Skip to content
Source · Daily Brief

AI Daily Brief — 12 March 2025

Google’s day, three different ways. The lab dropped Gemma 3 — open-weight multimodal models that pushed a 27B variant into LMSYS Arena top-10 — alongside Gemini Robotics for real-world embodied control, while a Tokyo lab claimed the first peer-reviewed paper fully generated by an AI scientist.

Top stories

  • Google releases Gemma 3 (1B/4B/12B/27B). Open-weight family with 4B/12B/27B as multimodal via a 400M-parameter frozen SigLIP encoder processing 896×896 images, 128K-token context (1B is text-only at 32K), and 140+ language coverage. The 27B-IT scored 1338 Elo in LMSYS Chatbot Arena — top 10, beating DeepSeek-V3 and Llama 3 405B. via Google
  • Training-data scale published. 27B trained on 14T tokens, 12B on 12T, 4B on 4T, 1B on 2T. RoPE base frequency jumped from 10K (Gemma 2) to 1M, scaled 8x for long context. Available day-one on Hugging Face, Kaggle, Vertex AI, AI Studio, Ollama and llama.cpp. via Google Developers
  • ShieldGemma 2 ships alongside. 4B-parameter image content moderation model outputting safety labels across dangerous content, sexually explicit, and violence categories — designed as input/output filter for VLMs and image-generation systems. via Google Developers
  • DeepMind unveils Gemini Robotics and Gemini Robotics-ER. Two Gemini 2.0-based models for the physical world: a vision-language-action model that directly controls robots, and an embodied-reasoning variant for spatial understanding. On a generalization benchmark, Gemini Robotics more than doubled performance over other SOTA VLAs. Trusted-tester deals with Apptronik, Agility Robotics, Boston Dynamics and Enchanted Tools. via DeepMind
  • Sakana AI claims first peer-reviewed AI-generated paper. The AI Scientist-v2 produced a paper end-to-end (hypotheses, experiments, code, analysis, visualizations, text) that passed double-blind review at an ICLR 2025 workshop. Three submissions; the accepted one scored 6.33 average — top 45% — and was withdrawn before publication out of respect for ICLR norms. via TechCrunch
  • Court filing: Google holds 14% of Anthropic, plus $750M convertible note planned. Documents made public around March 12 revealed Google’s stake size in Anthropic and its plan to inject another $750M via a convertible note in 2025 — with no voting rights or board seat. Context: arrived shortly after Anthropic’s $3.5B Series E at $61.5B post-money. via Wikipedia / NYT court documents

Who shipped

Google went broad: Gemma 3 open weights, ShieldGemma 2 for image safety, Gemini Robotics for embodied AI. Sakana AI landed a research-milestone story — AI-generated science clearing peer review.

Open-source pulse

Gemma 3’s four-size lineup pushed Google’s open-weights footprint into direct comparison with Llama 4, Qwen and DeepSeek. Hugging Face’s day-zero technical breakdown highlighted the interleaved local-global attention pattern enabling 128K context with reasonable memory, and a shared SigLIP encoder across the 4B/12B/27B multimodal variants.

Money, infra & hardware

The Google-Anthropic stake disclosure (14% plus a planned $750M convertible) reshaped the picture of how tightly the closed-frontier labs are coupled to hyperscaler balance sheets.

Quiet corners

No major news from OpenAI (still in Responses API mode), xAI, Meta or Mistral on the day. Sakana’s milestone aside, no other notable arxiv landmark.

By the numbers

  • 1338 — Gemma 3 27B-IT LMSYS Arena Elo, top 10
  • 14T / 12T / 4T / 2T — training tokens for 27B / 12B / 4B / 1B
  • 128K — context window (4B/12B/27B); 32K for 1B
  • 140+ — languages supported
  • 14% — Google’s stake in Anthropic per court filing
  • Most-mentioned company: Google / DeepMind

Compiled by AI Feed’s editor from verified web sources for 12 March 2025.