Comment by chr15m
2 days ago
Although LLMs are not deterministic in normal operation (because of the temperature values used) they are still strongly biased in their outputs. Bias in their outputs is literally how they function.
So you get a different type of conformism where people who interact with LLMs a lot are going to be influenced towards the median, as encoded in those token probability look up tables.
The effect is basically the same as what Heidegger and Illich were worried about, and I think we're seeing it happen already everywhere.
A solution(?) is to try to be your true weird authentic self when you act in the world, rather than bending your words and behaviours to the machine and LLM probability shape.
I would like to state here, that LLMs are actually entirely deterministic. The temperature is just as much of an input as any other parameter and instructions and prompts. Deterministic does not mean "explainable" or "reasonable" or "understandable". Unless LLMs somehow have broken out of the confines of our very deterministic computers/calculation machines without us noticing, they are very much as deterministic as anything else we run on these machines.
What people perhaps mean by non-deterministic on a computer is, that we often do not control all inputs and therefore cannot tell what the output will be. That's however actually more about our own inability, rather than the processes within the machines.
If the temperature dependency is implemented via a deterministic pseudo-random algorithm, then they are deterministic. If, on the other hand, the output is at least partially causally dependent on a source of environmental randomness, then, even though everything 'inside the box' is deterministic, the process as a whole may or may not be, depending on whether the source of environmental randomness is, and that is ultimately a question about fundamental reality.
Aren't we also biased in our outputs? Many of our choices are predictable, even if not perfectly consistent. We just call it "personality".
Authentic being (Dasein) is absolutely unpredictable. You cannot live in that state all the time, obviously, we shuttle through life on autopilot but still, we catch glimpses of it now and then when interacting with a human being (or being one:).
And I distinctly feel AI outputs lack some of that variance and unpredictability. The fractal nature of natural things also is not there with AI. It's an eery feeling, akin to the uncanny valley.
I think that any good writing is a reflection of that Dasein and reading AI outputs I am convinced there is no 'soul' in AI yet.
I've always struggled with Heidegger, so I'll have to ask for more clarification.
If authentic being-in-itself is truly random, then how can it have identity? It can't have memory and it can't be influenced by events.
> (or being one:)
Good save, fellow human!
The gamut of human experience is breathtaking in its diversity but individual humans are generally quite predictable. You can identify someone by their mannerisms, by how they walk, by the sound of their breathing. Humans are creatures of habit, and I'd venture to say have an even lower expression space than an LLM on average.
5 replies →
saw a meme graph lately showing the distribution of 3 different sources of a number from 1 to 10. random, human, ai. random is flat of course, human is a hilly landscape with low hills on all odd values above 1, with 7 the highest, ai is all 7.
appreciate the humor in the meme, but I did become curious about what the latest models would actually produce, and interestingly after asking for 10,000 numbers between 1 and 10, the distribution was a hilly landscape with low hills on odd values and 7 the highest...
Tried it out.
Prompt: “Pick a random number between 1 and 10. Do not use a script or program.”
LLM: “7”
2 replies →
A human can be biased toward extinction, like all who are living out of connection with the conditions that sustain life and/or no longer living in relationship with land as kin.
It's those humans who would make and justify the use of thinking machines that are biased in the same way without accounting for the bias.
This isn't an extinction path I've seen noted in most circles.
Indeed we are, but it takes very little to train a human being. In fact, several thousand unique ones finished pre training as I typed this comment! That allows for a considerable natural diversity compared to a chatbot.
Yes but we're all biased differently. It's a rich palette.