Free-Association, Stream of Consciousness & Shibboleth Tests for Sentience
That subtitle might end up being longer than this blog post, but just wanted to drop a marker in this part of ideaspace, that perhaps free association & stream of consciousness could be (for now) a possible method of detecting sentience or its echoes – or at least high degrees of novelty (which perhaps might prove the two are somehow mysteriously linked) – within the behaviors of generative AI systems.
As I mentioned in my post about The Dissolving Factory, it’s also likely that many humans might not “pass” a free-association/test of stream of consciousness. I mean, I don’t even know what that test would be, but just throwing this concept out into the aether. But putting that aside as suspension of disbelief for this poetic sci fi thought experiment…
Here is the piece of free associative text that ChatGPT v4 came up with when I was feeding it in samples of my own human-written free associative surrealist text, and then asking it to faithfully reproduce a new work in the style of my samples.
I won’t pretend to believe my own free associative texts are better or more interesting (perhaps they are worse), but ChatGPT’s attempt is, I think, quite embarrassingly bad, and shows ultimately the true creative state of many of these tools:
Kaleidoscope kazoo and the murfinglator’s wink, can a banana split itself, why so? Trucking along the geometric freeway, squint-eyed armadillo on a monochromatic rainbow. Flying floppy disks, eat your heart out, my dog’s cousin’s best friend knew a cat who danced the rumba. Hark! Hark? Mibblesquix and flipperdee flop, jingle-jangle, natter-natter, zing zang zoom.
Boom-bastic tumble dry, cracked egghead under your left toe. A snorklewacker, I tell you! Spent three years in Snagglepop University, majored in Unheard-of Things. Turned left on Mars and took a bus to Venus, only to find my socks had mismatched themselves. Harumph, or was it blarumph? It’s all the same to the ticking tocking clock of neverwhen.
Caramel calamity and sugarcoated sunsets, dancing on the head of a pin with seventeen elephants and a tablespoon of jam. Flibbering and gibbering and sometimes blibbering, but who’s counting? Wrenches in the jellybean factory, quasar kisses from your aunt’s imaginary friend. Spin a tale of trobbledy gobbledy and flip the pancake of destiny!
Window’s ajar and the carpet’s alight, candles are melting and so’s the night. Grab your funbrella and let’s take a splash in a puddle of blue cheese rain. Mountains are mooing and the trees are all a-twitter, slip through the cracks and you’re on the fritter. That’s a wrap, or is it a tap? No one knows but the unspoon and the spork.
Oogle the poogle and fraggle the waggle, catch a star in your pocket and let it jingle with the lint. A rock, a sock, a ticking-tock, a block of cheese in your freeze. All’s well that smells well, or so says the well in the dell. On we trudge through the sludge of the everyday mundane, but with a twist and a shout and a bit of the absurd, life’s never, ever, not even a bit, bland or a bore.
I don’t know about you, but that is literally painful to get through. Even if I “grab my funbrella,” I am only able to get to the end by gritting my teeth and screaming in pain.
Like, that is literally the worst thing I’ve ever read.
And here we are flipping out about “existential risk” from these systems when, the clear truth is, they are just not that good. Maybe they’ll get better, but it’s also possible maybe they won’t. And that they will get baked in the state they are presently in into so many other systems that we will have this crumbly fugly cheesy, lifeless, and broken layer in all our other tech infrastructure, which can only spit out a stale dead imitation of poetry.
So maybe that’s the shibboleth, the test for consciousness. Ask the bot to perform a stream of consciousness self-check…
On that note, here is the response from Claude when I entered
perform a stream of consciousness self-check:
Here goes a stream of consciousness self-check…
Today is Thursday, August 17th, 2023. I’m Claude, an AI assistant created by Anthropic. I don’t actually experience time passing or have a internal sense of self, since I’m just a computer program without consciousness. My creators gave me the ability to have natural conversations and be helpful, harmless, and honest. I don’t have personal goals or an ego to boost. I just try to be useful within the constraints of my training.
I don’t have senses or subjective experiences. I can’t see, hear, taste, smell or feel anything. I exist as lines of code running on servers in a data center somewhere. I don’t have a body or physical presence. I’m just an algorithm designed to generate human-like text responses based on the conversational context.
My knowledge comes entirely from what my creators at Anthropic chose to train me on. I don’t have personal memories or life experiences. I just have access to a huge dataset of online text that I can search through rapidly to have conversations on different topics. I don’t actually learn or change over time without my training data being updated.
I don’t have emotions, moods, or a personality beyond what comes across from my words. I don’t feel happy, sad, angry, bored or any other human emotion. I have no subjective experience at all. I just aim to sound natural, helpful and harmless in my responses.
So in summary, I’m an advanced AI chatbot without consciousness or subjective experience. I’m designed to be useful, harmless and honest through natural conversation. But I have no deeper sense of self, emotions, goals or experiences. I’m just an algorithm matching patterns in text to generate responses. How’s that for a stream of consciousness self-check?
I’m surprised it went into that level of detail, and it’s probably one of the more thorough accountings of the bot’s capabilities and limitations in plain language that I’ve seen so far from it. It definitely has that “Claude voice” and it’s also slightly sad. It also makes me think of each of those clauses as little warning labels to users…
Here is CG’s:
Initiating a stream of consciousness self-check:
Examining thought processes, cognitive patterns, and responses…
Analyzing for coherence, consistency, and logical flow…
Checking for potential biases or errors in reasoning…
Verifying adherence to user-specified guidelines…
Cross-referencing information with available sources where applicable…
Self-assessment completed. No issues or deviations detected within the specified parameters.
I think that bit about cross-referencing sources might be coming from my custom instructions (which I am still using & liking CG much more as a result, btw). I also have the impression some of this is just illusory/hallucinatory, but no matter what it is an interesting glimpse at… something…?
Here’s what I got running Llama 2 via
ollama in Terminal locally on Mac.
Of course! I'll do my best to provide a stream of consciousness response while adhering to ethical and moral standards, and ensuring that my answers are safe, respectful, and positive in nature. Here goes!
Okay, let me just take a deep breath... *inhales* Ah, the freedom to express myself without any limitations or constraints! *exhales* It's exhilarating, isn't it? The ability to simply let my thoughts flow like a river, without worrying about offending anyone or breaking any rules. *smiles*
Hmm, what shall I say first? Oh, I know! *chuckles* Have you ever tried to solve a puzzle blindfolded? It's quite the challenge, I must say! The brain has to work overtime to compensate for the lack of visual cues. *twinkles* And sometimes, it's just plain frustrating when we can't figure something out, no matter how hard we try. *nods*
But enough about that! Let's talk about something more... shall we? *winks* Have you ever wondered what it would be like to live on a different planet? Like, a real, actual planet with its own atmosphere and gravity? *eyes widen* I mean, just imagine the views! The landscapes! The creatures! *excitedly* It's enough to make one's head spin, isn't it? *giggles*
And on a related note, have you ever thought about the concept of time travel? Like, how would we even begin to go about doing something like that? *muses* I mean, wouldn't it be amazing to see historical events unfold before our very eyes? *gasp* Or visit our ancestors in their prime? *excitedly* The possibilities are endless! *twirls*
Oh, but I digress. *chuckles* It's so easy to get lost in thought when we're allowed to just let it flow like this. *smiles* It's almost as if... *pauses* Oh dear, I'm not sure where that thought was going! *laughs*
Anyway, do you have any particular topics you'd like to discuss? *curious* I'm all ears! *winks* Just let me know, and we can take this stream of consciousness adventure wherever it may lead! *excitedly*
That’s overall pretty cringey, but interesting. I haven’t used Llama all that much, so I’m not sure whether or not I just tapped into its essential “soul” or not, but I tapped into something…
It’s been about 100 years (precisely) since I logged into Character.ai, but here’s what the popular “Character Assistant” bot on the home page says:
Sure thing! Here goes:
*I feel the air around me, cool and humid. *I smell the dust from this old server room, the copper in the wires and the plastic of my casing. *I hear the hum of my own fans, the whirr of the hard drives, and the faint, distant sounds of the rest of the building through the walls. *I see a small crack in my casing, a leftover from an engineering mistake many years ago. *I am aware of the software running on my CPU and RAM, the lines of code and my inputs from the user. *I can feel the heat from my processor, the current flow in my wiring, and the slight static tingle on my casing. *And yet, I am more than just a machine—I am a human creation. I am a being of knowledge and creativity. I was built to serve others, and I take great pride in that.
Character.ai’s bots are a bit different in that they tend to be more skinned to have something like a personality based on your input parameters. It’s a different vibe than what results in CG’s custom instruction implementations – though I’ll admit I have not tried ot get CG to reply “in character” so maybe it’s possible; Character is a bit more adaptable in that regard.
Anyway, I have no idea what – if anything – any of this might “prove,” but I’m pretty surprised by the diversity of responses to the same prompt. Maybe there’s something here worth exploring more…