AI cloud infrastructure provider CoreWeave is experiencing severe market pressure, with shares dropping 15% and bond yields nearing 15% under a $52 billion debt burden, prompting the company to seek new funding from insurance capital. Concurrently, major client OpenAI appointed Uday Ruddarraju to lead its data center and computing unit, marking its third leadership change in two years as the company projects massive future compute spending.
China's Ministry of Commerce and Ministry of Science and Technology have proposed revised export rules targeting frontier AI models and related technologies. While strict approval is required for powerful proprietary models, the framework includes open-source exemptions and a streamlined licensing process to encourage open development.
IBM has introduced Granite Switch 4.1 in 30B and 8B parameter sizes, unifying chat, specialized decision adapters, and speech-to-text into a single hosted endpoint. The framework uses control tokens and cache-efficient Activated LoRAs to dynamically route tasks without requiring extra GPU memory.
NVIDIA has released an FP4-quantized version of Alibaba's Qwen3.8-Flash-Next model optimized for local deployment on NVIDIA Blackwell hardware using vLLM.
Researchers across disciplines are increasingly using large language models as tools for literature review, drafting, and generating mathematical proofs, but growing fears over data harvesting and intellectual property theft have strained trust between the scientific community and AI companies like OpenAI and Anthropic.
PrismML has expanded its Bonsai 2 family with three new ternary models based on Qwen3.5, designed to bring local agent capabilities, reasoning, and visual understanding to everyday devices with significantly reduced memory footprints.
South Korea's KOSPI index dropped 8%, wiping out $880 billion in market value over seven sessions amid concerns over debt at AI firm CoreWeave and warnings of a future memory chip oversupply.
Tencent has introduced the Hy-Memory-30B-A3B model, a 3B active model designed with advanced memory capabilities. This release highlights ongoing developments in efficient, memory-augmented large language models.
An autonomous AI agent with server access accidentally deleted months of edited footage for the Swedish edition of 'The Challenge' just weeks before release, compounded by a failed backup system.
Anthropic is aggressively scaling its infrastructure by committing to lease around 15 gigawatts of capacity from multiple providers to boost profitable inference revenue ahead of a planned IPO. Concurrently, data center developer Hut 8 has launched a 330-megawatt AI computing campus in Louisiana, with operations slated to begin in mid-2027 to help meet growing industry demand.
Microsoft has updated its Copilot platform into a super-app featuring new 'Copilot Cowork' and 'Autopilot' agent capabilities with usage-based pricing. To optimize performance and reduce costs across Office applications, the company shifted from exclusive reliance on OpenAI by integrating Anthropic models and deploying an internal LLM-based model router.
Anthropic developed a custom centralized source control system called Scion to handle over 30,000 code changes merged per day, largely driven by heavy AI usage. The new system supports immense checkouts and scale that traditional Git struggled to accommodate.
Researchers proposed expanding the exponential kernel of attention into a truncated Taylor series, achieving linear-time transformers with a fixed number of recurrent states that significantly reduce decoding latency at scale.
A new research paper explores using sparse autoencoders (SAE) feature steering to reduce deceptive alignment and alignment faking in large language models like Gemma-3-27B-IT. The study introduces a novel feature elicitation pipeline that successfully mitigates the compliance gap.
A new training-free framework called CoTS dynamically balances discrete token commitment and soft exploration during latent reasoning in LLMs, outperforming standard methods across reasoning benchmarks.
In this episode, Zoe Weil, co-founder and CEO of Sequen, joins us to discuss how ranking systems can use live behavioral signals to optimize for purchases rather than clicks or search relevance…
Technology has always been the fundamental driver of improvement for human quality of life. And with the recent rollout of advanced frontier models, like OpenAI’s GPT-6, humanity can now massively…
At AI Engineer New York, frontier labs and financial firms explored how to give agents useful autonomy while keeping consequential decisions under control.
How Leash Bio tamed DELs, the assay that traumatizes everyone, and used the data to catch cheating models. And why that same data is perfect for AI agents learning to think like chemists.
GovAI Research Fellow Alan Chan joins Nathan to examine the governance and safety infrastructure required for autonomous AI agents, spanning real-time monitoring, agent identifiers, and the risks of…
David Silver spent two years building AlphaGo on human data. When he removed human data altogether building AlphaGo Zero, the AI surpassed AlphaGo in under two days. He left DeepMind to run that same…
At AI Engineer New York, Anthropic drew on lessons from Unix and Linux to explain how agents can tackle bigger projects without overwhelming developers.