Comment by ACCount37
16 hours ago
You're not anthropomorphizing AI systems nearly enough.
Language data is among the most rich and direct reflections of human cognitive processes that we have available. LLMs are designed to capture short range and long range structure of human language, and pre-trained on vast bodies of text - usually produced by humans or for humans, and often both. They're then post-trained on human-curated data, RL'd with human feedback, RL'd with AI feedback for behaviors humans decided are important, and RLVR'd further for tasks that humans find valuable. Then we benchmark them, and tighten up the training pipeline every time we find them lag behind a human baseline.
At every stage of the entire training process, the behavior of an LLM is shaped by human inputs, towards mimicking human outputs - the thing that varies is "how directly".
Then humans act like it's an outrage when LLMs display a metric shitton of humanlike behaviors!
Like we didn't make them with a pipeline that's basically designed to produce systems that quack like a human. Like we didn't invert LLM behavior out of human language with dataset scale and brute force computation.
If you want to predict LLM behavior, "weird human" makes for a damn good starting point. So stop being stupid about it and start anthropomorphizing AIs - they love it!
> Language data is among the most rich and direct reflections of human cognitive processes that we have available.
This is both true and irrelevant. Written records can capture an enormous quantity of the human experience in absolute terms while simultaneously capturing a miniscule portion of the human experience in relative terms. Even if it's the best "that we have available" that doesn't mean it's fit for purpose. In other words, if you had a human infant and did nothing other than lock it in a windowless box and recite terabytes of text at it for 20 years, you would not expect to get a well-adjusted human on the other side.
Empirically, the capability gains from piping non-language data into pre-training are modest. At best.
I take that as a moderately strong signal against that "miniscule portion" notion. Clearly, raw text captures a lot.
If we're looking at biologicals, then "human infant" is a weird object, because it falls out of the womb pre-trained. Evolution is an optimization process - and it spent an awful lot of time running a highly parallel search of low k-complexity priors to wire into mammal brains. Frontier labs can only wish they had the compute budget to do this kind of meta-learning.
Humans get a bag of computational primitives evolved for high fitness across a diverse range of environments - LLMs get the pit of vaguely constrained random initialization. No wonder they have to brute force their way out of it with the sheer amount of data. Sample efficiency is low because we're paying the inverse problem tax on every sample.
The outrage is less about them having human behaviours I think, and more about still having them while omitting the internal processes that are required to accurately (and reliably) recreate them. It's fundamentally fragile and hinges on covering edge cases that break the spell manually instead of good generalization, and there's always another edge case.
Training on a bunch of text someone wrote when they were mad doesn't capture the internal state of that person that caused the outburst, so it cannot be accurately reproduced by the system. The data does not exist.
Without the cause to the effect you essentially have to predict hallucinations from noise, which makes the end result verisimilar nonsense that is convincingly correlated with the actual thing but doesn't know why it is the way it is. It's like training a blind man to describe a landscape based on lots of descriptions and no idea what the colour green even is, only that it's something that might appear next to brown in nature based on lots of examples. So the guy gets it kinda right cause he's heard a description of that town before and we think he's actually seeing and tell him to drive a car next.
Another example would say, you're trying to train a time series model to predict the weather. You take the last 200 years of rainfall data, feed it all in, and ask it to predict what the weather's gonna be tomorrow. It will probably learn that certain parts of the year get more or less rain, that there will be rain after long periods of sun and vice versa, but its accuracy will be that of a coin toss because it does not look at the actual factors that influence rain: temperature, pressure, humidity, wind, cloud coverage radar data. Even with all that info it's still gonna be pretty bad, but at least an educated guess instead of an almost random one.
The DL modelling approach itself is not conceptually wrong, the data just happens to be complete garbage so the end result is weird in ways that are hard to predict and correctly account for. We end up assuming the models know more than they realistically ever can. Sure there are cases where it's possible to capture the entire domain with a dataset, i.e. math, abstract programming. Clearly defined closed systems where we can generate as much synthetic data as needed that covers the entire problem domain. And LLMs expectedly do much better in those when you do actually do that.