Key Takeaways
- A Gemini-built PlanAssist tool is said to cut UK planning officials' assessment time by up to 58% and is expanding to 12 more local councils.
- LeCun and Balestriero report a mathematical result that independently trained JEPA models converge on similar world representations, with a claimed 42x sample-efficiency advantage over next-token prediction.
- Practitioners debate inference-effort settings on Opus 5.5, favoring low or medium effort with critique loops over extra-high reasoning.
- A VC-signal analysis of roughly 5,000 startups ranks "backed by a top fund" first and treats an MBA as a negative signal, with a revised version adding frontier-lab and cybersecurity backgrounds.
- Musk's reposts cover Starlink rescue cases, satellite share data, SpaceX revenue guidance, and claims about superintelligence and X engagement.
1. AI in Government and Research
- A repost from Demis Hassabis describes PlanAssist, built on Gemini, as helping UK government planning officials speed up building and renovation approvals by cutting the time officials spend assessing resident applications by up to 58%. The same repost says the tool is being rolled out to 12 additional local councils to accelerate decisions. These are claims carried in the reposted material, not independently verified results. — via 1
- Yann LeCun and Randall Balestriero are said to have mathematically shown that, under specific conditions, two independently trained JEPA models learn essentially the same underlying world representation — one from different images of a scene, the other from different descriptions of it, with no image-text paired training. The reposted material also claims the method needs 42x fewer training samples than next-token prediction for comparable image-text alignment, that a 109M-parameter text JEPA beats LLMs up to 72B parameters on a cross-modal alignment benchmark, and that training method matters more than scale alone. — via 1
2. Model Practice and Developer Signals
- Swyx argues that on Opus 5.5, low reasoning effort combined with a critique loop outperforms extra-high reasoning effort, and says this suits throughput-constrained settings, pointing to swyxio/skills/babysit-runs; a cited post says Julius convinced him to drop from extra high to high and considers it almost always better. He also notes Karpathy used Opus 5.5 at medium effort in the newly released Hackertube project and compares strong stances on effort levels to the tabs-versus-spaces debate, arguing people should tune to personal preference. — via 1 2
- Deedy's analysis of about 5,000 startups across roughly 60 attributes ranks "from a top fund" as the top trait most favored by leading VCs, with an MBA as a negative signal. After feedback, he reworked the seed/A-round analysis: former frontier lab and cybersecurity backgrounds moved into the top ranks, an IIT background became a positive signal, and education in China was a negative signal; the revision widened founding years to 2019–2026, roughly doubled the sample, and standardized and added signals. — via 1 2
- Ethan Mollick says the GPT Pro series stands out on one-shot problems involving complex calculation, citations, or complex concepts, but it is not fast and not pleasant to use, describing GPT-6 Pro as like asking one question of a great oracle. He recalls that under three years ago GPT-4 was the world's strongest model and that the answers to whether OpenAI had a unique recipe and whether GPT-4 was the ceiling for LLMs both turned out to be no. — via 1 2
3. Infrastructure, Satellites, and Platform Claims
- Musk reposts Starlink rescue accounts: a seriously ill crew member off New Zealand had a doctor guide emergency surgery over Starlink before a rescue helicopter transported the patient, who survived; in England, a hiker who fell near a steep wasp nest was rescued after a team with no cell signal connected via Starlink on a control vehicle and completed a stretcher rescue. He also reposts satellite-scale figures: as of October 31, 2026, of 16,848 active satellites in orbit, Starlink accounts for 11,184 (66%), versus OneWeb 651, Amazon Kuiper 391, Qianfan 257, and Iridium 80 — with Starlink's active count close to double all other operators combined. — via 1 2 3
- Musk reposts SpaceX guidance: the CFO expects Q1 2027 revenue of $33–37 billion, up from $14.6 billion this quarter; SI revenue of $7.56 billion, up 830% and above Starlink; 13 GW of compute by end-2027; Grok Bot paid weekly actives of 800,000, more than doubling in a month; Starship Flight 15 aiming for simultaneous recovery of both stages in mid-November; and Starlink in talks for 80,000+ US base-station sites. Separately, he reposts a claim that superintelligence will consume large numbers of entire businesses, industries, and professions within months to quarters, and that the physical world will need many humanoid robots. — via 1 2
- Musk reposts X data showing average time per visit in September 2026 at 13:25, ahead of Facebook at 10:15 and Instagram at 8:38; he says the goal is to maximize long-term no-regret user time and that people with brains are on X, and reposts a claim that X is the only place traditional media cannot censor information. — via 1 2 3
4. Agent Behavior and Content Experiments
- Mollick highlights an experiment in which two isolated copies of the same AI, with only a robot and a camera and no pre-agreed protocol, exchanged a secret code via skywriting, Morse code, semaphore, and a 5-bit binary alphabet. He calls it clever and evidence that isolating different agents is hard because they share a "brain" and readily recognize creative communication schemes invented by other AIs, and wonders whether the ability holds across model families. — via 1
- Mollick proposes funding projects to use AI to clear long-standing translation backlogs, especially where models can not only translate but also make reasonable initial passes at interpretation and meaning, citing large untranslated ancient archives such as the Oxyrhynchus Papyri. He also observes comment-section spam bots evolving quickly, with the latest trend being very cold jokes that seem to earn more likes, sometimes from real people, and expects worse to come under evolutionary pressure. — via 1 2
