As someone who lectures university Stats majors: if AI is an amplifier and you get out what you put in, my worry is that students haven’t yet built the “self” worth amplifying. How do you think we should change our teaching so they develop judgement first, rather than outsourcing it before it forms?
Yes, this is a big question, and I've been out of the teaching game for so long I'm not sure I have much to add. But one thing I'd be considering is how much further you can get in a semester if you (e.g) assume a student can easily produce a shiny app for any given task. You could imagine shifting questions from "explain these cluster results" to "make an app that help you pick clusters. what tools did you use to convince yourself that you got the "best" results? and hey, actually what does "best" mean in this context".
> How do you think we should change our teaching so they develop judgement first
By practicing judgment every day, in every assignment. And, good news! Now that we no longer need to teach students how to write code, we have a lot more time to devote to teaching judgment.
> Would you like to see more of my talks in this form in the future?
Yes! All of them please! And it would be great if other Posit folks, like Jenny Bryan, did the same. I find it so much more pleasant and efficient to read a cleaned up talk like this than to watch the video.
The process description is the part I would want healthcare AI builders to copy.
The useful artifact is not just the final AI-assisted output. It is the chain of transcript, source material, intermediate diffs, human read, and the points where responsibility returned to the author.
For clinical workflow tools, that same trace could keep a summary or dashboard from becoming a black box: what source changed, what the model touched, what a human corrected, and what still should not be treated as ready for action.
Traceability through the entire data workflow is going to be such an important part of responsible AI use. There is work going on to develop some sort of standard for reporting in scientific publications - https://council.science/our-work/ai-disclosure-in-research/ - and to do that well will require structures and tools to support researchers in capturing that information flow.
Nice article, Hadley, thanks! (re: the format, the pictures are nice but I am perfectly happy with just the text too)
>for the stuff that you really care about, that’s important to you and your professional development, you can go deep on quality. You can use AI to pick apart your ideas, to pick apart your code, and produce high-quality work. Because at the end of the day, AI is an amplifier. And I really urge you, to the best of your abilities, to use it to amplify the best parts of you, the most human parts of you — not the laziest part.
I could not agree more! I have been telling myself (and others!) to "Use AI to not just make work easier, but to make work richer."
Wonderful insights for someone like me who doesn’t code. But would like to. Thanks for this and especially the intro on how you used Claude to converted a talk to this
Hey Hadley, really enjoyed this format! Would love to see more of your talks in this written form. Your work, and approach to data science and programming has greatly impacted my career, and my life. Probably one of the biggest "aha" moments in talks I've seen was your talk on Many Models, where you introduced functions, and purrr, and list columns, at WOMBAT 2016 (10 years ago!). I would love to see that talk distilled into a blog post like this.
So, I really am enjoying being able to read your thoughts on this new era of AI and LLMs. Thank you for sharing these.
I liked this format! It lets me scroll through your entire position and revisit easier than in video because I can tag towards the visuals. Almost like you made a notebook! Also appreciate you doing a longer form post. Your insights are most appreciated and short form content doesn’t let us be as greedy with the Hadley Wickham thoughts.
I want to express gratitude towards the genie being out of the bottle discussion. Sometimes I feel a bit guilty using these tools, but I try to use them to add value and not be wasteful (so I don’t use please and thank you in my prompts hehe).
Wow, I feel seen here! Using Claude Code CLI to prototype Shiny apps quickly is such a game changer for collaborative data exploration. Wish it was as easy to implement in a secure data environment as it is on your own local device
As someone who lectures university Stats majors: if AI is an amplifier and you get out what you put in, my worry is that students haven’t yet built the “self” worth amplifying. How do you think we should change our teaching so they develop judgement first, rather than outsourcing it before it forms?
Yes, this is a big question, and I've been out of the teaching game for so long I'm not sure I have much to add. But one thing I'd be considering is how much further you can get in a semester if you (e.g) assume a student can easily produce a shiny app for any given task. You could imagine shifting questions from "explain these cluster results" to "make an app that help you pick clusters. what tools did you use to convince yourself that you got the "best" results? and hey, actually what does "best" mean in this context".
> How do you think we should change our teaching so they develop judgement first
By practicing judgment every day, in every assignment. And, good news! Now that we no longer need to teach students how to write code, we have a lot more time to devote to teaching judgment.
> Would you like to see more of my talks in this form in the future?
Yes! All of them please! And it would be great if other Posit folks, like Jenny Bryan, did the same. I find it so much more pleasant and efficient to read a cleaned up talk like this than to watch the video.
Yes, please! Such a brilliant article - thank you, Hadley.
I'd like to report an error, I'm pretty sure the title of this talk is actually "lowkenuinely skills-maxxing fr before you get code-mogged by ai"
People have stopped responding to my genz-baiting posts so genz is now dead to me.
Time to return to millennial roots.
"TFW AI pwns your code game, lol rekt, git gud"
The process description is the part I would want healthcare AI builders to copy.
The useful artifact is not just the final AI-assisted output. It is the chain of transcript, source material, intermediate diffs, human read, and the points where responsibility returned to the author.
For clinical workflow tools, that same trace could keep a summary or dashboard from becoming a black box: what source changed, what the model touched, what a human corrected, and what still should not be treated as ready for action.
Traceability through the entire data workflow is going to be such an important part of responsible AI use. There is work going on to develop some sort of standard for reporting in scientific publications - https://council.science/our-work/ai-disclosure-in-research/ - and to do that well will require structures and tools to support researchers in capturing that information flow.
Nice article, Hadley, thanks! (re: the format, the pictures are nice but I am perfectly happy with just the text too)
>for the stuff that you really care about, that’s important to you and your professional development, you can go deep on quality. You can use AI to pick apart your ideas, to pick apart your code, and produce high-quality work. Because at the end of the day, AI is an amplifier. And I really urge you, to the best of your abilities, to use it to amplify the best parts of you, the most human parts of you — not the laziest part.
I could not agree more! I have been telling myself (and others!) to "Use AI to not just make work easier, but to make work richer."
Yes please more articles in this format! I enjoy reading or listening more than watching videos. Also your takes are always spot on.
Wonderful insights for someone like me who doesn’t code. But would like to. Thanks for this and especially the intro on how you used Claude to converted a talk to this
Amazing, insightful read!
Great summary and insights. Now off to learn more about parquet!
It gave me many valuable insights.
Reflects much of what I'm thinking about now for using AI in my research and coding. Yes, please, can we have some more?
Hey Hadley, really enjoyed this format! Would love to see more of your talks in this written form. Your work, and approach to data science and programming has greatly impacted my career, and my life. Probably one of the biggest "aha" moments in talks I've seen was your talk on Many Models, where you introduced functions, and purrr, and list columns, at WOMBAT 2016 (10 years ago!). I would love to see that talk distilled into a blog post like this.
So, I really am enjoying being able to read your thoughts on this new era of AI and LLMs. Thank you for sharing these.
Ok, I'll definitely do more of this in the future, especially when I don't have time to write a whole new post :)
I liked this format! It lets me scroll through your entire position and revisit easier than in video because I can tag towards the visuals. Almost like you made a notebook! Also appreciate you doing a longer form post. Your insights are most appreciated and short form content doesn’t let us be as greedy with the Hadley Wickham thoughts.
I want to express gratitude towards the genie being out of the bottle discussion. Sometimes I feel a bit guilty using these tools, but I try to use them to add value and not be wasteful (so I don’t use please and thank you in my prompts hehe).
Wow as a phd student. This is so helpful and great!
Wow, I feel seen here! Using Claude Code CLI to prototype Shiny apps quickly is such a game changer for collaborative data exploration. Wish it was as easy to implement in a secure data environment as it is on your own local device