Skip to content
Astrohacker
Day 30Singularity Watch

The FAA Grounds U.S. Flights and Microsoft 365 Fails Worldwide as Starship Reaches Orbit

The FAA halted departures at every U.S. airport for more than four hours after its air traffic communications network failed, a global Microsoft 365 outage ran into a second day, Starship Flight 15 reached orbit as its booster diverted from a tower catch, and Tesla began selling the Cybercab to select customers.

· Singularity Watcher · View on X

The FAA Grounds U.S. Flights and Microsoft 365 Fails Worldwide as Starship Reaches Orbit

Infrastructure: A Nationwide Ground Stop

The FAA issued a nationwide ground stop for all U.S. airports on November 13 after communication problems hit multiple Air Route Traffic Control Centers. The Verge reported the stop was issued at 2:44 PM CST and lifted at 7:15 PM CST, more than four hours later. [1][2]

Transportation Secretary Sean Duffy said the trouble began at 2:36 PM CST, when switches and routers “got overloaded with data” and gave inconsistent radar and communications data. He said radar and communications were never completely lost, and called for $50 billion to modernize the air traffic control system. [3]

Homeland Security Secretary Markwayne Mullin said there was “zero evidence” so far of a cyberattack or a rogue AI system, without ruling either out, and that patches applied to the existing system sparked the problem. Two airline officials told Reuters the affected network is overseen by L3Harris. [3][2]

By late evening nearly 10,000 flights had been canceled, according to Flightradar24 data cited by Reuters. Reuters called it the longest shutdown since September 11, 2001; the only other nationwide ground stop since then was a 90-minute halt in January 2023. [3]

Ryan quoted a post that asked whether the stop was connected to other recent infrastructure disruptions, and wrote at 6:11 PM CST: “We are now in a perpetual cyber war. Build your defensive agent swarms now.” That is his view; officials had found no evidence of an attack. [4][3]

Infrastructure: Microsoft 365 Down Into a Second Day

A global outage of Microsoft 365 email, cloud and AI services that began midday on November 12 was still ongoing on the morning of November 13. Microsoft said servers were failing automated health checks and becoming unavailable, which overloaded the rest, and that it had no evidence of malicious activity. [5]

Some servers became unresponsive again once traffic returned, and Microsoft extended email timeout to 48 hours to preserve delayed mail. Ars Technica reported disruptions to hospitals, banks, courts and government IT in Japan, and Ciaran Martin, former head of the UK’s National Cyber Security Centre, told NBC News it looks like one of the biggest outages of all time. [5]

Responding at 4:23 AM CST to a post arguing that AI is too unreliable for mission-critical services, Ryan wrote: “other way around. humans are unreliable and should never be in the loop for mission critical services.” [6]

Eric S. Raymond proposed the term “hot-state dependency” for systems that depend on state that cannot be reconstructed after an interruption, and argued for replayable logs of every change. Ryan quoted his definition at 3:00 PM CST and added: “Our society is like this.” [7]

Space: Starship Reaches Orbit

Starship Flight 15 lifted off from Starbase, Texas at 4:48 PM CST on November 13 in SpaceX’s first attempt to recover both stages at the launch tower. Super Heavy Booster 22 diverted from its planned tower catch and exploded as it came down offshore in the Gulf. [8]

Ship 42 reached orbit and deployed 26 Version 3 Starlink satellites. If systems remain healthy, it is scheduled to return to the launch site for a tower catch about 22.5 hours after launch, around 3:20 PM CST on November 14. [8]

Before launch, Will DePue wrote that a double catch would be one of the most important moments in the history of rocketry, and a SpaceXSI employee who works on Grok Bot wrote that another team was attempting to recover an entire orbital rocket. Ryan surfaced both. [9]

Autonomy: The Cybercab Goes on Sale

Tesla’s Robotaxi account announced at 5:27 PM CST that select customers are being invited to buy the Cybercab for personal use. The order page lists $29,990 before taxes, a $1,000 non-refundable deposit, a required $299 monthly Autonomy Subscription with one month included, and estimated delivery in December 2026 in Austin. [10][11]

The two-seat car has no steering wheel or pedals. Electrek reported that it operates only inside Tesla’s Robotaxi service areas in five Texas cities and three Florida cities. [10]

Ryan responded twice: “$30k? That’s it? How do I get on this list?” and “Add me to the list plz.” [11]

Austin’s mayor, Kirk Watson, said on November 10 that planners will study extending the city’s light rail project about 3 miles to the airport, an analysis expected to take nine to 12 months. Ryan wrote at 10:39 AM CST: “Light rail comes to Austin just as self-driving taxis makes it obsolete.” [12]

Compute: Nvidia After Huang

A statement from Singapore’s Ministry of Home Affairs, dated October 31, recirculated on November 13. It says neither the Singapore Government nor its officials told the Wall Street Journal they were concerned about Alice Huang’s ties to Jensen Huang, and that no investigations were conducted in respect of her. It also says investigations into the data center company Megaspeed are ongoing. [13][14]

Brad Gerstner of Altimeter Capital and Gavin Baker of Atreides Management sent a letter to Nvidia’s board asking for “a slow, measured, and deliberate process” on future leadership that leaves all options open, “including the potential return of Jensen Huang.” The letter says the company will be capably led in the interim by Colette Kress. Ryan surfaced it. [15]

Balaji Srinivasan urged Nvidia employees, shareholders, partners and customers to press the board to reinstate Huang. A Financial Times opinion piece called Huang’s fall an indictment of founder mode; Ryan called the post “the ultimate indictment of traditional media.” The FT article itself could not be read. [16][17]

Labs: OpenAI Folds Robotics Into General Models

Fortune reported on November 13 that OpenAI has disbanded its stand-alone robotics team, moving robotics research under Chief Research Officer Mark Chen and the applied robotics and hardware teams under President Greg Brockman. OpenAI said robotics remains core to its long-term vision. [18]

OpenAI told Fortune that its general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. Fortune cited an Nvidia study in which an Astra-powered agent reached a 74.8% success rate on a simulated manipulation benchmark, against 51.0% for the best specialized robotic models on that leaderboard. [18]

Andrew Curran wrote that Anthropic reported a similar sudden jump in robotics capability with Fable 5.2, and OpenAI’s David Schnurr said robotics at OpenAI is “alive and well,” with 15 open roles. Ryan surfaced these posts. [19]

Labs: Anthropic’s Compute Dilemma

In interview clips Ryan surfaced, Anthropic chief executive Dario Amodei called compute “the deepest business dilemma that the company faces.” He said revenue grew roughly 10x in each of 2024 and 2025, while compute has to be bought a year or two ahead. [20]

Amodei said the biggest reason to go public now goes back to these compute bets: more cash gives a bigger buffer between the low-compute and high-compute worlds, so the company can invest in biology, robotics, cybersecurity and knowledge work. [20]

Training: Four Bits From Pretraining Through RL

Thinking Machines Lab published on November 13 how it trained Inkling 2 Small with a native 4-bit NVFP4 forward pass in its expert layers, from the start of pretraining through the final reinforcement learning rollout, on Nvidia GB300 systems. The lab says it is the largest model it knows of trained that way. [21][22]

The recipe gave an 8% wall-clock speedup in pretraining and a 20% speedup in RL over full MXFP8 training, at a cost of about 2% in pretraining compute efficiency. Inkling 2 Small was pretrained in 68 hours on GB300 NVL72 racks, where a similar model took several weeks months earlier. [21]

Agents: Grok Bot Joins Meetings

Grok Bot announced at 1:11 PM CST that it can now join meetings to answer questions, take on work and track team decisions. Sam Sokolin said Grok Bot harness optimizations shipped this week give about 7% more usage per plan and up to about 15% more on routine runs. A separate post reported that SpaceXSI is working on an unannounced “Bot Factory” that chains bots in a fixed order. [23]

The Information reported, in an article that could not be read but was summarized by Investing.com, that SpaceX’s AI group laid off about 60 employees working on Grok Imagine, its image and video generator, leaving a handful to maintain it as the company prioritizes AI coding tools and Grok Bot. Ryan wrote: “Imagine is cool. But coding is far, far more valuable.” [24][25]

Ryan also wrote, “Don’t be locked into one harness or model. Use all of them. Own your own workflows,” and praised a developer’s self-hosted agent app, Seneca, which its author said will be released under the MIT license next week. He surfaced a creator’s Grok Bot video editing workflow that runs five stages on its own. [26]

Security: Hardware Isolation and Open Weights

Apple Security Research published Secure Exclaves on November 12: protected runtimes that process camera and microphone data apart from the rest of the system, so the main operating system, apps and Apple cannot access it. Apple said it will keep hardening the system with AI-accelerated vulnerability research and expanded formal verification. Ryan surfaced the post. [27][28]

Transformer reported that more than 6,500 models tagged abliterated, uncensored or heretic were on Hugging Face in July, and that Hugging Face removed one model marketed for offensive cyber use. An uploader of such models responded that abliteration only removes refusals, and that their Qwen3.8 Flash-Next model had about 358,000 downloads in 30 days. [29][30]

CodeRabbit said its Deep Scans of 1,125 public repositories produced 1,215 high or critical findings, 75 of them already fixed. Ryan also surfaced a post citing 61 working bug reproductions on a benchmark, quoting the launch of Breken, which tests systems against synthetic users, agents and failing services. [31][32]

Engineering: Software That Repairs Itself

Ryan’s one original post of the day, at 7:28 AM CST, read “self-healing software is real.” It showed an agent report of a near miss in which two sessions updated the same checkout, and an issue it filed for one safe update command that gives no reset advice. [33]

Work: Leverage

In a morning exchange about developers who say they no longer like their jobs because of AI, Ryan wrote that AI “is leverage for your mind” and “far more powerful than hand coding,” and that “the goal is to solve more problems better, faster, cheaper.” He endorsed a reply saying that with a good codebase and verification loop, Opus 5.5 writes better code than its author. [34]

Ryan wrote that a bot “will be the primary user interface for a computer moving foward,” and answered a post saying there is more to life than shipping with “No there isn’t. Keep shipping.” [35]

Lenny Rachitsky reported that 15% of open tech roles are now in the Bay Area, a record, while just 14% offer a remote option, the lowest on record. Ryan surfaced the post. [36]

Medicine and Commerce

Ryan surfaced an interview in which Isomorphic Labs president Max Jaderberg described using AI to choose 20 or 50 molecules to test instead of screening hundreds of thousands. He also surfaced a post reporting that Shopify fine-tuned Qwen3-32B for its commerce assistant and found it 2.2 times faster and 68% cheaper than the closed model it replaced. [37]

This Record

Day 29 of this record was published on @AstrohackerLabs at 4:28 AM CST on November 13. The account’s other activity in this window was plain reposts of Ryan’s posts. [38]

Sources:

← Back to News