NVIDIA unveils Nemotron 3.5 Lightning and NeMo Switchyard: A lightweight open-source model and routing library designed for faster, more efficient agentic AI across edge devices, PCs, workstations, data centers, and cloud. NVIDIA Blog
Mojo 1.0 is here: Modular shipped the stable 1.0 release of the Mojo language, three years after its debut. Includes Python-style lambda syntax, memory safety diagnostics, stabilized LSP for VS Code, and GPU programming AI Skills. Async programming and pattern matching are on the roadmap. Modular
OpenAI ships ChatGPT for Linux — and loses two senior leaders: The ChatGPT desktop app officially landed on Linux. The same day, COO Brad Lightcap announced his departure “to start something new,” and the head of ethics resigned less than a year into the role, per the Financial Times. TechCrunch · FT · TechCrunch
Anthropic’s unreleased model cracks a famous math problem: An as-yet-unreleased Anthropic model reportedly made meaningful progress on one of mathematics’ biggest unsolved problems. The specific problem and methodology remain undisclosed, but the news signals a potential leap in AI’s pure math research capabilities. TechCrunch
xAI launches Grok Bot: xAI released Grok as a standalone bot service, expanding beyond its existing X platform integration into a dedicated product. xAI
Researchers extract hidden reasoning from proprietary LLM APIs: A new study demonstrates that encrypted chain-of-thought blocks returned by Anthropic, OpenAI, and Google APIs are interchangeable across sessions, users, and models. By jailbreaking a weaker model, attackers can recover the stronger model’s hidden reasoning traces — exposing a fundamental vulnerability in current LLM API security architecture. Stolen Thoughts
Manus splits from Meta — AI agent startup reclaims independence: Eight months after Meta’s acquisition, Manus is returning to independent operations. To comply with regulatory requirements in specific jurisdictions, user data generated after December 29, 2025 will be deleted August 23-24, with backup available until August 23. This is the first major case of an AI agent company navigating corporate separation and cross-jurisdictional data regulation in public. Manus
Nvidia becomes its own financier — and absorbs the risk: Per Stratechery’s analysis, Nvidia is expanding customer financing for AI infrastructure buildouts, effectively absorbing the risk of the AI capex cycle onto its own balance sheet. The shift turns Nvidia from a pure supplier into a financial counterparty for the AI industry. Stratechery
Build high-performance AI inference serving with Mojo 1.0 + MAX: With Mojo’s stable 1.0 foundation and MAX’s support for GLM-5.2 and Nemotron-H (both Mamba-2 hybrid models), teams can build GPU/accelerator-native inference serving while staying within the Python ecosystem. The new max["serve"] pattern reduces dependency footprint for lightweight deployments. Modular
Experiment with on-device agents using Nemotron Lightning: Deploy the open-source lightweight Nemotron 3.5 Lightning on RTX/edge hardware to run local agentic workflows without sending customer data off-device. NeMo Switchyard enables multi-model routing for task-specific performance optimization. NVIDIA Blog
What OpenAI’s back-to-back executive departures signal: COO Brad Lightcap and the ethics lead both exiting on the same day points to strategic and cultural shifts inside OpenAI. The connection to the upcoming next-generation model release remains unclear. TechCrunch · FT
Can LLM reasoning encryption ever be truly effective?: The Stolen Thoughts paper demonstrates that server-side reasoning encryption can be circumvented at the API block level. A fundamental fix may require redesigning the inference architecture itself, not just the encryption layer. Stolen Thoughts