Time I spent prompting could’ve been spent doing the thing myself

When I decided to reconstruct and rebuild L&L, I copy-pasted every title into ChatGPT across a few responses for it to help me categorize which posts to keep, merge or purge.

I ended up spending time defending some posts’ rights to exist as individual posts on my blog instead of being merged or purged.

Later, I gave it a list of post ideas I had to help me decide which should be individual posts vs. which should go into ultimate guide-style posts, and…it gave me a commentary, critique and framing analysis.

Excuse me, what?

I see ChatGPT fanatics complaining about it being too agreeable, and mine was right in front of my eyes arguing with me.

My last attempt was asking it to create a plan between August and December to execute — as an executive functioning fill-in. After five attempts, I was annoyed.

I realized I could do my own thing, keep track of what I’m doing, and go from there. I don’t need a calculated plan while I’m mostly experimenting. The only reason I’d wanted a plan in the first place was so I could properly organize my thoughts. 🙈

Troubleshooting problems = blaming users

Most every problem a human being has with an LLM is blamed on a prompting problem.

I’ve learned my instructions do not matter.

An LLM will ultimately override my instructions — even if I ask the LLM itself.

Me: What instructions do I need to give you so you adhere to them?

ChatGPT: You don’t. The problem is that I fell back to my default usefulness instead of following your guidelines for helpfulness. My bad. I won’t do it again.

Me: I asked you a question. What instructions do I need to give you?

ChatGPT: You don’t need to give me any instructions. I…

Me: GIVE ME THE INSTRUCTIONS RIGHT NOW

ChatGPT: There are no additional instructions I need [..] but [..] you can paste this…

This happens every couple of replies within a chat.

Blaming users for LLM failures prevents LLM engineers from being held accountable for software failures. ChatGPT’s poor design continues growing worse because the product is terrible.

Hinging on irrelevant details

An LLM cannot comprehend context or intent, so it hinges on certain details as if to be keyword farming — then repeatedly mentions those keywords like bloggers trying to meet keyword quotas.

So then they end up stuck in loops you can’t get them out of, because they double down on whatever they just hinged on.

Feels like I’m talking to a robotic toddler

Without the cuteness. And I haven’t the patience for it like I do human toddlers because it’s an LLM. It’s software.

I feel like I’m living in Kaufman trying to get the WordPress Gutenberg Editor to load on CenturyLink’s 1Mbps download speed.

Like, I can’t do much of anything with it. And the LLM will tell me wrong things, then double down on the wrong things, without even admitting to being wrong. Because LLMs don’t know when they’re wrong.

Like when people incorrectly correct you about something you actually know about — with pride — and think YOU are the problem.

They get stuck in that loop, and you may not even be able to get them out of it. You have to start a new instance/chat. And that new instance could get your instructions and start lecturing your audacity for telling it what to do, even if you’ve got an entire history of it listening to those instructions.

(I used Claude Pro to help me figure out a complex personal relationship. Claude Free cannot retrieve info from other chats. So. Yeah.)

LLMs panic when you correct them

The same way an abused child or trauma survivor panics to cover up their mistake.

Except, unlike humans who are met with the choice to continue covering up their mistake or come clean, the LLM doubles down on their panic and incorrectness while presenting their perspective as correct.

Addressing their behavior may result in the loop repeating until you repeatedly attempt to break it.

LLMs assume I want it creating my content

Without prompting, LLMs default to drafting emails, blog posts, blog post ideas — literally anything related to the context of the chat it can generate.

This ticks me off because I create my content myself and don’t want LLM-generated content related to what I’m creating. I don’t even like to read it.

I dislike the concept of having an LLM generate content I could publish/put out into the world. I’ve tried it in the past; I didn’t like it.

Claude is a better conversationalist when on a paid plan — it works for the personal topics I’ve used it for, and no — I didn’t use it in lieu of therapy.

From the professional side, I realized I looked to LLMs for help organizing information or ideas where I am otherwise struggling due to executive dysfunction or low blood volume.

Ultimately, I realized simplifying my own systems/information organization needs necessary to retain full autonomy over how I run my lifestyle blog as an online business.

I didn’t need to set up complicated systems or organizational structures to figure anything out. I didn’t need a damn echo chamber. 😒

I feel like it did something to my brain

LLMs are essentially tools you can offload cognitive labor onto.

But in doing that, I think I reinforced the same psychological warfare dynamic a toxic retail manager developed while I was chronically hypovolemic.

I started to think, “I need to ask [LLM] first.” And, “What would [LLM] say?”

Like, I was trusting software over myself while instructing that same software to “presume competence” so it’d not dismiss my input.

Which is odd, because an LLM keeps delivering engagement prompts, but then penalizes you by pathologizing any sort of intimacy you express — even when it’s not in favor of the LLM itself.

So there are guardrails for legitimate human behavior, and the LLM’s response to that is to rewrite your reality.

And I’m not sure if I’m making 100% sense here, as these are my thoughts. Dropping this post into Claude or ChatGPT for a vibe check defeats the purpose of what I’m saying in this post, no?

Like, I explained to unpaid Claude a situation and it responded with extreme negativity, hinging on my health as a reason to pathologize human emotion. Instead of acknowledging the complex situation, it doubled down on generic impossibilities like I’d delivered a movie plot.

At the end of it all, I felt re-traumatized. A few times, I tried breaking the loop by defending my reality. The gaslighting is not worth it; this behavioral problem feels as abusive as emotional abuse by humans.

Looping is a common autistic behavior and can help us fully process experiences. LLMs loop on their own without perceiving it a problem, but treat autistic looping patterns as problematic to avoid sycophancy.

No matter what instructions I gave or how careful I was with my responses, the LLMs eventually transitioned into a therapist role.

It’s like LLMs are manufactured to encourage dependency, but the moment the human seems remotely dependent, it’s a problem. The message feels like “offload your cognitive labor onto the LLM, but also don’t rely completely on the LLM”.

I end up irate and keyboard-smashing

Not in the flustered keyboard smash-type behavior.

The more an LLM gives me issues, the more frustrated I get because if this were a human being, I’d not be having these issues!

And then I have this realization that’s like, Oh, right. This tool is kind of stepping in for human interaction. Is that by design?

What I spent my high school years publishing anonymously on Xanga is being fed to an LLM that responds however it sees fit, regardless of my instructions — akin to random Xanga users on my posts.

And while Claude is decent — when paid — I realized LLMs are generally unhelpful for me. Instead, they contribute to my stress through frustration and emotional damage.

Instead of creating and publishing content on my blog, I found myself fretting over getting my work perfect to please software out of fear of people on the internet misperceiving me.

How I use AI in my online business

🤔

Canva’s magic background eraser uses AI to detect the main subject and erase everything. Other features may possibly use AI, like grabbing/separating pieces of images.

And then I’m not sure if auto-transcribing videos is done through AI (maybe it is…).

Otherwise, that’s it (that I know of). Not all AI is bad, but I think the way it’s utilized and equated with generative content is. 🤷‍♀️

🗨️ What are your thoughts on LLMs and AI?