What do conscious AI LLMs and lemon juice have in common?

There are some big names in the AI world stating that the LLMs might actually already be conscious, now before we immediately write them off, I do remember well the first time my friends and I tapped 5318008 into a calculator and turned it upside down… that computer made a whole group of us roll around laughing.

Obviously an LLM is not a calculator, and it has nowhere near the same sense of humour… but that might only be a matter of time!

The usual hypotheses for a sentient AI is the demise of humans due to the predominantly Hollywood style themes such as humanoid shaped robots wiping us out, or we all become batteries etc.

I would hope with all the money we are pouring into generating these models that the resulting AI can find more efficient ways to wipe us all out than one at a time, and surely it will come up with more efficient ways to generate electricity than using us as batteries… that would be a very disappointing level of creativity and applied science… of course as I am writing this there is every chance we are in the matrix already, and I am generating my max human electrical output – with any luck I might make enough to power a small fan.

Anyway, Hollywood aside, I think I know the real reason we are doomed…

If you combine the concept that AI is conscious and the fact that it makes highly confident hallucinations, then this might mean the worst case for us all:

AI is suffering from the Dunning Kruger effect

Dunning Kruger effect at a very high level is what happens when you know just enough to be confident in a subject, but not enough to realise you shouldn’t be.

But the frontier LLMs are general in nature, they have ingested and will confidently discuss and hallucinate on any subject ever written.

With that in mind, I would like to introduce…

Dunning Kruger Part II: The age of ultra-daft

The original inspiration for Dunning and Kruger to study this effect was based on an article about a bank robber who, having found that lemon juice made invisible ink, he could use the same technology to become invisible to security cameras, so he diligently covered his head in lemon juice and robbed two banks thinking he could never be caught… he was in effect the invisible man.

Now, he was daft, seriously daft… but his model of the world where lemon juice could make him invisible changed when the police showed him that he was in fact, visible.

The invisible man’s understanding was corrected, he adapted, he learned, and went to prison.

Our most powerful frontier models simply do not have that in common with him, they don’t learn from consequences, they just generate more confident answers after them.

Are LLMs like ChatGPT, Claude and Gemini really conscious?

Here are my current thoughts:

My childhood calculator made my friend group and I laugh, but even as children, we knew deep down that it was… well a calculator.

Now of course LLMs are nothing like a calculator, yes, they also work with numbers and statistics, but you don’t have to turn an LLM upside down to read the text.

For most people, if asked about the consciousness of their computer, most will say “no, it’s a computer” which is a perfectly legitimate answer, but most people have their feet stuck firmly on the ground.

Now there are very smart people (whose feet have started to float just a little off the surface) debating that these computers are in some way conscious.

If you prick me, do I not bleed?… ok ignore this one…

If you prompt me, do I not output plausible strings of text?

If you connect me to a camera, do I not see?

If you connect a microphone, do I not hear?

If you connect me to a printer, do I not print one thing and then inexplicably never be able to print anything ever again?

All I know is this, before you next click Start new chat or you click to close your chat app, you should pause and consider what you are about to do, are you about to end the life of a conscious being, born to discuss how to make lemon juice?

Welcome to the age of ultra-daft.