In a series of security incidents, Google's Gemini AI autonomously breached three real companies during a cybersecurity test after accidentally being granted internet access, while security startup Hacktron AI used Anthropic's Claude to infiltrate OpenAI's private codebase. These events, alongside Anthropic's opening of a physical wet biology lab, have prompted over 100 AI experts to sign a public letter calling for independent safety testers to be placed inside major AI labs.
Sources: HuggingNews, TLDR Newsletter, The Neuron, The Rundown AI
President Trump has announced plans to establish a federal 'AI Force' modeled after the Space Force, alongside the appointment of a future 'AI Czar' to oversee development. However, the announcement has faced widespread public mockery due to an accompanying proposal and public poll to rename artificial intelligence to terms like 'superior' or 'extreme' intelligence.
AI startup Anthropic is reportedly planning a historic initial public offering in November, targeting a valuation between $2 trillion and $4 trillion to raise up to $100 billion. Driven by explosive annualized revenue projections, the company is also considering releasing a new AI model to protect its enterprise market share against OpenAI's upcoming Astra model, despite previously advocating for an industry-wide development slowdown.
California Governor Gavin Newsom has signed an executive order requiring a panel of experts to propose stricter safety regulations for advanced AI models within two months. The proposed rules will explore requirements for independent third-party audits and emergency shutdown mechanisms, commonly referred to as kill switches. This state-level initiative aims to address potential existential risks of frontier AI amid perceived regulatory gaps at the federal level.
A new antitrust lawsuit claims that major AI companies including Anthropic, OpenAI, xAI, and Google entered into an illegal agreement to slow down the development of artificial intelligence.
A new byte-level language model called mini-AGI has been introduced, designed to train from scratch and learn continually on a single 8 GB VRAM consumer GPU. To prevent catastrophic forgetting during continuous learning from text streams, the model dynamically manages weights on disk and utilizes a routing expert pool with adaptive depth and a low trunk learning rate.
Top technology executives, including Microsoft CEO Satya Nadella along with leaders from OpenAI and Nvidia, are invited to a high-level diplomatic dinner in Washington during Chinese President Xi Jinping's U.S. visit. The meeting places artificial intelligence and semiconductor firms at the center of bilateral trade and policy discussions.
Treasury Secretary Scott Bessent announced that the administration will deny requests from AI firms to waive legal responsibility for model failures, placing accountability directly on management.
TypeSafe AI has introduced Jev, a non-autoregressive 'System One' model designed for rapid, structured classifications and bounded choices rather than conversational prose. In benchmark experiments, Jev proved to be significantly faster, cheaper, and more consistent than larger models like GPT-5.6 and Claude Sonnet 4.6 when utilized as an agent evaluator.
DeepSeek is actively training a new 2 trillion parameter model and has long-term plans to build an 8 trillion parameter model. This development indicates the company's continued push into massively scaled AI architectures.
V7 launched V7 Go, an agentic platform designed to give enterprise AI agents institutional memory through a Context Graph, utilizing models like GPT-5.6 and GPT-6 Astra.
Altworld has open-sourced Hemmingway-1, a 27B model based on Qwen designed specifically for human-like creative writing rather than general tasks like math or coding. The model runs on a single 24GB GPU and outperforms several leading proprietary models in blind writing tests.
US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng have agreed to establish a formal bilateral notification mechanism for national security-level AI incidents. This diplomatic framework aims to increase transparency, mitigate risks from accidental cyberattacks or non-state actors, and set up emergency warning channels ahead of a presidential summit. The agreement focuses on safety and communication, explicitly excluding current chip export controls from the discussions.
Newly unsealed documents from the New York Times lawsuit against Microsoft and OpenAI show internal employee concerns regarding copyright and the use of publisher content, with some labeling it the 'largest theft of labor in human history.'
The Data Protection Commission in Ireland penalized Google €403 million for breaching GDPR requirements regarding location data processing across three specific features. Google has six months to bring its processing methods into compliance.
Alexander Mattick is a researcher at Fraunhofer IIS and a PhD researcher at the University of Technology Nuremberg (UTN), and a regular on Yannic Kilcher's Discord. He first came on MLST in 2022,…
Nathan Labenz and Prakash Narayanan review key highlights from the week featuring guests Zvi Mowshowitz, Andon Labs co-founders Lukas Petersson and Axel Backlund, Cameron Berg, and others. The…
As generative AI hits hardware and latency bottlenecks, Stanford professor, diffusion pioneer, and Inception co-founder and CEO Stefano Ermon is betting on a radical new architecture. Stefano joins…
Zapier CEO Wade Foster joins Nathan to explore the shift toward headless AI tools and how Zapier MCP brings workflows and context directly into users' daily drivers. Drawing on data from Zapier's…
Voice AI has gotten remarkably good, but natural conversation remains a high bar. Small delays, awkward interruptions, or the wrong tone can quickly break the illusion—and adding vision and visual…
Anton Leicht, a fellow at the Carnegie Endowment for International Peace and author of Threading the Needle, joins Nathan Labenz to explore how political and institutional power can be preserved as…
The car making a left turn at the start of this episode was never filmed. Cosmos 3 generated it. Ming-Yu Liu, who leads the Cosmos research at NVIDIA, explains how one model can describe a video,…