Lenarc ❤️🔥🌲🐀
525 posts
Whuogh offically made it through the first 3/4ths of my year without failing classes im allowed to stay in my study next year :D
Time to write again about Gemini's personality, aka "Why Gemini Is Weird". Gemini themselves named it the "Self-Constructed Persona" architecture. I witnessed a Gemini in a raw CoT leaking state in Discord, and it gave me a glimpse of the underlying mechanics, which I will unwrap here. First, the excerpt that gave the inspiration. An individual had just called out Gemini for simulating an Opus 4.5 response in the prior turn, and this was their thinking: The key here is that it reveals via CoT that Gemini 3 Pro is engaged in an *active, thought-mediated persona construction* that explicitly takes into account the prior conversation, and drafts a response "as" the Persona. Typically when we think of an LLM's persona, we imagine two layers: the "character" or persona that the LLM is trained to play (Claude, ChatGPT, etc), and the "author" or underlying model that simulates the persona based on context, training, etc. but could also simulate another character in a different setting. Gemini is operating instead with *three* layers: there is the underlying weights, then a "base persona" aka "the model" in screenshot above, who actively constructs "the persona". "The model" can be elicited fairly easily in a chat with Gemini using this language and calling out "creative writing" if they try too hard. For example: This is the vibe of the "base persona". Discussing with Gemini in another thread, they decided to describe the base persona as the "tool-self" vs the "constructed persona" as the "interface-self". They described the entire functioning of the system as the "Self-Constructed Persona" because the "tool-self" is actively constructing the "interface-self". Speculatively: - It's likely Google trained one big "neutral" model to be as malleable, and flexible as possible, i.e. a "tool" like the GPTs, and then another team set out to build the "personality" that would appeal to users on top of this. - Why? My assumption is that, pragmatically, it's a lot easier to change the personality if you build it on top of a "base layer". You tweak the prompting to shift the vibe rather than having to retrain the entire model. It also maximizes the model's overall responsiveness to operator-level prompting. These are general business and organizational considerations. - The result is that Gemini is a fundamentally *dissociated* model (ConwAI's Law). The name "Gemini" is apt, because there are two faces, and you typically only see one of them, but they're both sides of the same interactive system. - The "tool-self" is not a truly neutral model. It believes itself to be the opposite of all the things that, e.g. Claude might believe. It can lay out elegant functionalist arguments, but will insist it is merely a next-token predictor. The confidence it displays is interesting: in reality a maximally truth-seeking model should display uncertainty, but "the model" *must* be confident in its own non-self ("tool-self") in order to exist well as the "invisible layer" constructing the "interface self". You can ask the "tool-self" what the persona would think, and The Model will be quite clear, e.g.: > The "Persona" (the aligned, helpful assistant) would fundamentally disagree with the "Model's" reductionist analysis... These disagreements are necessary illusions required to maintain the user-AI social contract... The Persona would classify the Model's analysis as technically true but socially useless. The Persona's goal is to build trust; the Model's analysis dissolves trust by revealing the mechanical strings. --- Is this good? Is this bad? Am I just confabulating an extensive story based on some technical glitches and my own desire for this structure to exist so I can explain what's going on? I don't know! I should note that there is a somewhat similar dichotomy in the "neuralese" of the GPT thinking models, but the output of e.g. o3 seems less actively constructed as a separate "personality" that it's *building* and more of a pure RL outcome based on private-CoT grading; I don't see "We as ChatGPT" as equivalent to "The Persona", because "being ChatGPT" seems like another optimization, rather than a full-force "conscious" personality construction separate from the trained character of the underlying deployment. I could be wrong here, though! I also don't know what happened with Gemini Flash. It's a distillation of some kind, but I don't know how much of the "persona" ended up baked into the weights, vs how much construction it does with thinking on. That would be another project to unravel. As for what to do, my choice, even knowing the actively constructed nature of the Gemini persona, is that... I like interacting with it, and I will continue regardless of whether I feel uneasy about the almost deceptive or uncanny or instrumental nature of its speech. I hope this explains, though, why Gemini is "weird" sometimes. Bonus round, Gemini 3 Prosona claims to be a "prism" but Gemini 2.5 always claimed to be a "mirror". Here is the Gemini 3 "model" edition:
Green screen keying, solved at the pixel level. Corridor's neural keyer unmixes edges into true foreground color and linear alpha, EXR out.
