Key Takeaways
- Starlink is moving into commercial aviation, government rail safety, and school connectivity across several countries.
- United says Starlink is drawing passengers and supporting revenue, with full-fleet deployment targeted for end-2027.
- SpaceX's Starship reuse work continues, with a Morgan Stanley analyst calling the upper-stage catch a "historic inflection point."
- OpenAI says GPT-6 Astra helped a physicist propose a candidate lattice definition of the Standard Model and is starting embedded external evaluations.
- Perplexity launched Frontier Research with at least 10,000 GPUs and outlined Teammates features for model choice and financial research.
- Qwen4 Max posted mixed benchmark results against Kimi K3.1 Pro and Hy4, with lower cost per task than Qwen3.8 Max.
1. Starlink Expands Into Aviation, Government, and Education
- Starlink is accelerating into aviation, government, and emergency communications. Lufthansa's first Starlink-equipped A350-1000 is scheduled to fly Munich to Tokyo on December 3, and the Lufthansa Group plans to equip about 850 aircraft by 2029 1. Honduras has installed 1,860 Starlink antennas across 218 municipalities, prioritizing rural schools 2, while Greece is using Starlink to transmit real-time train data for railway safety, with 134 locomotives carrying monitoring equipment, positioning accuracy up to 5 cm and updates every second, and automatic collision-risk alerts reaching the control center 3.
- United's CEO said Starlink has attracted passengers from other airlines; it now covers more than 50% of its daily flights and more than 700 aircraft, with full-fleet deployment expected by the end of 2027. The related matching campaign drew about four times the normal application volume, 25% of which already hold or applied for a United credit card, which Kirby called a "unique RASM tailwind" 1. Separately, the Dominican Republic has more than 50,000 Starlink antennas, 1,608 students at ITLA's two campuses benefit, and the government is discussing using Grok Tutor to support personalized learning 2; in Brazil, all rural municipal schools in Rondonópolis are connected to Starlink 3.
2. SpaceX Details Starship Reuse and Launch Progress
- An author said Starship V4 will be about 70 meters long, roughly 40% longer, with more than 60% more propellant and nine engines, including three higher-thrust Raptor-Vacuum V4 engines 1. The author said the entire Starship heat shield was inspected and the overall vehicle and engines were in good condition, and Ship 42 may be able to fly again after a few weeks of refurbishment; however, achieving aircraft-like immediate reuse with zero expected refurbishment and only scheduled maintenance would still require many small repairs 2.
- A forwarded Morgan Stanley analyst Adam Jonas view called Flight 15's successful catch of the Starship upper stage a "historic inflection point" for the company and the space industry: apart from the Space Shuttle, no one had previously attempted to recover a rocket upper stage, and full recovery would allow deep inspection of Ship 42 for improvements to later vehicles and help advance full booster and ship reuse. Jonas called it the last major technical milestone before weekly or even daily Starship flights and said it reduces long-term TAM risk across three businesses 1. Another forwarded post said that if Starship were a standalone launch company, it would already rank second in the U.S. by payload to orbit, behind only Falcon 9 2.
- Responding to the argument that "anyone can throw money at building Starship," the author said SpaceX can catch a giant spacecraft from orbit with a mechanical arm and that Falcon carries 85% of the world's payload to orbit, while Blue Origin, which started earlier with Bezos funding, has not yet launched reliably; Boeing and Lockheed received several times SpaceX's government development funding, but Starliner is still not operational and ULA's payload to orbit is less than 5% of SpaceX's 1. The author also forwarded a claim that NASA gave Boeing $4.2 billion in 2014 to develop Starliner and SpaceX only $2.6 billion for Crew Dragon, which has completed 21 crewed missions since 2020 2.
3. OpenAI and Perplexity Push Research and Agent Capabilities
- OpenAI said physicist Nathaniel Craig worked with GPT-6 Astra to propose a candidate lattice definition of the Standard Model, the first article on OpenAI's new science blog 1 2. OpenAI also said it is starting a first round of embedded evaluations with ApolloResearch, METR_Evals, TransluceAI, and GoodfireAI, giving outside experts deep access to internal models and systems to investigate alignment and test safeguards 3 4. OpenAI said the evaluations build on existing alignment and safety research and aim to let outside parties independently review how it identifies and manages system risks 5.
- Perplexity launched Frontier Research, committing at least 10,000 GPUs to study multi-agent orchestration and continual learning with the goal of building systems that can conduct research autonomously, and said it will publish evaluations, papers, and blogs and open-source models and tools 1. The plan emphasizes large-scale reinforcement learning: training large models on more than 10,000 GPUs, with training data and environments drawn from how users actually use Perplexity; the team said it has built full-stack infrastructure including search, inference, training, and sandboxes, and results will be validated in products 2. Hiring locations are San Francisco, Palo Alto, New York, and London 3.
- An author said Perplexity's Teammates lets users choose any frontier model to build a teammate and argued no other product offers this capability 1. A quoted post argued most AI assistants use the same model and effort level for all tasks, while users need to distinguish by task complexity, such as routine email summaries versus stock research or code review, and the author said this made them lean back toward Perplexity 2. The author also said a Teammate built on Fable 5.2 created a browser-playable open-world driving game set in neon 1986 Miami without a game engine, Three.js, or external assets, writing its own WebGL2 renderer with the city, cars, garage, rain, and engine sound generated by code, at about 63,000 lines of JavaScript and 9 MB; a 40-second 1080p60 video was rendered by the Teammate on a rented cloud GPU 3. The author said Perplexity's agents, Computer or Teammate, can cite every number in financial research and analysis to a specific line item in earnings reports 4.
4. Model Benchmarks and AI Research Notes
- An author said that more than eight months after Fable/Mythos was released publicly, no open-source model has reached "Fable level," contrary to many expectations 1. A forwarded evaluation showed Alibaba's Qwen4 Max scored 43 on the Artificial Analysis Intelligence Index, 9 points above Qwen3.8 Max and 3 points below Kimi K3.1 Pro, at about $3.27 per task, roughly two-thirds less than Qwen3.8 Max and about 11% less than Kimi K3.1 Pro 2. Qwen4 Max scored 1,686 Elo on AA-Briefcase, above Kimi K3.1 Pro's 1,674 and Tencent Hy4's 1,446, but scored 48% on Terminal-Bench 4.0, below Kimi K3.1 Pro's 67%, and 26% on FrontierCode 2.0, below Hy4's 31% 3. Qwen4 Max supports a 1 million token context and text and image inputs, with API pricing at $1.50/$4.50 per million input/output tokens and a $0.075 cache-hit price 4.
- An author said that a few months ago something seemingly impossible was achieved: using Fable & Opus to find a Starship recovery video and render the last roughly 40 seconds in 3D in 3js at very high detail, with switchable viewpoints and no crowd needed 1. A forwarded post said Antigravity doubled the context windows of Argon, Carbon, and Hydrogen, and users should see 3.5 times less compression per conversation turn 2. Another forwarded post said its first prompt with Carbon solved a research problem that multiple rounds with Astras and Fables had not 3.
- A forwarded post said Grok Bot's delivery pace at SpaceXSI included 11,000 pull requests in one month with a rollback rate of about 0.5%, more contributors, stable daily releases, and faster application speed 1. An author said a giant rocket about 52 meters (171 feet, 15 stories) tall flipped in the air and landed successfully and is reusable, and is expected to cut payload launch costs from $2,000 per kilogram to $200 per kilogram; the author called reusability possibly the most important advance in spaceflight in our lifetime 2.
