Fuck AI

Ohhh. I think we’re both defending different hills! I’m not against the use of generative AI for purposeful creation. What I’m against is the delegation of critical thinking.

It’s the difference between:

• “Implement this specific feature this specific way. Never disable type checking or relax type strictness, never solve a problem using trial and error, consult documentation first, don’t make assumptions and stop and ask for guidance if you’re unsure about anything”
• “Paint me a photorealistic depiction of a galaxy spinning around the wick of a candle”

(That last one is admittedly my own guilty contribution to the slop soup and favourite desktop background of at least a whole year)

Versus:

• “build me an e-shop”
• “draw me a cat”.

The difference is oversight and vision. The first two are asking AI to execute well-defined tasks with explicit parameters and rules, the first example in particular offers the LLM an out if it finds itself at an impasse.

The latter examples are asking a prediction engine to predict a vague concept. Don’t expect originality/innovation from something that was forcibly constrained to pick from a soup made of prior art then locked down, because that’s what gradient descent essentially does to the neural networks during training: reduce the error margin by restricting the possible solutions for any given problem to only what is possible within the training set, which is also known as plagiarism.

Edit: a slight elaboration on the last part:

Neural networks trained with gradient descent will do the absolute minimum to reach a solution. That’s the nature of the training process.

What this essentially means is that effort scales with prompt complexity! A simple/basic prompt begets you the most generic result possible. Because it allows the network to slide along the shortest path from the input token to a very predictable result.