9 topics covered
Runway Praxis-1 Open-Source Robotics Model
What happened: Runway released Praxis-1, a robotics model that uses knowledge from video training to control robots, with open weights and source code.
Key details:
- Released with open weights and source code under the Apache 2.0 license.
- Leads the World Action Model category on Nvidia's RoboLab-120 benchmark, per The Decoder.
- The model cannot yet plan complex action sequences on its own, according to the article.
Why it matters: It is an open-license example of a video-trained model being applied directly to robot control.
Practical takeaway: Developers interested in robot control can inspect the Apache 2.0 weights and code to evaluate its limits on multi-step tasks.
Nvidia Blackwell Chips in Chinese AI Models Claim
What happened: Former OpenAI researcher Will DePue says he found technical evidence that several Chinese AI developers used Nvidia's Blackwell GPUs despite US export controls.
Key details:
- Names DeepSeek, Moonshot AI, MiniMax, and Ant Group as the developers involved.
- The method reads characteristic chip signatures directly from published model weights.
- DePue estimates the hardware behind the five training runs cost more than half a billion dollars.
- The Decoder describes this as the first technical evidence of this kind, per DePue's claim.
Why it matters: If confirmed, it would indicate restricted-chip use by named Chinese labs, bearing directly on export control enforcement.
Practical takeaway: Treat this as a single researcher's forensic claim; look for independent verification or responses from the named companies.
Anthropic Claude for Manufacturing
What happened: Anthropic launched Claude for Manufacturing, a plugin that connects Claude to factory and business software.
Key details:
- Connects directly to Autodesk Fusion, Oracle NetSuite, and Infor.
- Includes 17 specialized skills that can draft quotes, compare suppliers, and place backup orders when deliveries are delayed.
- Actions occur after human approval, per The Decoder.
- The plugin does not control physical machines yet; Anthropic is separately developing the Model Hardware Standard for that.
Why it matters: It extends Claude from office tools into industrial software workflows, though it stops short of controlling equipment.
Practical takeaway: Manufacturers using Autodesk Fusion, NetSuite, or Infor can evaluate the plugin's approval-gated workflows for quoting and procurement.
Public-Sector Unions and AI Guardrails
What happened: US public-sector unions are increasingly winning protections against AI surveillance and automated firings, as nearly 70 percent of US government agencies already use generative AI.
Key details:
- In Oakland, a strike vote forced retraining programs instead of layoffs.
- A transport union in Ohio won veto power over autonomous vehicles.
- California banned so-called bossware.
Why it matters: Labor contracts and state law are becoming a practical check on how government agencies deploy AI.
Practical takeaway: Public-sector employees and managers can track these union and state protections as templates for AI workplace terms.
llama.cpp v1.0.0 Release
What happened: The llama.cpp and ggml projects released version 1.0.0 libraries, with new rules for how API changes are versioned.
Key details:
- Released as libllama (llama.cpp) and libggml (ggml) v1.0.0, per the Hugging Face post dated November 13, 2026.
- Going forward, API and ABI incompatibilities will be signaled with major-version bumps.
- Supports more than 150 architecture variants, per the post.
- Distribution channels include Docker, Debian, Homebrew, conda-forge, MacPorts, Winget, and XCFramework.
Why it matters: A 1.0.0 stability commitment gives downstream applications clearer expectations for API compatibility.
Practical takeaway: Developers depending on llama.cpp's C API should review the new semantic-versioning rules before upgrading across major versions.
ChatGPT Wallet for Dots Agent Purchases
What happened: OpenAI launched ChatGPT Wallet, a payment feature that lets its Dots AI agents make purchases for users.
Key details:
- Rolling out starting today for ChatGPT Pro subscribers in the US aged 18 and older with a US billing address.
- Supported cards: 'eligible US-issued American Express, Chase, Mastercard and Visa cards.'
- Every purchase requires a biometric check such as Face ID or a fingerprint scan before checkout.
- OpenAI says it substitutes card details with a payment token, so the actual card number is not shared with the Dots agent or the merchant.
- OpenAI says it expects users to eventually be able to pre-approve purchases using spending limits, purchase categories, or time periods.
- Some retailers restrict or block AI agents, and OpenAI says websites must ensure agents are not inadvertently blocked.
Why it matters: This is OpenAI's second attempt at agent-driven shopping, after The Decoder reported it scrapped its first approach, Instant Checkout.
Practical takeaway: Pro users in the US can add an eligible card and approve each agent purchase individually; expect some merchants to block agent checkouts.
Amazon World Product Graph
What happened: Amazon reportedly plans a 'World Product Graph,' a database of products sold anywhere online, including items Amazon does not sell, according to Reuters.
Key details:
- The project is meant to keep shoppers on Amazon for product research instead of Reddit or Google, per The Verge's summary of Reuters.
- An internal Amazon document says 'Many Rufus answers are factually wrong or incomplete' and that running searches with the AI is expensive.
- The document includes an October 1st, 2027 release date, but Reuters said it is unclear 'if or when this plan may be implemented.'
- The project 'depends on discovering how brands organize their products on their own websites,' which could involve crawling other retailers' sites.
- Amazon said in a statement that 'Draft, outdated documents often paint an incomplete picture' and that the document does not necessarily reflect current plans.
Why it matters: If implemented, Amazon's product index could shape which products AI shopping agents surface, since agents may rely on the largest product data source.
Practical takeaway: Treat this as an unconfirmed plan; watch for whether Amazon ships the index and how it affects agent-based shopping.
Anthropic IPO Demand and Investor Restrictions
What happened: Anthropic's IPO is reportedly drawing about six times more orders than available shares, while banks are blocking buy orders from Chinese and Hong Kong investors.
Key details:
- The Decoder reports orders exceed available shares by a factor of six, and the IPO is on track to be the largest public offering in history.
- Bloomberg reports banks are rejecting buy orders from mainland China and Hong Kong, citing U.S. export control rules known as ITAR.
- The Decoder notes similar restrictions applied during SpaceX's IPO, and that Anthropic's products are not available in either region.
Why it matters: The ITAR-based restriction shuts a region out of one of the largest expected offerings, and it comes amid strong reported demand.
Practical takeaway: Watch for official allocation details and final pricing from Anthropic's IPO before drawing conclusions about demand.
Meta Muse Business Calling Feature
What happened: Meta's AI agent Muse can now call businesses on behalf of paying US subscribers.
Key details:
- Handles tasks ranging from calls to cable providers to restaurants.
- Muse already has over 10 million users, according to Meta's AI chief Alexandr Wang.
- Google Gemini, Instinct, and xAI's Grok Bot offer the same feature, per The Decoder.
- The Decoder reports many businesses hang up as soon as they detect an AI on the line.
Why it matters: Outbound AI calling tests whether businesses will accept AI agents as customers, a question that also affects agent-driven shopping.
Practical takeaway: US paying Muse subscribers can use the calling feature, though success may vary depending on how businesses handle AI callers.