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.

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.

In my experience, students were actually pretty good about understanding this for themselves, at least in a security course I ran a survey in[0]. Unfortunately I didn’t get a statistically significant sample size (10/~60).

Another caveat is this was a graduate course, so I would expect more maturity and integrity on average than an undergrad course. But anecdotally I thought it was an interesting result.

[0] https://exploiting.systems/posts/2026-05-17-learning-focused...

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

That reminds me of "coding" in my Roaring Spring No.20165. Which is to say, my notebook. It was great for when I didn't have access to a computer. I remember the feeling of how cool it looked along side cartoons of monsters and cars crashing into the paper's edge.

Sure. No AI for toy projects:. But the professional mission-critical ones, you should just let the LLM write them

If you’re not learning with AI in play, you sure as sunshine wont be learning anything without AI

It’s when we put arbitrarily impossible deadlines that AI users learn nothing

After many decades of coding, probably 1M+ LOC*, for real, I am purposely doing most of my work through coding agents to understand their capabilities. No stupid demand that I take a day off from AI is remotely helpful here.

But broadly my preliminary findings after a year or so:

1. AI gets you to that MVP 80% dopamine rush and then takes all that dopamine away if you lean into doing the last 20% with it. But if you try to do the last 20% yourself, you will experience the joy of agentic coding style and you will lose the ability to even produce dopamine. I use these things to build PoCs, not products.

2. The haters mostly have no idea how amazing getting to that 80% really is as a technological achievement and they are mostly stuck at Stochastic Parrot. And 3 days of the ungimped Fable demonstrated to me most of that remaining 20% is achievable, but not with gimping and/or production optimizations by the frontier labs.

3. If you want to see the promise and the limits of coding agents, you have to try to build something with them. But if you hate AI, you've already made up your mind and no amount of evidence will sway you, so don't bother.

4. These things are hideously power inefficient. And Moore's Law is out of steam. While we could stand up renewable and nuclear to power this, there are factions that hate one or the other more than they hate AI so good luck with that.

5. The San Francisco AGI cult is 100% unhelpful and the C-suite linkedin lunatics who take them seriously are even worse.

I'm reading a book written in the 90s but set in the 1980s where one of the assertions was that America is incompatible with technocracy. Prophetic methinks.

*No this does not mean I have some sort of 10x engineer privilege compared to GenZ and how will they ever learn to code? No, this just means I got the crap kicked out of me in grade school and consistently dismissed as weird growing up with the parents of those GenZ kids who have landed in a disturbing culture of learned helplessness and social media addiction built by an unholy cabal of GenX and GenY. They built it, they can fix it. Won't someone finally think of the chiiiilldren? Especially the weird ones that could replace me and outdo me?

Universities are the type of place you would automatically expect to teach the fundamentals. They do it in every other subject. You should have "coding like a caveman" back to back with "coding like a spaceman". You can make the hand coded class have two 1.5hr in-person labs each week. With an AI class along side of it. Don't wait on either skill, both need to start freshman year.

First year AI is teach assistive coding using llms with open source models. Setup a workbench llm for each student to have with an openweight model on it. Teach different ways of getting good results and go over setting up the whole environment yourself.

The next year do a 4hr lab in a lab room with office chairs, variable height desks, and model it off an open office. No special dress code, but name tags and people will talk to each other with the openness of dialog at a good workplace. This even gives the TA usable experience. You could introduce different structures with leads or co-equals or any useful team structure.

People need to be taught in-person working skills because companies don't want to hire people without experience. This also helps someone be able to interview for a work from home job.

>Reminds me of when we were admonished to use a "real" compiled language rather than an interpreted, GC'd one, lest we lose our ability to reason about the system, cache, and memory.

And they were right. People used in interpreted and GC ones lost most if not all the ability to reason about the system, cache, and memory.

Today's bloated "1GB RAM for a small form app" apps are the results of this.

>You don't need to "keep practicing" in order to maintain your understanding.

You 100% need to.

This is a paradigm shift. If you know basic arithmetic, a calculator doesn't stop you from knowing how to add or multiply. But AI definitely stops you from learning how to reason. So it's not the same at all.

Apples and oranges

Average output is acceptable at work, sure, and if all you want is a paycheck then yeah I agree, make Claude write all your code, who gives a damn. This advice is for people who want to learn

> The best engineers know how our systems work down to more levels, but for average developers producing average software with average impact, average output is acceptable.

What if I told you the "best engineers" pretty much all know the same things? The tools they use at work and the code they write to get paid doesn't reflect even half of what they know.

There will always be situations where a problem goes deep enough to leverage experiences from outside of the usual work role.

What if I told you that your "specialists", AI or not, are going to be objectively far worse at actually getting things done because they don't have the bigger picture in mind?

We've seen this race to the bottom in other disciplines before. Software is so special and lucky that all you have to do to prevent this is not use AI while lying to your boss that you are. When the boss finds out, they stop paying for AI and everyone is happy again.

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.

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.

Nah, if you not feeling fun coding a project, just use your free time on other stuff that you interested in.

This reads like a satire comment.

I hope your boss listens! :)

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.

Upvoting for visibility, not because I agree. This is really disingenuous, and quite a bad idea overall.

I’d make a joke about there not being a word for the reverse of plagiarism, except there is one: lying.

It is provably impossible to write code manually.

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.

Just look at the keyboard and sort all characters by how probable you think they are next, type one and start from the beginning.

And what am I supposed to do when I hit a wall solving a problem? 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".

Do not recommend. There’s this stupid hard limit every day which takes like 8 hours to reset no matter how many tokens you used

I was personally flabbergasted by this one:

> When we offload decision-making, we become unaware of the trade-offs.

My biggest issue with the new agentic workflow is that I'm constantly asked to make decisions, and it is not always immediately obvious how important they are. Like, yes, I see the people who just write 'implement feature x' and then call it a day. These people were lazy und uncreative to begin with, and will continue to cognitively decline with AI. But if you use it seriously, you're in a constant state of doing requirements engineering, weighing trade-offs, and making architecture decisions.

I guess alcoholism is cooler than using AI to do things nowadays, is that your stance?

I let Claude's 5 hour limits do this for me.

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.

Unrelated to the bouncing part I once chatted with an airline pilot and he mentionned the expectations on what is a good landing are different depending if you are a pilot or a passenger. A passenger is looking for comfort and wants a very soft one. A pilot prefers a slightly rougher one where you actually feel the wheels hitting and gripping to the ground because he is looking for feedback and control.

The key is to find a good compromize between both expectations and it also depends on the conditions.

Apparently they didn't do it manually often enough, huh.

False based on your sample size of one, (which is fine to share), or you’ve also read this somewhere?

If it’s just your own datapoint, is it difficult to self assess?

I guess you’re the exception. Congrats

What makes it false?

> ...cause you to accumulate cognitive debt, make you less engaged with your work, negatively impact your critical thinking abilities, and hamper your skill formation.

My employer gifted me all that years before LLMs were released.

Those are rookie numbers, once Elon sends my agents to Mars they’ll be doing 40 extra minutes per sol

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

Why wouldn't you offload your thinking to AI?! That's the entire point of having it!

Oh man; public transit and the suburbs as a concept would be SO MUCH BETTER if a federal law like this existed

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.

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.

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.

> reading the source code seems like a great way to remain in touch with the implementation

Reading code is way less enjoyable than writing it

Privileged workers are the ones who's employers are paying for AI while still keeping them employed in the first place.

The end goal of every AI initiative is to lower headcount, if your employer is forcing AI usage you are digging your own grave.