Deep Learning Weekly: Issue 421

This week in deep learning, we bring you Claude can now create and use files, Measuring Thinking Efficiency in Reasoning Models: The Missing Benchmark, and a paper on Parallel-R1: Towards Parallel Thinking via Reinforcement Learning.

You may also enjoy The UAE Showcases Its Abilities In AI Reasoning With K2 Think Model, Why language models hallucinate, a paper on On the Theoretical Limitations of Embedding-Based Retrieval and more!

As always, happy reading and hacking. If you have something you think should be in next week's issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Claude can now create and use files \ Anthropic

Claude can now create and edit Excel spreadsheets, documents, PowerPoint slide decks, and PDFs directly in Claude.ai and the desktop app

The UAE Showcases Its Abilities In AI Reasoning With K2 Think Model

The Mohamed bin Zayed University of Artificial Intelligence and the G42 technology group have just announced the open-source K2 Think reasoning model.

A new generative AI approach to predicting chemical reactions

A new generative AI system developed at MIT could provide realistic predictions for a wide variety of chemical reactions, while maintaining real-world physical constraints.

Atlassian acquires AI browser developer The Browser Company for $610M

Atlassian is buying The Browser Company, a startup that develops browsers with embedded AI features, for $610M.

Mistral AI raises $2B led by semiconductor equipment maker ASML at $14B valuation

Mistral AI announced that it has raised €1.7 billion, about $2 billion, in a Series C funding round led by Dutch semiconductor equipment manufacturer ASML.

MLOps & LLMOps

Nano Banana + Milvus: Turning Hype into Enterprise-Ready Multimodal RAG

A blog post demonstrating how to integrate the Nano Banana image generation model with the Milvus vector database to build an enterprise-ready multimodal RAG system.

Learning

Measuring Thinking Efficiency in Reasoning Models: The Missing Benchmark

A critical report about a systematic investigation into the token efficiency of Large Reasoning Models, comparing open-weight and closed-weight models across various reasoning domains.

Why language models hallucinate

An explanatory research publication from OpenAI arguing that language models hallucinate because current evaluation methods reward guessing.

Welcome EmbeddingGemma, Google's new efficient embedding model

An article introducing Google's EmbeddingGemma, a new efficient, state-of-the-art multilingual embedding model with 308M parameters, designed for diverse natural language applications and on-device use cases.

Accelerating scientific discovery with AI-powered empirical software

An article about an AI system ...