Fourteen years ago, Scott Alexander wrote of the Whispering Earring. It's a cautionary tale of an earring that advises you, and to which you increasingly defer as you realize its suggestions are better than you'd do otherwise. You become very successful, but in the end there's nothing of you left. Ten days after Scott published the story, AlexNet won ImageNet 2012 by a huge margin, kicking off the deep learning revolution.
The earring whispers advice as "better for you if you...", and it's always right. While AI today would have seemed like distant science fiction in 2012, it's still far from infallible. Looking ahead, I expect AI to continue getting better quickly, but I don't expect it ever gets to "always right". A Whispering Earring dynamic with AI, however, doesn't need infallibility. Instead, the AI just needs to be capable enough that you can't improve on its advice. [1]
That's not where we are today. I use AI heavily, both in my professional and personal life, but we're working together. The AI makes mistakes and I catch them; I make mistakes and the AI catches them. It does much more of the work, but overall it's a brilliant—if uneven—collaborator.
I like this stage a lot, but it's clear it's not stopping here. AI is now speeding up the process of making AI (and I think we should slow down). But even without much recursive self-improvement, Fable 5.2 needs me much less than Opus 5.5, which already needed me much less than Opus 4.5 did. A year ago collaboration was like working with a very fast junior staff member who could take a well-scoped problem and provide something I needed to check carefully. Today the best models still do dumb things, but probably less often than I do.
When today's AI gets things wrong, or isn't able to provide much value, it's usually because it doesn't have enough context on what I'm trying to do and why. A lot of my effort to use AI more effectively recently has been trying to get more into it. A good analogy here is an executive assistant or a (in the tech company sense) chief of staff: they're supporting someone, and to do that well they would ideally know everything their principal knows. For me this has recently included moving my todo list from paper to digital, and using AI note-taking (with consent) even for in-person meetings.
This feels a lot like piercing my ear for a whispering earring. I'm not handing my thinking over to an AI, but if the capability increase over the next year is as large as we saw over the previous one, I'm not sure I'll be able to add value over the AI a year from now.
This is not something I look forward to. I find my job engaging, and I like solving problems. Collaborating with a brilliant partner is great, but I want to be engaged in the work and do my own thinking. Even more than that, however, I want my non-profit to achieve our mission: I don't want us all to die in a pandemic. Just like it would be wrong for me to let my ego get in the way of taking advice from an excellent staff member, once I stop being able to contribute I fear the right thing for me to do professionally is to accept the earring.
One thing that gives me hope is that, at least through 2027, there will be critical information my AI doesn't have. There's a lot in my head that I've never written down and haven't yet said in a transcribed meeting, internally we have a low bar for turning off note-taking, some external partners will be unconditionally opposed to notes, and there's too much exfiltration risk in giving information hazards to an AI chief of staff. I'll need to consider what it is that my AI doesn't know, and notice when it means we should act or prioritize differently. Initially this will keep me solidly engaged in my work, attentive to the places AI is imperfect, and alert to cases where it might be going off the rails. I'm also thankful that AIs are not yet very good writers: to be the writer in a collaborative relationship I have to engage deeply and really understand the material. But I expect these gaps will close, as I share more of my understanding of the world and models become better writers, and I fear that we're only perhaps a year from when I can no longer add enough value to remain engaged. [2]
The mythical earring is worn constantly, but in my personal life I'm not planning to defer to AI. Professionally I'm pushing on an important mission, and if fully deferring most advances the mission then that is what it is, but my personal life is for me and my loved ones. I want to be myself with my family, even if an AI-whispered husband and father would be better at saying and doing the right things. I enjoy collaborating with AI on blog posts, but handing over the whole process would miss the point. AI got better than me at Go years before it beat Lee Sedol, but that's not relevant to how my dad and I should play Go. I like playing music live, and even if pressing play would make the best music for dancing, that's not something I want to do and my dance community (doesn't want it either).
In the story, the wearers would live "an abnormally successful life, usually ending out as a rich and much-beloved pillar of the community with a large and happy family" but their "neocortexes had wasted away." In my work the mission is the point; outside work it's not. Being appreciated as an empty shell is not something I want, and to the extent I or others end up in this sort of place I expect it to be through the incredibly powerful pull of convenience.
The earring's first piece of advice to each wearer is "better for you if you take me off." When it comes to the good of the individual, I think the earring is correct that you should not let yourself fall into giving your thinking over to an advisor, no matter how aligned and competent. But if indeed we get to a point where in my professional life my intellectual work is no longer useful, I expect to choose to put the mission's good over my own and put the earring in.
[1] The story has the key line, "it's advice is always better than what the wearer would have come up with on her own." But a weaker earring that only had this property (even if "it's better for you if X" was at times incorrect), would still give the same dynamic.
[2] Another frame here that may be helpful as AI gets better is thinking of yourself as on a non-profit board. Holden Karnofsky's perspective is a clarifying one: the board has fundamental authority, but has also fully delegated operating the organization to the ED. The main job is evaluation, which relies on seeking out independent sources of information that let them assess whether the ED is doing their job well and push back when needed. But my impression is that, even with a human ED, boards tend not to engage to the extent necessary to do this well, and this would be even harder when you'd delegated something to an AI that not only is much closer to the work but is also much smarter than you. This also reminds me a lot of writing on gradual disempowerment.
Comment via: facebook, lesswrong, the EA Forum, mastodon, bluesky