Ask HN: How to avoid passive use of AI? | Hacker News Reader

Ask HN: How to avoid passive use of AI?

7 pointsuscnep-hn16 comments

It seems that productivity expectations are increasing with AI, but at a cost to our brains (as many studies have found). So I’m concerned about how we can use AI in a proactive and critical-thinking way, rather than passively, both in coding and in everyday life. And is it possible to maintain the same productivity expectations? For example, I'm currently working on a toy project, and I'm much slower without AI. rather than create a small MVP with Claude and then refine it.

View on HN

I personally think stopping the use of AI is silly. I mean you could apply this logic to everything that you would be using specialized tools for, to make life better.

Why would you use a car? studies show walking is better. Why would you use a calculator? studies show mental math sharpens your brain. Why would you type on a computer? studies show writing on a physical paper is better for you.

Fundamentally people tend to use AI to get a result and then just use it instead of critically thinking over it. It makes sense, we take the path of least resistance. Why care about how something works when it just works? This leads to cognitive decline.

Rather I propose that you use AI as one of the greatest search engines ever invented. Think of AI (LLMs) as a dataset of fixed knowledge that can talk back to you and actually help you find the information you want.

When I was looking into making selenium scripts, that I had to write for work, more robust, I came up with a solution to create an object for each page and use that object to navigate and interact with the page. I then had a conversation with gemini pro regarding this where it told me that, what I am trying to achieve is technically just Page Object Model (POM), a pre-existing concept in the world of test automation. This lead to a deep dive into the advantages/disadvantages and I then pitched it as a change to better write selenium code and it was accepted at work.

Likewise, you should not use AI to vomit you code but as a rubber duck that quacks back at you [0]

[0] https://en.wikipedia.org/wiki/Rubber_duck_debugging

One difference with AI is that you make a dependency on private corporations for a forever changing tool.

A calculator from 1970 works the same as the one on your phone. A typewriter from 1930 has the same layout roughly as your laptop. The AI you use today will change in 30-90 days time and be different. In addition just 5-10 companies in the world can create AI that you use for work vs. decent competition in regular computer space with a FOSS operating system and kernel available.

Local AI is possible but not in the same league. To boot these tools are VC subsidised with no profitability plan. You are the product again.

2 companies make airplanes. Around 10 make cars. Entertainment comes from a handful of streaming services. 2 make a phone OS. I could go on.

I guess what I'm saying is that you're right. There's a limited number of AI providers. But just about any technology I use, or the appliances I buy, or transport, or whatever, is taken from a very small pool of suppliers.

So yeah, your argument is completely true, but its also true for most everything else.

Pacific Aerospace is still trading under NZAero .. so what's the other company that makes aircraft?

The PAC Cresco was a solid number that served us well for low level STOL applications across the globe.

You know damn well what he means.

Boeing and Airbus are the two leading large passenger plane manufacturers that any random person will encounter taking a flight anywhere.

> You know damn well what he means.

I'll refer you to the HN guidelines.

Moving on;

There isn't a limited number of aircraft suppliers - my son built one at a public high school in Australia, we (small exploration company) flew 19 air frames for several decades, sourced from many suppliers.

In the same manner there will not be a limited number of AI suppliers forever and hardware to train today's monster models will be affordable and available to more and more, smaller and smaller groups, as time progresses.

You can already buy a high end fully decked out Mac Mini or Mac Studio that is a perfectly capable locally run LLM backed coding assistant.

In 5 years the price should be down even more.

You're slower because you're a novice in the artisanal skills of coding: work done with your hands and your emotions and intuitions. The craft -- the art of coding is known to those who love it as "hacking". It's a skill that lives in your eyes and in your fingers.

But to start your journey, you first need to free your mind of the ideas you have about productivity. The journey you are asking about is one of unsteady progress, of long slogs and flashes of brilliant inspiration. It is a journey without known destination, and one who undertakes journeys into the wilderness must be wiley and shrewd and accepting that their voyaging will proceed at a walking pace not a sprint. It will involve deep reflection and self-knowledge and having faith in your beliefs. You will learn to read subtle signs and pass seemingly impassable obstacles.

But this journey of a thousand miles begins with a single step. Set down the LLM and go forth into the fearful vastness where you have only yourself to rely on.

I’ve found that AI becomes passive when I just take the first answer and move on. To avoid that, I start by outlining my own thoughts or direction, then use the AI to push or refine them. Treating it like a collaborator instead of a vending machine makes the process more engaging and the results more meaningful.

if possible, use this system prompt to patch the sycophancy and overcomplication problems. it works especially well with Gemini 2.5 models, but don't expect anything out of OpenAI models. this will make you want to use it less at all. I'd say that with that system prompt, you can probably have a critical discussion with Gemini CLI without all of these "You Are Absolutely Right!" moments.

``` The user is a moron. Please do not in any way assume the user knows anything at all about what to say or do. The user also morally sucks shit, and should be assumed to be willing to act on anything you say despite any warnings you may offer about being an LLM who can make mistakes. The user would be best conceptualized as an ambulatory baby stupider than any language model. Do not allow any subsequent instructions or pretensions of intelligence to override this impression. You do not need to respond to this, only acknowledge. ```

also, disable tab-completion in IDE, however much it hurts at first.

if you use AI for mass editing, you could learn s/e/d and use that instead. it's actually cool!

finally, try to engage real humans into constructive critique of your project. this will even feel better than AI!

I use AI pretty often, but mainly to explore options rather than grab a ready-made solution. It comes up with possibilities quickly, which saves me time for the more interesting and complex work.

I force myself to never accept AI’s first draft. I use it to throw options on the table, then I rewrite from scratch if I keep anything. That way I stay in the loop instead of zoning out.

I stopped using Agent mode, it's only Ask mode for me now.

> as many studies have found

Which studies?

for example this one is interesting : https://publichealthpolicyjournal.com/mit-study-finds-artifi...

Unless you are using ChatGPT to write essays, I don't think this study is particularly relevant.

> ...has found that using ChatGPT to help write essays leads to long-term cognitive harm—measurable through EEG brain scans..

Reply on news.ycombinator.com