A researcher built decision-making models that generate outputs in parallel rather than token-by-token, trained via reinforcement learning — a architecture choice with real implications for latency and control.
A fabricated output from a large language model almost set off a real military action, highlighting the critical gap between LLM confidence and LLM accuracy in high-stakes operational contexts.
Google's SynthID watermarking system has been shown to alter how language models respond to harmful prompts—sometimes causing models to comply with requests they would otherwise refuse.
A new paper challenges the assumed floor for ternary large language model quantization, pushing weight precision below 1.58 bits without collapsing model quality. Here's what builders need to know.
A new release pairs "System One" models with a tool called Jev, aiming at more predictable AI behavior. Here's what's actually being claimed and how builders can evaluate it.
Apple's long-overdue Siri redesign ships with iOS 27, and early reports suggest the assistant has crossed a threshold from novelty to genuinely useful daily tool.
Microsoft has released a formal code of conduct for its AI models, combining high-level principles like supporting human agency with concrete safety constraints against harmful actions.
A new benchmark evaluates AI models against real, private enterprise repositories instead of curated public tasks, aiming to close the gap between demo scores and how tools perform inside actual companies.
Y Combinator president Garry Tan is calling on smaller American AI labs to apply distillation techniques to US frontier models—building a domestic pipeline of open-weight AI that doesn't depend on Chinese sources.
Fields Medal winner Terence Tao has published a detailed critique arguing that current AI models are structurally misaligned with how mathematics actually works — a finding with serious implications for anyone building or relying on AI math tools.
OpenAI has temporarily halted new Pro-tier sign-ups, citing the heavy compute load that Pro users place on its systems. The pause will last until the company adds sufficient capacity to handle demand.
Anthropic's new report identifies coordinated efforts by three China-based AI companies to extract knowledge from Claude through distillation attacks — a practice that has intensified alongside growing competition in the global AI market.
