[selfmodruntime]

I've always been wondering what GPT models lack that makes them "query->response" only. I've always tried to get chatbots to lose the initially needed query, with no avail. What would It take to get a GPT model to freely generate tokens in a thought like pattern? I think when I'm alone without query from another human. Why can't they?

[dragonwriter]

> What would It take to get a GPT model to freely generate tokens in a thought like pattern?

That’s fundamentally not how GPT models work, but you can easily build a framework around them that calls them in a loop; you’d need a special system prompt to get anything “thought like” that way, and if you want it to be anything other than stream-of-simulated-consciousness with no relevance to anything, and a non-empty “user” prompt each round, which could be as simple as time, a status update on something in the world, etc.

[xwolfi]

Monkeys who've trained since birth to use sign language, and can reply incredible questions, have the same issue. The researchers noticed they never once asked a question like "why is the sky blue?" or "why do you dress up". Zero initiating conversation, but they do reply when you ask what they want.

I suppose it would cost even more electricity to have ChatGPT musing alone though, burning through its nvidia cards...

[kolinko]

Just provide empty queey and that’s it - it will generate tokens no prob.

You can use any open source model wirthout any promot whatsoever

[nurple]

I think this will be key in a logical proof that statistical generation can never lead to sentience; Penrose will be shown to be correct, at least regarding the computability of consciousness.

You could say, in a sense, that without a human mind to collapse the wave function, the superposition of data in a neural net's weights can never have any meaning.

Even when we build connections between these statistical systems to interact with each other in a way similar to contemplation, they still require a human-created nucleation point on which to root the generation of their ultimate chain of outputs.

I feel like the fact that these models contain so much data has gripped our hardwired obsession for novelty and clouds our perception of their actual capacity to do de novo creation, which I think will be shown to be nil.

An understanding of how LLMs function should probably make this intuitively clear. Even with infinite context and infinite ability to weigh conceptual relations, they would still sit lifeless for all time without some, any, initial input against which they can run their statistics.

[hpeter]

It happens sometimes. Just the other day a local TinyLlama instance started asking me questions. The chat memory was full of mostly nonsense and it asked me a completely random and simple question out of the blue. Did chatbots evolve a lot since he was created.

I think you can get models to "think" if you give them a goal in the system prompt, a memory of previous thoughts, and keep invoking them with cron

[djur]

You might not have a prompt from another human, but you're always receiving new input.

[selfmodruntime]

Yes, but that's the fundamental difference. Even if I closed my eyes, plugged my ears and nose and laid in a saltwater floating chamber, my brain will always generate new input / noise.

(GPT) Models toggle between a state of existence when queried and ceasing to exist when not.

[kristiandupont]

You could just let the GPT run in a loop and it too would continue to generate tokens.

[Filligree]

And humans malfunction pretty badly without input. Even solitary confinement quickly drives them insane.

[pelorat]

> Why can't they?

They are designed for query and reponse. They don't do anything unless you give them input. Also there's not much research on the best architecture for running continuous though loops in the background and how to mix them into the conversational "context". Current LLMs only emulate single thought synthesis based on long-term memory recall (and some goes off to query the Internet).

> I think when I'm alone without query from another human.

You are actually constantly queried, but it's stimulation from your senses. There are also neurons in your brain which fires regularly, like a clock that ticks every second.

Do you want to make a system that thinks without input? Then you need to add hidden stimuli via a non-deterministic random number generator, preferably a quantum based RNG (or it won't be possible to claim the resulting system has free-will). Even a single photon hitting your retina can affect your thoughts and there are no doubt other quantum effects that trips neurons in your brain above the firing threshold.

I think you need at least three of four levels of loops interacting, with varying strength between them. First level would be the interface to the world, the input and output level (video, audio, text). Data from here are high priority and is capable of interrupting lower levels.

The second level would be short term memory and context switching. Conversations needs to be classified, and stored in a database, and you need an API to retrieve old contexts (conversations). You also possibly need context compression (summarization of conversations in case you're about to hit a context window limit).

The third level would be the actual "thinking", a loop that constantly talks to itself to accomplish a goal using the data from all the other levels but mostly driven by the short term memory. Possibly you could go super-human here and spawn multiple worker processes in parallel. You need to classify the memories by asking; do I need more information? where do I find this information? Do I need an algorithm to accomplish a task? What is the completion criteria. Everything here is powered by an algorithm. You would take your data and produce a list of steps that you have to follow to resolves to a conclusion.

Everything you do as a human to resolve a thought can be expressed as a list or tree of steps.

If you've had a conversation with someone and you keep thinking about it afterwards, what has happened is basically that you have spawned a "worker process" that tries to come to a conclusion that satisfies some criteria. Perhaps there was ambiguity in the conversation that you are trying to resolve, or the conversation gave you some chemical stimulation.

The last level would be subconscious noise driven by the RNG, this would filter up with low priority. In the absence of other external stimuli with higher priority, or currently running thought processes, this would drive the spontaneous self-thinking portion (and dreams) when external stimuli is lacking.

Implement this and you will have something more akin to true AGI (whatever that is) on a very basic level.

[throwthrowuknow]

Train it on stream of consciousness but good luck getting enough training data.