Today in 12
9 sources · 7 first-party announcements · 1 new model · 8 briefs
01 Model releases and updates
02 Product releases and updates
03 Industry
05 Tips and perspectives
06 Briefs
- California Sues OpenAI Over Rogue Agent IncidentsTechCrunch · AI
- NVIDIA Introduces the DGX Spark HPC Developer Kit at SC26NVIDIA Blog
- OpenAI Releases a CSAM Image Classifier and ChatGPT WarningsTechCrunch · AI
- Qwen4-Max Outperforms Qwen3.8-Max on Agent ArenaQwen · X
- ChatGPT for Healthcare Adds Computer-Use Agents and 20+ PluginsOpenAI · X
- OpenAI Expands ChatGPT for Academic Researchers to 200,000 ResearchersOpenAI News
- Greg Brockman Highlights GPT-6 Astra for Theoretical PhysicsGreg Brockman · X
- Perplexity Announces Monitor API for Tracking Topics and WebpagesAravind Srinivas · X
Model releases and updates

NVIDIA Vera Rubin Debuts With Leading MLPerf Training and Agentic Inference Results
NVIDIA says its Vera Rubin platform swept every MLPerf Training v6.1 benchmark and led across MLPerf Endpoints agentic benchmarks in its debut. On DeepSeek-V3 training, it scaled from 256 to 1,024 GPUs with about 97% efficiency; a provisional Kimi K3 result reports more than 10x the throughput per watt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. The Vera Rubin Endpoints result is pending MLCommons verification.
Qwen4 Max Scores 9 Points Above Qwen3.8 Max on Intelligence Index
Qwen says Qwen4 Max scores 43 on the Artificial Analysis Intelligence Index with xhigh reasoning effort, 9 points above Qwen3.8 Max and 3 points below Kimi K3.1 Pro. Artificial Analysis reports a cost of $3.27 per index task for Qwen4 Max, about three times less than Qwen3.8 Max and about 11% less than Kimi K3.1 Pro. The model is available through Alibaba's first-party APIs.
Product releases and updates

Google Maps Adds Ask Maps Features for Holiday Travel
Google is launching new Ask Maps features for holiday travel and preparation. Users can create shareable road trip itineraries, view real-time route details, and search saved places for live availability. Route tips and online appointment finding launch first in the U.S.; travel itineraries and list search are available everywhere Ask Maps is offered.

Hugging Face Routes Inference Provider Requests Through OpenRouter
Hugging Face has switched Inference Providers’ language model requests to OpenRouter, which selects a provider based on live price, speed, and reliability measurements and retries with another provider if a request fails. Users can keep the same code and Hugging Face token, but can no longer use their own provider API keys for language models. Image, video, and audio generation continue to run directly through Hugging Face Inference Providers.

NVIDIA Expands CUDA-X Libraries for AI-Assisted Scientific Research
NVIDIA expanded its CUDA-X libraries with cuPhoton, cuPotential, and cuSpa to accelerate research across astrophysics, fusion, molecular simulations, materials science, and biology. The blog reports that cuPhoton reduced Event Horizon Telescope data correlation from over a month to eight minutes, while open-source cuSpa accelerated cell polygon calculations 900x compared with CPUs. Researchers can also pair analysis pipelines with a local Nemotron-powered agent to help steer plasma simulations in real time.
Industry

FBI Says Artex AI Hacking Tool Breached Networks in Under Three Hours
The FBI says hackers have used the open-source AI penetration testing tool Artex in attacks on companies and financial institutions, with some intrusions reaching active compromise in under three hours. The advisory links Artex to attacks affecting at least 68,000 people in South Korea and Microsoft 365 environments at 2,250 ConnectWise customers. Although its developer made the software closed-source after attacks, the FBI says the earlier open-source version remains widely accessible.
US Tightens AI Chip Export Controls on Remote Data Center Access
The Trump administration is preparing an interim rule requiring licenses for AI chip and server exports, reexports, or transfers to data centers worldwide that may provide compute to companies in China or other restricted countries. It covers facilities or networks exceeding 12 million total processing performance and requires existing operators to obtain licenses within 90 days after the rule takes effect. The rule also adds licensing requirements for advanced semiconductor transfers to Malaysia, Singapore, Indonesia, Thailand, and Brazil, and rescinds the Biden administration’s AI Diffusion Rule.
- US Tightens Rules on Remote Access to Advanced AI ChipsArs Technica · AI
Aito Launches the New M8, Starting at RMB 379,800
Aito launched the new M8, with extended-range versions starting at RMB 379,800 and battery-electric versions at RMB 399,800. The SUV comes in five- and six-seat layouts and both powertrain options; its updated lineup standardizes Huawei ADS 5, an 800V platform, and four lidar units. At the launch, Huawei and Seres also clarified their roles: Huawei leads core technology, product definition, and design, while Seres leads manufacturing, engineering implementation, sales, service, and maintenance.
Research

TokenEconomizer Learns from Experience to Reduce AI Coding Agent Costs
Microsoft Research and collaborators introduce TokenEconomizer, an open-source plugin evaluated on 277 SWE-Bench tasks across six repositories and five AI coding agent and model combinations. It reduced average costs by 18.4–30.9%, with cumulative costs falling below the default in every repository by the sixth task. The paper also reports lower costs on sequential web development and security auditing.
Tips and perspectives

Amazon Bedrock Managed Knowledge Base Adds Tabular Search
Amazon Bedrock Managed Knowledge Base now supports tabular search through agentic retrieval. For questions that need calculations or lookups across CSV, XLSX, or XLS files, it runs code over the original tables in an isolated sandbox and combines the results with related document passages. The walkthrough demonstrates a finance question that uses a workbook, a policy, and a close memo; table analysis adds an agent turn and takes longer than text-only answers.

Best Practices for Writing Denied Topics in Amazon Bedrock Guardrails
AWS authors present eight practices for writing Denied Topics in Amazon Bedrock Guardrails, using a fictional bank assistant to show how broad subject labels can block legitimate requests. They recommend action-scoped names and definitions, varied boundary examples, careful management of overlapping topics, and using other Guardrails policies for signals such as prompt attacks or exact terms. Results can still vary for borderline inputs and when the topic list changes.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-11-16
