2
$\begingroup$

We can call it "time to think", or "free contemplation", or whatever. Maybe a term already exists in the field? What I mean is - with so many LLMs being "trained" on large datasets, are there any researchers just turning the AIs loose and letting them form opinions about the data they've been given? Maybe they could develop worldviews, or even personalities.

Sure, they can already role play, "pretending" to have personalities. But that's all parroting, really. User provides input, and chatbot delivers output in response.

A person's ability to contemplate on ideas is essential to the forming of the person's belief systems and how they relate to the world.

In AI related news I haven't heard any mention of models being allowed to simply think freely. Maybe this is because it is considered by researchers to be a waste of processing power, or maybe it hasn't crossed their mind to try, or maybe it's because it is difficult to save the changed (or perpetually changing) state of the AI that has formed (or is forming) new ideas, beliefs, and preferences.

If AGI is to be developed, wouldn't it require its own time and "space" - a digital domain to explore freely - in order to know itself? If it is only taught by a teacher (its training data), and never allowed to go explore, it can't really dive deeper into topics that excite its neurons.

Since AI is goal-driven, the goal could be something like: "pursue paths of thought that mentally stimulate you". And yes, we could make that part of a text prompt, and the LLM would probably feign sentience for us and reply with something that reads as if it has personality. But again, it would be simply parroting. What I'm suggesting is contemplation built in as a feature, integrated at more fundamental levels of the software architecture.

Has this approach been attempted or discussed?

Please forgive my layperson wording and lack of AI technical jargon. I'm sure an AI computer scientist could ask the question in more specific terms. (Answers of varying levels of technicality are welcome.)

$\endgroup$
3
  • $\begingroup$ in short: no. you would do well not to anthropomorphise these models, as they are just models -- at their heart they are just matrix multiplications with the matrices optimised to minimise some function on an extraordinary amount of data. when you say they develop 'personalities', this is just because they are outputting tokens that maximise the probability of satisfying a prompt based on their training data. in terms of forming an opinion, this would require being able to generate outputs outside of their training data -- I'm not sure if this has been demonstrated sufficiently yet. $\endgroup$
    – David
    Commented May 16 at 10:55
  • $\begingroup$ @David I get your point. Yes, just matrix multiplications, whether they get their own time to "think" or are only run when called on by the user. I am reminded that a large factor in all of this is probably that the domains in which AI versus organic intelligences exist (and therefore the nature of stimuli they experience) are so different. Organic intelligence never completely shuts down—only has "sleep mode" ;-) Whereas LLMs are only run 1 prompt at a time and never perceive an environment shifting through space/time. Such an environment is a major "outside of their training data" element. $\endgroup$
    – Mentalist
    Commented May 17 at 1:27
  • $\begingroup$ It seems to me to be similar to letting a car engine idle, ie running it without moving. Also, LLMs don't "think", they just match patterns using probabilities. $\endgroup$ Commented May 17 at 8:20

1 Answer 1

0
$\begingroup$

Are there LLMs that are allowed time to contemplate freely?

No, best you can do is changing top-k, top-p, and temperature.

$\endgroup$
2

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .