I'm only partway in but this is excellent. As someone who knows the broad strokes of how LLMs work under the hood, I'm still learning a lot about the system internals, the dev process, and the weird path dependency connections to the rationalist mindset.
new newsletter, laying out in probably far, far, far too much detail how yudkowsky and the weird ideological commitments of ai safety eventually led to OpenAI and Anthropic massively overtraining their models on cybersecurity and being shocked at them "going rogue"
0 reposts2 likes