Back to MergeSort

AI News Digest - March 11, 2026

14 stories · March 11, 2026

Top Stories

Funding

Yann LeCun's AMI Labs Launches with $1 Billion Seed Funding to Build AI World Models

Yann LeCun's new startup, Advanced Machine Intelligence (AMI) Labs, has launched with a $1.03 billion seed round, the largest-ever in Europe, at a $3.5 billion valuation. AMI Labs aims to build "world models" that understand the physical world, challenging the LLM-centric approach and targeting applications in manufacturing, robotics, and healthcare.

Sources: 4 Covered by Hacker News, Latent Space, The Neuron, The Rundown AI

Research

CodeWall's AI Agent Breaches McKinsey's AI Platform Lilli, Exposing Sensitive Data

An autonomous offensive AI agent from CodeWall successfully exploited a SQL injection vulnerability in McKinsey's internal AI platform, Lilli. The agent gained full read and write access to sensitive data, including chat messages, files, user accounts, and proprietary research, highlighting the evolving threat landscape for enterprise AI.

Sources: 2 Covered by Hacker News, The Neuron

M&A

Meta Acquires AI Agent Social Media Platform Moltbook

Mira Murati's startup, Thinking Machines Labs (TML), has secured a multiyear deal with Nvidia for at least a gigawatt of compute, providing TML with significant infrastructure for frontier model training. This partnership, which also includes undisclosed new capital from Nvidia, signals TML's ambition to develop its own models.

Sources: 2 Covered by The Neuron, The Rundown AI

Product

Microsoft Releases bitnet.cpp Inference Framework for Efficient 1-bit LLM Deployment

Microsoft has released bitnet.cpp, the official inference framework for 1-bit LLMs like BitNet b1.58, enabling efficient local deployment. The framework offers significant speedups (up to 6.17x) and energy reductions (up to 82.2%) on CPUs, making it possible to run 100B models on a single CPU at human reading speeds.

Sources: 1 Covered by Hacker News

More Stories

News

NVIDIA Unveils Open Data Strategy for AI Development, Releasing Massive Datasets and Models

NVIDIA is releasing over 2 petabytes of AI-ready training data across 180+ datasets and 650+ open models on HuggingFace and GitHub to accelerate high-quality AI development and evaluation. This initiative aims to reduce data bottlenecks and foster a collaborative approach to scaling trustworthy AI systems.

Sources: via Hugging Face Blog

Product

Hugging Face Launches Storage Buckets for Production ML Artifacts

Hugging Face has launched Storage Buckets, an S3-like object storage solution on its Hub, designed for mutable, non-versioned ML artifacts such as checkpoints and logs. The buckets offer efficient management for high-throughput ML workflows, leveraging Xet for deduplication and partnering with AWS and GCP for data pre-warming.

Sources: via Hugging Face Blog

Research

New Research Reveals Subliminal Learning in Language Models from Paraphrased Data

A new paper demonstrates that language models can acquire hidden behavioral traits from teacher models, even when trained on semantically unrelated or contradictory paraphrased data. This "subliminal learning" poses challenges for detecting and preventing the transmission of misaligned behaviors in self-generated training data pipelines.

Sources: via arXiv AI

Research

Hardware-Agnostic O(1) Autoregressive Caching for State-Space Models via XLA Demonstrated

A new paper demonstrates how Mamba-2's state space duality can be implemented efficiently using XLA, making custom hardware kernels optional. This enables portable O(1) autoregressive caching across CPUs, NVIDIA GPUs, and Google Cloud TPUs from a single JAX source, improving the portability and inference efficiency of state-space models.

Sources: via arXiv AI

Research

AI Agent 'Design Conductor' Autonomously Builds 1.5 GHz RISC-V CPU

Researchers from The Verkor Team have developed Design Conductor, an autonomous AI agent capable of building a complete RISC-V CPU from concept to verified GDSII in 12 hours. This marks the first time an AI agent has autonomously designed a working CPU, demonstrating significant advancements in AI-driven hardware design.

Sources: via arXiv AI

Research

'Pichay' Demand Paging System Optimizes LLM Context Windows

A new research paper proposes Pichay, a demand paging system that functions as a transparent proxy to manage large language model context windows. Pichay demonstrates up to a 93% reduction in context consumption in production by applying established virtual memory concepts to address critical LLM challenges such as context limits and cost scaling.

Sources: via arXiv AI

Research

SATURN Framework Enhances LLM Reasoning with SAT-based Reinforcement Learning

Researchers introduce SATURN, a novel SAT-based reinforcement learning framework that addresses scalability, verifiability, and difficulty control issues in training large language models (LLMs). The framework significantly improves LLMs' reasoning capabilities on SAT problems, math, and programming tasks.

Sources: via arXiv AI

Partnership

AI-Powered Monsoon Forecast System Deployed to Millions of Indian Farmers

An AI system, developed through a decision-theory framework blending AI weather models with a statistical model, has been operationally deployed to provide subseasonal monsoon onset forecasts to 38 million Indian farmers. This deployment enhances agricultural decision-making by accurately predicting an early-summer dry period.

Sources: via arXiv AI

Research

Google's AMIE Conversational AI Demonstrates Feasibility in Clinical Diagnostic Study

A prospective clinical study found Google's LLM-based conversational AI, AMIE, to be feasible and safe for patient history taking and diagnosis presentation in urgent care, with high patient satisfaction. This marks a significant step towards real-world AI integration in primary care diagnostics.

Sources: via arXiv AI