Gen "AI" disclosure

I feel it is necessary for me to define how I use the so-called Generative Artificial Intelligence so others know what to expect from my work.

Note

Hopefully no one here needs to read this, but what some refer to as "AI" is not Artificial Intelligence. Machine learning models such as LLMs or Stable Diffusion are blobs of data generated from content scraped off of the internet. They may be able to generate human-like output, but they are incapable of reason, and therefore mustn't be treated as "intelligent" in any way.

What the fuck is Gen AI

The term “AI” is so vague it’s outrageous. If someone told me they use “AI”, I wouldn’t know whether they mean they chat with LLMs, generate images, create music with vocaloids, or even just use upscaling or frame generation in video games. The term is a buzzword meant to confuse everyone, so let me first define what I consider to be “Generative AI”.

Generative AI is:

Generative AI isn’t:

In short: Everything machine-learning-related that can produce human-like or realistic output is Gen AI. Not a very precise definition for sure, but you get the idea.

My opinion on Gen AI

I am of the opinion that LLMs and some of the other machine learning model types that have cropped up since 2020 do have their genuine use cases. Period.

Having said that, I despise the AI companies for stealing copyrighted content, building their stupid little data centers, inflating hardware prices, and so much more. Which is why I refuse to pay for AI model subscriptions, and I do what I reasonably can to limit the amount of data I share with the AI corpos.

The technology itself is fine. How it’s produced and utilized is what matters.

How I use Gen AI

I use LLMs as assistance with coding. I never outsource my thinking to them, and I never “vibe code” software meant to be used by others.

I may let them complete trivial or repetitive tasks so I can use my (limited) time on something much more productive. For instance, I use LLMs for writing large, repetitive chunks of code (some trait implementations in Rust) and creating big data structures from JSON1.

In the future I might also use them for reviewing my code and finding bugs - tasks LLMs don’t completely suck at and where the code isn’t directly altered by them.

I don’t want AI nowhere near my creative work - I will never intentionally2 include AI-generated content in my work unless it is necessary to do so for the context.

How I run my models

My hardware isn’t powerful enough to run any useful models at a reasonable speed, so I primarily use free online tools offering LLM chats such as duck.ai3. I avoid services openly stating that they retain chat logs off-device. But, like with any kind of cloud processing, I kinda just have to trust them on that. Which is less than ideal.

If I do decide to use a local LLM for something specific, I run it in the Alpaca app on my laptop.

  1. To me honest, there’s a bigger chance that I would mess up transforming JSON than an LLM ↩

  2. I may include AI-generated media in some of my work if I don’t realize that it is AI-generated (in which case please point it out and I will gladly remove it!) ↩

  3. At the time of writing this, in their “Chat Protection” menu they state:

    All model providers are prevented from training their AI on your conversations.

    and

    This model provider does not store any data associated with this chat. Other model providers follow a limited data retention model.

    For Anthropic models, the above text changes to:

    This model provider deletes data associated with this chat within 30 days. Other model providers follow a zero data retention model.

    Which is why I favour duck.ai over other LLM chat providers. ↩