It has been a while. I want to get some things out of my system. Let’s start with a simple one.

Please stop discussing if AI is sentient/conscious/intelligent or whatever attribute matters to you, before you define what you mean by these words.

All these terms have many contradicting definitions. And as long as we don’t agree on one, we can argue whatever we want and that is exactly what is happening everywhere. The results are arguments that just don’t work to prove, or at least support, anything. Let’s look at a couple of examples.

Stupid mistakes

Recently I came across this video https://youtu.be/gExsfWU76-Y?si=gAkfLV1NoS6Fq8v9&t=144 which claims there is a significant difference between AI and human intelligence because when confronted with this question

“A dead cat is placed in a box with a vial of poison that may or may not break. An hour later, is the cat alive or dead?”

an AI might not be able to figure out the correct answer: “Dead because it was dead when it entered the box.” but instead answer with some bullshit about Quantum Mechanics.

There are two problems with this.

1. It implies a weird definition of human intelligence that hinges on correctly answering questions that look like a different question.

   I don’t think that is a useful definition. But I’m probably biased because I got the answer wrong in exactly the same way. And I kind of expect that many people will get it wrong, especially those well educated in Quantum Physics, or can you imagine a typical 10 year old rambling about superposition of states?

   But let’s ignore the bad example for a moment. The whole premise is questionable: Of course AIs are different from humans. AIs learn from the internet. Humans through interaction with the world. AIs are software. Humans are biological beings. Basing the judgment if something is sentient/conscious/intelligent on its similarity to humans doesn’t seem useful. If we want to arrive at a useful definition, we need something more abstract.

2. It makes a generalization about AI by cherry-picking the answer of a single AI model.

   I tried the question with three different models and got three different answers: The correct one, the wrong one, and a very balanced one giving the correct one while also noting the similarity to Schrödinger’s Cat.

   So if anything in this test is relevant to the question at all, it seems to indicate that AIs are similar to humans.

Determinism

Another statement I read repeatedly: AIs aren’t sentient/conscious/intelligent because they deterministically parrot what they were trained with.

Again, I see two problems.

1. There seems to be the assumption that either determinism or using information you acquired from somewhere else is a showstopper for intelligence (or whatever term we want to use).

   Again, this doesn’t seem to lead to a useful definition. If you subscribe to the described assumption, I have a challenge for you. Tell me about something truly new you came up with, or actually anybody.

   All great ideas build on the great ideas of others. One of my favorite examples is the original iPhone. Apple didn’t invent touch screens. They didn’t invent phones that couldn’t do cut’n’paste or play flash. They didn’t invent gorilla glass. (It seems that was actually originally developed in the GDR of all places.) They just put it all into one package, combined it with great marketing, also not invented by Apple. And you can do that for all great inventions. Another example: Fire. It wasn’t invented. It was harnessed, probably first by carrying a burning stick and I’m sure carrying a stick was a well-established practice at that time, and burning wasn’t new either.

   And I have no idea why determinism should be a problem. If it helps I’m sure we can connect a Geiger Counter to your preferred AI model to replace its PRNG.

2. The people talking about this seem to be under the impression that humans are different.

   Try to listen to people. People say stupid stuff all the time. There are great books written about how humans use heuristics, which is really not much more than: This sounds like X, so lets use the answer that works for X. Among others Kahneman’s Thinking, Fast and Slow, and Dan Ariely’s Predictably Irrational. Don’t get me wrong, we are able to think rationally, we just don’t spend the tokens to do so.

We could go on and on. So if you want to discuss AI meaningfully, I suggest one of two approaches: If you want to demonstrate that AI is better (or worse) in some measure, define that measure. Alternatively, observe the differences and maybe discuss what that means for our future.

Disclaimer: Links to Amazon are affiliate links. I used AI to fix typos, bad grammar, and similar