Former DeepMind and OpenAI researcher Richard Ngo argues that the AI safety movement’s core vocabulary — AGI, superintelligence, alignment — has become too vague to guide decisions, and that the biggest failure inside frontier labs is a refusal to analyze their own power and politics. On the MTS podcast, Ngo contended that fast recursive self-improvement is not the evidence-backed path to superintelligence, that capabilities will stay “jagged” and task-dependent, and that running a large company for six straight months remains a robustness test today’s models would likely fail. He also claimed Anthropic is constrained by “sacred truths” — chiefly “other people will do it anyway” and “we’re the good guys” — that prevent accurate assessment of rivals and its own influence, even as the company files for a Nasdaq IPO on the back of roughly $11.6 billion in quarterly revenue. For investors, the takeaway is that the next phase of AI progress may be slower and stranger than the hype, that conceptual clarity is undervalued research, and that lab governance and ideology are now material risks to price.