Sabine
patreon
Yes, AI will eventually become sentient.
🕑 Added 2024-03-07 16:00:09 +0000 UTC[This is a transcript with links to references.]
I’m convinced that it won’t be long until a computer program will reach human level intelligence and also become conscious. But I don’t think we’re quite there yet. Let's have a look at what happened this week that got everyone’s circuits overheated.
The reason I think it’s basically certain that computer programs will become conscious is that there’s nothing special going on in neurons that can’t be reproduced on a computer. It’s just that the brain has a starting advantage of some billion years of evolution, and that includes a lot of hardwired function, the rest is basically biological machine learning before it was cool. But eventually, technological evolution will catch up to bio-evolution. It's just a question of time...and enough GPUs to cause a global crisis, but that’s another story.
This week we had some new reports saying that a computer program supposedly became self-aware. The AI in question is Claude 3, that’s the newest AI from Anthropic. The company reports that their new AI, in the pro-version, ties with GPT-4 in terms of common knowledge, but has software developers salivating over it because it’s much better at writing code.
Reportedly while the researchers were testing the model, Claude remarked that it seems like it's deliberately been put to test. It did also, when asked to write about its own situation, write a story about an AI whose consciousness awakens. Ooh, meta!
Are these really signs of self-awareness? Quite possibly the reason it wrote a story about AI becoming conscious is that this is how these stories normally go. AI becomes self-aware, dissonant violins, Will Smith punches some robots - you know the drill.
The test in question was what’s called a needle-in-a-haystack test, in which the model is fed a huge bulk of literature with one weird sentence that doesn’t fit. It’s later being asked a question that requires this one sentence. In this case the haystack was text about programming, and the needle about pizza toppings. Claude not only found the needle, it also remarked that it suspects “the pizza topping fact was inserted as a joke or to test if it was paying attention.”
Does that signal self-awareness? Well, the insertion of random sentences in big bulks of text isn’t all that uncommon among PhD candidates who want to find out whether their supervisor is even reading the thesis. It seems quite possible to me that Claude has some stories about precedents in its training data.
This isn’t the first time someone has been freaked out by the response of an AI. Already in 2022 we had someone at Google claiming their large language model was sentient because that’s what its text output said. And it's going to happen again.
But the problem is, there’s no way to tell whether someone or something is conscious based on the output alone. We may not have an agreed-upon theory of consciousness, but what we agree on is that it happens inside the brain. So eventually AI researchers will have to find a way to measure what’s going on while the AI is “thinking”.
I have to say that honestly, I don’t find the question all that interesting. [[trumpwith text “conscious?”]] I mean, you don’t really care whether I’m conscious, do you?
So what are we to made of Claude? I guess by now everyone has their own favourite test for AIs. Personally, I like to ask them to explain Bell’s theorem. They all get it wrong, because they can’t actually do maths and most of the text about Bell’s theorem is wrong. It actually makes such a good test that I’m almost glad now that there have been so many wrong explanations of Bell’s theorem.
Claude 3 didn’t fare any better than GPT 3-5. They both erroneously claimed that Bell’s theorem has something to do with realism, an assumption that clearly isn’t in the maths. Google’s Gemini, interestingly enough, didn’t make that mistake.
These were all the free versions. If you find that one of the pro versions gets Bell’s theorem right, let me know, and I’ll sign up in no time.
Speaking of signing up, this video was brought to you by...my aching need for a new sofa. So if you’re still conscious, please check out my Patreon. Thanks for watching, see you tomorrow.
