ebben.bsky.social
Interested in the future of society, long time student of the mind sciences, and enjoy discussions.
68 posts
262 followers
438 following
Regular Contributor
Active Commenter
comment in response to
post
I don't know any details, but the above is from Shannon Sands speculating about LLMs learning function vectors like backtracking and forking tasks as part of their chain of thought. That's a sort of "thinking" like process, maybe?
comment in response to
post
"we've had an informal theory for a while now that just as LLMs learn latent representations or features for different behaviours like writing style, honesty etc, and they also learn function vectors for tasks"
comment in response to
post
Ok, fair enough. You report that you observe that "academic people [tend to] believe philosophy is about categorization and not about critical thinking."
comment in response to
post
Ok, but you seemed to do what you were critical of, categorizing instead of critical thinking. It is a big leap (seemingly judgmental) to say academic people are foregoing critical thinking for simple categorization. That should call for some evidence, no?
comment in response to
post
Yes, sorta. But the problem is motivating critical thinking more than the problem of people offloading it onto AI. That is, don't confuse the problem of lack of motivation to do critical thinking with the availability of AI.
comment in response to
post
Why do you say it is clear? What evidence do you have that suggests academics prefer categorization over thinking?
The whole point of philosophy is to formulate arguments for and against an idea
Aren't you doing this yourself in this post? Where's your argument?
comment in response to
post
Also, "representation" has to include action patterns. An action pattern, like a horse's various gaits, are representations, in this case of how to move in terrain. This theoretical aspect of memory over time ("abilities") is probably the most contentious issue and part of the confusion.
comment in response to
post
I think the confusion stems from old views of symbolic processing: "cognitive processes are the rule governed manipulation of internal symbolic representations"
The term "representation" has a much broader, non-symbolic meaning now, and thus not necessarily in conflict with dynamical theory.
comment in response to
post
I'm wondering if these articles about disinformation and social networking are missing the elephant in the room - the extreme right wing having taken over the major power nodes of society. For example, Fox news, the supreme court, other media outlets sane washing trump all are major factors.
comment in response to
post
Probability is largely based on combinatorial math, the number of different ways an outcome can possibly occur. So, it is objective that a coin will have a 50/50 chance of landing heads, given a fair coin. So in that sense probability is as objective as anything.
comment in response to
post
A chair, number 5, color blue, energy, information are constructs that refer to objective properties of the world. "Food" is an objective property of the world, but the article would have it that "food" is non-objective because it's only a mental construct. Does this clarify my initial reaction?
comment in response to
post
Author says, "probability ... is not an objective property of the world, but a construction based on personal or collective judgements". My claim would be that most concepts are like that, but because they, or probability in particular, is a construct, doesn't make it non-objective.
comment in response to
post
Any good book on applied probability (statistics) will discuss the difficulty in applying probability theory which this article depends on. But the issue here is more philosophical or a matter of conceptual analysis - the question of objective existence of something.
comment in response to
post
I read the article a while back. The title is misleading. The analysis leaves a bit to be desired. There are a couple of uses - frequency and likelihood. But that doesn't make the concept of probability an illusion. I guess i prefer more straight forward writing.
comment in response to
post
I claimed LLM's can't "understand". But, the latest ARC test used a LLM that could output "language programs" which are fed back into the system, enabling the LLM to rewrite its knowledge of the state of the world, sub-divide problems, and chain "thoughts" together. Some "understanding" maybe?
comment in response to
post
A second point concerns benchmarks. A good theory requires both empirical data points and a structural/dynamic model that helps interpret results. I have not seen much in the way of explanatory models of cognition to guide development of benchmarks. Any references you can suggest?
comment in response to
post
Actually the article isn't that far off the mark. But only because cognitive scientists are still trying to articulate the nature of cognition. What are some recent articles that describe theories of cognition? Melanie's Dec 2021 paper on understanding is a start, but what is there more recently?
comment in response to
post
You can imagine there are lots of wealthy employers who would love to mess with Bluesky's success. So, i'm guessing this current situation could be a reconnoiter and training ground operation.
comment in response to
post
LLM's don't "understand" anything. Nothing in the architecture to support "understanding". No internal world model to compare generated responses to the state of the world, no mechanism to support truth semantics.
But yes, image and speech and other pattern recognition abilities have been achieved.
comment in response to
post
This is a CEO talking, it is a sales pitch. "AGI as traditionally known" is a buzzword. "Superintelligence" is a buzzword.
The actual research is much more sober about AGI, about what cognition really requires to qualify as intelligence. LLM's and diffusion models are a tiny step forward.
comment in response to
post
Oh, i wanted to mention the notion of the necessity of an agent needing to learn by intervening on the world in order to get causality. So, interventional learning that llm's lack is another way of seeing the limitations. Which sort of what you said.
comment in response to
post
So, anyway i'm coming at the puzzle from the point of view of identifying the system of parts of cognition necessary to be intelligent. Anyway, just batting these ideas around, not arguing.
comment in response to
post
from language "no meaning from form alone". Since LLM's can often speak meaningfully, there is a paradox. My resolution is to say llm's do capture some semantics in the foundation model by compressing input down to a neat vector based semantic model. ("distributional semantics")
comment in response to
post
Ok, i was responding to "Causal relations can’t be extracted from data because..." by saying the larger problem is they can't model anything, causal relations and, like you say, many more things. But i'm now reminded Emily Bender was outspoken about LLM's not being able to capture any meaning ...
comment in response to
post
Quick response, what you say maybe/is probably true, but LLM's have a deeper problem, they don't have a model of the world, no way to represent states of the world with which to use truth semantics. They don't "know" what is true or false, only what has been expressed.
comment in response to
post
Probably lots of references to a naturalistic view of mind, but i'll just mention www.nbi.dk/~natphil/sal...
Oh, btw, early on Pattee referred to the "epistemic cut", but he seemed to quit using that term.
Also, Naturalising Agent Causation www.mdpi.com/1099-4300/24...
comment in response to
post
I would say this quote is an obsolete (romanticized) view of mind. At least it is wrong, but ignorance is bliss. A more modern view is to start with "process", not matter. Then you can trace the emergence of the mind from physical processes. An early account of this idea comes from Howard Pattee.
comment in response to
post
Regarding a mathy term for what LLM's are doing, I'd say 2 things: The foundation model is based on "auto-regression"
and the inference engine is based on matrix algebra that does a sort of multidimensional optimization search ("attention") over the input prompt+foundation model. (Loosely speaking).
comment in response to
post
What? Am I missing something? Are you literally saying our feeds are sketchy? Or is this some humor going over my head?
comment in response to
post
comment in response to
post
Yes, that's why high dimensionality is an interesting new phenomenon. Good luck figuring out "interpretability", it has a lot of people puzzled.
I'm reminded to suitcase words:
alexvermeer.com/unpacking-su...
comment in response to
post
Yes, exactly. Didn't the politico article dismiss the relevance of this? Anyway i expect there will be a ton of analysis about how the masses were manipulated now that everyone is online. Welcome to the new era, lol.
comment in response to
post
We can say the brain is probability based too at the implementation level, but we don't talk like that. Maybe that's the conceptual confusion with the term statistics in a computer vs "thinking" in humans. I know this is unclear, I'm just brainstorming how to describe this possibly new phenomenon.
comment in response to
post
Good question. Yes it's numbers based because the model was implemented by a computer. But, from a Marr levels perspective, you can look at it from a higher level of abstraction. In a very loose sense LLMs develop a sort of semantic model of language.
comment in response to
post
That argument doesn't make sense. People's political minds are almost totally shaped by news and social media they consume. I'm pretty disgusted about this topic so i won't dig into it. But the politico article would benefit from a more scientific perspective.
comment in response to
post
Yes, I couldn't get into it, seemed kinda scatterbrained. Article didn't seem to step back and consider the larger picture, larger concepts. But i didn't study the article, may have missed something.
comment in response to
post
About the novel concept of high-dimensional semantic space in millions, sometimes billions of dimensions. We don't have concepts for this. Language models compress all they have read into a neat vector-based space where semantic distances between elements can be calculated along all the dimensions.
comment in response to
post
Secondly, note that it is the foundation model that captures the statistics. But the new thing is a level of abstraction above that. Turns out the model captures a semantic space in many dimensions where distances in this large space between words can be easily calculated. So, more than statistics.
comment in response to
post
Yes, good way to put it. Looking at the architecture of a LLM bot there is a "foundation model" and an "inference engine". The inference engine can be developed to incrementally improve the "reasoning process" - break down problem to steps and write language programs in turn operate on those steps.