Key Takeaways
- SpaceX completed its 15th Starship test flight, with Ship 42 becoming the first ship caught by the launch tower's mechanical arms after 14 orbits; Musk called it an "epic day."
- Musk said Ship 42 looked good after inspection and traced the catch concept back to his December 2020 push for Mechazilla-style arm capture.
- Google described multi-agent workflows based on Gemini 4 solving open math, science and engineering problems, and opened related tools in Antigravity with free and paid access tiers.
- Perplexity's Aravind Srinivas reported rapid growth in agent usage and described a search-as-code SDK, millions of concurrent sandboxes and an Apple Silicon inference engine.
- Ethan Mollick's SatBench found 63% of 173 benchmarks cited by labs since 2023 are now saturated, with saturation accelerating.
- The "No AI Slop Manifesto" launched at the AI Engineer Code Summit with 300-plus signatures, arguing AI amplifies rather than replaces human judgment.
1. SpaceX Starship
- SpaceX's 15th Starship test flight ended with Ship 42 being caught by the launch tower's mechanical arms for the first time, after the vehicle flew 14 orbits; Musk confirmed re-entry burn was normal and called it an "epic day," thanking the SpaceX team. The mission launched Nov. 13 from Starbase in Texas and returned Nov. 14, with forwarded content calling it a "holy grail" moment for rocket technology. 1 2 3 4 5 6 7 8 9 10
- Musk said he inspected Ship 42 and considered it in good condition, adding that he hugged it upon its return; forwarded images and video showed the capture process. He also noted that in December 2020 he suggested and insisted on using the Mechazilla tower arms to catch returning Starships, and forwarded content said this was the first ship catch after three booster catches on Oct. 13, 2024, Jan. 16, 2025 and March 6, 2025. 1 2 3 4 5 6 7
2. AI Models and Infrastructure
- Google described multi-agent workflows using /teamwork and DeepReview based on Gemini 4 that reportedly solved more than 10 open math and CS theory problems, including a 50-million-line Lean 4 proof of the Courtade–Kumar conjecture, quantum query complexity and the Tarjan deque conjecture; it also reported autonomous Lean formalization of the Weil, Birkhoff and 3D Bochner–Riesz conjectures. In chemistry, NMR molecular structure parsing reached 95% top-1 exact match, nearly double a single agent, while systems work included clustering trillion-edge graphs up to 180x faster, security Rust rewrites of 800,000-line Zircon and 170,000-line libgav1 running 20–30% faster than C++/ASM, a 4.09x JAXBench TPU kernel speedup, over 1 PiB of cluster memory saved and a 24% reduction in Borg TPU preemptions. 1
- The tools are open in Antigravity: Hydrogen is available to all users, Carbon and Argon to enterprises and Ultra/paid users, 10,000 academic researchers can apply for six months of free top-tier usage, and all users get two free DeepReview calls per week. Forwarded first results for Gemini 4 Hydrogen on 105 planted bugs across two real repositories said it was stronger and cheaper than Haiku 5.5 (max), stronger and faster than GPT-6 Luna (max), slightly behind Sonnet 5.5 (xhigh) and clearly above GPT-6.1 Terra (xhigh), making it suitable as a fast sub-agent that can do real work. 1 2
- Perplexity CEO Aravind Srinivas relayed notes from a roughly 15-minute SGLang Summit talk, "Building the Next Computer": Perplexity built a Slack agent in January that employees now message more than each other, exposes search capabilities through a Python SDK so agents can build their own pipelines, runs millions of concurrent sandboxes daily, and reported a fine-tuned Qwen 27B scoring 59.6% locally on 89 tasks, rising to 73% with on-demand Opus 5. The notes also said an intern-turned-full-time employee built the Apple Silicon inference engine Lily, which was about 22% faster on prefill and 30% faster on decode than MLX-LM on an M5 Max. 1
- Srinivas said Teammate drove an 11x week-over-week increase in weekly agent hours, up from 7x the prior week, with week-over-week growth above 50%; since September, the number of employees gaining an extra workday (8+ hours) and an extra workweek (40+ hours) more than tripled in a week. By his calculation, about $10 million in token costs corresponds to roughly $80 million to $180 million in annual labor value, or 9x to 19x depending on the role. 1
3. AI Research and Practice
- Ethan Mollick released SatBench, which measures how fast major AI benchmarks are being "maxed out": of 173 benchmarks cited by labs since 2023, 63% are now saturated, and the pace is accelerating, which he said shows AI progress outpacing measurement. He added that a similar benchmark and analysis in STATA would have taken days a few years ago and been time-consuming to make presentable, while this took about four prompts to Fable 5.2. 1 2
- Mollick also described using Fable 5.2 to reconstruct W. H. Auden's 1941 poetry course, called "the hardest course in the humanities," from Auden's work and scattered textual evidence in surviving notes; the result was highly readable, and the cited post said the agent repeatedly searched full texts and read a rare but relevant book in the Internet Archive line by line. 1
- Deedy tested Fable 5.2 on tasks including an Anthropic S-1 video explainer, a physics-accurate mobile cricket game, an animated story explaining the Bible from a PDF, a highly optimized Rust crawler at 40,000 QPS, algorithmic art and a visualization of US healthcare funding flows; 3D face rendering failed, while anime character drawing, rigging and scene performance were average. 1
- The "No AI Slop Manifesto" launched at the AI Engineer Code Summit in San Francisco, reached the Hacker News front page and gathered 300-plus signatures from companies including Baseten, SpaceX, OpenAI, PostHog, Shopify, Glean, ZoomInfo and others. Its core claims are that AI can only amplify rather than replace human judgment, people remain responsible for the intent, taste and consequences of deliverables, and even if models get strong enough to skip reading most code, plans, designs or results still need review or output will look like everyone else's. 1
- swyx argued that current AI interfaces remain largely turn-based and that AI UX is severely under-optimized, especially given models cheap enough that cost is negligible. On "software factories," he said AI output should be slowed to human speed or it will overwhelm players, accountants, merge queues and machines; software factories are extremely expensive to maintain, and he has torn down five previous factories because rewriting was cheaper than fixing, so he does not recommend buying or building one hastily; next year will bring widespread human/agent and agent/agent workplace friction, with nearly everyone having a personal chief-of-staff agent and bringing it into work. 1 2
4. Industry Signals
- Yann LeCun argued that extinction and jobs-catastrophe scenarios are "completely absurd" and that p(doom) estimates are meaningless because humans have full agency over the relevant events; he said he differs with Geoff on "faith in humanity" and the priority of intelligence, that intelligence does not equal power domination, and that superintelligent AI will obey humans under safety guardrails without a desire to take over, citing a proposed "goal-driven AI architecture." He also said that despite being saddened by something, he will work harder to advance open-source AI because it matters and cannot slow down, warned against rushing to follow the "encoder-free" trend without careful thought, and said everyone in a democracy bears partial responsibility for steering technology's benefits and harms, with AI's gains real and set to expand and negative effects so far small. 1 2 3
- Musk forwarded claims that low-orbit satellite volume is necessary for global bandwidth gains and that the achievement "doesn't sound that good"; he expects SpaceX communications satellites to eventually exceed 100,000, mostly V4 or V5, with the system potentially carrying about half of global traffic. He also forwarded praise for Grok Bot, including a user saying it handled a flight rebooking and hotel and meal requests in 14 minutes, another saying it could join video meetings with low token use, and a claim that Grok Bot has the best retention among comparable products while ChatGPT Dots was paused over bugs and frequent refusals and Claude Duet had limited features; one forwarded view claimed Grok Bot combined with Opus/Fable and Argon has achieved AGI. 1 2 3 4 5
- NVIDIA showcased with partners how cities apply AI in public services, transportation and infrastructure at the Smart City Expo World Congress 2026 in Barcelona. 1
