Feeling Smart
LLMs want to make me feel smart, but make me feel dumb. When I mentioned this to someone else the other day, they argued that I shouldn't be so insecure, which made me realise that the whole "feeling smart" thing is something I always assumed to be widely known, when it apparently isn't.
Let's start with semantics. I am firmly talking about feeling smart here, which is not necessarily the same as being smart (Hello, Mrs. Dunning and Kruger!), or as being perceived as being smart by others.
People generally want to feel smart, and not feel dumb, in that order, and if you want to stay on the good side of people, you should avoid making them feel dumb, and start making them feel smart - again, in that order.
This is something that is well-used in media. A well-written whodunnit leaves the reader feeling incredibly smart because they, too, could have figured it out themselves. Doesn't matter if this is true, though. A great adventure game rewards experimentation. It doesn't even need to further the plot - if you try something obscure and the game reacts to it - some remark, a non-standard game over - then it's a payoff that makes you feel good. It's also something pretty common in Sales - a good salesman wants you to feel smart about your decision to buy something.
Applying this rule - stop making people feel dumb, start making them feel smart - is simple and impactful. Let's say someone made a mistake - a co-worker, a direct, your kid during homework. You can point out that mistake in a way that makes the other person feel dumb, or you can point it out in a different way. In my experience, people usually learn more from their mistakes if they are not busy feeling dumb.
Here's an example: One of my kids is currently learning to write, and she has trouble staying within the lines - her letters are consistently too big and misplaced. Every time when I check her homework, I start by asking her something like "If Laura Letter, inventor of letters, would come in and look at your homework - would she weep with joy because it's so beautifully written?" Usually, my kid then snatches the homework out of my hand and revises it a bit.[1] When she hands it back to me I do point out a few things where Laura would frown upon (pointing out mistakes without making her feel dumb), but I also point out things where Laura would be happy (making her feel smart).
There are ways to do this wrong, though, which will lead us back to LLMs in a minute. You can, in an effort to not make someone feel dumb, fail to actually point out the mistake, or downplay the severity of the mistake.[2] Or you can seem insincere.
LLMs are fine-tuned to appeal to you (which can go wrong, as discussed here), and they, too, want to make you feel smart. Google's Gemini goes completely overboard with this and makes me feel dumb all the time to a point where, all other things being equal, I would simply use another LLM. Around 2-3 sentences into a conversation, it states something like "You've hit a common frustration/issue/blocker with [the topic we are discussing]!"
This is clearly designed to make me feel good, but this backfires. Sometimes, I ask an LLM something incredibly obviously stupid, simply because I forgot and Google's search has become unusable. In this case, I have not hit "a common frustration with cooking eggs", I just forgot some very basic thing and now I feel dumb about it. Or, I ask an LLM something extremely sophisticated, an edge case about some low-level stuff that I think only very few people have looked at before. In that case, being told that I have hit something people apparently wonder about all day makes me again feel dumb.
Zooming out, my mild annoyance about the LLM being tweaked in the wrong way seems unimportant. Why do I care what my tool pretends to think about me?
I don't. However, LLMs are currently forcibly inserted everywhere. Anthropic, Microsoft and Google want their LLMs to become everyone's day-to-day companion, including people who do not think of it as a tool, who are emotionally vulnerable and/or insecure. How these models are tweaked and the emotions they evoke (accidental or by design) deserve that we, as the society, keep an eye on it.