenEvolve: Automated GPU Kernel Discovery Outperforms Human Engineers by 21%

- **OpenEvolve** utilizes **evolutionary programming** to optimize Metal GPU kernels for transformer attention, achieving an **average decode speed improvement of +12.5%** across 20 inference scenarios, with peak gains of **+106%** in specific tasks.

## Sirius: A GPU-native SQL engine

- **Sirius is a GPU-native SQL engine** that integrates seamlessly with existing databases like DuckDB, achieving a **~10x speedup** over traditional CPU engines for TPC-H queries, making it ideal for analytics and ETL tasks.

## A New Kind of Computer (April 2025)

- **Lightmatter's photonic processor** represents a significant leap in computing, capable of executing advanced AI models like ResNet and BERT with accuracy comparable to traditional systems, marking a shift from theoretical promise to practical application.

## LLMs bring new nature of abstraction – up and sideways

- **LLMs** represent a **fundamental shift** in programming, introducing a **non-deterministic abstraction** that contrasts sharply with the deterministic nature of traditional programming languages like Fortran and Ruby.

## [R] Enigmata: Scaling Logical Reasoning In LLMs With Synthetic Verifiable Puzzles

- **Enigmata** introduces a **comprehensive suite** of 36 tasks designed to enhance Large Language Models' (LLMs) puzzle reasoning skills, featuring a generator for unlimited examples and a rule-based verifier for automatic evaluation.

## Life of an inference request (vLLM V1): How LLMs are served efficiently at scale

- **vLLM V1** serves large language models efficiently by deploying multiple instances across GPUs, utilizing a **continuous batching** algorithm to maximize GPU utilization and throughput while processing requests asynchronously.

## Theoretical Analysis of Positional Encodings in Transformer Models

- **Positional encodings** are essential for transformer models, influencing their **expressiveness** and **generalization**; this study introduces new methods based on **orthogonal functions** like wavelets and Legendre polynomials to enhance performance.

## [D] NVIDIA acquires CentML — what does this mean for inference infra?

- **NVIDIA's acquisition of CentML** signals a strategic shift towards owning both **hardware and software** for AI inference, enhancing efficiency through techniques like **batching** and **quantization**.