News archive.
Earlier AI news and analysis, kept for reference. Current coverage lives in News & Analysis.
PE firms ditch long-term deal theses for execution-first playbooks
McKinsey reports private equity is moving away from static investment theses toward adaptive models that evolve during ownership. Here's what that shift means for deal selection and value capture.
Geopolitics tops investor concerns for 2026, AI disruption close behind
McKinsey survey finds investors now prioritize geopolitical risk above all other concerns, with AI disruption and capital discipline also reshaping corporate strategy for the year ahead.
McKinsey survey finds geopolitical risk gap—leaders unprepared
McKinsey data shows executives overestimate their ability to handle geopolitical shocks. Five concrete actions can close the readiness gap before crisis hits.
European private banks face margin squeeze without profit-focused strategy
McKinsey warns private banking profits are under pressure. Banks must tighten business focus and monetization discipline or risk further decline. Here's what the industry faces.
NVIDIA TensorRT 11 Splits Inference Across 8 GPUs Without Losing Speed
TensorRT 11.0 adds multi-GPU inference with context parallelism, letting you run large generative models across multiple devices while keeping optimizations like kernel fusion intact. Video and image benchmarks included.
Vulkan Descriptor Heaps Cut GPU Resource Binding Overhead
NVIDIA and Khronos released VK_EXT_descriptor_heap, a Vulkan extension that simplifies how shaders access GPU memory and textures. Supported in driver 610+, it mirrors Direct3D 12's model and reduces boilerplate for dynamic texture indexing and ray tracing.
Meta's Privacy Classifier Keeps LLMs Out of Production Decisions
Meta built a hybrid system that uses LLMs to interpret ambiguous data assets, then distills decisions into deterministic rules that run without AI. The approach routes 85% of requests through logic-based paths in under 40ms.
Brain scans confirm AI-written stories target specific regions
Microsoft researchers used LLMs to write synthetic stories that activate predicted brain regions, closing a gap between predictive models and neuroscience theory. A new method turns black boxes into testable hypotheses.
Hybrid models beat transformers on meaning, lose on copy-paste
Allen AI's token-level analysis reveals hybrid architectures excel at predicting nouns and verbs, but struggle when text repeats verbatim. Here's where each architecture wins.
Spin up a vLLM server on Hugging Face in one command, pay per second
Hugging Face Jobs now lets you launch a private, OpenAI-compatible LLM endpoint with a single CLI command. No infrastructure setup required — useful for evals, batch runs, and testing before committing to production.
360 Launches Frontier AI Model to Challenge Anthropic Claude
360 has released a new frontier-class AI model positioning itself as a direct alternative to Anthropic's Claude. Details on capabilities and availability remain sparse.
OpenAI builds its own AI chip to cut inference costs
OpenAI has designed its first custom silicon for running AI models, aiming to reduce reliance on NVIDIA and lower per-token inference expenses. Details on performance and timeline remain unclear.