At 'neolabs' every model is secret — no evidence at all on their safety properties
sjgadler · x · 2026-11-11
sjgadler observes that at OpenAI and Anthropic, the most powerful models are secret and used mainly internally — but released models still let outsiders observe safety traits. At various 'neolabs', however, every model is essentially secret, leaving basically no evidence about their safety properties. The quoted post from HumanHarlan provocatively suggests Safe Superintelligence Inc being a Theranos-style investment scam might be less unethical than secretly racing toward an uncontrollable 'digital god', since at least OpenAI/Anthropic can be watched by policymakers and the public.
More from AGI Musings
- Top forecaster Eli Lifland moves expert-level AI timeline from 2034 to 2028, citing METR report — birchlse · 2026-11-11
- Sayash Kapoor, Rishi Bommasani, Arvind Narayanan launch nonprofit STABL on AI societal bottlenecks — soumitrashukla9 · 2026-11-11
- Neuroscientist Doris Tsao unveils framework to crack consciousness and test if AI is conscious — soniajoseph_ · 2026-11-11
- Pope Leo's AI encyclical: don't let a few people decide AI's morality — an argument echoed by model abliteration advocates — willknight · 2026-11-11
- WaPo: RISJ and OpenAI reports show uneven global shift in AI usage and trust — _FelixSimon_ · 2026-11-11
- HBR: Employee AI Hesitation Is About Trust and Agency, Not More Oversight — rseroter · 2026-11-11