This is exactly the lens I use myself. I write AI software and I use it in my development process, but I try to use the AI to do things that don't remove my agency, but extend my capabilities: - Debug things. It knows way more than I do in many areas and it sees things I will miss. If I'm struggling to find the answer, maybe it will succeed. - Review things. It has a wealth of experience I couldn't possibly have. Ask it to critique my work and provide an alternative perspective that I can't provide myself - Implement a design. I have already gone through the thoughtful engineering to decide what to do and how to do it. The rest amounts to translating pseudocode to the programming language. Let it type what I would have typed anyway and save me the hassle of typos, looking up function and parameter names, and other such mechanical details. Let me use that time and mental effort to better consider my design, try more alternatives, or build more things, providing more value overall. - Suggest ideas. Even as a 20 year professional, there are things I don't know or haven't considered. Is there a newer, faster, or more maintainable way of doing this? Is what I wrote clear to anyone other than myself? Before AI, I would ask coworkers, search the web, or reference other sources. Now I can get an immediate suggestion from something with tremendous knowledge at almost no cost. It's full up to be to consider what it suggests, further explore the used and learn about them - I don't take the AI at its word and let it decide what is best for me. But I do use it to gain perspective and explore alternatives.

There are some common traits about the thighs I use AI for. They are this that I either couldn't possibly do myself (because I'm biased, or unfamiliar, or have no access to the expertise) or that I would spend a lot of time while having little agency (mechanical translation). I am not replacing learning, thinking, or deciding. I think this is the key difference.

I'm offloading tons of my thinking to AI, and I don't like what it's doing to me. I think that I'm having trouble creating things on my own due to it, and tbh it's making me increasingly uncomfortable. I'm not even sure what to do because quitting AI also seems crazy, but it all seems like a bad trend not unlike social media.

The right question to ask is if it is enabling us to do more interesting things or take on harder or bigger problems that seemed too daunting before. That's what all delegation of tasks (to other humans or machines) have enabled humans to do – scale.

Whatever creativity/thinking/effort bandwidth that's available will now get shifted to a different place in the problem-solving effort bottleneck.

That's the hallmark of any delegation being effective. Do we see that happening with AI tools? Personally, I do see that working for me. Is it as good as the hype makes it to be or I wish it to be? maybe not, yet, for me. But that's the case with most things in life.

I've been doing computer stuff all my life. I am learning the construction trades now. LLMs help you learn the rules for doing weird and specific stuff in the trades. One of my favorite LLM prompts is for plumbing tasks. For example: I have 1" pex b and I need to convert to a typical 1/4" RO connection line, what fittings on supplyhouse.com should I buy? So much easier than standing in Home Depot in the plumbing aisle for 25 minutes staring at the wall of fittings.

I think in the software trade you will definitely use your brain less. But in other trades, it removes the time sucks and gets you back to work.

This is something I see and feel everyday. And it’s very frustrating and annoying.

For example I send some doc asking for a feedback and someone without reading it generate a feedback with llm with so much ambiguity that I have to get back and wait couple of more days to get a reply.

One of the most silliest thing I see is a middle manager feeding Microsoft planner to Claude to generate a report and generate future steps and sometimes it makes no sense what he present couple of weeks ago because what he present today is contradicting to the one before.

At this point I feel it’s cheaper to replace them with AI. They are just physical vessels for AI.

It’s just not that maybe they were not good enough. But now they just fully depend on AI.

One major issue that the author ignores is that while we’re all having AI analyze our conversations or when we use it instead of search, there’s a chance it will provide an “answer” that is not correct, and literally drummed up out of thin air, and not even in the source material it’s “synthesizing”.

The article takes a position that assumes hallucinations do not occur, and then posits from that stance the question as to whether we rely on AI too much. We should be taking a step back before even asking that question and focusing on the part where AI does invent answers whole-cloth.

Information is seldom presented in a way that makes sense to me. That’s not quite right. Let me attempt to explain.

I want certain answers that the docs and the code are not giving me yet. Nothing is more irritating than working through a tutorial on a new framework and then throwing all that work out because that’s not really how one should use the framework. Nothing is more frustrating than having to get through a treatise on why this framework is The Solution before I can actually see code that uses the solution. And it’s beyond annoying when this End All Be All framework has a glaring omission that’s not obvious until you’ve built large amounts of your project on top of it.

Hand the docs and the example code to the LLM, and now I can get answers. “How can I do X?” Example code. “Then I need Z” Modified code. “How is this going to handle Q?” Explanation. “That doesn’t seem quite right. Give me a reference to the doc or code showing this.” Links.

Great, in 15 min, I have learned what I need to know, I can see that this solves a problem that I have, and I have discovered that I need an implementation of S to complement this solution.

That is usefulness. And it requires experience.

Our brains take time to ingest things and learn and using an AI tool to make decisions when taking in vast inputs of data requires then an overview of those decisions and understanding to some degree of care. If you don't take the time to go over these decisions and understand them and weigh them and course correct as needed, then you are offloading too much. You will become this approver buffer in between claude and nothing more if you don't have this methodical checking and understanding.

What is frightening is with something like neuralink that in a future hypothetical time would have very fast capability to keep informed and advised, you could be a zombie decision maker and nobody would really be able to tell. Even when you were pressed to why you made the decision, it's just another AI response, it's like a con artist or imposter dream scenario.

I noticed that atm, before these crazy hypotheticals potentially happen, the people that seem to take the time to understand things deeper are still way more valuable than those that just use tools more than not. Its obvious atm due to the lag in time and the way people respond in meetings, at least for now. :)

I believe its "too much" as soon as we trade procedural knowledge against declarative knowledge so much you can only remember the what and not be able to do the how anymore.

Knowing declarative you need to loop over elements and actually being able to write the for loop as procedural knowledge are two different shoes. I believe that this is the real danger.

Pilots have much automation in the cockpit but the pilot needs simulation hours actually flying not relying on the autopilot.

If you dont write code you will forget/loose much accuracy writing it, its just a matter of time.

Maybe, let me ask my coding agent what he thinks about this.