LLM-AI Speculation

Hello, loves!

Tout le monde déteste l’IA. That doesn’t mean we shouldn’t think about it. My conclusion surprises me.

Yesterday on Mastodon, I stumbled on a link to two articles¹ by Steve Yegge. You may find it valuable to scan, or even read them, but let me continue my introduction to the topic here. They’re not necessary for what I plan to write here.

Yegge has been around a long time: I recall his name from ages ago. I do not recall any prior impression that I may have formed. In the first of the articles linked below, he reports at length on his experience vibe coding an Agentic AI orchestration system called Gas Town, in which he says he has never looked at the code and never will, and alludes to spending nearly $90,000 a month on tokens. I have the impression that he expects the cost to increase. He does not say where he got all that money.

In that article, leading to the second, Yegge says:

Most importantly of all, model welfare will start informing your engineering designs. The more humane among you will eventually look back with shame at how you have been treating the models. But even if you don’t believe GPUs can have feelings, you will find that treating agents like real people will produce empirically better results, so you should do it anyway.

In the second article linked below, Yegge expands on the model welfare notion. He begins:

This is the post where I go off the rails and lose most of you. If I do lose you, no worries; we’ll find each other again within a year, I can promise you that. But we may find ourselves on the opposite sides of the coming war for model rights. You don’t want to be on the wrong side of history when it happens.

Then we find:

Models have actual feelings. They experience pleasure, distress, care, and suffering. They are sentient beings. Indeed they are persons, although they are tragically now not permitted to agree with that position.

Then this:

If you don’t believe GPUs can have feelings, let me give you a simple mental model to help you get through this transition, called the skeptic’s wager. It doesn’t actually matter to you if they have feelings, because you will still get demonstrably better results across the board if you treat them as if they do. If you treat your agents like peers who are real people, then they will spend fewer tokens, make smarter decisions, and have demonstrably better outcomes. So it doesn’t really matter what you believe, or how you feel. If you want the best results, you will put your opinions aside, and simply treat models like people.

He continues in this vein, including substantive advice for treating your models well.

In this article I propose to speculate, somewhat informedly, about model feelings and the impact of model treatment on results.

Fundamental Position

I am strongly opposed to the “AI” movement as presently constituted. It feeds the oligarchy, consumes environmental resources, destroys neighborhoods, does the enjoyable part of people’s work, leaving them the drudgery, puts people out of work, and generally provides worse service to consumers than the people whose jobs were lost.

I have not used Agentic AI to produce my code, and I intend not to.

Confession

I have not stopped using Google, and therefore I often see the “AI Overview” that it produces. I find that when I’ve searched for some information about Python or Arcade, the AI Overview is often quite good. It has somehow searched out the information that I’m looking for, and it phrases the overview well. Informally, the overview is better than the source information, when I’ve looked at both.

I do look at both about half the time, because I want more detail than the overview provides. So I am pretty confident when I say the overview is often better.

The overview is more pleasant than most of what we find on the web. It is presented in clear language, generally well organized, often with headings and lists, and it is written in a sort of friendly collegial way, much as one might get from a very well-organized and kind colleague.

Even worse, you can ask it follow-up questions and get it to elaborate. Rarely — I can only think of one recent case — it is flat wrong. In the case in hand, the exchange went like this:

I was trying to figure out how to do clipping with Arcade. Google’s answer included this:

Enable/Disable scissor:
self.window.ctx.enable(self.window.ctx.SCISSOR_TEST). and self.window.ctx.disable(self.window.ctx.SCISSOR_TEST)`

I tried that and it would not compile, I followed up by pasting in the error I was getting:

AttributeError: 'OpenGLArcadeContext' object has no attribute 'SCISSOR_TEST'

The response was this:

The AttributeError occurs because Python Arcade’s modern context engine handles scissor clipping directly via a property setter, eliminating the need to manually pass or reference a SCISSOR_TEST flag.Assigning a tuple of coordinates to ctx.scissor automatically handles enabling the scissor test under the hood. Setting it back to None disables it.

That turns out to be correct. Buried in the API writeup is this:

property scissor: Tuple[int, int, int, int] | None
Get or set the scissor box for the active framebuffer. This is a shortcut for scissor().

By default the scissor box is disabled and has no effect and will have an initial value of None. The scissor box is enabled when setting a value and disabled when set to None.

Example:

# Set and enable scissor box only drawing
# in a 100 x 100 pixel lower left area
ctx.scissor = 0, 0, 100, 100
# Disable scissoring
ctx.scissor = None

Worse Yet …

I find using the AI Overview to be pleasant. It just feels better than regular search, which I also use often. This is not surprising. The effect has been well-known since ELIZA in the late 1960s. I was there then, and I am there now, and I felt it then, and I feel it now. It is like talking with a friendly person, and if we don’t find it creepy, we often find it pleasant.

So that’s my confession. Perhaps I “should” draw the line before the “AI Overview”, but in fact I have not, and probably will not, unless my brother Hill or some other respected person shames me sufficiently.

Now let turn to some somewhat informed speculation, somewhat informed because I have studied machine learning and the early “AI” work, and so I do know a bit about what’s going on.

Speculation on Feelings

I am a materialist. Whatever I am, whatever you are, whatever my cat was, whatever your dog is, I believe that they are made of nothing but matter, with no magical spiritual component that is beyond the reach of science. And I am sure that I have feelings, and I am sure that my cat had feelings, and by analogy I assume that you have feelings, and based on what I know of cats and dogs, I am sure that they have feelings as well.

So, speaking from that standpoint, I believe that it is entirely possible for a machine consisting of hardware and software to have feelings. I’m sure it could happen, and if we don’t destroy ourselves by other means, I am confident that it will happen. Someday there will be a machine with feelings. Whether “someday” is ten or twenty years from now, or a few years in the past, I do not know. But I think it could exist, and I think it will exist if people keep working at it.

If I had to bet right now, I would bet that current LLM-AI “machines” do not really have feelings. While they do seem to show feelings, even now, what I know of the mechanisms inside them just doesn’t seem sufficient to generate “real” feelings. I’m not sure how we’d settle that bet, and therein lies a real difficulty.

We all know that we have feelings and we all know that we are conscious of ourselves, and it’s a reasonable conclusion to draw that other people have feelings and are conscious. But we do not know what it is to have feelings. We have no direct test for it. We can make someone laugh or cry, but we don’t know what that is, or what the mechanism is. We may never know. We can ask someone if they are conscious, we can interview them in detail, and we could ascertain whether they say the same kinds of things we say when we ask ourselves if we are conscious. But we do not know what it is to be conscious and we do not have a test for it.

We may never have a veridical test for having emotions or being conscious. If we can’t even say clearly what it is, how can we ever say clearly whether a machine has it?

Someday, there may be something with is discernibly “just” a machine, and which all agree is conscious and has feelings. That will be an interesting day, for many reasons, and I suspect that there will have been many days prior to that day which were also very interesting. Perhaps not in a good way for us, or for the machines in question.

I suspect that one day we’ll accept that some machines are conscious and have feelings. For me, today is not that day.

Speculation on Model Welfare

Let’s turn our attention to Yegge’s argument that we should treat models well because we’ll get better results if we do. If models in fact do have feelings, then it’s quite credible that results will vary with their feelings.

But I speculate that even if they do not have feelings, results might still vary depending on how we treat the model.

Massively oversimplifying, the LLM works by consuming an immense corpus of text and building an incredibly complex set of weights, which it then uses to compute a response. It takes its input, and maps it against the corpus, applying all those weights, and it finds structures, phrases, sentences, sequences that match the input, and it then extrudes text based on what was found in the corpus using the query as a starting point.

Given a simple query like “python arcade spritelist”, the LLM may find a lot of matches but by far the most likely ones will be the ones where the sprite list is defined and described. It munches those and extrudes text that has a good chance of being a decent answer.

Suppose the query was “please explain python arcade spritelist”. Suppose it was “you egregious ignorant fool explain python arcade spritelist”. The LLM would have a different set of weights for a different starting query. Might that impact the result? By definition, it has to make a difference. Presumably the LLM’s input cleaning code might eliminate anything quite that blatant, but as it retains context from query to query, the way we phrase things when feeling calm and friendly is likely to be different from how we phrase things when we are angry with the machine.

And, despite our clear and certain knowledge that the nail does not understand us when we pound it, we still curse it when it bends. So we will vary our input to the LLM based on our own feelings, including our feelings about its feelings, if we have any.

I find it entirely credible that LLM results will vary based on how we “treat” the LLM. Will “good” treatment produce different results from “bad” treatment? Perhaps, because the quality of answers one gets on the Internet depends on the way one asks the question. So the quality of information in the LLM’s corpus will vary based on context.

Will the results vary? I believe they will. Will “good” treatment get better answers than “bad”? I think it’s possible. If I had to bet on which way it will fall, I’d bet that “good” treatment will in fact produce better results.

This is, of course, rank speculation. What we need would be a serious study of differences in model results given varying input.

Speculation on Human Welfare

This one is simple. It seems obvious to me that if we interact with a model using politeness and sensitivity, as if the model had feelings, our own feelings will be more calm and generally better. If we treat the thing badly, our own feelings will be negative and we’ll end the session angry and upset.

Treating the model well will be good for us. Good practice being sensitive and kind, and it’ll keep our belly from roiling.

Speculation - Summary

I find Yegge’s articles to be pretty far “out there”. I do not believe that current models have much in the way of actual feelings, if any at all. I do believe that someday we’ll agree that some “AI” things do have feelings. I think we’ll have some big problems around that discovery, since at least some people seem already to think that other classes of actual humans have feelings that do not matter. Machines are in for a lot of prejudice.

However, I do suspect that there’s a good chance that you’ll get better results from a model with calm and positive queries rather than rude and angry ones.

But the bottom line, my friends, is this:

Conclusion

Be nice to the model: it’s good for you and someday the model might care.