I've been thinking about emergent personality in AI. I imagine such a phenomena exists, and we'll see it come to bear in the near future. One prerequisite: Long context windows. LLMs need tokens to think (hence one advantage of Chain-of-Thought), but extended self-dialogues come with flavors that are hard to discern. LLMs start conversations from scratch, so short conversations don't give it time to develop extended conversational subtexts, but extending their context window will potentially enable this to occur. Some of the below reading is my initial effort to learn more about where this stands. There are many philosophical ideas I want to revisit in this vein (read: Grice, RSA, QUD, SCA), but for now, I want to focus on catching up to the current state of affairs.
-
(https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1484260/full)
-
(https://arxiv.org/pdf/2503.14499)
-
(https://mistral.ai/news/mistral-ocr) This is critical for the SAT/SHSAT question LLM-generator project. Doing OCR on Math problems (and the English problems even) was a major hassle and, frankly, the primary impediment to getting the project off the ground. This API should enable an easy pipeline from PDF to Vector Store!
-
(https://www.nature.com/articles/s41586-025-08705-9) The mathematical aspect of conversational subtext should mirror that of some brain region or network (we can have conversations with ourselves, and quite readily imagine having conversations with others - it is, at least, a pasttime of my own), and dmPFC might serve that function. Obviously, ToM - a requisite network for playing the role of the conversee - is necessary and mediated by the DMN, but I can see a joint between thinking of another person's intention and inner cogitation, and the coding and transmission of that intention. A radical claim (that I quietly hold) is that the language mechanism has some dimension of personality embedded in it. The source of this personality might lie in the embedding mechanism, either Broca's or some other prefrontal region that mediates between Frontal Task Network and DMN (dACC might serve as a gating mechanism for this neural negotiation). Certainly, personality is diffuse but if it lies anywhere, I wouldn't place it in the association cortex or basal ganglia lol. Though dogs have personalities and their prefrontal cortices are like peanuts compared to ours. Obviously I should go back and actually read the Gazzaniga chapter on personality! To the todo list.