At the very least, I do encourage people who are working on personal projects for explicitly learning and studying purposes to purposefully not use AI. Of course, if the purpose of the project is to learn AI, then that rule is allowed to be broken.
Anecdotal example but I know plenty of students (Including my own brother) who graduated recently who did it almost entirely with AI (unfortunately at the request of their own professors) who can't answer the most basic of questions that I would have expected of sometimes even first year CS students in the days before AI was a thing, which is why I encourage them to build projects without it.
Personally, I'm starting to think that in university and education systems, all projects should now be done in class and in computer labs/monitored study halls because we can't trust anyone to not use AI anymore.
I've been doing this on my latest personal project. What I've realised I missed the most is feeling in the zone or in a "flow" state as some call it. It really feels amazing. Coding agents take that away from you. I really think we need some coding agent implementation which takes less agency away from programmers and helps people be in the flow even more.
For my project I use Gemini/Google's AI mode to great effect. Quick answers to be reminded about syntax, and to get help when stuck. It actually enhances the flow state. Plus it's all free.
For my personal “project” I’ve been using chatgpt to teach me from math (linear algebra) to quantum computing - asked it to create a big list of prompts optimal for single threads, for me to learn with Socratic method.
I plan to write my notes as a blog for each thread in my own words.
I think this is a good use of AI for personal goals, it’s the best learning experience for me so far. I’ve never been able to focus on lecture videos
I got out of tech about 5 years ago, before AI really took off. I still do a lot of personal projects and now never use AI. I tried it a bit (only free plans) but noticed how much it caused me to no longer be able to think through a solution myself. I thought I was learning a lot from it but for some reason whatever I learned from it didn't stick. So now I'm no longer using it at all. I do realize that this is very specific to my situation of course.
There was a well known phenomenon people called ‘tutorial hell’ where when learning to program you could understand all the code from tutorials, but when it came time to actually create your own solutions for stuff you got stuck. I think AI is basically enabling eternal tutorial hell, you never actually need to break out of it anymore.
Difference between taking your own notes and skimming someone else's.
I have found in the last year I "get more done" but I learn less. Whats even worse is if I burn though my token quota I am less likely(and less able) to "switch to manual". LLMs make an ocean of decisions and assumptions very quickly and if you don't slow it down very intentionally (and lose the promise of the AI providers are pitching) you will quickly not have a good idea of how systems work.
It’s funny, I’m the exact opposite. I’ve tackled projects at scales I could never have done before, and learned more than I would have.
My actual coding ability is probably lower, but I’ve learned so much about music theory and production, CAD, and statistics. All because my hobby projects can go so much further than before.
I really do not get all of y'all claiming this causes you to "not be able to think".
I worked on six nines active active payments infra that handled billions in daily transaction volume. I've built laser projection systems to play interactive games on skyscrapers. That ability is not atrophying.
I'm able to do 10-20x what I was able to do before AI, and I'm able to work at a breadth, depth, and level of polish that was previously impossible without mastering other diciplines (CSS, design, etc).
With AI, I've built several video games, my own SRS software, my own productivity software, and a business making $5M ARR and growing 70% month on month.
I've now done more hardware projects in the past eight months than I have in the last ten years. I might even start doing physics and bio (I did my undergrad in biochem) if I can find the time.
I'm also back to making movies. I was a major hobbyist filmmaker before the pandemic. I've shot dozens of films. They're extremely logistically complex and take so much time, and I was never capable of working in the high fantasy or sci fi genres I wanted to. Seedance is scratching so many of my itches it isn't even funny. It's like a gift from the gods.
I'm working more now than ever before, and I'm turning those hours of writing pointless glue code or testing new libraries and frameworks into actual productive work. Getting real work done.
2010-2020 felt like tech dystopia to me. Cellphone incrementalism, walled gardens, adtech, crypto, locked down devices. It was a pale shadow of the excitement of the early web. Personal websites, kids making Flash games, the indie web, p2p, people building stuff being 70% of the internet. For the first time in a long time, I feel extremely hopeful about the future and excited to work with tech again.
tl;dr - I can still write in cursive, but it turns out I just don't need to as I write my movie.
I do not get you folks who claim this is bad. This is magical.
I believe you. I'm just telling my own personal experience. What works for you doesn't seem to work for me. Maybe I'm doing something wrong? But also: I'm purely doing things for my own pleasure now, so whatever I enjoy more is what I'm free to do. Again though, that's just me.
Hear hear. I do not understand the people who bemoan the loss of slide rules and books of trig tables in the era of graphing calculators.
I’m with you — I’m producing more and better art, with a deeper understanding of the art which comes from scale. My music is more innovative, my simulations are richer and more fun, I can wake up in the morning with a new idea and have a prototype in hours. It is magical.
I would actually really recommend people who are using coding agents full time to make AI brakes for your weekend projects. Think of it like a small break to take the time to reconnect to the code and keep your brain muscles intact, a bit like a professional cyclist going for a casual ride in the forest to experience what they don’t experience anymore in their professional life.
I see where you're coming from, but I can't help being reminded of my youth when teachers would admonish me "YoU wOn'T aLwaAYs haVe A CalCulator in yOuR poCKet!"
Now the entire world has a calculator in their pocket, and all that wrote memorization not only made me hate math as a child, it was a giant waste of time.
I suspect that one day we'll look back on advice like no-AI Fridays exactly as modern readers would react to "Punch-card only Tuesdays" in a modern enterprise.
> Now the entire world has a calculator in their pocket, and all that wrote memorization not only made me hate math as a child, it was a giant waste of time.
I dunno, I think this example kind of works in the opposite direction. If we're talking about memorising tables of logarithms or something, then yeah that was a complete waste of time. But a lot of the stuff kids would like to skip because calculators can do it is the absolute basics. IMO it is still important to have a solid grasp of arithmetic, useful to have enough trigonometry that you intuitively understand where it can be useful, and so on. (Much more so if you want to leave open the possibility of studying further mathematics.) And for people to whom these things don't come naturally, you really need to drill it a bit in order to develop the conceptual understanding.
I don't know exactly where this does and doesn't apply to programming/CS, though. I would like to think that a lot of the fundamentals will remain useful, but maybe that will only be true for a few specialised workers and self-motivated hobbyists.
How much rote memorization did it take to harm you? the addition of 1 digit numbers is kind of mandatory to function in everyday life, the multiplication table is at most (8*9)/2 36 facts.
More like: take some time without your phone. Something I assume everybody here would agree is a sane thing to do.
Not using a calculator as a kid for math is pretty important, because the goal is to develop intuition. Same for software, playing music, learning a language. You have to keep some level of friction to develop and keep your intuition
For me its quite the opposite. My employer is too afraid to commit to any of it. Meanwhile, I've built more stable software for myself and my personal computing needs of higher quality and in faster turnaround times.
I've also seen how insanely good at doing a total rewrite / modernization the models can be.
Oh man, I help host a monthly AI meetup called AI Friday: https://aifri.day
My heart sank when I saw this post. I thought we’d created enemies!
No conflict here, though.
Personal example... A few weeks ago, I realized that with all these agents running and all the things I could be working on, I was spending my days managing shallow pings instead of going deep.
I have AI agents managing SEO and ad spend. I used to spend a little time every day reading what each had to say. Now each gives me one detailed report per week, with links to the underlying reports and specific things I need to look at. I block 90 minutes for each one, on different days, and use that time to go deep. It’s helped a lot.
It takes an agent on the order of 100 billion operations to produce a single token. A person (generously) taking just a second for a single mental addition or multiplication would take tens of thousands of years to write just a single line of code.
And what am I supposed to do when I hit a wall solving a problem? What am I supposed to do, just sleep on it?
Sarcasm aside, I think there's a real phenomenon of AI removing the experience of getting stuck on a problem (where often I'll wake up the next day with a perfect solution in my head ready to implement). The same way social media removed the experience of "being bored".
When I was 9 or 10 I flew somewhere with my family for the first time. On landing, the pilot bounced the plane off the runway. While taxiing he came on the intercom and apologized for the "slightly rough" landing, explaining that they usually use <fragmented memory> some autopilot to land/smooth the landing and that to keep qualified they have to do it manually every so often.
So anyways, I fully embrace and support the idea of bouncing your code off the compiler every Friday. It's what I do and I can really tell those skills need practice even after just a week.
The technique I'm following is: If it's easier to explain it to an LLM, have the LLM do it. If it's easier to do it myself, I do it.
This is useful when iterating on UI, where I need to make a few minor changes; or when I need to "explore" to understand how a query or data model needs to work.
Otherwise, I don't need to dedicate a day to doing things by hand; I just do a few things by hand every day.
My old workplace did the opposite: we had All AI Fridays where you had carte blanche to build out AI tooling or just work on AI stuff as long as it's vaguely related to your job.
Giving yourself time to be idle and let a problem marinate is often where the breakthroughs and good ideas come from. If taste and human judgment are what separate useful AI output from slop, we need to give our brains time to cook instead of constantly bouncing between agents.
I don't know about Fridays but I like spending some significant amount of time full-vibing then some significant amount of time full-manual, coming to terms with what has been done-but-not-pushed, figuring out how I want the code to actually work and look, with roughly zero regard to how much transformation or starting-anew it will take from what has been vibed - easy come, easy go. Then once I really know what exactly I want, either transform the vibed code into it, or delete and start over (maybe manually, maybe prompting).
i inadvertently do this at work anyway. i burn through my monthly token budget on boring-but-valuable tasks, and then i spend the rest of the month doing stuff by hand.
Nobody should be in a position where they need to clear something like this with their boss. Now, I'm certain people are but that just means those of us with seniority and in leadership have failed those people.
We need No AI Hours everyday. We need this kind of time every day to clear our minds, reflect on what AI has brought us, and consider whether it's necessary to address the issue of technical debt.
100%. Making it a whole day makes it easier to sell and stick to, that's it. It's just brutal to have a whole day when you need to focus and not just mindlessly prompt.
Sounds good. They're encouraging others to join in but it'd be great to hear if this actually has an impact so far and how its worked out for HTMX. I imagine there are some interesting anecdotes about what actually happens when teams do this.
[Edit] just seen that OP is pitching this rather than practicing it. I wonder if any team actually do have a practice like this...
> Study after study shows that using LLMs can cause you to accumulate cognitive debt, make you less engaged with your work, negatively impact your critical thinking abilities, and hamper your skill formation.
False
I've learned more, I am more engaged, it has positively impacted my critical thinking abilities and has increased my skill formation
I find it interesting when people say, “I need time away from the AI to think.” I use AI for development but I would not say I’ve offloaded my thinking to AI. AI is great for pushing text but I haven’t found its system design skills to be that good.
To me the design has always felt like the thinking though.
I felt the same until I actually tried to stop for a week or so. And I could feel the constant temptation and anxiety of not having the coding agents as a sort of safety net. Taking a clean brake made me realize how much I was losing my skills and connection to the actual implementation, not just the high level pieces of it. I also didn’t realize at the time how much of my design thinking is done while working on code, and all the constant small decisions that are actually impactful
This is as silly as saying you shouldn’t drive your car on Friday’s and walk to work instead.
Why only Friday? What does “doing things the old way” one day a week prove or benefit when you spend 90% of your time working with AI?
Tech people who have their entire identities wrapped up in being smart or the best engineer in the room have suddenly learned that intelligence is now a commodity and a cheap one at that.
Objectively, it actually would be beneficial for most people to walk to work once a week…
Also, being intelligent is still important. Enabling the dumbest people in the room to dump their tidal wave of slop on others isn’t really a positive.
Great initiative. I was expecting something like this to happen. I think a good addition to the website can be success stories people share from this approach. Sending a link this to a manager is not very convincing, but managers respond very well to social proof.
"Not too much AI" or "Just enough AI" also, if you feel the need for a No AI Friday it could mean you also need to just use less AI, or try to only use it for the correct things so you don't lose visibility, for example don't let it make decisions for you, have it present the decisions.
Choosing from a set of decisions the agent presents you might feel like you are making a decision, but you’ll never make the one that wasn’t presented.
I’m not going to go that far (you do you), but reading the source code seems like a great way to remain in touch with the implementation. On my hobby project, I do a lot of reading code, asking the AI pointed questions about how things work, and asking it to clean up messes.
So you’re left with the least interesting tasks? Not trying to be too snarky, or too much of a luddite, but that would be a killer for me, motivationally.
This will be the pattern only the "privileged" workers will be able to afford.
Not using AI is similar to giving employees 20% of their time for any non-sprint work.
This privilege of No AI usage a day will lower the productivity, and rarely any companies will allow it (if they track your usage, which I am sure they do)
Anecdotal example but I know plenty of students (Including my own brother) who graduated recently who did it almost entirely with AI (unfortunately at the request of their own professors) who can't answer the most basic of questions that I would have expected of sometimes even first year CS students in the days before AI was a thing, which is why I encourage them to build projects without it.
Personally, I'm starting to think that in university and education systems, all projects should now be done in class and in computer labs/monitored study halls because we can't trust anyone to not use AI anymore.
For my project I use Gemini/Google's AI mode to great effect. Quick answers to be reminded about syntax, and to get help when stuck. It actually enhances the flow state. Plus it's all free.
I have found in the last year I "get more done" but I learn less. Whats even worse is if I burn though my token quota I am less likely(and less able) to "switch to manual". LLMs make an ocean of decisions and assumptions very quickly and if you don't slow it down very intentionally (and lose the promise of the AI providers are pitching) you will quickly not have a good idea of how systems work.
My actual coding ability is probably lower, but I’ve learned so much about music theory and production, CAD, and statistics. All because my hobby projects can go so much further than before.
I worked on six nines active active payments infra that handled billions in daily transaction volume. I've built laser projection systems to play interactive games on skyscrapers. That ability is not atrophying.
I'm able to do 10-20x what I was able to do before AI, and I'm able to work at a breadth, depth, and level of polish that was previously impossible without mastering other diciplines (CSS, design, etc).
With AI, I've built several video games, my own SRS software, my own productivity software, and a business making $5M ARR and growing 70% month on month.
I've now done more hardware projects in the past eight months than I have in the last ten years. I might even start doing physics and bio (I did my undergrad in biochem) if I can find the time.
I'm also back to making movies. I was a major hobbyist filmmaker before the pandemic. I've shot dozens of films. They're extremely logistically complex and take so much time, and I was never capable of working in the high fantasy or sci fi genres I wanted to. Seedance is scratching so many of my itches it isn't even funny. It's like a gift from the gods.
I'm working more now than ever before, and I'm turning those hours of writing pointless glue code or testing new libraries and frameworks into actual productive work. Getting real work done.
2010-2020 felt like tech dystopia to me. Cellphone incrementalism, walled gardens, adtech, crypto, locked down devices. It was a pale shadow of the excitement of the early web. Personal websites, kids making Flash games, the indie web, p2p, people building stuff being 70% of the internet. For the first time in a long time, I feel extremely hopeful about the future and excited to work with tech again.
tl;dr - I can still write in cursive, but it turns out I just don't need to as I write my movie.
I do not get you folks who claim this is bad. This is magical.
I’m with you — I’m producing more and better art, with a deeper understanding of the art which comes from scale. My music is more innovative, my simulations are richer and more fun, I can wake up in the morning with a new idea and have a prototype in hours. It is magical.
Now the entire world has a calculator in their pocket, and all that wrote memorization not only made me hate math as a child, it was a giant waste of time.
I suspect that one day we'll look back on advice like no-AI Fridays exactly as modern readers would react to "Punch-card only Tuesdays" in a modern enterprise.
I dunno, I think this example kind of works in the opposite direction. If we're talking about memorising tables of logarithms or something, then yeah that was a complete waste of time. But a lot of the stuff kids would like to skip because calculators can do it is the absolute basics. IMO it is still important to have a solid grasp of arithmetic, useful to have enough trigonometry that you intuitively understand where it can be useful, and so on. (Much more so if you want to leave open the possibility of studying further mathematics.) And for people to whom these things don't come naturally, you really need to drill it a bit in order to develop the conceptual understanding.
I don't know exactly where this does and doesn't apply to programming/CS, though. I would like to think that a lot of the fundamentals will remain useful, but maybe that will only be true for a few specialised workers and self-motivated hobbyists.
Not using a calculator as a kid for math is pretty important, because the goal is to develop intuition. Same for software, playing music, learning a language. You have to keep some level of friction to develop and keep your intuition
I've also seen how insanely good at doing a total rewrite / modernization the models can be.
My heart sank when I saw this post. I thought we’d created enemies!
No conflict here, though.
Personal example... A few weeks ago, I realized that with all these agents running and all the things I could be working on, I was spending my days managing shallow pings instead of going deep.
I have AI agents managing SEO and ad spend. I used to spend a little time every day reading what each had to say. Now each gives me one detailed report per week, with links to the underlying reports and specific things I need to look at. I block 90 minutes for each one, on different days, and use that time to go deep. It’s helped a lot.
How is this supposed to work? Like emulate an agent with my brain and write down the output?
It takes an agent on the order of 100 billion operations to produce a single token. A person (generously) taking just a second for a single mental addition or multiplication would take tens of thousands of years to write just a single line of code.
Sarcasm aside, I think there's a real phenomenon of AI removing the experience of getting stuck on a problem (where often I'll wake up the next day with a perfect solution in my head ready to implement). The same way social media removed the experience of "being bored".
So anyways, I fully embrace and support the idea of bouncing your code off the compiler every Friday. It's what I do and I can really tell those skills need practice even after just a week.
This is useful when iterating on UI, where I need to make a few minor changes; or when I need to "explore" to understand how a query or data model needs to work.
Otherwise, I don't need to dedicate a day to doing things by hand; I just do a few things by hand every day.
I also support No GenAI every day ending in a Y, but I support this, too.
Giving yourself time to be idle and let a problem marinate is often where the breakthroughs and good ideas come from. If taste and human judgment are what separate useful AI output from slop, we need to give our brains time to cook instead of constantly bouncing between agents.
I’d make a joke about there not being a word for the reverse of plagiarism, except there is one: lying.
TODO: insert something about open source, beer, free beer, someone else's beer, trading beer, subgenius, slack etc etc
[Edit] just seen that OP is pitching this rather than practicing it. I wonder if any team actually do have a practice like this...
And its four preprints and one published one with an extremely weak experiment.
But seriously who offload decision-making to LLMs? Thats the main thing they can't do proper.
> Study after study shows that using LLMs can cause you to accumulate cognitive debt, make you less engaged with your work, negatively impact your critical thinking abilities, and hamper your skill formation.
False
I've learned more, I am more engaged, it has positively impacted my critical thinking abilities and has increased my skill formation
For me it's AI 24/7 even on sundays
If it’s just your own datapoint, is it difficult to self assess?
To me the design has always felt like the thinking though.
Could be another case for four-day work weeks, especially if AI is supposedly improving productivity during those four days a week.
Why only Friday? What does “doing things the old way” one day a week prove or benefit when you spend 90% of your time working with AI?
Tech people who have their entire identities wrapped up in being smart or the best engineer in the room have suddenly learned that intelligence is now a commodity and a cheap one at that.
Also, being intelligent is still important. Enabling the dumbest people in the room to dump their tidal wave of slop on others isn’t really a positive.
Reading code is way less enjoyable than writing it
Not using AI is similar to giving employees 20% of their time for any non-sprint work.
This privilege of No AI usage a day will lower the productivity, and rarely any companies will allow it (if they track your usage, which I am sure they do)