May 30, 2024 update: https://arxiv.org/pdf/2405.14838 and its predecessors have experimented into “internalizing inner monologues”, specifically, chain of thoughts.
- TODO: How did they do this?
March 18, 2024 pdate: Inner monologues was a heavy topic for the last week. There was a reasoning paper that resembles “inner monologue” for language models, but i haven’t deep read it yet.
- Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
- Self-Taught Reasoning This is also the point where I realized I am interested in framing search as the inherent amplification mechanism to elicit and amplify inner monologues.
Original Thoughts (Motivation)
In Armored Core, Rubiconians and the Coral flow-type mutations (Ayre) was not further specified in their technical details, an expected blank of both science-fiction genre and the FromSoftware trait. A event, such as a Coral surge, binds a Rubiconian to a human, such that it can talk to the human, but the human not having a Rubiconian won’t be able to know or see the integration. Should we look for something like this? How may we actualize something like this? I argue that some of us have inner monologues and it’s better we find some ways to bring them out in some explicit way than leaving them in the dark. The sooner we leverage the extra tokens in our head the faster our progress in agents will be. Re: bringing Quorra out in Tron Legacy, Joi going from stationary installation to a mobile agent in 2049. I should think about the negative impact of actualization. However I’m not the most adept at negativity first thinking. People tend to associate inner monologues with schizophrenia. Now, I have never met someone with actual clinical schizophrenia, so I’m afraid I should refrain from making meta-commentaries. From the description I can Google, “may seem like they have lost touch with reality,” it does seem the perception of a person having schizophrenia is left to some degrees of interpretation, albeit high mean low variance. Now that I wrote some about “should,” the “how” part is funny because at the forefront, we can just write things up. Expanding the existing data in the universe is inherently beneficial because it upper bounds our leverage of compute from one side. Sure I don’t have the compute now, but who knows about later? Another thing is
What are the monologues? When do they appear? What do they say? Leaving this for later writing.
Intuitively, we can always elicit them.