October 31 2025
- X Platform, Algorithm Changes, and Grok
- New feature: "Disgraziad Corner"
- First recipients: Jason Calcanis and Pete Buttigieg for virtue signaling about past work at Amnesty International
- Algorithm (the "torrent" issue)
- Previously a bug hid posts from followed accounts
- Overcorrection added too much gain: any interaction produced a torrent of similar content
- Mostly fixed now
- Benefit: interest-based algorithm provides a 360-degree view on subjects
- Legacy code and deletion
- Gradually deleting legacy Twitter heuristics
- Removing one bug can reveal another that was previously covered
- Future updates
- Add a curated Following tab
- Grok will surface most interesting posts from friends/followed to cut through prolific posters
- Grok's scale and role
- Platform handles ~100 million posts per day
- Grok to read, understand, categorize, and match top 10 million posts to users daily
- Requires substantial compute, estimated on the order of 50,000 H100 servers
- New search
- Replace keyword search with semantic search to find matching text, images, or video via plain language questions
- Truth seeking on X
- Tap Grok on any post for analysis of truthfulness, context, and due diligence
- Grok acts as a "truth-seeking missile" alongside Community Notes
- Goal: X becomes the best source of truth by hearing all views and correcting with Notes + Grok
- New feature: "Disgraziad Corner"
- Grokipedia
- Definition: built with latest, maximally truth-seeking Grok
- Training and method
- Grok trained on critical thinking: true axioms, non-contradiction, valid conclusions
- Iterated over 1M top Wikipedia articles, researched broader internet, then added/modified/deleted based on accuracy
- Advantages over Wikipedia
- Intended to be more neutral and accurate
- More comprehensive information
- Fixes errors and counters propaganda where technically true facts mislead; even v0.1 gives richer descriptions of people, events, and technical subjects
- Future
- Add images and video
- Use Grok Imagine to generate explanatory videos from Grokipedia text
- Three Years of X (acquisition legacy)
- Acquisition context: takeover 3 years ago to save free speech and make truth exist again, fighting the "woke mind virus"
- Free speech reforms (three key issues)
- Banning: reinstated accounts of diverse figures like Trump, Jordan Peterson, Andrew Tate
- Shadow banning: exposed extensive deboosting by prior trust and safety group using elaborate tools
- Government collusion: Twitter Files showed FBI with ~80 agents submitting takedown requests
- Current policy: strictly follow laws of each country; going beyond law "puts a thumb on the scale"
- Global pushback: worldwide moves to suppress speech under "hate speech," notably Europe, UK, Australia; UK imprisonments for posts with reports of releasing violent criminals to make space
- Acquisition environment (anecdotes)
- Occupancy: one HQ empty; the other ~5% occupied
- Waste: subsidized lunch effectively ~$400 per person at low occupancy
- Bizarre practices: weekly menstrual products delivered to men's bathroom in unused building; millions paid for unused SAS software; paying for pedestrian traffic software in an empty building
- Culture: room of merch like "#Staywoke" shirts and "I am an engineer" buttons
- Political activists go where the power is
- They go to ISS and Glass Lewis - who influence half the power of the public market
- Fear he might be fired for political reasons after he helps create the army of robots
- AI Development and OpenAI lawsuit
- AI timeline: "supersonic tsunami" moving fast and massive
- OpenAI history: founded as open source nonprofit counterweight to Google's AI dominance due to safety concerns
- Lawsuit: corporate governance case expected to jury trial in February or March
- Incorporation docs: open source nonprofit, no founder financial benefit
- Shifted to "closed for maximum profit" AI; released broken models as a "fig leaf"
- AI compute efficiency
- Human benchmark: brain ~20 W total, ~10 W for higher functions → huge headroom vs current AI
- Tesla compute scale: 100M vehicle fleet × 1 kW per car = ~100 GW of inference compute potential
- Tesla Autonomy and Cybercab
- Current vehicles: Model 3/Y are "hyper intelligent" and capable of unsupervised full autonomy
- Cybercab: dedicated product with no steering wheel or pedals
- Production targeted for Q2 next year
- Volumes expected to reach millions per year over time
- Safety timeline: aim for cars operating without a safety monitor before year-end, likely December in Austin
- Fleet management learnings (Austin experiment)
- Scale number of cars: 500+ in Austin, 1,000+ in Bay Area by year-end
- Prevent car clustering at chargers or intersections
- Handle corner cases: illegal parking spots, complex regulations at airports
- Energy and climate change - Bill Gates memo
- “Billy G is not my lover” - Elon
- Scientific reality: CO2 rising 2–3 ppm per year; moving carbon from underground to surface cycles shifts atmosphere and oceans chemistry
- Concern horizon: serious concern over ~50 years
- Recommended policy: lean toward sustainable energy, especially solar and batteries; remove subsidies across industries, including long-standing oil and gas tax write-offs
- Solar magnitude
- Sun holds 99.8% of solar system mass
- China produces ~1.5 TW of solar panels annually; US steady-state power ~0.5 TW → capacity could power US grid in ~18 months with batteries
- Commercial panels ~25–26% efficiency
- Nuclear and fusion
- Nuclear fission generally very safe (US Navy uses it)
- Fusion feasible; larger tokamaks make net gain easier
- But fusion on Earth seen as a "fun science project" vs using the Sun, the giant free reactor
- Material feasibility: no material shortage for solar + batteries at Earth scale; silicon abundant; LFP cells use common elements like iron, phosphorus, carbon, lithium
- Analogy for Grok's impact
- Grok on X with Community Notes acts like a fact-checking immune system that analyzes incoming information and provides corrective data at the source
Comments