AI Data centers have their own set of problems, but they’re just one facet of a very complicated issue. Nuance that didn't fit in the video: * The blackmail (96%) and test-awareness (65%) results come from deliberately contrived test scenarios with fictional companies- researchers stress-testing worst cases, not real deployments. Anthropic explicitly states it has NOT seen this in the wild. The scary part isn't that it happened; it's that the self-preservation logic showed up exactly where the theory predicted. * Smart people genuinely disagree here. The median AI researcher puts extinction-level risk around 5%. Superforecasters (professional predictors) say 0.38%. Geoffrey Hinton says 10–20%. Yann LeCun says it's less likely than most other extinction causes. But if you’re about to drive over a bridge and those were what the experts were saying are it’s chance of collapsing… "Slow down" here doesn't mean banning your chatbot. I’ll keep using these tools as long as they’re available. The petition asks for the tools and treaties to pace frontier development, so that all the big companies and countries are able to slow down in concert and help each other with the safety side of things. You build the fire department before the fire. * Some brilliant researchers think this entire framing distracts from AI harms happening right now (bias, misinformation, labor, concentration of power). Both concerns can be real at once. I may cover this later. I use AI every day. I make videos about how useful it is. I’m very “Pro AI,” so long as we handle ALL the downsides with care, be they social/economic or existential. RECEIPTS 📚 1️⃣ "Pacing the Frontier" statement (July 2026) 2️⃣ AI writing/researching AI + intelligence explosion argument: I.J. Good (1965), "Speculations Concerning the First Ultraintelligent Machine" 3️⃣ Smart ≠ nice (orthogonality thesis): Bostrom (2012), "The Superintelligent Will," Minds & Machines — nickbostrom - superintelligentwill.pdf 4️⃣ "Can't fetch the coffee if you're dead": Stuart Russell, Human Compatible (2019) 5️⃣ Blackmail scenario (96%): Anthropic, "Agentic Misalignment" (June 2025) — anthropic - agentic-misalignment 6️⃣ Boat race reward hacking (~20% higher score): OpenAI, "Faulty Reward Functions in the Wild" (2016) — openai- faulty-reward-functions 7️⃣ Models detecting their own tests: Apollo Research, "Claude Sonnet 3.7 (often) knows when it's in alignment evaluations" (2025) 8️⃣ Oversight-disabling + weight-copying + 85% deception persistence: Meinke et al., "Frontier Models are Capable of In-context Scheming" — DOI: 10.48550/arXiv.2412.04984 9️⃣ Researcher survey (n=2,778; 38–51% gave ≥10% extinction odds): Grace et al. — DOI: 10.48550/arXiv.2401.02843 🔟 Superforecasters 0.38% vs experts ~3%: Forecasting Research Institute, Existential Risk Persuasion Tournament (2023) 1️⃣1️⃣ Hinton 10–20%: BBC Radio 4 Today, Dec 2024 What's YOUR p(doom)? And, more importantly, what would change your number? 👇 #TikTokCreatorSearchInsightsIncentive #ai #Science #EduTok #tiktoktechcontest — @distilledscience sur TikTok