People vs AI processing
There are some fundamental differences between how most people think and how AI LLMs work.
In most conversations, and especially in narrative writing, people start out with a particular focus or context, but as the conversation progresses, the focus often narrows down from that larger focus. But that change of contextual scope is not something that needs to be explicitly mentioned, except perhaps for people with autism that can miss the context change because there are no obvious cues. Thus people can talk and the focus can narrow, broaden or pivot to something related, even conceptually by analogy or metaphor.
Conversely, LLMs are adding to the same context as the interaction proceeds, so some answers will include information that is no longer relevant to a person who has narrowed their focus. This can be frustrating for people to have to tell the LLM to ignore some aspect that it brought up.
But LLMs do progressively weigh more recent discussions as more relevant, degrading earlier ones, perhaps by paring them to salient points. This can also cause frustration for people who expect the larger context that was part of earlier discussions in the session to still be relevant and be able to be referred to at will.
These two LLM patterns are intrinsic to their mechanical operation, so they cannot be changed by programming, tagging or special prompting. This may help to understand the real limitations of LLMs, besides that they do not actually know whether the information they have is actually true, but may just be statistically consistent in their training data to be perceived as probably being true. Armed with this understanding, better but realistic use of LLMs can be had.
Computers raise the lower bar in any endeavour, making it easier to produce a better standard of results, and so a lot of what are called bullshit jobs can easily be done by LLMs. But it takes human creativity to raise the upper bar, providing new ways to perceive and take action. But we do not need to wait for other people or LLMs to do that for us, as we can use our own creativity and our hands to build something novel that someday may be implemented by LLMs for all.
Smashing galaxies together
△Any human creative endeavour involves understanding how multiple different aspects can be made to produce a coherent whole.
- a.HTML for page structures.
- b.CSS for presentation of those structures.
- c.JavaScript for activating those structures in browsers.
- d.PHP for server-side processing of pages.
- e.User interface design to make sure the pages are visually integrated.
- f.Accessibility to make the interface work smoothly for audiences of different capabilities.
- g.How all these interact.
Many use various server and JavaScript frameworks to bypass much of the integration work as their designers have thought through those issues, perhaps by much trial and error. However, while they may short-circuit a lot of development time, they impose their own design philosophy as well as introduce their own idiosyncrasies that provide further integration issues.
This all means that changes to any one of them can potentially require changes to many others, if not all. Perceiving the consequences of such changes across this stack is essential to efficiently building something elegant and workable. That means that a person needs to have a working integrated model of how it all works in their consciousness so they can see the effects of changes they contemplate making. They need to merge existing and alternative multilevel structures, and see where there will be synergies and mismatches. It is complex, and is what smashing galaxies together is hinting at.
But LLMs do not work that way. They focus on one aspect to resolution, and then move to the next. And it is that, along with their lack of context retention and compression of previous information, that makes seeing cross-level consequences difficult, let alone them being able to find optimisation opportunities. This is why AI website tools use frameworks or other pre-built components that people have designed to plug into their designs. Basically, they are relying upon other people having already done the integration work.
People can hold a complex model in their consciousness and come back to it over time, expanding what they may be focussed upon at the time, even to years. They can see what can be improved, not just one aspect at a time, but changes that involve many. LLMs are just not capable of such visualisations. And their not having the accumulated experience to know what may or may not work means that their designs may have major flaws that can lead to hidden technical debt that is difficult to root out.
Unfortunately, many people seem to think that AI tools mean that they can stop thinking and just rely upon what they produce, without reality checks as to whether that output is valid. In fact, AI tools require us to think more, and perhaps far more than necessary, because we have to have to ensure that their outputs are fit for purpose. Given the nature of LLMs, that fitness is unlikely.
LLMs are excellent for whiteboarding with, as they force us to justify our ideas to them against the conventions that they have been trained upon. For solo developers, they are a boon, though the sheer volume of irrelevant information they generate in the process can be overwhelming. This requires us to skim it for inconsistencies or just plain errors, making several pages of their output wasted while we quiz them on the lone points. Over all, this interactive scenario is where we can gather many valuable thinking and evaluative skills that we would otherwise bore other people with.