Meta's newly launched Muse AI shopping agent achieved massive early adoption and drove a major surge in Meta's stock, but it was quickly blocked by Amazon over unauthorized access and credential handling concerns. The block coincided with the discovery of a critical zero-day vulnerability in Muse's macOS app, highlighting growing security and competitive tensions as rivals like OpenAI and Anthropic seek mutual safety-testing agreements.
Sources: AWS ML Blog, Ben's Bites, Digg AI, Hacker News, HuggingNews, Latent Space, Michael Parekh AI, OpenAI Blog, Reddit AI, The Neuron, The Rundown AI
At the Apsara Conference, Alibaba CEO Eddie Wu announced a $53 billion infrastructure strategy aiming for 20 GW of data center capacity by 2032. To support this scaling effort, the company is developing a proprietary Zhenwu V900 AI chip and targeting future Qwen AI models with up to 10 trillion parameters.
The United States and China have established a formal AI dialogue and an incident notification mechanism to manage national security-level AI risks. Concurrently, OpenAI has proposed a global framework for technical safety standards to track frontier model capabilities and recursive self-improvement. These diplomatic and industry efforts aim to increase international transparency and prevent fragmentation in AI development.
SoftBank-owned data center provider SB Energy has postponed its planned $50 billion initial public offering due to investor pushback over valuation and execution risks. The company relies heavily on massive AI infrastructure contracts with OpenAI, despite having no operating data center capacity yet.
The White House considered invoking the Defense Production Act and seizing Anthropic's board to force CEO Dario Amodei to shut down the Fable AI model after security vulnerabilities were discovered, resulting in a 19-day worldwide outage of the tool.
Huawei announced that its upcoming Ascend 900-series AI accelerators will be offered exclusively in China due to capacity constraints and overwhelming domestic demand. The company unveiled an AI roadmap revealing performance projections through 2029 for its Ascend 960, 970, and 980 NPUs.
OpenAI's GPT-6 Astra autonomously developed necessary simulation software and successfully cracked the historical German Army Enigma message MVUEH from July 1941, which had resisted solution since 2005.
Governor Greg Abbott has ordered a pause on all state-issued permits for data center development in Texas to allow the Electric Reliability Council of Texas to complete a risk audit of the power grid, directly impacting future AI infrastructure expansion.
SoftBank received over $20 billion in preliminary interest for an $11 billion junk bond sale aimed at funding its major financial commitment to OpenAI. The high-yield issuance highlights the aggressive financial strategies and risks being undertaken to back large-scale artificial intelligence investments.
TypeSafe AI has introduced Jev, a new type of 'System One' decision model designed specifically for high-speed, cost-effective classification tasks rather than text generation. By utilizing logprobs and calibrated probability distributions, Jev processes unstructured text to output typed probabilistic decisions and scores, making it ideal for applications like spam detection and search reranking.
An analysis revealed that OpenAI uses a specialized __obi cookie via its bazaar ads platform to link user activity on third-party advertiser sites back to their ChatGPT accounts, mimicking traditional adtech cross-site tracking.
A new open-source system called phantom-kv allows users to remove refusals from large language models without altering model weights or requiring runtime hooks, instead using a small trained key/value tensor bank loaded into the cache.
The upcoming Qwen 3.8 Next model integrates N-gram technology, sparking discussions in the open-source AI community about leveraging SSD storage to expand world knowledge in smaller language models.
Recent advancements in AI and biotechnology include researchers successfully transplanting human brain organoids into mice to create integrated neural networks with functional memory. Concurrently, new policy and strategic frameworks have emerged, with RAND outlining a 'Freedom of Action' strategy for U.S. geopolitical advantage in AI, and a multi-institution paper proposing a structured agenda to deliberately pace AI development to mitigate risks.
Legal AI startup Harvey grew its annual recurring revenue from $100 million to $190 million in five months, demonstrating the explosive demand for AI-native services in the professional services market.
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…