# Sep 16, 2025

## Microsoft Favors Anthropic over OpenAI for Visual Studio Code

- **Microsoft's Visual Studio Code** now features an **auto AI model selector** that prioritizes **Claude 4** over **GPT-5**, indicating a strategic shift towards Anthropic's AI capabilities for enhanced coding performance.

## IBM Technology Atlas

- **IBM's Technology Atlas** outlines six strategic roadmaps aimed at revolutionizing **performance** and **efficiency** in IT and business, focusing on areas like **AI**, **Quantum**, and **Hybrid Cloud**.

## GPT‑5-Codex and upgrades to Codex

- **GPT-5-Codex** is a fine-tuned variant of GPT-5, specifically designed for AI-assisted programming tools, enhancing code review capabilities and integrating with existing platforms like VS Code and Codex Cloud.

## Show HN: Pyproc – Call Python from Go Without CGO or Microservices

- **pyproc** enables seamless integration of Python functions into Go applications, allowing for **zero network overhead** and **true parallelism** by utilizing Unix Domain Sockets for inter-process communication, thus avoiding the limitations of traditional methods like CGO and microservices.

## Building towards age prediction

## 🤗`LeRobotDataset`: Bringing large-scale datasets to lerobot

- **LeRobotDataset:v3.0** introduces a **file-based structure** that consolidates multiple episodes into single files, enhancing scalability and efficiency for large datasets in robot learning.

## AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models

- **AMQ** (Automated Mixed-Precision Weight-Only Quantization) optimally balances **model quality** and **memory usage** for Large Language Models (LLMs) by assigning layer-wise quantization bit-widths, enabling broader deployment under strict constraints.

## RAGs to Riches: RAG-like Few-shot Learning for Large Language Model Role-playing

- The **RAGs-to-Riches** framework reformulates LLM role-playing as a **text retrieval problem**, enhancing few-shot learning by incorporating curated reference demonstrations to improve response authenticity and character consistency.

## SpecVLM: Fast Speculative Decoding in Vision-Language Models

- **SpecVLM** introduces a novel approach to **speculative decoding** in vision-language models (VLMs), achieving **1.5–2.3x speedups** over traditional autoregressive inference through an innovative elastic visual compressor that optimizes performance based on input characteristics.

## HARP: Hallucination Detection via Reasoning Subspace Projection

- **HARP** (Hallucination detection via reasoning subspace projection) introduces a framework that effectively separates **semantic** and **reasoning** information in LLMs, enhancing hallucination detection by projecting hidden states onto a reasoning subspace.