Taiwanese prosecutors detained an Nvidia employee and raided the company's Taiwan offices as part of a probe into the alleged smuggling of AI accelerators to China — the first time the investigation has reached the chipmaker itself. The probe has also ensnared Super Micro executives in a $2.5 billion diversion scheme.
Microsoft unveiled its first cybersecurity-specialized AI model alongside Project Perception, an agentic platform that deploys teams of AI agents for continuous monitoring, patching, attack simulation, investigation, and remediation at roughly half the cost of larger general-purpose models.
NVIDIA introduced the Open Secure AI Alliance, bringing together Hugging Face, Nous Research, LMSYS, and other industry leaders to advance AI security through open models, shared vulnerability research, and collaborative defense tooling.
Proposes a method for constructing VLM pretraining mixtures that decouples ratio search from convex allocation, enabling attribution of model performance to specific data sources at scale.
Systematic study of how long-horizon planning ability is acquired in foundation model agents, proposing single- and multi-teacher on-policy agentic distillation that significantly improves planning across the full training pipeline.
Introduces page-local compact key summaries that replace full KV-cache reads during LLM decoding, achieving significant speedups for long-context inference while maintaining output quality.
Enables efficient MoE inference on consumer devices by staging routed experts on demand, dramatically reducing latency for single-user settings without sacrificing quality.
Introduces speculative sandbox scheduling for LLM agents using the Model Context Protocol, reducing memory overhead while maintaining low tail latency for sandboxed tool execution.
Nvidia is using AI agents to automate chip design for its next-generation Rosa CPU and Feynman datacenter platform, integrating EDA tools from Cadence and Synopsys to handle verification, physical implementation, and signoff tasks.
Italian deep-tech startup Generative Bionics revealed the fully functional GENE.01 humanoid platform with smart-skin technology, backed by research published in Nature and built by a team with automotive manufacturing expertise.
At the Design Automation Conference in Long Beach, AI-native EDA startups presented verified production metrics including a 70% first-pass debug rate, marking a milestone where AI agents are now used in production-grade semiconductor design.
Roblox launched Build, an AI-powered feature in its mobile app that lets users generate playable games from text prompts with zero programming skills required. Public alpha testing started July 28 in New Zealand.
Comms lets users deploy AI agents on a real iMessage line by describing the agent in plain English. Handles support, bookings, onboarding, and payments inside the messaging thread with a free tier of 3,000 messages per month.
Webhound builds AI research agents that scale quality with time and budget, generating structured datasets or fully cited research reports. Y Combinator S23 startup, placed #3 on Product Hunt.
Robynn AI's Website Studio uses AI agents that continuously audit, rewrite, and measure websites for SEO, AI search visibility, and conversion — connecting to live sites without rebuild or migration.
Kdeps is a new CLI coding agent built to run on open-weight models, offering an open alternative to proprietary coding agents like Claude Code or Codex.
Source:
Source:
