Richard Ngo warns prioritizing directional correctness over literal accuracy confuses AI debate

Richard Ngo warns prioritizing directional correctness over literal accuracy confuses AI debate · Digg

Commentary on X

Thanks. I think there is often some kind of trade-off between "what do I, Nathan think is true" and "what makes enough sense to other poeple that they can check my work" and "what can be conveyed to others". In this sense, when I think about a question like "how much good has McKenzie Scott done, I want to have the right answer, but to me that's composed of trying to hear a number of answers and then compare them. I can forecast very well on specific timebound question, but I don't rate myself on messy ones. Also I want answers that are going to spread faithfully. Building stable equilibria of understanding seems important to me. In some sense, who cares what I think? So where does that leave us with QALYs. Well first, I guess yeah it's a fair point that perhaps they don't quantify societal change well. What would you suggest? How might we compare impact in a way that works? But also, I think there is something useful here in terms of trying to pin down Scott's impact and discuss it in *any* terms. I'm not claiming it's directionally correct. I am saying it's a discussion worth having. And I sort of do think that most discussions require a few false starts. What work would you like Max to have done? What is the right frame here? And in that sense, I don't claim to have got the first tweet right, but I sort of don't regret attempting to point to work or start a discussion. Over time through many failures, I have found that to draw me closer to answers I later endorse.

8:31 PM · Jul 14, 2026

Richard Ngo@RichardMCNgo

Yepp, it’s helpful here. And I’m sorry this was emotionally draining for you. Let me say some things about my own emotional state. I feel like I’m wandering around the world looking for something I’ll call sparks of reasonableness. I get excited when I see people saying stuff that isn’t just derivative of the standard worldviews they’ve imbibed. I have been excited in this way about some of your tweets in the past. I also have a sense of frustration that it’s so rare for people to put different pieces together. This might mean believing “wokeness is bad” and “MacKenzie’s donations promote wokeness” while overlooking the hypothesis “MacKenzie’s donations are bad”. Or it might mean believing “QALYs are a bad way to measure systemic changes” and also “systemic changes (like to how the world handles AI) are the most important type of impact” without propagating that to “QALYs are a bad way to measure impact”. You might then say that it’s not important to propagate these updates when simplifying for a general audience. But actually it’s very easy for “directionally correct” to make discourse more confused overall. So one threshold I use for “trying to make intellectual progress” is “how much do you care about saying literally correct things even when simplifying for a general audience?” Because without that one can do a huge amount of optimization for the political effect of one’s words, then retreat to “but it was directionally correct” when challenged. Whereas people who care about such correctness can more easily get over the antimemetic hump described below, where they selectively omit from attention hypotheses that would be politically inconvenient to say. Obviously there’s something that’s not working very well in my attempts to convey this, and in hindsight it’s a bit weird how frustrated I was in the tweets above. I think maybe I am blasting this energy across the timeline in a way that seems kinda random. Whereas probably the right move here is to write up a canonical explainer of what I’m doing and why, and link that instead of getting into protracted disagreements.

7:57 PM · Jul 14, 2026

Richard Ngo@RichardMCNgo

7:57 PM · Jul 14, 2026

Combined views

Commentary on X

Combined views

Sentiment

Positive Read

Digg Deeper

Commentary on X

Sentiment

Positive Read

Digg Deeper