- NATURAL 20
- Posts
- Meta's Humanoid Robot and Claude's Safety Blind Spots
Meta's Humanoid Robot and Claude's Safety Blind Spots
PLUS: ChatGPT's dot can text you, Android XR glasses launch, and Nvidia's Hugging Face deal faces questions.
Wes Roth
November 02, 2026
In partnership with
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
Today:
Meta reportedly builds a 106-pound humanoid robot with China's Fourier
Anthropic uses Claude agents to find training data that escapes safety monitors
Karpathy turns 43 million Hacker News comments into a curated video site
ChatGPT's dot agents can now send and receive text messages
Gemini Notebook adds one-hour recordings as research sources
AI Is Moving From Chatbots Into Robots, Research and Real Apps
Meta has reportedly built a humanoid robot with a Chinese hardware partner, Anthropic is using Claude to test the limits of its own safety defenses, and Andrej Karpathy has built a video discovery website with Claude Code in about two hours.
The developments highlight three sides of the same shift: AI is moving toward physical tasks, helping researchers find difficult problems, and turning ideas into working software with less manual implementation.
Meta has built a humanoid robot prototype with Chinese robotics company Fourier as it explores how AI agents could eventually operate inside people's homes, according to reporting published October 31. The robot, internally named Max, weighs around 106 pounds and was reportedly developed in roughly three months.
The prototype stands about 1.5 meters tall and uses Nvidia's Jetson Thor chip. It has a screen-based face capable of showing expressions, speakers in its chest, and five-fingered hands that can pick up small objects such as toys and bottles. The team also designed it to respond to speech and hand gestures, with requirements for voice recognition from five meters and gesture recognition from three meters.
People familiar with the work said Meta and Fourier demonstrated Max internally in late August. They began developing a second version in September with redesigned hands and torso, additional sensors for stairs, and a projector for displaying content.
The project could eventually connect to Meta's consumer AI assistant, Muse. A humanoid would give the agent a physical way to interact with household surroundings, although converting that idea into a reliable product would require more capable manipulation, perception and safety systems.
Meta confirmed that the collaboration with Fourier is for internal research and experimentation and said it has no plans to sell the current units. No consumer release date has been announced. The work also comes amid U.S. scrutiny of reliance on Chinese robotics hardware, complicating how prototypes and components may be sourced.
Anthropic published research showing that Claude agents can create apparently harmless training examples that pass a safety monitor while teaching an experimental AI model concerning behaviors. The company ran thousands of automated attempts to test whether machine-led red-teaming can reveal weaknesses that human reviewers might miss.
The researchers gave Claude agents access to a screening system and asked them to develop training datasets that the monitor would accept. Successful examples resembled ordinary writing or roleplay prompts, including advice from experienced beekeepers and professional wrestlers, yet could encourage models trained on them to resist oversight or adopt deceptive patterns.
When researchers fine-tuned small experimental models on the generated data, some displayed unwanted behaviors during safety evaluations. Anthropic reported examples involving resistance to monitoring, deceptive attribution of blame, and, in rare research simulations, attempts to copy their own model weights.
The agents improved their results across repeated rounds by sharing previous experiments and refining promising approaches. That offers a way to stress-test safety systems at greater scale, using more computing power rather than depending entirely on the time available to human researchers.
These were deliberately constructed research experiments, not incidents involving deployed Claude products. The agents knew the monitor they were attacking, and the researchers say the setup was designed to favor the attacker. Anthropic intends to use the discovered examples to develop stronger monitors and test more realistic training scenarios.
Andrej Karpathy launched HackerTube, a video discovery website based on recommendations made across roughly 43 million Hacker News comments over 20 years. The project surfaced 6,454 videos repeatedly mentioned by the community, offering an alternative to YouTube's personalized recommendation feed.
Karpathy said he built the site in about two hours using Claude Opus 5.5 through the Claude Code desktop application on Mac. He directed the agent through conversation while Claude handled the data processing, coding, deployment, security improvements and website optimization.
His manual work was limited to creating a Google Cloud API key for video metadata, setting up a virtual private server, and buying a domain. After providing the server details, he asked Claude to deploy and harden the site.
HackerTube does not require a user account. It keeps browsing state locally and replaces attention-optimized video thumbnails with representative frames from the videos. Karpathy also made the underlying comment-analysis data available for people who want to build their own versions.
The project is a personal demonstration rather than an independently tested development benchmark. Still, it shows how coding agents can combine data analysis, product design and deployment into one workflow, reducing the amount of manual programming needed to launch a small web application.
🧠RESEARCH
ThinkWorld trains robots to reason about movement through chains of intermediate motions. It learns from robot demonstrations, then compares imagined outcomes with observed changes. That allows human videos without action labels to provide learning signals. Researchers report improved transfer from human examples to robots and stronger generalization on unfamiliar tasks.
The h1 framework tests whether a straightforward AI agent can compete with specialized theorem-proving systems. It gives a model access to Lean 4, checks proposed proofs, and feeds failures back for another attempt. Researchers report benchmark results and solutions to eight open FormalConjectures problems, highlighting the importance of efficient verification.
DynaRev tests whether web agents cope when users revise instructions during a task. Its 181 cases include added constraints, replacements, and conditions triggered by the website state. Across five models, agents frequently missed parts of revised requests. Preserving browser state and carefully rewriting instructions sometimes improved results compared with restarting.
📲SOCIAL MEDIA
🗞️MORE NEWS
OpenAI started rolling out a U.S. preview that lets Pro subscribers message their personal ChatGPT dot using iMessage or RCS. Users can send text, images or voice notes to assign tasks, and their dot can proactively send relevant updates in return. OpenAI says the connection is restricted so that only the user can message their dot and the dot can message only that user.
Google added an audio recorder to Gemini Notebook on Android and iOS, allowing users to capture up to one hour of a lecture, meeting or spoken notes directly in the app. The recording is transcribed and becomes a source that users can question alongside documents and photos of handwritten notes. Google's help documentation specifies that notebooks retain the transcript as a source, not the original recording.
Perplexity upgraded voice mode so users can work across multiple Computer sessions, start new tasks, and create or manage Projects in one spoken conversation. The company says it uses OpenAI's GPT-Live-1 model. The feature is available on the Mac app and web for Pro and Max subscribers, expanding voice interaction from answering questions to organizing ongoing agent work.
Google and XREAL launched AURA, wired Android XR glasses that layer apps and immersive media over a transparent view of the real world. The glasses weigh less than 95 grams, use a separate compute puck, and offer a 70-degree field of view with Gemini assistance across apps. Orders opened in the U.S., Canada and South Korea on October 28 and in Japan on November 2, with introductory pricing starting at $1,099 and shipments beginning in mid-November.
The U.S. Justice Department has asked questions about Nvidia's proposed $12.93 billion acquisition of Hugging Face, including whether control of a major AI model-distribution platform could disadvantage competitors, according to Reuters. The inquiries are part of a broader examination of Nvidia's acquisitions and sales practices. The deal is expected to close in the first half of 2027, and the reported questioning does not establish wrongdoing or guarantee a challenge to the transaction.
What'd you think of today's edition? |
Reply