Boring AI News

A selection of the latest DS/ML/AI news from the past 3 days from Boring AI News channel, a curated and opinionated collection of news for pros, enthusiasts, and anyone eager to learn. Why?

OpenR1 Update 2

learning

Hugging Face successfully runs distillation on R1 to create 800k reasoning traces in open-source project.

DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL

learning

This study presents robust methods for enhancing AIME performance efficiently. Researchers found that gradually increasing context length during RL training improves outcomes without excessive compute.

Competitive Programming with Large Reasoning Models

learning

papers

Reinforcement learning enhances large language models’ performance on complex coding tasks. General-purpose model o3 outperforms specialized ones, achieving gold at IOI 2024 and elite Codeforces ratings.

OpenAI Updates Model Spec

misc

OpenAI’s updated Model Spec prioritizes transparency, customizability, and safety under a CC0 license.

SQLite Page Explorer

code

A visual tool displaying SQLite databases as stored on disk and viewed by SQLite.

OpenAI’s Roadmap for GPT-5

news

OpenAI will release GPT-4.5 before GPT-5, merging its o-series and GPT-series models with chain-of-thought technology.

The Illustrated DeepSeek-R1

learning

Jay Alammar’s latest guide uses graphics to highlight DeepSeek-R1’s innovative approach to language models.

Julia for R users

learning

Discover Julia’s speed, multithreading, and type system. Plus, run R within Julia for a seamless transition.

RLHF Book

learning

Nathan Lambert’s concise guide to RLHF and post-training methods for language models.

Agency Is Frame-Dependent

learning

papers

New research suggests that agency, a system’s ability to direct outcomes towards a goal, is frame-dependent. This means that measuring agency must be done relative to a reference frame. The study argues that key properties of agency are also frame-dependent, impacting our understanding of agency in fields like biology, AI, and cognitive science.

AI ‘godfather’ predicts another revolution in the tech in next five years

misc

Yann LeCun anticipates a major AI advancement, highlighting current limitations. AI pioneers, including Yoshua Bengio, discuss safety risks at the global AI summit.

QuEST: Stable Training of LLMs with 1-Bit Weights and Activations

code

QuEST enables stable training using 1-bit weights and activations by enhancing quantization accuracy and speed, and introducing a trust gradient estimator to minimize quantization error.

User-friendly system can help developers build more efficient simulations and AI models

news

MIT researchers automate system for deep learning developers to exploit data redundancy and enhance AI model efficiency.

Designing monochrome data visualisations

learning

Use shades and patterns to create clear, accessible data visualizations with a monochrome palette. Improve your graphs and meet publication standards.

Simpson’s Paradox in a Logistic Regression

learning

Nina Zumel illustrates how Simpson’s paradox, where trends in individual groups reverse when data is aggregated, can occur in logistic regression using a weight loss study example. Unbalanced groups and hidden variables are key factors.

AIs and Robots Should Sound Robotic

misc

AI voices mimic human speech too well. Adding a robotic sound can help users distinguish AI interactions, addressing concerns about deception.

Audiobox Aesthetics

code

This repository offers models to score and rate audio files using various metrics, aiding in retrieval and reinforcement learning rewards.

Better Programming Assistance

code

A conversational approach that integrates coding history, current code, and user instructions for better programming support.

Reinforcement Learning for Long-Horizon Interactive LLM Agents

papers

A method using RL to train interactive digital agents within their intended environments.