# Oct 29, 2025

## EuroLLM: LLM made in Europe built to support all 24 official EU languages

- **EuroLLM** is a **9 billion parameter** multilingual model trained on **over 4 trillion tokens**, designed to support all **24 official EU languages** and enhance Europe’s digital sovereignty through open-source access.

## Nvidia takes $1B stake in Nokia

- **Nvidia's $1 billion investment in Nokia** aims to bolster the latter's AI initiatives and support the development of **next-generation 6G technology**, marking a significant strategic partnership in the tech landscape.

## Our LLM-controlled office robot can't pass butter

- **LLMs struggle with practical tasks**, achieving only **40% completion** on Butter-Bench compared to **95% for humans**, highlighting significant limitations in spatial intelligence and task execution.

## Mapping the off-target effects of every FDA-approved drug in existence

- **EvE Bio's dataset** offers a **comprehensive mapping of interactions** between 1,600 FDA-approved drugs and human cellular receptors, aiming to illuminate off-target effects that are often overlooked in traditional drug development.

## Cursor Composer: Building a fast frontier model with RL

- **Composer** is a **reinforcement learning**-based agent model that achieves coding results **four times faster** than similar models, optimized for real-world software engineering challenges.

## Glyph: Scaling Context Windows via Visual-Text Compression

- **Glyph** innovatively scales context length by transforming long text into images, leveraging visual-text compression to enhance efficiency in vision-language models (VLMs) while maintaining semantic integrity.

## Beyond RaspberryPi: What are all the other SoC vendors up to

- **Q4 2025** marked significant advancements in the **embedded SBC market**, with **Qualcomm's acquisition of Arduino** and the introduction of competitive **Dragonwing SoCs** reshaping the landscape.

## Continuous Nvidia CUDA Profiling in Production

- **Continuous profiling** of NVIDIA CUDA applications is now possible with the **world's first open-source low-overhead profiler**, integrated into the [v0.43.0 release of the parca-agent](https://github.com/parca-dev/parca-agent/releases/tag/v0.43.0), enabling real-time performance insights without significant performance penalties.

## Extropic is building thermodynamic computing hardware

- **Extropic** introduces **thermodynamic computing**, a groundbreaking approach that leverages thermodynamic principles to enhance computational efficiency and performance, as showcased in their [launch video](https://www.youtube.com/watch?v=Y28JQzS6TlE).

## Powerful and precise multi-color lasers now fit on a single chip

- **Researchers have developed a compact chip that generates a powerful frequency comb**, enabling the simultaneous transmission of multiple data streams, which enhances the efficiency of data centers and other applications.

## Tongyi DeepResearch Technical Report

- **Tongyi DeepResearch** is an advanced **agentic large language model** designed for **long-horizon research tasks**, utilizing an innovative end-to-end training framework that enhances autonomous deep research capabilities.

## The Continual Learning Problem

- **Memory layers** enable continual learning by allowing models to update parameters selectively, achieving only **11% forgetting** compared to **89%** with full finetuning and **71%** with LoRA when learning new facts from TriviaQA.

## AgentFold: Long-Horizon Web Agents with Proactive Context Management

- **AgentFold** introduces a **proactive context management** approach, enhancing long-horizon task performance by dynamically sculpting context rather than passively accumulating it, inspired by human cognitive processes.

## [D]NLP conferences look like a scam.

- **NLP conference papers often lack theoretical justification**, with 90% of reviewed papers failing to provide solid foundations, while the remaining 10% mislabel findings as _theorems_ under unrealistic assumptions.

## A Year of Fast Apply – Our Path to 10k Tokens per Second

- **Relace Apply 3** achieves **10k+ tokens per second** throughput while maintaining **state-of-the-art accuracy**, thanks to innovative training methods and dataset curation that focus on real-world coding tasks.