Key Takeaways
- SpaceX recovered an orbital-class Starship second stage with the launch tower's “chopsticks,” a first for a returning orbital vehicle, according to Musk and a reposted analysis.
- OpenAI expanded free frontier-model access to up to 200,000 scientists, citing a study in which 1,800+ researchers reported near-doubled productivity.
- Perplexity launched a scheduled-change Monitor API and a Multiagent mode, while its CEO said ARR has passed $1 billion.
- Hugging Face and OpenRouter integrated inference, shared tokens and endpoints, and added model statistics on the Hub.
- NVIDIA released LidarVerse for LiDAR simulation and Nemotron for Science, and reported Cosmos at 20 million Hugging Face downloads.
- Tesla plans a roughly €3 billion Berlin expansion for Cybercab capacity, with construction targeted for the first half of 2027.
- Isomorphic Labs raised $5 billion to accelerate AI-first drug design, according to Demis Hassabis.
1. Space and Infrastructure
- SpaceX completed the first capture of Ship 42, a second stage returning from orbit, by the launch tower's “mechanical chopsticks,” Musk said 1 2 3 4 5. A reposted analysis said an orbital second stage must reach orbital velocity, deploy payloads, coast, perform a precise deorbit burn and survive reentry heating, making the feat harder than booster recovery; it said no system, including the Space Shuttle, had recovered an orbital-class rocket this way 6. NASA Administrator Isaacman called it something that “really changes everything,” with the goal of near-hourly Starship flights supporting the orbital economy and commercial space stations 7 8.
- Tesla plans to invest about €3 billion to expand its Berlin Gigafactory with a building for up to 1 million Cybercabs a year, according to the supplied material 1 2 3. Construction is planned to begin in the first half of 2027, Cybercab production is targeted for 2029, and the permit application is due by the end of 2026; the plan includes ancillary buildings, a new wastewater plant and hundreds of added construction and production jobs 4 5 6. The plant currently makes about 7,000 Model Y vehicles a week and targets 7,500 7.
- Starlink deployments cited in the material include 36 North Star offshore wind and maritime vessels replacing roughly 1 Mbps VSAT links 1, 435 kits for rural schools and health centers in Burundi under a World Bank-funded project 2, and free Wi-Fi on Jin Air Boeing 737-8 aircraft next month, with Korean Air and Asiana already having 42 equipped planes 3. Separately, SpaceX added $30 million to Brownsville water infrastructure, bringing its deposited total to $70 million of a $220 million commitment for South Texas water and wastewater facilities 4.
2. AI Models, Agents and Research
- OpenAI expanded free frontier-model access to up to 200,000 scientists worldwide with training and support; applications open November 23, and data shared in the program is not used to train models by default 1. A cited post said the program started with 10,000 researchers and aims to reach 100,000 by 2027 for scientists, mathematicians and engineers 2. In the ChatGPT for Academic Researchers study, more than 1,800 participants reported near-doubled productivity, with privacy-preserving methods used to examine work from ideation and data analysis to coding 3 4.
- Perplexity launched a Monitor API that tracks any event, topic or web page change on a schedule and pushes updates via webhook, priced at $1 per 1,000 runs or fetches and driven by its search API and a relevance-and-timeliness decider 1 2. It also launched Multiagent in Computer and Teammates, orchestrated by Fable 5.2 across Gemini 4 Carbon, GPT 6.1 Sol and Opus 5.5 parallel agent streams for complex tasks; it is available to web Max users first, with mobile later 3. CEO Aravind Srinivas said ARR has passed $1 billion with strong margins and healthy growth 4.
- Hugging Face and OpenRouter integrated OpenRouter directly into Hugging Face with a shared token and endpoint, live model statistics on the Hub and redesigned inference components 1. A reposted description said Hugging Face handles storage and collaboration while OpenRouter handles inference, supporting OpenRouter inference on HF, open model weights from OpenRouter, and OpenRouter agent traces stored on HF for inspection, training or sharing 2. A separate reposted account said Claude implemented trellis.2 in CUDA C++ at 5.3 times the speed of the original PyTorch version, with the unexpected bottleneck in glb export rather than the neural network; the zero-Python, all-C++ implementation was tested on H200 via Hugging Face Jobs at about $7 in Claude cost and $60 in compute, retained a 2.2 times gain with bitwise-identical output, and was open-sourced 3.
- NVIDIA Research released LidarVerse, described as the first model that can simulate scans for different LiDAR sensors, reconstruct scenes from new viewpoints, generate scenes from maps and predict a sensor's next observation in real time; the paper and code are public 1. NVIDIA also released Nemotron for Science at SC26, with open models, datasets, tools and post-training recipes for research AI agents that plan tasks, run simulations, check results and adjust 2. NVIDIA Cosmos reached 20 million Hugging Face downloads, positioned for developers and researchers building robots, autonomous driving and visual AI agents 3 4.
- Isomorphic Labs raised $5 billion to accelerate its mission to “solve all diseases,” Demis Hassabis said, adding that the team continues to use AI to rethink drug design from first principles and sharply cut the time and cost of new drug development 1.
3. Product and Industry Signals
- Grok Bot 1.0 launched with team Bot improvements, meeting participation, proactive task execution, learning and faster, smarter behavior, Musk promoted it as “Try Grok @Bot” 1. Users described a 90-minute AT&T support call handled by the Bot, with the agent downloading it after learning what it was 2, and a case in which it read a wife's medical records, researched cancer treatments and trials, found a Stanford specialist and drafted an email, with Stanford calling hours later to start intake 3. Another user said Grok completed a task in two hours that OpenAI's Dot had not progressed on 4.
- Gamma's large-scale testing of new LLMs found OpenAI leading on efficiency and vision, Anthropic strong in nuanced understanding, Google still competitive and open models clearly behind its benchmark performance, according to a reposted account 1. The same account said Gamma's spending on OpenAI exceeded Anthropic for the first time in years last month, but called OpenAI's lead unstable and flagged Gamma's own model as a possible next lab to watch 2.
- Noeri, the startup from Yejin Choi and Carlos Guestrin, emerged from stealth with $110 million in funding from investors including Menlo Ventures, Madrona Ventures, NVIDIA, NEA, Kindred Ventures, Wing VC, AIX Ventures, Principal VC, Cisco and Snowflake 1. It calls itself a “neolab” advocating “plural intelligence,” an ecosystem of models of different sizes and capabilities rather than one giant model, and says its approach includes learning continuously from real work, using other AIs as tools and choosing appropriately sized models at far lower cost than frontier models 2.
- Mistral AI said its Vibe tool would be free in Europe for the first three months with code LECHONK, calling Vibe and Le Chonk the “world's best open-source model-tool combination” and claiming Vibe is “genuinely better than GLM / DeepSeek” for coding, a one-sided claim without evaluation evidence in the material 1. Mistral later said the day's codes were gone and pointed to more news the next day 2.
