# Jan 10, 2026

## Extracting books from production language models (2026)
- **Substantial amounts of copyrighted text can be extracted from production LLMs**, with findings indicating that models like Claude 3.7 Sonnet can output entire books near-verbatim, achieving an nv-recall of **95.8%** in some cases.

## AI solves Erdos problem #728 (Terence Tao mathstodon post)
- **Terence Tao** discusses the **impact of AI tools** on solving problems related to **Erdős**, highlighting their potential to enhance mathematical research and discovery.

## Show HN: EuConform – Offline-first EU AI Act compliance tool (open source)
- **EuConform** is an **open-source tool** designed to ensure compliance with the **EU AI Act**, enabling users to classify risk levels, detect algorithmic bias, and generate compliance reports entirely offline, adhering to **GDPR** and **WCAG 2.2 AA** standards.

## GPU memory snapshots: sub-second startup (2025)
- **GPU memory snapshots** enhance cold start performance by capturing the entire GPU state, enabling workloads to start up **up to 10x faster** than traditional methods, significantly improving user experience in serverless environments.

## [D] deepseek published a new training method for scaling llms. anyone read the mhc paper?
- **Deepseek's new method, manifold constrained hyper connections (MHC), aims to enhance performance in large language models (LLMs) by allowing internal information sharing while maintaining stability**, as detailed in their [arXiv paper](https://www.arxiv.org/abs/2512.24880).

## How we made v0 an effective coding agent
- The **v0 Composite Model Family** enhances coding efficiency through a **multi-step agentic pipeline**, integrating a dynamic system prompt, **LLM Suspense**, and deterministic autofixers to improve code generation reliability significantly.

## [P] LLM Jigsaw: Benchmarking Spatial Reasoning in VLMs - frontier models hit a wall at 5×5 puzzles
- **LLM performance declines sharply** as puzzle complexity increases, with **solve rates plummeting from 95% on 3×3 grids to nearly 0% on 5×5 grids**, highlighting a significant challenge for frontier multimodal LLMs.

## Warp Specialization in Triton: Design and Roadmap
- **Warp specialization in Triton enhances GPU kernel performance** by creating tailored code paths for each warp, effectively minimizing control flow divergence and optimizing hardware utilization, which is crucial as kernel complexity increases.

## Supercharging LLMs: Scalable RL with torchforge and Weaver
- **torchforge** is a **PyTorch-native RL library** that simplifies large-scale reinforcement learning for LLMs, enabling seamless scaling from single-node to multi-node clusters without infrastructure complexity.