SpaceX Plans First Full Starship Catch on 15th Flight as AI Benchmark and Agent Product Claims Collide

Key Takeaways

1. SpaceX plans its 15th Starship flight for Nov. 13, aiming to catch both the Super Heavy booster and the ship with the launch tower for the first time.
2. The flight profile includes 26 Starlink V3 satellites, a 22-hour orbital stay, and multiple contingency splashdown zones before any return attempt.
3. A benchmark variant called MaxxBench shows models clustered near 74% on DeepSWE diverging by 23 percentage points, raising questions about industrial benchmark reliability.
4. Yann LeCun distinguishes engineering-scale JEPA models from academic method papers, warning against conflating the two.
5. Perplexity's Teammate is claimed to reach product-market fit in two days, with automated run volume up 3.7x from launch day to the next day.
6. Musk says Cybertruck order backlog exceeds pre-Cybercab capacity, while Alphabet's SpaceX stake is reported at 555 million shares worth $83.7 billion.

1. SpaceX Starship Flight 15

1. SpaceX plans to launch Starship on Nov. 13 at 17:45 ET, attempting for the first time to recover both the Super Heavy booster and the Starship vehicle with the launch tower's mechanical arms in the same mission; a success would mark the first complete recovery of an orbital launch system. The ship is to reach orbit at about 275 km, fly more than 22 hours over roughly 14 orbits, and deploy 26 Starlink V3 satellites, three of which carry cameras to inspect the ship's heat tiles; customer service could begin within weeks if on-orbit checks go well. 1 2
2. The vehicle first enters a passively safe suborbital trajectory, with the control team assessing system health before deciding whether to enter orbit. During about 21 hours on orbit it will check engines, power and communications, and tank structures, with contingency splashdown zones in the Indian Ocean, north of Hawaii, off Chile, and in the Gulf of Mexico; a return-to-site catch is attempted only if FAA-approved conditions are met. On reentry the ship will use flap control to fly over sparsely populated areas of northern Mexico and southern Texas, avoiding major Rio Grande Valley cities, and can deliberately tumble or divert to the Gulf of Mexico if the vehicle or tower has problems. 1 2

2. AI Benchmarks and Research Claims

1. Hamel Husain highlighted MaxxBench, which contains 89 DeepSWE tasks with only small changes to instructions or starter code, keeping the engineering problem the same while varying requirements slightly. Models that cluster around 74% on DeepSWE diverge by 23 percentage points on MaxxBench, raising the question of whether they are good at programming or good at DeepSWE; a quoted author said the "industrial benchmark complex is on thin ice," expressing doubt about industrial benchmark reliability. 1 2
2. Yann LeCun distinguished image and video self-supervised learning methods by reconstruction use, joint-embedding architecture with or without a predictor (JEA versus JEPA), anti-collapse method, and data augmentation. He said DINO is technically a JEA and a kind of JEPA, differing in its contrastive-clustering plus stop-gradient/EMA anti-collapse criterion and task-specific image transformations, and noted that V-JEPA 2.1 and DINO both come from Meta-FAIR with engineering teams, large or proprietary datasets, substantial compute, and fine-tuning. He said many JEPA papers are applications or fine-tuned versions of DINO, V-JEPA, or I-JEPA, while LeJEPA and LeWM-style papers are mostly method research by academic students with little compute and small or public datasets, tested on toy problems, simulations, or small benchmarks and not directly comparable to engineered models. 1

3. Agent Products and Industry Signals

1. Aravind Srinivas said Perplexity Teammate acts as a chief of staff or executive assistant, covering scheduling, email follow-up, task triage, vendor coordination, daily briefings, meeting prep, and next-step planning after a one-time setup, and claimed immediate product-market fit after two days of use. He said automated run volume grew 3.7x from launch day to the next day, which was a Friday, and argued Teammate is a tier above similar products because of its code-based execution framework for agent trajectories, model neutrality that does not force weaker non-frontier models, and Perplexity's search and accuracy quality. A quoted user said AuthorBot autonomously retrieved 92 papers from 1956 to 2026, wrote an 8-chapter monograph with 64 formulas and 13 figures, produced an EPUB3 and a 226-page print edition, and submitted it to KDP, notifying the user after Amazon accepted it. 1 2 3 4
2. Musk said Cybertruck order backlog is far beyond capacity before a line is removed for Cybercab expansion in June or July next year, and that an extra shift will be added to try to deliver on time; he said Cybercab can use a 40 kWh battery pack and that smart, focused people are limited, so keeping the complex Cybertruck line would take engineering and factory resources from Cybercab expansion. A reposted claim said Alphabet added 3.85 million SpaceX shares in the third quarter, holding 555 million shares worth $83.7 billion as of Sept. 30, nearly 95% of its disclosed stock portfolio; Google invested $900 million in January 2015 and the stake is now about 93 times the original investment. 1 2 3
3. Deedy described CXMT as about 10 years old, the world's 32nd-largest company and fourth-largest DRAM maker with about 10% global share, a $430 billion valuation, and briefly China's most valuable company. He said revenue grew more than 6x year over year, with $40 billion in the first nine months of the year and a 50% gross margin; about two-thirds of the business comes from phone memory and AI computing is a small share. Founder Zhu Yiming studied at Tsinghua, earned a doctorate at Stony Brook, worked in Silicon Valley semiconductors, returned to China in 2005, and founded CXMT in Hefei in 2016; the Hefei city government holds about 37% after providing physical city resources and absorbing 10 years of losses, which Deedy contrasted with the US venture model. 1
4. Demis Hassabis said Google DeepMind projects are undergoing a "true phase transition" in the speed of scientific discovery and that releases in the coming weeks will reflect it. Separately, swyx said Cursor has performed strongly since being acquired, calling it possibly one of Musk's best acquisitions and better than Twitter itself, and cited Cursor at $1 billion ARR at the end of 2025 with enterprise business up 300% quarter over quarter. 1 2