Cal Newport

That's what it is.

I wrote about this talking to like a well-known business school professor years ago for my book, Deep Work.

And he talked about, he just realized, oh, being a business professor, publishing papers is about data access.

I have to spend most of my year talking to people, building relationships, trying to set up an agreement with a company where I can get good data that I can get three papers out of.

In all of that work, there's one day in there where you're crunching the numbers and making a plot.

And it's nicer if you could do a little bit faster, but it's not a productivity bottleneck.

It's a marginal efficiency.

I think there's a lot of that going on right now with AI and productivity as we look at what the AI can do and then try to make that thing into somehow being the key to getting things done.

Well, they tried.

I thought that was going to be... This is what I was excited about earlier in the Gen AI revolution.

I was like, okay, here's the real value prop.

Is natural language interfacing to advanced features on software?

Where I can just say...

all right, I want you to go take this column in the spreadsheet and get rid of all the rows that have values before this, and then I want to make a pie chart, because I don't want to learn how to do all that in Excel.

I don't know how to do that.

And they tried it.

I mean, this was Microsoft Copilot, but it turns out we...

underestimated the degree to which when we as humans are interacting with a chatbot that we're incredibly gracious, we're able to adjust and kind of get the gist of what it means and filter out the part of the chatbot response that's not really relevant or ask the follow-up question.

And when they tried to just use LLM responses to automate

actions within programs, it's just not accurate enough.