🧵 Announcing @open_phil's Technical AI Safety RFP! We're seeking proposals across 21 research areas to help make AI systems more trustworthy, rule-following, and aligned, even as they become more capable.

@MTabarrok Asking earnestly: why is this not true when it comes winning chess games?

Thanks to @jakemendel99 and @PeterFavaloro for developing these directions with me, and a big thank you to the many researchers who provided invaluable feedback on this RFP. 10/10

@jasoncrawford

@kamilelukosiute @bindshell_ That's convincing. Do you know if anyone has asked people who do phishing attacks why they don't use LLMs more?

@kamilelukosiute @bindshell_ In that case, should I read your paper as being addressed to the subset of people who are "framing successful task completion as direct evidence of risk", and not as a response/reaction to the usage of DC evals for RSPs/if-then-style policies?

I'm a bit lost. In your outlook, what's the point of having a red line if you expect guardrails to fail? Just in case this wasn't clear: in my tweet, I meant the term "safeguards" to include security measures, monitoring of internal and external usage to catch safety spec violations, heavy internal usage of highly capable models to patch vulns in the wider world, and plenty of other "safeguards" that aren't just safety-training.

This post from Steven Byrnes claims that established, traditional economic models/heuristics give mutually incompatible forecasts for what will happen with AI. Is this true? What economic models/heuristics seem most applicable to AI progress, what models are mostly commonly brought up by their critics, and to what extent do any pair of economic models lead to contradictory forecasts?

@dwarkesh_sp @EgeErdil2 @tamaybes Chess seems like an interesting case study for rhetoric about comparative advantage, Amdahl’s Law, etc. Humans are net negative to games of chess, and AIs can do all the work alone. What other domains will/won’t be like this?

@dwarkesh_sp @EgeErdil2 @tamaybes How can traditional economic metrics like GDP adapt to a world that includes relatively self-reliant AIs that engage in trade w/ each other and w/ human civilization? How should we expand defn of “final goods and services” for this world?

Pros and cons for various summary statistics used in discussions of explosive growth, eg GWP, growth, energy capture, energy capture growth, multiplier on speed of R&D progress (broadly or just AI R&D), other metrics that economists have invented Which of these actually reliably track the most important questions about AI impacts? Which devolve into debates about technicalities of forecast resolution? Ways to make these predictions of massive increase in output of intellectual labor more concrete? Eg maybe we should instead all be talking about how much progress we’ll make on highly desirable, intellectual-labor-bottlenecked tasks like healthspan extension, personalized art creation, math/science breakthroughs.