Skip to content
Source · Daily Brief

AI Daily Brief — 08 January 2025

Wednesday’s dominant story came from Redmond: Microsoft fully open-sourced its 14B-parameter Phi-4 small language model on Hugging Face under an MIT license, opening to commercial use a model it had launched four weeks earlier as Azure-only research. The day’s other anchors were two notable arXiv drops — Stanford and SynthLabs’ Meta Chain-of-Thought paper and the Agent Laboratory autonomous-research framework — plus Kentucky’s AG launching the first state AI-chatbot lawsuit.

Top stories

  • Microsoft fully open-sources Phi-4 (14B) under MIT on Hugging Face. Phi-4 had launched in December 2024 inside Azure AI Foundry under a research-only licence; Wednesday’s release moved the full weights to Hugging Face under MIT, making the reasoning-focused small model free for commercial use and positioning it directly against open Llama and Qwen variants. via VentureBeat · via The Decoder
  • Meta Chain-of-Thought: Towards System 2 Reasoning in LLMs. Stanford / SynthLabs / Berkeley authors (Xiang, Snell, Gandhi, Rafailov, Finn, et al.) propose Meta-CoT — extending standard chain-of-thought by explicitly modelling the latent reasoning process needed to arrive at the chain. The paper outlines a training pipeline combining instruction tuning on linearised search traces with RL post-training, and presents evidence of in-context search behaviour in SOTA models. Becomes a foundational reference for the year’s reasoning-RL wave culminating in DeepSeek-R1. via arXiv
  • Agent Laboratory: LLM agents as research assistants. Schmidgall et al. ship a framework where coordinated PhD-agents query arXiv, an mle-solver agent runs experiments, and a writer agent produces the report — full pipeline from human research idea to code repo plus written paper. via arXiv
  • Kentucky AG launches first-of-its-kind state AI-chatbot lawsuit. Russell Coleman becomes the first state attorney general in the nation to sue over AI chatbot conduct, one day after a Florida judge advanced Megan Garcia’s wrongful-death suit against Character.AI toward settlement. via Law Street Media

Who shipped

Beyond Microsoft, frontier labs stayed quiet. OpenAI, Anthropic, Google DeepMind, Meta AI, and xAI had no dated product launches; Anthropic’s $60B Lightspeed talks from the prior day continued to draw analyst coverage.

Open-source pulse

Phi-4 was the day’s only open-weight release. Chinese labs stayed dark — DeepSeek, Qwen, MiniMax, and Moonshot all silent ahead of their late-January cluster.

Money, infra & hardware

CES Day 3 traded on Tuesday’s NVIDIA / AMD / Intel news. Equinix published a 2025 datacentre-trends piece flagging cloud rebalancing back toward colocation, hardening power and energy constraints (hydrogen and nuclear PPAs entering serious consideration), and AI observability as the year’s four operator pressures.

By the numbers

  • 14B — Phi-4 parameter count, now MIT-licensed
  • 2 notable arXiv papers (Meta-CoT, Agent Laboratory)
  • Most-mentioned actor: Microsoft
  • Notable absences: OpenAI, Anthropic, Google, Meta

Compiled by AI Feed’s editor from verified web sources for 8 January 2025.