SpaceX Completes Three Falcon Launches in Under 13 Hours as Starship V3 Deploys 26 Satellites

Key Takeaways

1. SpaceX completed three Falcon launches and four booster landings in under 13 hours, plus a Dragon ISS docking for the 21st crewed mission.
2. Starship Flight 14 reached orbit for the first time and deployed 26 Starlink V3 satellites.
3. Musk said about 90% of SpaceX revenue this year will come from commercial business, with government revenue under 5% in Q4.
4. Tesla reported Q3 2026 production of 464,391 vehicles, 486,532 deliveries, and 13.7 GWh of energy storage.
5. Perplexity released multiple open-source projects including pplx-decider-v1-27b, Lily, PII-Tracer, WANDR, Numbat, and Bumblebee.
6. NVIDIA's Nemotron 3.5 ASR cut Arabic dialect word error rates from 55% to 30% after fine-tuning.
7. Trillium Labs launched as a new nonprofit for open frontier AI science, with initial backing from Halcyon Futures and Schmidt Sciences.

1. SpaceX Launch Cadence and Starship Progress

1. SpaceX completed three Falcon launches and four first-stage landings across four landing zones in Florida and California in under 13 hours, while also docking a Dragon spacecraft to the ISS for the company's 21st crewed mission. A separate summary listed five missions in six days: USSF-385 on September 26, Starship Flight 14 on September 28, and Crew-13, Transporter-18, and NROL-97 on October 1, with three of those occurring within 13 hours. 1 2 3 4
2. Starship Flight 14 completed its first orbital flight and deployed 26 Starlink V3 satellites. Musk described the hot-staging sequence as "science fiction in real life." A retweet claimed each launch could add more than $700 million in annual incremental revenue, growing as SpaceX AI enters AI software monetization. 1 2 3
3. Musk said about 90% of SpaceX revenue this year will come from commercial business, with government revenue falling below 5% in the fourth quarter. He contrasted this with federal spending at roughly 25% of the economy, calling SpaceX "far more commercial" than the average US company. He also cited a claim that SpaceX has saved the US government and taxpayers about $14 billion, adding that the actual figure is now higher. 1 2

2. Tesla Operations and AI Infrastructure

1. Tesla reported Q3 2026 production of 464,391 vehicles, 486,532 deliveries, and 13.7 GWh of energy storage deployment. The company said it will update on October 21 at 16:30 CT, streamed on X. 1
2. Musk said Robotaxi operating hours in Austin were extended from 10 PM to 11 PM, with the main issue being difficulty seeing pets at night to avoid running over a gray kitten in the dark. Tesla's Supercharger network now has more than 85,000 charging stalls across over 9,000 stations. 1 2
3. NASA certified the Cybertruck as an emergency evacuation vehicle for astronauts, allowing SpaceX to replace aging MRAPs. A NASA spokesperson said both parties evaluated vehicle capabilities across emergency scenarios and determined the Cybertruck meets NASA requirements. 1
4. A retweet described Terafab's target as 1 terawatt of compute capacity per year, roughly 50 times current global AI compute output and far beyond existing semiconductor industry scale. Musk commented that "a sufficiently large quantity is itself a quality," comparing the difference between a single-celled bacterium and a human with 35 trillion cells. 1

3. Perplexity Open-Source Releases and Agent Infrastructure

1. Perplexity released pplx-decider-v1-27b, a multimodal decision model averaging 85.7% across 11 benchmarks and leading Jev; pplx-embed-v2-context-9b-preview, a context embedding model that performed best on ConTEB and turbopuffer context-bench; and Lily, a local inference engine for Apple silicon built in Rust with custom Metal kernels and no PyTorch or MLX dependency, achieving 1.23x faster prefill and 1.35x faster decode than MLX-LM on M5 Max. 1
2. Additional open-source projects include PII-Tracer, a 0.6B on-device PII classifier for deciding whether mixed-compute tasks stay local on a Mac, which outperformed OpenAI's Privacy Filter on all five public benchmarks and shipped alongside the PII-TRACE benchmark with 13 languages and 13,000 conversations; WANDR, a benchmark for breadth and depth research agents with 500 tasks requiring 170,000 sourced records; Numbat, an agent detection and response tool for laptops and workstations with 52 rules in a single Go binary supporting macOS, Linux, and Windows; and Bumblebee, a read-only supply chain scanner for developer machines covering packages, MCP configurations, and editor and browser extensions. 1
3. The author said more open-source contributions are coming and described a goal of vertically integrating agent infrastructure by building its own sandbox and optimizing it for the best chips, calling Vera far superior to x86 and expecting to share more as Perplexity Computer begins deploying on Vera. The author also retweeted anticipation for working with NVIDIA Vera CPU on SPACE. 1 2
4. The author cited a Computer project building a 3D map of nearly 26,000 restaurants and cafes across New York City, saying that on Computer creativity is the only limit and such projects can let agents run for hours and produce excellent results. The author also retweeted five examples of Computer generating inline visualizations directly within a conversation thread, the first being an inline 3D cutaway showing how a jet engine works. 1 2

4. NVIDIA Speech Models and Recognition

1. NVIDIA Nemotron 3.5 ASR, after fine-tuning, reduced word error rates for Saudi Najdi and Hijazi dialects from 55% to 30%. NVIDIA said speech models can support Arabic but still struggle with local dialects, and released a tutorial on adapting to other dialects and languages. 1
2. NVIDIA said its partnership with Korea has exceeded 25 years and thanked The Korea Society for awarding Jensen Huang the 2026 Van Fleet Award and for recognizing the US-Korea partnership. 1
3. Nemotron 3 Diarization trended on Hugging Face, with NVIDIA thanking downloaders and commenters and noting that @sabbassi_11 had answered some questions. The model tracks who speaks when in multi-person conversations, handles overlapping speech and up to 8 speakers, has 100M parameters, and is available on Hugging Face. 1

5. Open AI Science and Industry Debate

1. Trillium Labs launched as a new nonprofit aiming to advance open science for frontier AI, initially building open post-training recipes and later expanding to open infrastructure, with research directions including RSI, reward hacking, and multi-agent systems. The organization said its theory of change is that more eyes are needed to solve difficult technical problems and expressed concern that AI is becoming too closed to benefit from the scientific method and community. 1
2. Trillium Labs was co-founded by the tweet author and longtime friend and collaborator Tom Zick. It is hiring full-time and student collaborators/interns, raising funds, and seeking compute, with offices in the Bay Area and Cambridge, Massachusetts, and remote options. It has initial support from Halcyon Futures and Schmidt Sciences, with more funding on the way, and advisors including Thom Wolf, Hanna Hajishirzi, Gneubig, and Ctnzr. 1
3. Ethan Mollick argued that self-congratulation about AI being early is overdone: many executives are smart, motivated, business-savvy, and technically skilled and are embracing AI, but transforming a company is a longer process. A quoted tweet said AI has gone mainstream, executives are using AI to code, adopting Claude, and looking for automation workflows, AI capex is supporting the S&P, and US enterprises are ready. 1
4. Mollick cited current best evidence as supporting only a narrow conclusion: AI may already be affecting hiring margins for entry-level white-collar jobs most exposed to AI, but that attribution is disputed and broader labor market disruption has not yet appeared in the data. 1
5. Yann LeCun highlighted a split in AI's employment impact: routine office roles such as data entry and customer service representatives are currently harmed, while data scientists and skilled trades such as data center construction workers and electricians benefit. A retweeted view said a two-year-old video remains relevant and human-level AI is still far away. 1 2
6. The same view noted AI already has superhuman performance in math, programming, and question answering with known answers, but questioned why Level-5 fully autonomous cars are still absent, noting Tesla FSD is Level 2 and Waymo is Level 4. It asked why no self-driving car can learn to drive like a 17-year-old through 20 hours of practice, said imitation learning has failed to train autonomous driving even with billions of hours of training data, and questioned why home robots and robots capable of 8-year-old-level learning do not exist, arguing AI still struggles with the complexity and messiness of the physical world. 1

6. AI Products and Platform Changes

1. Sam Altman called dot his current favorite OpenAI product, saying it visibly improves every day because it continuously learns his workflow and style, and that letting dot handle tasks he dislikes and that usually accumulate into a "gravity well of dread" makes him happy. He also responded to speculation about a Cerebras partnership, calling Cerebras a close partner with deep collaboration pushing the frontier of speed. 1 2
2. Altman said he is very disturbed by attempts to give AI models religious-style power or to ask people to abandon human judgment, calling it a real safety issue. 1
3. Runway shared how a fashion brand uses Runway Dev from initial sketches to final marketing campaigns, rendering sketches into real garments, exploring colors and patterns, generating e-commerce product images, and producing model photos and videos in studio or on location. 1
4. YouTube announced changes to its short-video algorithm to prioritize original content and reduce distribution of re-uploads without added content, meaning channels focused on re-uploads will see reduced distribution in the Shorts feed. The platform also said it does not encourage "minor technical edits" or "template" changes. A retweeter said this will hit editors who build on others' content but helps address content quality issues. 1
5. Hamel Husain will hold his last public class of the year, "How to Nail the AI Evals Interview," on Tuesday, October 6 at 8:30 AM Pacific, aimed at people with basic evals knowledge. It will cover traps in evals interviews, what companies care about, and how to structure a take-home or report. A retweeter said applied AI evals interview preparation is a common need, recommended the speaker, said the lecture was effective, and noted it is free and open to all. 1 2 3
6. Husain criticized a demo video as hard to understand and "visual garbage," questioning why a demo cannot be made understandable while saying it is not disrespectful and the feature itself may be cool. He also expressed confusion about workflows that require keeping a laptop open, suggesting SSH into the computer or using Codex remote features with a Mac mini, and said continuing that way is unreasonable if the problem has trapped you more than three times. 1 2

7. Other Notable Items

1. A retweet described the death of 93-year-old bedridden disabled woman Marie-Louise in Toulouse, France, found at home with blood on her body and a naked 33-year-old Sudanese illegal immigrant lying on top of her. Investigators suspect sexual assault and beating; the suspect had consumed large amounts of alcohol, violently resisted arrest, and injured a police officer. 1
2. A retweet said all testing frameworks should remove queues, just as plan mode was removed. 1
3. A retweet said academia is suited for high-risk research bets: PhD students lack unlimited compute but have the freedom to explore ideas outside the mainstream and weigh different tradeoffs, which may produce new paradigms. The related podcast also discussed kernels, harness design, and multi-agent systems. 1
4. A quoted tweet said dots is powered by GPT-6 Astra and is an always-on agent designed to handle various tasks; the quoting account added that the team has been developing dots for some time and teased "one more thing." 1
5. Mollick said his upcoming book Co-Existence may be the first to include endorsements written specifically for AI readers, with a recommendation from Tyler Cowen. The book website also has a page for AI and offers carefully designed preorder rewards. 1
6. Mollick questioned whether X Money involves some kind of scam: he has not enabled the feature, but occasionally his posts are mass-retweeted by bots with messages like "lets send him X Money," making him suspect a精心 designed (crypto?) operation. 1
7. Mollick imagined superintelligence using the Dungeons & Dragons approach to a 25-Intelligence villain: do not try to outplan the players, but act as if the villain foresaw everything no matter what the players do or roll. 1
8. Ben Tossell said his first tech job was working with someone he called the "GOAT." 1
9. A retweet previewed that @swyx will have new content coming. 1
10. A retweet thanked @swyx and @vibhusapra for an invitation, saying participating in a project was "very fun" and looking forward to others seeing it. 1