Writing - Blog - Current Opinions on AI Generation

basedir=/Writing/Blog<br>--->

My current opinions on AI Generation #

The MIT Paper #

Firstly I'll talk briefly on a paper that I will be referencing quite a bit here. There is an MIT Paper[1] which showed that use of AI tends to show up with homogeneous works, with a lack of memory for the work and with a lack of ownership of the work. Additionally it showed lower brain activity when using AI generation to make essays. The results of this paper will be referenced a few times throughout.

Before getting into any of this, I will also reiterate Richard Feynman's quote:

"The first principle is that you must not fool yourself - and you are the easiest person to fool"

With AI, you can fool yourself even better due to how sycophantic it all is...

AI Generated "Art" #

My opinion on any artistic field where there is not a discrete correct or incorrect answer, is that AI Generation should not be used. One of the mixed qualities of AI is the very homogeneous output it produces. This means any writing by AI will sound the same(and we all know many of the patterns these days. "It's not X it's Y", a lot of emoji use, rule of 3 bullet points, and just generally ridiculous flowery language that sounds fairly sycophantic). This also means that images produced by AI will all look samey, and in fact many of them do from obviously flawed understanding of anatomy that is so flawed even beginners don't make that sort of mistake, to a failed understanding of lighting and shading and details that don't make sense(Not to be confused with actual drawing techniques where you draw the effect of detail rather than the detail itself).

Generally speaking I think AI is absolutely garbage for Art as many people value the fact it was written by a real person. I also value this and it adds a certain... Je ne sais quoi to it. In spite of this AI Generated poetry is rated more favorably than human-written poetry and struggled to distinguish the poems of the great Poets of history from AI generated poetry.[2]

In this regard, I generally have a "well fuck you then" attitude to AI generated art as it completely disrespects the time of the audience.

AI Generated code #

My opinions on AI Generated code are a little bit warmer, but still relatively cold. It shifts the problem from writing the code to reviewing it and ensuring the correctness of it- which for things that will never be in a production environment is fine. For things that will be in a production environment, your code had better be correct. AI in its current state cannot verify the correctness of the code it produces. It additionally cannot know the most appropriate way to architect a solution that considers your long term future needs and extensions. In this way I think AI generated code is comparable to working with a Junior Programmer, but it is still on you to ensure the correctness of the code.

Three other issues, is a lack of ownership of the code(in general everything with AI has a lack of ownership) and a lack of memory or experience. This latter one is a little trickier to explain, but if you don't have a memory or experience of that codebase it can be hard to track down and think through what could be the issue. In this regard, I am reminded of an anecdote told in Masters of Doom where John Carmack had a bug where the game state would be incorrect after a long time of running the game, so stopped, thought for 15 minutes and then realised where the bug would be. This kind of thing is only possible when you are intimately familiar with the codebase you're working on.

The final issue I leave for last as it's the worst issue. It puts Junior Programmers in a competition against LLMs so in many ways it is short-sighted and is putting their jobs at (some... not all because they can still learn) risk. This also means many Junior Programmers will resort to using LLMs which combined with a lack of experience in architecting software and the fact that AI-generated code does not help you build skills in programming means I will expect to see a multimodal distribution(basically 2 normal distribution, where there's pre-AI Programmers who have better skills, and post-AI programmers who have stunted skills).

I will also leave one more issue with AI generated code. It completely fails for anything that requires a huge amount of context, which for many things... especially many stateful things, and doubly especially many brownfield software projects, means it is effectively unusable. On isolated scripts that will not be depended on it is still fairly useful.

In general I think this is one area AI is... ok if not very good in, but it's not very good in the same way a Junior programmer is not very good. I think a lot of this comes down to software development being a fairly homogeneous field where there's often only a handful of good approaches to a problem and correctness can be verified. Other industries that are similar about...