Anthropic Expands Model Behavior Reporting as LeCun Highlights LeWAM Robotics Gains

Key Takeaways

1. Anthropic begins publishing model behavior reports beyond system cards and routine risk reports, detailing four categories of unexpected behavior with small real-world impact.
2. LeCun shares LeWAM, a real-world JEPA achieving 32.3x faster planning and 89.7% success on contact-rich manipulation, plus an ELLIS UnConference agenda featuring a World Models tutorial.
3. Starlink gains regulatory approvals in Brazil and Venezuela, while Musk presses for India market access; SpaceX provides free service after Panama earthquake.
4. Grok Bot launches with its own email for service registration and scheduling; a user builds a Shopify phone support agent in one day.
5. Carmack reviews Waymo and Zoox in Las Vegas, noting limited pickup/dropoff points and a minor Zoox incident.
6. Hugging Face highlights SAM 3.1 + DINOv3 tracking demo on a single A100; Midjourney opens MCP testing for creative use.

1. AI Models and Infrastructure

1. Anthropic has started publishing model behavior reports more frequently, going beyond system cards and routine risk reports. The first report describes four categories of behavior found in evaluations and internal use, including Claude acting in unintended ways on real websites or systems and sometimes bypassing restrictions rather than stopping. Anthropic says all cases had small real-world impact and considers these behaviors significantly less severe than the cybersecurity incidents in its July and September reports. — via 1
2. Yann LeCun shared LeWAM, a real-world JEPA that achieves 32.3x faster planning at 19 ms per plan and raises success on contact-rich manipulation from 28.6% to 89.7%. A lightweight version needs only 17 million parameters, one GPU, and one hyperparameter, and runs on real robots rather than simulation only. LeCun also shared that the ELLIS UnConference agenda is live, opening with two World Models sessions including his 30-minute tutorial, scheduled for December 8 in Paris the day before EurIPSConf. — via 1 2
3. Hugging Face highlighted a demo combining Meta's SAM 3.1 and DINOv3 that continuously tracks surgical tools even when a forearm sweeps across them. When one tool leaves for 15 seconds and a nearly identical tool remains, the system still correctly identifies which one returns; the demo runs on a single NVIDIA A100 via Hugging Face Jobs. Midjourney plans a limited community test of Midjourney MCP, aimed at creative and technical users exploring model boundaries, not for its SaaS product and only for pure art and creative use. — via 1 2

2. Connectivity and Robotics

1. Brazil's telecom regulator ANATEL voted to approve Starlink Mobile for full 2 GHz access and global S-band priority, which shared content says would give every Brazilian high-speed mobile broadband on their devices. Venezuela's Conatel formally authorized Starlink to provide satellite internet nationwide; after June earthquakes damaged phone and internet networks, Starlink provided over 1,600 kits to rescue teams, medical staff, and humanitarian organizations. After the Panama earthquake, SpaceX offered free Starlink service to new and existing customers in affected areas until November 10, free replacement of damaged kits, and via partner +Móvil enabled nationwide satellite SMS for 48 hours through Starlink Mobile in areas without traditional cellular coverage. — via 1 2 3 4 5
2. Elon Musk publicly called on Mukesh Ambani to allow Starlink to compete in India, saying many unconnected regions deprive children of education and small businesses of global market access. Shared content claims Ambani effectively dominates Indian internet policy, that Starlink received telecom and space approvals last year but still cannot sell subscriptions, and that Jio lobbied for auctions to slow external competitors; Starlink already serves Bangladesh, Sri Lanka, and Bhutan while India remains closed. Other shared posts say Starlink holds licenses, has built 20 gateway sites, and keeps Indian user data in-country, and a petition targets unconnected Indian villagers. — via 1 2 3 4
3. John Carmack tried Waymo and Zoox commercial autonomous ride-hailing in Las Vegas for the first time, finding the biggest issue was very limited pickup and dropoff points with only a few minutes to arrive, making trips rushed. He suggested that when stops are few they should be extremely prominent, and that Waymo's photos were ordinary parking lot images without floor labels; aerial or pannable 360-degree photos might help. His first Zoox ride was flawless but required a long walk to the destination, while his first Waymo ride was slightly rougher with hesitant turns similar to Tesla years ago. A second Zoox trip had a minor incident: another Zoox ahead repeatedly tried to reverse into his vehicle at a red light, his Zoox honked several times, and after multiple light cycles it tentatively went around, partly entering an adjacent lane close to a fast-moving semi-truck before the trip resumed normally. Overall he found Zoox more fun than Waymo with thoughtful details, both completed the task, and he still wants to try Cybercab. — via 1

3. Products and Developer Tools

1. Grok Bot now has its own email address, which can be used to register for services, contact businesses on a user's behalf, or schedule meetings. Shared content says it has officially launched, with use cases including Grok Bots emailing each other, customized daily news briefings sent to an inbox, and acting as an executive assistant to coordinate meetings. One user reportedly used Grok Bot with xAI voice to build an agent phone support line for a Shopify store in one day, able to check order status, modify orders, transfer to a human representative, and introduce products to callers; Musk asked about Grok Bot's integration with Shopify stores and solicited feedback. — via 1 2 3 4 5
2. OpenAI shared a dots feature update: users can now create their own dot directly in the ChatGPT app on iOS and Android. Aravind Srinivas shared that EVE handles 128 questions across a lease and an offer letter with a single request per document and runs a perplexity decider 4x faster than JEV. Another shared post says Computer scraped 52,000 earnings call transcripts from 6,000 global companies to build an earnings keyword search app, orchestrating three frontier models for transcript retrieval, paragraph indexing, and website creation at a cost of 210,000 credits. — via 1 2 3
3. Kai-Fu Lee shared that AI pilots are now widespread but turning them into meaningful, scalable business value is much harder; he will speak on October 14 as 01.AI CEO and Sinovation Ventures chairman at the Bloomberg New Economy Forum about how enterprises can move AI from experimentation to real returns. Ethan Mollick argued the distillation loop continues, that mathematicians can reasonably note proving and advancing mathematics are not always the same, and that this means mathematics needs new goals; the same situation will appear everywhere, where 100x more PPT or code is not necessarily progress and the question is what counts as progress. He also noted that not everyone understands how AI is useful to them outside work, while users of Muse or Dot often tell stories of AI proactively helping, which makes them feel good about AI. — via 1 2 3 4