AI Daily — 2026-03-10

English 中文

AMI Labs Raises $1.03B Seed Round to Build World Models · Gemini Embedding 2: SOTA Multimodal Emb...

Covering 37 AI news items

🔥 Top Stories

1. AMI Labs Raises $1.03B Seed Round to Build World Models

AMI Labs, led by Yann LeCun after his Meta departure, announced a seed round of $1.03B (~€890M), among the largest ever for a European startup. The company aims to build world-model-based AI systems with persistent memory, reasoning, planning, and safety controls, signaling strong investor appetite for universal intelligence. This funding could accelerate research directions toward generalizable AI while elevating scrutiny around safety and governance. Source-x

2. Gemini Embedding 2: SOTA Multimodal Embedding Space

Gemini Embedding 2 is presented as a state-of-the-art multimodal model that maps text, images, video, audio, and documents into a single embedding space. This enables unified representations across modalities for downstream tasks, potentially speeding cross-modal tasks from search to robotics. The update positions Gemini as the leader in multimodal embeddings today. Source-x

3. Gemini-Powered Docs, Sheets, Slides, Drive Get AI Overviews

Google rolls out Gemini-powered updates across Docs, Sheets, Slides, and Drive, featuring AI Overviews, editable AI-generated slides, and grounded sources for context-aware writing. The enhancements aim to boost productivity with AI-assisted creation and reliable sourcing, and are available today to G1 Pro and Ultra users. Source-x

📰 Featured

AlphaGo/AI Milestones

AlphaGo’s Move 37 Ignited Modern AI Era — Move 37 is recognized as a watershed moment that demonstrated AI could tackle real-world problems and catalyzed modern RL and AGI thinking; its influence continues to shape safety and governance discussions. Source-x

Hardware & Infrastructure

Mira Murati thanks Nvidia for 1GW Vera Rubin systems — OpenAI plans to deploy at least 1GW of Vera Rubin systems, hinting at a Nvidia partnership to scale adaptable collaborative AI and signaling hardware-scale commitments as a core AI enabler. Source-x

On-device AI / Edge

RunAnywhere Launches Fast AI Inference on Apple Silicon (YC W26) — On-device AI inference engine optimized for Apple Silicon with MetalRT claims outperforming competing stacks; open-sources RCLI, a fast end-to-end voice AI pipeline on-device, with benchmarks on M4 Max showing noticeable speedups. Source-github

Open Source AI Agents / Data Tools

BettaFish to MiroFish: Open-Source Multi-Agent Public Opinion AI — BettaFish (data collection) and MiroFish (panoramic forecasting) enable automated public opinion analysis across 30+ platforms, highlighting an AI-enabled data-to-decision loop with a Wuhan University case study. Six advantages are touted: AI-driven monitoring, a multi-model analysis engine, strong multimodal capabilities, a forum-style Agent collaboration mechanism, and seamless private-public data integration. Source-github

LLM Benchmarks & Open Source Leaderboards

Topped HuggingFace Open LLM Leaderboard on Two Gaming GPUs — A claim of beating the HuggingFace Open LLM Leaderboard using only two consumer GPUs, with details on hardware, software stack, and optimizations provided. Source-rss

Open LLM Leaderboard with 2x RTX 4090s

Topped Open LLM Leaderboard with 2x RTX 4090s, no weights — Reddit post describes duplicating a circuit-sized block of about seven middle layers in Qwen2-72B without changing any weights to top the leaderboard; notes ongoing work with newer models and GH200 rigs. Source-reddit

Open Source TTS

Fish Audio Releases Open-Source S2 TTS with Expressive Control — S2 TTS enables expressive voices via natural language emotion tags, supports multi-speaker dialogue in one pass, 100ms time-to-first-audio, 80+ languages, and claims superiority on Audio Turing Test and EmergentTTS-Eval versus closed-source rivals. Source-reddit

Generated by AI News Agent | 2026-03-10

⚡ Quick Bites

OpenAI acquires Promptfoo to bolster Frontier security testing — Acquisition to strengthen security testing for frontier AI initiatives. Source-x
Thinking Machines partners with Nvidia for frontier model training — Collaboration with Nvidia to accelerate frontier model training. Source-x
ConStor Reveals Consistency Bugs in Long-Form LLM Storytelling — Findings about storytelling consistency in long-form LLM outputs. Source-huggingface
Holi-Spatial: Evolving Video Streams into 3D Spatial Intelligence — paper proposing 3D spatial understanding of evolving video streams. Source-huggingface
LoGeR Enables Long-Context 3D Reconstruction with Hybrid Memory — approach for long-context 3D reconstruction using hybrid memory. Source-huggingface
Distribution-Guided Confidence Calibration Improves Answer Selection — calibration method to improve answer selection in LLMs. Source-huggingface
Nanochat: Minimal, cheap LLM training harness — lightweight toolkit for training LLMs. Source-github
Amazon requires sign-off by senior engineers for AI changes — AWS policy requiring senior-engineer sign-off after outages. Source-rss
Qwen3.5-35B-A3B Uncensored Aggressive GGUF Release — Release of an uncensored GGUF for Qwen3.5-35B-A3B. Source-reddit
Happy Birthday, llama.cpp: Open-Source LLaMA Toolkit Expands AI — Anniversary coverage of llama.cpp and its impact on open-source AI tooling. Source-reddit
Qwen3 ASR Outperforms Whisper in Real-Time and Offline — ASR results favor Qwen3 over Whisper in both modes. Source-reddit
1 Million LocalLLaMAs — Milestone celebration for LocalLLaMA models. Source-reddit
Kimi AI Most Token-Efficient for Coding, Easy Setup — Kimi AI emphasizes token efficiency for coding tasks. Source-x
AI agents to become economic actors, buying services and data — Vision of agents acting as market participants and consumers. Source-x
AI Agents Produce but Cannot Transact Economically — Analysis of limitations preventing AI agents from performing autonomous transactions. Source-reddit
Unsupervised RLVR Poised to Scale LLM Training — RLVR approach for scaling LLM training unsupervised. Source-huggingface
I built a programming language with Claude Code — Personal engineering effort integrating Claude Code into a new language. Source-rss
Claude Skills Library Offers 170 Production-Ready AI Coding Plugins — Claude-enabled plugin library for AI coding tasks. Source-github
PgAdmin 4 9.13 Adds AI Assistant Panel — New AI assistant features in the SQL admin tool. Source-rss
Anthropic Claude Code pricing clarified: no $5k per user — Clarity on Claude Code pricing, debunking $5k per user claim. Source-rss
Regret Over Discovering LocalLLaMA and Local LLM Obsession — Personal reflection on LocalLLaMA journey. Source-reddit
Dense attention reduces DeepSeek V3.2 reasoning — Report on dense attention reducing reasoning capabilities in DeepSeek. Source-reddit
Gated Delta Net Enables Vulkan Acceleration in Llama.cpp — Gated Delta Net enables Vulkan acceleration for Llama.cpp. Source-reddit
Anthropic: Claude code reviews cost $15–25 per review — Claude code review cost range disclosed. Source-x
T3 Code Does Not Support Local Models — T3 Code does not support local models. Source-x
Debian decides not to decide on AI-generated contributions — Debian stance on AI-generated contributions. Source-rss
Learnings from Paying Artists Royalties for AI-Generated Art — Case study on royalties for AI-generated art. Source-rss

Generated by AI News Agent | 2026-03-10

━━━━━━ End of Template ━━━━━━