The Taste of Ideas
Philippe Beaudoin · September 26, 2026
Have you visited a foreign country recently? Have you been surprised by the food there? By the fact that everyone seems to like something that almost nobody eats where you come from?
It’s as if things tasted different to different people, no?
I’d like to suggest things actually taste different. I’d like to suggest that a taste is not something we’re born with, it’s something we learn. Your taste buds are able to capture the chemical signals from umami food, but unless you repeatedly practice to observe the flavor, the taste itself remains elusive to you. And unless you hang out with people who find sushi “yummy” you may never come to find it delicious yourself.
Tastes are literally acquired.
And, through the process of living together, people slowly align on the taste of food. They slowly align on what is yummy and what is yucky.
Now let’s jump from the realm of food to the realm of ideas. We’re used to thinking of ideas as things that are “good” or “bad”. We’d never tell someone they are wrong for not liking sushi, but we’ll do so if they pretend the Earth is flat.
In short, we believe tastes are something you develop while ideas are something you choose to hold or not. We see the process of adopting an idea as instantaneous: “Just believe the Earth is a sphere already!”
The fact that we see food and ideas differently dictates different interventions. If someone doesn’t like sushi we ask “Would you like to try?” and may start them off with a single bite of a mild tuna maki. If they make a disgusted face and spit it out, we laugh it off. Maybe they will try again later, maybe they won’t. But the world keeps turning and we move on.
For ideas it’s different. If a friend believes the Earth is flat we’ll be surprised. We probably won’t smile. We probably won’t encourage them to try a bite of a different idea. Instead, we’re likely to become mildly judgemental: “How can you believe something like that?” For strangers we may be even harsher and start using divisive labels. If you’ve been on social media recently, you know what I’m talking about. 😄
But what if we came to see that ideas have a taste too? That some of them taste yucky to Alice and yummy to Bob? That the process of changing one’s taste-for-an-idea required them to try it in a relaxed setting? To take a bite out of their own volition, not because someone forced them to? And — maybe the hardest move — what if we were patient when someone’s taste for an idea didn’t change right away?
Maybe, then, we’d realize that the slow evolution of our tastes is the best way for humanity to march towards a better future.
I’m not suggesting that all ideas are equally good.
It’s desirable for your kids to develop a distaste of poisonous mushrooms. In the same way, it’s useful for them to develop a distaste for the idea that the Earth is flat. But once we see ideas as subjectively yucky or yummy, it becomes easier for us to stay away from harsh judgements and divisive labels.
Let’s shift gears and consider this in the context of AI. Maybe you’ve heard of the alignment problem. In its overly-simplified version, it points at the difficulty of training an AI behaving in a way that leads to a positive future for humanity.
I’d like to suggest that the process through which populations slowly converge towards a set of shared tastes is precisely how humans solve the alignment problem together. When we solve human-human alignment well, we converge towards “yummy ideas” that help us evolve long into the future.
Said otherwise, solving the human-human alignment problem is about finding the best way to help people develop a taste for healthy mushrooms and a distaste for poisonous ones.
This observation suggests that solving human-AI alignment requires a different approach to AI development than the one we’re currently taking. Our belief that there are “objectievely good ideas” is leading to the industrialized AI training approach of frontier labs. It’s what justifies the breakneck speed at which we deploy new AIs. We do this because we genuinely believe each new LLM version is objectively better. That they hold objectively better ideas.
If we instead saw the march of humanity as a slow alignment process on the taste of ideas, we’d likely develop AI to be compatible with that. AI with “taste for ideas” evolving at a pace compatible with humanity. We may deploy a collection of diverse AIs, each with their own “tastes”, and each slowly aligning with the people they are in contact with.
We’d choose to slow down AI not because we’re afraid of the current speed, but because we realize that AI evolving at a human-compatible rhythm makes us genuinely happier. In the same way seeing your friend try their first tuna maki made you happy. In the same way their disgusted face led both of you to burst out in laughter.