> Many LLM maximalists say they use the tech to learn new things, but to what effect? Are you going to apply that knowledge of physics or computer science yourself, or will you just prompt the LLM again?

We had some contractors at the house doing some work. The contractors had been doing this work for a long time. I could tell they were good at what they did because I had some experience with that type of work when I was younger and they were way better than me.

We ran into an unusual situation (old house, old building techniques, overlapping old renovations) that they had never seen before. The contractors pulled up ChatGPT and started chatting with it. Eventually they came up with a solution with some good materials and techniques to make it work without turning it into a bigger tear out and complete rebuild project, and they got quick links to confirm it matched code.

This worked because they had the experience to judge the accuracy of the answers. They still checked the claims against the description of the materials we needed to order. After they did the work they had the experience to do it again, and it didn’t take them any guessing, trial and error, or searching through forums or Facebook groups (which are very popular with contractors) to get advice.

This is the right way to benefit from LLMs in learning something: Get the information then put it into practice so you learn it.

I worry when I see someone claim to learn something from an LLM but then the next turn is having the LLM do the work. It’s like skimming a math textbook but then skipping the exercises and never doing any quizzes. Every student in school learns that you have to practice to really understand something because reading the words is not enough. You have to work through it to internalize it.

LLMs can be very powerful tools in this way. They can also trick people into thinking they’re learning when they’re not. I think everyone has to go through that learning process about how to learn all over again like we all did with calculators in elementary school. There will be a lot of people who get trapped into being LLM promoters, whose output ceiling is limited by what LLMs can do as limited by their ability to recognize what’s right and wrong.

By now, almost everyone can think of someone they know who is happy to turn their brain off and press enter in Claude until it says the task is complete. I worry how this is going to intersect with the Reddit generation who grew up consuming doomer content about how jobs are just “bullshit jobs” and how your goal should be to do as little as possible at work, because there will come a time where those who spent their careers offloading their thinking to LLMs without doing the hard work of learning will not have much compelling reason to be hired over any other intern who can prompt Claude.

On the long term view I can see a lot of the LLM prompters getting replaced by LLMs doing the prompting. On social media like Reddit you can already see a lot of people gloating about how they just copy from Jira into Claude, push a PR, and then have Claude handle code review from their peers while they do nothing. Not hard to see that someone like that isn’t doing much that couldn’t be done by going one level deeper with Claude and letting it automate that person away. You have to put in some work to stay valuable.

Similar to your contractor example, I like to think of the useful way to use LLMs as doing "grep -C" (show lines surrounding the match).

You provide some context you are looking to expand on and let the LLM autocomplete provide you with relevant terms and concepts that have historically come up in that context and may or may not be actually useful. Then you do due diligence to learn about those related concepts, determine their applicability and distill it all down into a solution.