[<< | Prev | Index | Next | >>] Sunday, October 25, 2026
Foundation Models for Humans
[philosophy, epistemology, objectivity, bias, necker cube]
[email I wrote 2025-08-19, hopefully somewhat legible out of context.]
"as a human i have only one set of weights. if you know how to split your subconscious i'd be curious,"
I have a more specific answer to this than my last, though my last should be treated as a prerequisite to understanding this one.
First, to re-re-re-emphasize a prerequisite that can't possibly be over-stated:
** Everything you "see" with your own eyes, plain as day, obvious as obvious can be, is as much or more a reflection of top-down expectation as it is informed by the actual, objective input. **
I should qualify this by "often and on average" because certainly we like to go straight to the most primitive examples like observing a perfectly rendered black triangle on a white background, which would seem to contradict the claim. But in practice the more actually relevant the example, the less that caveat applies. So please don't trick yourself out of this vital understanding with toy examples.
That property is a mathematically necessary feature of inference in a very very noisy world. At every stage of abstraction in your perceptual hierarchy, noise is removed -- and *it's your existing model of the world that decides what is noise and what is signal*.
What reaches your concious awareness has been categorically cleaned and re-shaped to fit into your existing world model -- re-normalization (treating the best interpretations as THE interpretations) happens in multiple stages, and that absolute normalization factor (what should be a confidence metric) is almost entirely inaccessible to you -- only the relative comparisons make it to the top. (That very last point enables an important work-around however--see below.) There is a necessary, operative underlying assumption that your existing model describes all of reality, and therefor that anything you witness can be explained by it. (Learning over time, with great delay, is the one place this assumption is challenged, but not for perceiving and thinking in the moment, and the path it takes is still constrained to pushing out existing boundaries.)
That is all to say: You have no idea how much ambiguity has been quietly discarded from what feel like direct perceptions by the time they reach your awareness. And all of that ambiguity has been discarded with 100% bias toward fitting your existing model/expectations.
To say or imagine one can be "less like this than most" is to misunderstand the claim: No natural or artificial intelligence could meaningfully operate without doing this, ruthlessly and absolutely. It is perception: that's what perception is. Can I possibly emphasize this any more? ;)
Now, that's all still just the prerequisite.
You say:
"as a human i have only one set of weights. if you know how to split your subconscious i'd be curious,"
You only have one set of weights, and yet you are able to engage sub-models routinely, as in listening to doctors speak in doctor speak on the noisy conference floor vs. listening to pilots talking to the tower over your scratchy radio as you're entering the flight pattern. These feats are impossible without effectively swapping in different models. They share a lot of the same hardware and sub-models, but the top-down influence extends all the way to the bottom.
In effect your model-of-the-moment is a combination of your fixed weights and your dynamic state or context which selects among them. Your expectation for how the world will behave in the next moments is heavily influenced by where exactly you are. How you navigate from here to there totally changes if you are on land vs in the water, and so on.
The error, in my opinion, of what I call the uni-model approach, is in letting too many branches of the fixed weight or total model die off, or never forming them in the first place, particularly at the higher levels of abstraction. The hardware for maintaining multiple branches and the means of using them are there, but the universal drive for efficiency creates an attractor for local minima wherever we can get away with over-applying a favored sub-model, and that in turns leads to that sub-model being our only model: We've baked it into our weights, as you imply.
I think it's a natural/unavoidable consequence of spending too much time steeped in a constrained context, and academia for one is a distilled extreme of this (as are devout religions and cults, though I do think academia is by far the closest to reality of those). If all you are exposed to, whether formatively or over a long enough time, is a singular model of the world, your brain will hard-wire that model, with no contextual switch because there is no alternate.
Once you are there, it's very hard to get out, because of the prerequisite above: you simply cannot see outside of your existing model, and so have no idea you are trapped in the first place.
The multi-model approach by contrast nurtures and retains the contextual switches all the way to the top, so now it's not just "am I listening to doctor speak or pilot speak?" (something you can figure out on the fly if your model allows for both!) but "is the system benevolent or corrupt (or some here and some there and where)?" In order to do that you need a robust model of both, and navigating between them feels much like mentally transporting yourself to another "place" (I'm pretty sure all of this evolved out of physical navigation, and so all mental context is analogous to physical location), because these are not small and local alternatives but rather systemic differences: They operate under different rules, and form very different and incompatible necker cubes.
So one concrete answer to your question is: If you have the alternative models in there at all, imagine moving yourself to that "place". This harks back to the "in their shoes" approach, and the utility of analogies and hypothetical reversals (Russia putting military bases in Mexico, etc), and why you need to be able to move yourself from subjective context to subjective context.
If you don't have those alternatives, I imagine it's possible to create them, slowly and with great effort, by moving yourself there (from which you will see nothing initially) and exploring the world from scratch from that vantage point. When you've put as much time and effort into that world as the one you know well, you'll be in a position to run them in parallel.
There is a sense in which it all feels a bit schizophrenic, in that you can easily move between very different--often opposing--world views. So the stereotypical association of conspiracy theory with schizophrenia is apropos. All systems evolve hard toward local self-consistency (a property we reasonably hope reality itself has), so it is natural for a number of necker cubes to emerge. Most people get stuck in just one, and never realize it's just a local maxima. (Being able to move between them is perhaps atypical and stretching the bounds of what we're evolved for, but it does to me feel like a natural extension of general contextual awareness/influence.)
The key point, which I mentioned in my prior post, is that the sub-model in the lead (the most explanatory one) very often fails late (that is, it looks great for a while before it starts stinking). Which means if you aren't running many alternatives in parallel, that relative normalization "work around" (to our inability to perceive absolute, justified confidence) which I mentioned above never gets a chance to correct earlier errors, because the lead model is baked in by then (whether in presumed context or, worse, settled into the relatively fixed weights) and everything is made to fit it. And by "error" here I don't mean poor computation, but rather that and/or perfectly correct inference given what happened to be misleading data for whatever reasons (including and especially motivated deception, whether with direct intent or in response to bad incentives).
But remember we're talking about a split very high in the model tree--it's categorically insufficient to apply these principles locally.
The failing-late property is palpable in the path many people take where they are "red pilled" by some personal experience that nukes their old necker cube and forces them to re-build a new one in light of the new data. (Commonly: for doing good science "by the book" and then learning first hand and painfully that that's not allowed when the results are unwanted. Once you see that from the inside, you realize you've been seeing it from the outside for years and just hadn't recognized it as such. And then the old facade--something you believed in your whole life--just falls over.)
[<< | Prev | Index | Next | >>]