Starship Flight 15 Reaches Orbit, Deploys 26 Starlink V3 Satellites; Booster 22 Lost During Turnaround

Key Takeaways

1. Starship Flight 15 reached orbit, deployed 26 Starlink V3 satellites, and is planned to attempt a return and capture after a ~21-hour coast; Booster 22 was lost during an offshore turnaround.
2. Zambia began deploying Starlink kits across 751 health facilities in all 10 provinces, with 90 kits delivered to Western Province.
3. Hugging Face released TRL v1.18.0 with a stable AsyncGRPOTrainer and introduced per-model export, quantization, and end-to-end execution for transformers models.
4. ChatGPT Wallet launched for US Pro subscribers with American Express, Mastercard, and Visa, letting agents make purchases with user approval without accessing real card numbers.
5. Perplexity processes 1 billion decision-model requests per month on OpenRouter and expanded Teammates to its native macOS app for Max subscribers.
6. NVIDIA released the open System One model GR00T 2 for robot research and reported that Thinky Machines' 594B-parameter Inkling 2 Small was pretrained in 68 hours on a GB300 NVL72 rack using NVFP4.
7. Gemini 4 Carbon ranked first on MathArena, replacing GPT-6.1 Sol, with Fable 5.2 second, according to a repost by Demis Hassabis.

1. Starship Flight 15 and Starlink Deployment

1. Starship Flight 15 entered orbit and began a roughly 21-hour Earth coast phase, planned for 14 orbits before deorbit and a return-to-launch-site capture attempt; the ship and heat shield were described as in good condition, and the launch pad and tower systems were said to be ready for capture. The flight deployed 26 Starlink V3 satellites over a deployment sequence of about 42 minutes, all reaching orbit, and the author claimed future Starship flights could deliver more than 50 Tb/s of Starlink bandwidth per launch. 1 2 3 4 5 6 7 8
2. Super Heavy Booster 22 was lost during an offshore turnaround. The author attributed the green flame to engine performance slightly below expectations, involving copper combustion in the regenerative cooling circuit and an oxygen-rich main combustion chamber, and said the fix is to improve engine performance and consistency with Raptor V3.2 in progress, while moving MECO1 about 1 second earlier to preserve propellant for return and landing. The author said the booster design was already obsolete and did not affect the timeline for a high-frequency, fully reusable Starship. 1 2
3. On Boca Chica Beach, the author denied that SpaceX is trying to buy the beach and said the company is investing heavily in protecting and improving it. A reposted SpaceX statement said the beach will remain publicly open, citing weekly cleanups with the City of Starbase that have removed over 50 tons of trash, a proposed multi-million-dollar effort to address erosion and restore the shoreline, and a plan to buy about 1,200 acres around launch facilities, excluding the beach and Texas coastline, with about 600 acres permanently protected as sensitive coastal habitat and sale proceeds going to Rio Grande Valley parks and recreation. 1 2 3
4. Zambia began deploying Starlink kits at 751 health facilities across all 10 provinces, with the first 90 kits delivered to the Western Province health office, targeting underserved and hard-to-reach communities and funded by the Global Fund through the Ministry of Health with UNDP support. A related repost noted that many of these facilities are in Zambian floodplains that are inundated during the rainy season and that Starlink only needs a view of the sky, bypassing ground fiber and base station challenges. 1 2

2. AI Models, Developer Tools, and Agents

1. Hugging Face reposted the release of TRL v1.18.0, where AsyncGRPOTrainer is now stable and generation and training can run simultaneously on different GPUs rather than taking turns; the release also documents how to run it, what can be trained, and how it compares with GRPOTrainer. Hugging Face also announced support for exporting, quantizing, and running every transformers model end-to-end on any backend with the same inputs and outputs, positioned as a drop-in replacement. 1 2
2. ChatGPT Wallet launched with American Express, Mastercard, and Visa, allowing dot to complete purchases securely. The quoted post said the wallet lets agents pay on the user's behalf with approval, cannot access real card numbers, and requires a security check for each transaction; it is rolling out starting today to US Pro subscribers. 1
3. Perplexity processes 1 billion decision-model requests per month on OpenRouter and is still growing steadily; a quoted post noted that decider v1.3 is free for one week on OpenRouter. Perplexity also brought Teammates to its native macOS app for Max subscribers, where Teammates can independently or collaboratively handle QA, customer support, research, and growth marketing. The company said it will launch several new products that reshape how people perceive Perplexity and is hiring for frontier product roles, and a repost said browser agents in Computer and Teammates can now show their work so users can watch in real time, take over, log in securely when needed, and hand control back. 1 2 3 4
4. NVIDIA released the open System One model GR00T 2 for robotics research, which it says can learn skills from human demonstrations and predict how its own actions will change the physical world, with skills transferable across robot embodiments to reduce the need to teach every task from scratch on every robot. NVIDIA also said visual AI and robotics are being used in recycling to sort old clothes, bottles, and mixed recyclables, keeping thousands of tons of waste out of landfills through several startups. 1 2
5. NVIDIA AI said its collaboration with the SGLang project and the Miles community now spans five generations of accelerated computing from Ampere to Vera Rubin, with more than two years of engineering collaboration, over 100 model zero-day releases, and more than 10 developer events. A repost showed a robot based on Gemma 4 E4B running real-time conversation, tool calls, animation, and vision models on a single 8GB Jetson, scanning a room and turning on a TV to play a movie after startup. NVIDIA also congratulated Thinky Machines' Inkling 2 Small, a 594B-parameter model pretrained in only 68 hours on a GB300 NVL72 rack using NVFP4; the quoted post said the model uses native 4-bit forward propagation from the first pretraining step through final RL rollout to reduce the effect of quantization error on learning. 1 2 3
6. Gemini 4 Carbon ranked first on MathArena, replacing GPT-6.1 Sol, with Fable 5.2 second, according to a repost by Demis Hassabis that emphasized it was the first time in a long while that a non-OpenAI model topped the leaderboard. A reposted weekly release list included two eighth-generation TPUs on Monday; Gemini 4 Carbon, an AlphaProof Code-verified Signal protocol, and Gemini Robotics 2.5 on Tuesday; AlphaGenome 2 for disease research and a Gemini Decisions API in AI Studio on Wednesday; and Gemini 4 Hydrogen plus smart glasses preorders on Thursday. 1 2

3. Safety, Evaluation, and Industry Debate

1. Anthropic published an engineering blog with 10 recommendations for designing and running safety evaluations and reporting results, arguing that safety evaluations are important for deciding whether a model can be deployed, what safeguards are needed, and where training and research should focus. 1
2. Hamel Husain reposted and commented on an OCR benchmark dispute, saying much of the dissatisfaction stems from benchmarks having too much influence on vendor selection while people look only at top-line scores and not the details. He argued that publishing benchmarks carries moral hazard even with good intentions and that model users who rely on benchmarks should inspect the trajectories of 3-5 failed tasks themselves rather than concluding from Twitter arguments, especially for large procurement decisions, and warned that benchmarks of benchmarks can also be low-quality content. A reposted post said flawed OCR benchmarks are numerous and added weekly and offered a guide to check any benchmark in 5 minutes, run evaluations on your own documents, and identify common problems. 1 2 3
3. Ethan Mollick said that after watching student presentations at Penn's Generative AI studio, he believes creative people can now push personal projects further with AI than a full team could three years ago. He also said that swarms of agents executing research tasks will generate substantial traffic and do not tire, citing Fable 5.2 finding an obscure book in the Internet Archive and reading it line by line through repeated full-text searches to reconstruct a 1941 W.H. Auden poetry class. 1 2
4. Yann LeCun reposted a talk preview on end-to-end learning of predictions from arbitrary data, covering why humans excel at such learning and why JEPAs are the next step. He also reposted that Courant launched a new Center for Mathematics and AI (MAIC) whose mission is to make frontier AI systems objects of scientific study so the mathematics community can dissect, improve, reproduce, and ultimately understand them. 1 2
5. A repost by Yann LeCun of a Chinese post alleged that Jensen Huang's resignation was a setup involving US politicians, media, the CIA, and Anthropic, and that a New York Times reporter acknowledged receiving a tip from a "tech contact" with open-source intelligence pointing to Anthropic co-founder Jack Clark; the post also listed claims that Anthropic donated $40 million to organizations supporting tighter China chip controls and that Senator Kramer received $51,000 in political donations. These are the post author's allegations and inferences, disputed/unverified. 1
6. Elon Musk said he signed a petition calling for Jensen Huang's reinstatement, and a repost said Altimeter and Atreides manage over $25 billion in assets combined and that the founders of Uber, Shopify, Replit, and Vercel, as well as Musk, publicly support it. Musk also said many of Huang's supply chain deals at NVIDIA are handshake agreements with him personally rather than written contracts, and that losing him would cost at least some key deals. 1 2 3 4 5