[bayarearefugee]

> Aggressively learn AI.

Unless the author is talking about learning how transformer architectures work, and I don't think they are (and if they are, it won't help the vast majority of people anyway) this is the dumbest advice I keep seeing.

You don't have to "learn AI". "Learning AI" will not be a moat, for anyone. The power of "AI" is that you prompt it in plain language. And it just goes and does the thing. Using AI is not really a skill. It arguably was a little bit when the models were a lot dumber, but now it isn't.

This "transition" is going to be way worse and way more disruptive than even people who think they are thinking about this problem assume.

[simonw]

Prompting in plain English doesn't mean learning to use AI tools is easy. I continue to believe that "AI is easy" is the single biggest misconception in the entire field.

I've been a daily user of LLMs since ChatGPT came out and I'm still figuring out new ways to use them on a daily basis.

[Wowfunhappy]

If you wrote a blog post about this, I would be very interested in reading it.

I'm certainly aware of fun things I can do with local models, which takes setup, and if you're into e.g. ComfyUI those workflows can get very complicated. But, that's more a hobby—I don't actually think I get better results this way vs naively prompting a SoTA model.

There are some more advanced workflows for e.g. Claude Code, but I feel like all of that is likely to go away once the underlying models get better (for example, longer context windows mean less need to manage context).

[simonw]

I've started putting together a whole guide! https://simonwillison.net/guides/agentic-engineering-pattern...

[dlivingston]

"Learn AI" doesn't mean "learn how to prompt." It means that intelligence will be a commodity, and the business value of that will be the integration of intelligence into business models. So learn that. Where would a business line benefit from integration of AI? In what form would that take? The software industry has their answer in the form of Claude Code and code autocomplete. That's a design and integration problem. What's the equivalent for, say, energy companies? Or hospital administration?

[K0balt]

People are painfully ignorant about how this is going to be with robotics. There is an expectation based on current performance… understandable but fails to incorporate the why, which is grossly inadequate training data. When that is solved (and it soon will be) robots are going to have their GPT3.5 moment.