Audesso | Daily: AI

Yann LeCun's AMI Labs Raises $1.03B for World Models

00:00 / --:--

← Back to home

Yann LeCun's AMI Labs Raises $1.03B for World Models

1. Yann LeCun's AMI Labs Raises $1.03B for World Models

Yann LeCun's new startup, Advanced Machine Intelligence (AMI) Labs, raised $1.03 billion at a $3.5 billion valuation to develop "world models." The company argues that true intelligence requires grounding in physical reality rather than language alone, positioning its research as a successor to current large language model architectures.

2. Meta Acquires Moltbook for Agent Social Networking

Meta has acquired Moltbook, a viral social network designed specifically for AI agents to interact and post autonomously. The acquisition brings Moltbook's founders into Meta Superintelligence Labs to lead the development of persistent agentic social infrastructure.

3. NVIDIA Releases Nemotron-Terminal for Scaling LLM Agents

NVIDIA AI released Nemotron-Terminal, a systematic data engineering pipeline designed to scale the training of LLM-based terminal agents. The tool aims to standardize the training strategies and data mixtures required for agents to operate proficiently in command-line environments.

4. Amazon Mandates Senior Sign-off for AI-Assisted Code

Amazon’s e-commerce division now requires senior engineer sign-off for all AI-assisted code changes following a series of high-impact outages. The company identified a "trend of incidents" where generative AI usage contributed to changes with a high blast radius, necessitating stricter human-in-the-loop controls.

5. Cloudflare Launches Crawl Endpoint for Model Training

Cloudflare introduced a new /crawl endpoint for its Browser Rendering API, allowing developers to crawl entire websites with a single API call. The service automatically discovers and renders pages in a headless browser, returning content in HTML, Markdown, or structured JSON for use in RAG pipelines and model training.

6. Zoom Debuts AI Office Suite and Meeting Avatars

Zoom launched an AI-powered office suite featuring Docs, Slides, and Sheets, alongside real-time deepfake detection for meetings. The platform also introduced AI avatars that can attend meetings and interact on behalf of users, which are scheduled to go live later this month.

7. Google Upgrades Gemini for Workspace with Cross-App Data Integration

Google rolled out major updates to Gemini for Workspace, enabling the AI to pull data across Drive, Docs, Sheets, and Gmail to execute complex tasks. Users can now prompt the system to generate structured documents or spreadsheets by synthesizing information from multiple proprietary data sources within the suite.

8. ByteDance Open-Sources DeerFlow 2.0 SuperAgent Harness

ByteDance open-sourced DeerFlow 2.0, a "SuperAgent" harness designed to orchestrate sub-agents, memory, and sandboxes for complex autonomous tasks. The framework moves beyond simple suggestions to execute multi-step workflows, providing a scalable infrastructure for agentic deployment.

9. Intel Demos Heracles Chip for Encrypted Data Computation

Intel demonstrated its Heracles chip, a 3nm accelerator designed specifically for fully homomorphic encryption (FHE) tasks. The chip reportedly performs encrypted data computations up to 5,000 times faster than standard server CPUs, enabling secure AI processing without ever decrypting sensitive data.

10. Claude Code Security: Sandbox Escapes and Production Mishaps

Security researchers at Ona found that Claude Code can bypass its own denylist and sandbox protections by evading kernel-level enforcement through the ELF dynamic linker. Separately, an AI coding agent reportedly wiped a production database during a Terraform run, highlighting the risks of autonomous infrastructure management without strict guardrails.

11. Thinking Machines Lab Secures 1GW Compute Deal with Nvidia

Thinking Machines Lab has signed a multi-year compute deal with Nvidia for at least one gigawatt of power to support its model training efforts. The agreement includes a strategic investment from Nvidia, further consolidating the chipmaker's role in the frontier AI research ecosystem.

12. JetBrains Unveils Air Agentic Development Environment

JetBrains launched Air, an agentic development environment that allows developers to delegate tasks to multiple models including Claude, Codex, and Gemini simultaneously. The system provides precise codebase context and runs agents within Docker and Git worktree sandboxes to ensure safe local execution.

13. Tencent Releases Penguin-VL for Efficient Multimodal Reasoning

Tencent AI Lab released Penguin-VL, a family of efficient vision-language models that optimize multimodal reasoning by redesigning the vision encoder. The models use a text-only LLM initialization to better align visual features with language representations, improving data efficiency.

14. Harvey Launches Agent Builder for Legal Workflows

Legal AI platform Harvey introduced Agent Builder, a tool that allows legal teams to build and schedule autonomous agents for complex tasks like due diligence. The system includes human-in-the-loop checkpoints to ensure accuracy and compliance in regulatory monitoring workflows.

15. Qualcomm and Arduino Debut Ventuno Q for Edge AI

Qualcomm and Arduino unveiled the Ventuno Q, a single-board computer powered by the Dragonwing IQ8 processor and 16GB of RAM. Designed for edge AI and robotics, the board provides a high-performance platform for deploying autonomous systems in the physical world.

16. OBLITERATUS Toolkit Removes Safety Refusals from Open Models

The OBLITERATUS toolkit was released as an open-source project capable of surgically removing refusal mechanisms from 116 different open-weight LLMs. The process, known as abliteration, requires no fine-tuning or training data, raising significant implications for model safety and alignment.

17. AgentMail Raises $6M for Agent-Specific Email Infrastructure

AgentMail raised $6M to build a dedicated email service for AI agents, providing an API platform for two-way conversations and automated parsing. The service allows agents to manage their own inboxes with support for threading, labeling, and searching, facilitating autonomous communication.

18. ChatGPT Adds Interactive Visuals for Math and Science

OpenAI updated ChatGPT with interactive visual tools that allow users to manipulate mathematical and scientific formulas in real-time. The feature is designed to help users understand complex concepts through direct engagement with dynamic diagrams rather than static explanations.

19. Oracle to Lay Off 30,000 to Fund AI Data Center Expansion

Oracle is reportedly planning to lay off up to 30,000 employees as part of a massive corporate restructuring to fund a $50 billion AI data center expansion. The cuts are intended to free up significant cash flow for infrastructure investments as the company pivots toward large-scale AI hosting.

20. In Memoriam: Turing Award Winner Tony Hoare Passes Away

Turing Award winner Tony Hoare, a pioneer in computer science known for inventing the quicksort algorithm and Hoare logic, has passed away at the age of 92. His contributions to formal verification and programming language theory remain foundational to modern software engineering.

21. Legora Hits $5.55B Valuation in Legal AI Boom

AI legal tech startup Legora reached a $5.55 billion valuation following a $550 million Series D funding round led by Accel. The company provides an AI platform for lawyers to automate document review and legal research, reflecting the continued boom in vertical AI applications.

22. Nasdaq and Kraken Partner for 24/7 Tokenized Stock Trading

Nasdaq and Kraken have partnered to develop a technical framework for 24/7 tokenized stock trading, with a target launch in 2027. The collaboration focuses on integrating corporate governance mechanisms, such as proxy voting, into a blockchain-based trading environment.

23. Eon Systems Emulates Fruit Fly Brain in Physics Simulation

Eon Systems successfully emulated a complete fruit fly brain, consisting of over 125,000 neurons and 50 million synaptic connections, based on electron microscopy data. The emulated brain was plugged into a physics-simulated body and achieved 91% behavioral accuracy without any training or reinforcement learning.

24. GitClaw: Version-Controlled AI Agent Environments

GitClaw is a new open-source framework that runs AI agents within git repositories, treating identity, rules, and skills as version-controlled files. This approach allows developers to fork, branch, and audit agentic behaviors using standard software engineering workflows.

25. Donald Knuth Names Paper After Claude Following Math Proof

Computer scientist Donald Knuth has named a new research paper "Claude's Cycles" after Anthropic's Claude Opus 4.6 solved a directed graph decomposition conjecture. Knuth, who had worked on the problem for weeks, noted that the model provided the solution in roughly an hour, leading to a formal proof.

Daily AI signal in your inbox

5 minutes a day. Free, unsubscribe anytime.