AI Daily · 2026-07-27

Today’s biggest story: NVIDIA released Cosmos-H-Dreams, a real-time generative surgical simulator that lets users or policies interact with surgical scenes in closed loop on a single GPU—paving the way for fast policy evaluation and synthetic data generation. Meanwhile, an investigation uncovered a major LLM token resale market centered in China, where resellers abuse free trials, stolen credit cards, and open-source proxies like 「one-api」 to sell API access at a heavy discount, raising serious concerns about API key security. In other news, OpenAI published research on AI reshaping job boundaries, NVIDIA is harnessing its Vera CPU to accelerate chip design and expanded its engineering agent toolkit, and LangChain added full-text search to agent trace data.

North America · First-hand

OpenAI

⭐⭐ [Research] How AI is expanding what people do at work

OpenAI's latest research shows AI is expanding what people do at work, with ChatGPT users taking on cross-role tasks and reshaping job boundaries.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics

NVIDIA introduces Cosmos-H-Dreams, a real-time action-conditioned generative simulator for surgical robotics. It distills the Cosmos-H-Surgical-Simulator into a causal student model and serves it via the FlashDreams inference library, achieving real-time interaction on a single RTX PRO 6000 GPU. The model enables closed-loop control by humans or learned policies, trained on dVRK tabletop suturing data and integrated with CMR Surgical's Versius controller, accelerating policy evaluation and synthetic data generation.

Product Update

⭐⭐ [Product Update] NVIDIA Harnesses Vera CPU to Speed Up Design of Next-Generation CPUs and GPUs

NVIDIA is collaborating with Cadence and Synopsys to optimize critical electronic design automation (EDA) applications for the NVIDIA Vera CPU. This collaboration aims to accelerate the design of next-generation CPUs, GPUs and AI systems, and Vera is already being deployed internally at NVIDIA.

⭐⭐ [Product Update] NVIDIA Expands NVIDIA Agent Toolkit With NVIDIA PhysicsNeMo and CUDA-X Libraries to Transform How the World Engineers, Designs and Builds

NVIDIA announced an expansion of its Agent Toolkit for engineering, integrating NVIDIA PhysicsNeMo and CUDA-X libraries as ready-to-use agent tools and skills. This move aims to transform product design and development by making physics simulation and high-performance computing libraries accessible to AI agents. The update reflects NVIDIA's strategy to shift traditional simulation software toward AI-driven workflows.

⭐⭐ [Product Update] Full Text Search in SmithDB: Designing an Inverted Index for Object Storage

SmithDB now supports full-text search and JSON filtering over agent traces, achieving a median (P50) latency of 400 ms despite the underlying data consisting of large, deeply nested JSON documents in object storage. The implementation revolves around an inverted index designed to work efficiently with object storage and complex JSON structures. This gives LangSmith users more flexible querying capabilities for debugging and auditing agent behavior.

Opinion

⭐⭐⭐ [Opinion] An Inside Look at the Relay Market Powering Token Resellers and Fraud

An investigation by Matt Lenhard exposes an LLM token resale market, primarily operating in China. Resellers pool API keys by abusing free trials, proxying through unprotected support bots, stolen credit cards, or chargeback attacks, and use the open-source proxy tools 'one-api' and its fork 'new-api' to load-balance requests, offering access at significant discounts. Buyers seek cheap tokens, evade geo-restrictions, and some collect data for model distillation. Simon Willison says this ecosystem makes him even more cautious about exposing LLM‑powered applications publicly and urges LLM vendors to offer strict per‑key spending caps.