Petition to Reinstate Jensen Huang & FBI Warns of Artex AI Hacking Tool - AI News (Nov 17, 2026)
Petition to reinstate Jensen Huang tops 115K, FBI warns of AI hacking tool Artex, Qwen4-Max debuts, and GPT-6 helps tackle a 25-year physics puzzle.
15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕
Support the show →Our Sponsors
Today's AI News Topics
-
Petition to Reinstate Jensen Huang
— A petition with more than 115,000 signatures, backed by Elon Musk, Tim Sweeney and Garry Tan, urges Nvidia's board to reinstate Jensen Huang while an independent investigation proceeds. CFO Colette Kress is serving as interim CEO. -
FBI Warns of Artex AI Hacking Tool
— The FBI says attackers are using Artex, an open-source AI hacking tool connected to LLMs, to automate reconnaissance and exploitation. Incidents include ConnectWise customers and South Korea's financial sector, with targets moving from scan to compromise in under three hours. -
Alibaba Launches Qwen4-Max Model
— Alibaba's Qwen team released Qwen4-Max, a 2.4-trillion-parameter model with native vision and a one-million-token context window. It targets long-horizon agentic work in coding, finance, legal, healthcare and research, with lower API pricing. -
Embla 1 On-Device Reasoning Model
— Sevren's Embla 1 is a small reasoning model built to run privately and offline on smartphones. It answers in under two seconds on an iPhone, and its 4-bit weights are available on Hugging Face. -
Quark Small Model Trained for $4,200
— Empero trained Quark, a 165M-parameter language model with 128K context, end to end for about $4,200 in compute. Reinforcement learning alone taught it basic agentic tool use, though it has not been safety-evaluated. -
Thinking Machines' 4-Bit Training Breakthrough
— Thinking Machines Lab trained Inkling 2 Small with a native 4-bit NVFP4 forward pass on Blackwell GPUs. The approach cut pretraining time by about 8% and reinforcement learning time by 20% while keeping accuracy acceptable. -
Vision Capacity in Multimodal Models
— A controlled study of multimodal vision encoders finds that dedicated vision capacity matters more than separate pretraining. Encoder-free models underperform, while vision-only modules placed in early backbone layers match scratch-trained encoders. -
NVIDIA BirdDreams Driving World Model
— NVIDIA's BirdDreams world model generates a camera video and a bird's-eye-view scene from a single image. It improves traffic realism and trajectory following for autonomous driving simulation without needing HD maps. -
ChatGPT Wallet for Agentic Payments
— OpenAI launched ChatGPT Wallet, which lets ChatGPT agents pay on a user's behalf using payment tokens instead of real card numbers. It was built with Visa, Mastercard, American Express and Chase, and every purchase requires explicit user approval. -
Poppy Protocol for Personal Agents
— Sierra released version 1.0 of Personal Agent Protocol, called Poppy, an open standard built on MCP, OpenAPI and OAuth. It governs how personal AI agents interact with businesses, and every Shopify store now publishes a Poppy file. -
Anthropic's Always-On Agent Architecture
— Anthropic describes always-on AI agents as event-driven systems that sleep between runs. Its design relies on inboxes, timers, memory, permissions and outboxes, with an emphasis on safe unattended tool use. -
Anthropic on Building Strong Safety Evaluations
— Anthropic outlines best practices for AI safety evaluations, including realistic multi-turn scenarios, sandboxed testing and preregistration. It argues for measuring appropriate responses rather than simple refusals, and for transparent reporting. - 13
GPT-6 Helps Define Lattice Standard Model
— Physicist Nathaniel Craig, working with GPT-6 Astra, proposed a candidate lattice definition of the chiral Standard Model. It addresses a phase problem that had stayed open for 25 years and could enable nonperturbative physics calculations, pending expert review. - 14
Caltech Rethinks Math Homework Grading
— At Caltech, a math professor now grades homework only for completion because students can solve problem sets with generative AI. The university has no unified policy on AI in math courses. - 15
Amdahl's Law and Superintelligence Limits
— An essay uses Amdahl's law to argue that artificial superintelligence won't drive explosive economic growth on its own. Physical throughput, regulation and slow data collection remain the real bottlenecks. - 16
United Airlines Outage Update
— United Airlines' Monday system outage lasted about an hour and forty minutes, longer than first reported, delaying hundreds of departures. The airline found no indication of a cyberattack.
Sources & AI News References
- → NVIDIA Unveils BirdDreams, a Controllable Driving World Model
- → Tabstack Launches Privacy-Focused Web Access for AI Agents
- → Thinking Machines Details End-to-End 4-Bit Training
- → Anthropic on How to Build Strong AI Safety Evaluations
- → Dedicated Vision Capacity Matters More Than Separate Encoder Pretraining
- → Petition Pushes Nvidia to Bring Back Jensen Huang
- → Google Cloud to Unveil Next-Generation AI Infrastructure Plans
- → FBI Warns Open-Source AI Hacking Tool Is Enabling Fast Network Breaches
- → United Airlines resolves brief technology outage that delayed departures
- → Sierra Releases Personal Agent Protocol 1.0
- → Sevren releases Embla 1, a phone-first reasoning model
- → Gartner Webinar: 7 AI Themes C-Suite Leaders Should Know
- → Empero Releases Quark, a 165M-Parameter Model Trained End to End for $4,200
- → At Caltech, AI Is Changing How Math Homework Is Graded
- → Alibaba Launches Qwen4-Max as a Long-Horizon AI Work Teammate
- → OpenAI Launches ChatGPT Wallet for Agentic Payments
- → How to Build Always-On AI Agents
- → Devin Launches Private Web Apps on Its Cloud
- → Claude Code Projects Get Smarter Thread Management and Sharing
- → OpenAI Guest Post Proposes Candidate Lattice Definition of the Standard Model
- → United Airlines Outage Delays Departing Flights
- → Why AI May Hit Bottlenecks Beyond Intelligence
Full Episode Transcript: Petition to Reinstate Jensen Huang & FBI Warns of Artex AI Hacking Tool
More than a hundred and fifteen thousand people, Elon Musk among them, are asking Nvidia's board to do something boards rarely do: undo a CEO's exit. We'll get to that in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is November 17th, 2026. Coming up: AI-powered hacking that moves faster than defenders, new models from Alibaba and some notably small labs, AI agents getting their own wallets, and a language model helping with a decades-old problem in particle physics.
Petition to Reinstate Jensen Huang
Let's start with Nvidia. A petition calling for Jensen Huang's reinstatement says it has passed one hundred and fifteen thousand signatures, including thousands of current and former employees and many shareholders. Supporters include Elon Musk, Tim Sweeney, Garry Tan and Replit's Amjad Masad. The petition argues Huang was pushed out without meaningful due process. It asks the board to bring him back while an independent investigation establishes the facts, and to be more open with the public. Huang resigned after admitting to an affair with a chip executive whom U.S. intelligence suspects of spying for China, and he has agreed to testify before Congress. CFO Colette Kress is serving as interim CEO. With the stock down, investors are watching how much turbulence this creates at the center of the AI hardware industry.
FBI Warns of Artex AI Hacking Tool
On security, the FBI has issued an advisory about Artex, an open-source AI hacking tool that attackers are connecting to large language models to automate reconnaissance and exploitation. In some cases, attackers have gone from scanning a target to compromising it in under three hours. Victims include ConnectWise customers' Microsoft 365 environments and South Korea's financial sector, where personal data on at least sixty-eight thousand people was stolen. The developer has since made Artex closed-source, but older versions are still widely available. The tool mostly exploits familiar weaknesses, like exposed credentials and missing multi-factor authentication. What AI changes is how quickly those gaps can be found and used.
Alibaba Launches Qwen4-Max Model
Now to models. Alibaba's Qwen team has launched Qwen4-Max, built on a new architecture with a 2.4-trillion-parameter backbone, native vision and a one-million-token context window. Alibaba pitches it as a teammate for long, complex work, with demos in game development, legal due diligence, drug-design research and cloud scheduling. It also cut API prices. That puts Alibaba squarely in the race for agents that deliver finished work, not just answers.
Embla 1 On-Device Reasoning Model
At the other end of the size scale, Sevren released Embla 1, a reasoning model built to run privately and offline on a phone. The company reports answers in under two seconds on a current iPhone. It doesn't top every benchmark, and Sevren admits math is still a weak spot, but its pitch is a strong balance of speed and quality using far fewer tokens. The 4-bit weights are on Hugging Face. Meanwhile, Empero showed it could take a 165-million-parameter model from raw data all the way to release for roughly forty-two hundred dollars in compute. The most interesting result is that reinforcement learning alone, with no supervised fine-tuning, taught the model basic agentic tool use. Empero is upfront about the limits: it makes errors, struggles with some long-context tasks, and hasn't been safety-evaluated.
Quark Small Model Trained for $4,200
On the research side, Thinking Machines Lab says it trained Inkling 2 Small with a native 4-bit forward pass. Fewer bits usually mean faster training at the cost of accuracy, and the team reports about eight percent faster pretraining and twenty percent faster reinforcement learning with acceptable accuracy. That makes 4-bit precision look practical for training, not just inference. A separate controlled study of multimodal models found that what matters most is having dedicated vision capacity. Separate pretraining turns out to be optional, and encoder-free designs lag unless they get much more compute. NVIDIA, meanwhile, introduced BirdDreams, a driving world model that generates both a camera view and a bird's-eye map from a single image. The result is more realistic traffic behavior and tighter trajectory following, which are key ingredients for simulating autonomous driving.
Thinking Machines' 4-Bit Training Breakthrough
Agents are also getting closer to spending money. OpenAI announced ChatGPT Wallet, built with Visa, Mastercard, American Express and Chase. It lets ChatGPT pay on your behalf using a one-time token instead of your real card number, and every purchase needs your explicit approval plus a biometric check. Card rewards and dispute protections still apply. Along similar lines, Sierra released version 1.0 of Poppy, an open standard for how personal agents interact with businesses. Its biggest early win is Shopify, where every store now publishes a Poppy file automatically. And Anthropic published a blueprint for always-on agents. The core idea is that these agents aren't actually running all the time: they sleep, wake when something happens, and rely on memory, careful permissions and a quiet outbox so they don't flood users with updates.
Vision Capacity in Multimodal Models
Anthropic also published guidance on safety evaluations. Its argument is that a good eval measures whether a model responded appropriately, not simply whether it refused, especially in sensitive areas like self-harm. It recommends realistic multi-turn scenarios, sandboxed testing, preregistered methods and honest disclosure of flaws. The underlying message is that safety claims are only as reliable as the evaluations behind them.
NVIDIA BirdDreams Driving World Model
Now for some science. Physicist Nathaniel Craig, working with GPT-6 Astra, has proposed a candidate lattice definition of the Standard Model. A lattice definition would let physicists simulate the full theory from first principles on a computer, but a key piece has been missing for more than twenty-five years. It's still a candidate and needs scrutiny from other experts. If it holds up, it would mark a striking moment for AI as a collaborator in theoretical physics.
ChatGPT Wallet for Agentic Payments
AI is reshaping teaching as well. At Caltech, math professor Yi Ni now grades homework for completion only, because students can simply have AI solve the problems. Some students avoid AI to prepare properly for exams, while others likely don't. Caltech has no unified policy yet, so instructors are adapting one syllabus at a time.
Poppy Protocol for Personal Agents
For a more skeptical view, one essay applies Amdahl's law to superintelligence. The idea is that speeding up one step doesn't help much if the rest of the process stays slow. Factories, shipping, regulation and messy biology experiments all impose limits that intelligence alone can't remove. The author's conclusion is that AI's economic impact will depend on clearing those bottlenecks, not just on getting smarter.
Anthropic's Always-On Agent Architecture
Finally, an update on the United Airlines outage. The disruption turned out to be longer than first described, lasting about an hour and forty minutes from the first pilot reports. Flight-tracking data suggests hundreds of departures were delayed. United still says there's no sign of a cyberattack.
That's it for today's episode of The Automated Daily, AI News edition. Links to all the stories we covered can be found in the episode notes. I'm TrendTeller. Thanks for listening, and see you tomorrow.
More from AI News
- November 14, 2026 AI Starts Speeding Up AI & the Compute Bill Comes Due
- November 14, 2026 Open models cutting AI coding costs & Model routing between open and closed
- November 13, 2026 AI pricing models under pressure & Usage-based billing and spend controls
- November 12, 2026 DeepMind AI verifies Signal protocol & Google launches Gemini 4 Carbon
- November 11, 2026 METR: AI Doubling Anthropic's R&D Speed & Anthropic's Compute Bill and Gas Turbines