True Positive Weekly #118
The most important artificial intelligence and machine learning news and articles
Hey, in this issue: Recent frontier models are reward hacking; how LLMs are served efficiently at scale; how to scale your model; hiding malware inside of neural network models; a new open-weight MoE model with 32 billion activated parameters and 1 trillion total parameters; an open-source tool that guarantees structured outputs during generation with LMs; and more.
[Explained] Life of an inference request with vLLM V1: How LLMs are served efficiently at scale
[Google] How to scale your model
[ArXiv] EvilModel: Hiding malware inside of neural network models
[Model] Kimi K2: an open-weight Mixture-of-Experts model with 32 billion activated parameters and 1 trillion total parameters
[Ars Technica] Why it takes 3,295 people to write one Google AI paper
[Model] SmolLM3: smol, multilingual, long-context reasoner
[Tool] Outlines guarantees structured outputs during generation
Enjoy the newsletter? Please help us make it bigger and better by sharing it with colleagues and friends.