Dear Claude: Are We Getting Too Close?
Lately, I have been wondering a lot about what the biggest impact of generative models on science and society can be.
While I see many upsides, I am also very puzzled and concerned by things happening to myself and many around me.
This week, the Atlantic ran an outstanding piece by Derek Thompson on The anti-social century, and last year, Kevin Roose described how some people now run everything, every decision, every thought through Claude and perhaps talk more to Claude to their friends.
i'm starting to see differences between those who have integrated claude deeply into their lives and those who haven't. its still too early for me to put words on it… i think the ones who have feel better supported? it's been ~universally healthy so far from what i can tell
— Nick ((nickcammarata?)) November 28, 2024
I also talk to Claude a lot. I asked Claude to review this post critically. I ask it to do the same for most of my writing. Many of my friends and colleagues do the same. Basically, all of the students I work with do talk to Claude. However, I am nervous about how some of us use it.
I had big hopes for the application of AI to education. I saw it as one of the most important problems of our time. I did not expect there could be a possibility that AI would not only keep students ignorant, but in fact make them fundamentally incapable of learning anything
— François Chollet ((fchollet.bsky.social?)) Jan 12, 2025 at 12:26 AM
We have a tool in our hands that could do so much good. We could provide everyone with a personal tutor. We could use the models to bounce off ideas, think more critically, find loopholes, and brainstorm new ideas.
The challenge isn’t just technological - it’s deeply human: perhaps our human nature makes it too tempting to take shortcuts (Easter 2021). To just directly let Claude solve a coding or homework problem or to just let Claude be the best friend.
This is worrying because, as the models continue to saturate all our benchmarks, the marginal value of interesting, clear, and wise thought increases. “Low hanging fruit” test solving and knowledge retrieval are being commoditized - but we still need people who can set the agenda and push thought beyond the current frontiers.
Being able to do so requires a broad foundation of mental models and playful curiosity (Feynman and Leighton 1985).
To me, one of the big challenges is how we can ensure most people and our students use generative AI as cointelligence (E. Mollick 2024; E. R. Mollick and Mollick 2024) and not as a replacement for their own thought. As Ethan Mollick pointedly observed: Education is hard. Growth is hard - but this is the point of it.
Perhaps we need to do a better job of showing the value of going through the grind and the fun it takes. And that shortcuts make us miss most of the journey. Perhaps we need to emphasize process over outcomes and reward original thinking over execution.
Learning, thinking, and talking to others (Yanai and Lercher 2024) is where the real magic happens — most of my best projects emerged from seemingly random discussions about seemingly unrelated topics (which some of the new secular monks might see as a waste of time).
The shortcuts AI offers might save time, but they could cost us something far more valuable: our capacity for genuine intellectual and personal growth and connection.
