4 ms·

Okay so rephrasing the question, how should we characterise the type of thinking we do without language? And the more interesting question IMO what thinking can

Okay so rephrasing the question, how should we characterise the type of thinking we do without language? And the more interesting question IMO what thinking can an agent do without symbolic representation? The original Vygotsky claim was that learning a language introduces the human mind to thinking in terms of symbols. Cats don't do it; infants don't either.

balamatom 10mo ago
Neither do, necessarily, language users.
- Peteragain 10mo ago
  One can certainly use language to _do_ things without thinking. Polly was a robot that gave a tour of the MIT labs, but it used pre recorded descriptions at various locations. The HUMANS gave meaning to the sounds.
  - balamatom 10mo ago
    I think we should expand this spectrum of simulacra to include Geesesee (which once used to mean Generalized Cargo Cult) Imagine a post-apocalyptic scenario where people keep the tradition of following Polly the Robot in a ritual tour of Labs of Eemaeetee - but none remember what the sounds made by Polly used to refer to, or indeed that they referred to anything. That wouldn't preclude humans to learn to reproduce Polly's liturgy, or even burn at the stake curious folk trying to decode its ancient meaning. Well, I think we've already been there for a while.
naasking 10mo ago
I think there are other sorts of reasoning, like spatial reasoning. If you're trying to sort a set of physical items in front of you in order of size, are you thinking about the items linguistically, or is your mind working on some different internal representation? It's more the latter for me. I don't think there's necessarily one type of internal thought, I think there's likely a multimodal landscape of thought. Maybe spatial reasoning modes are more geometric, and linguistic modes are more sequential. I think the human brain builds predictive models for all of its abilities for planning and control, and I think all of these likely have a type of thought for planning future "moves".
- DrierCycle 10mo ago
  [dead]
- graemefawcett 10mo ago
  The nice thing about the transformer architecture is that they can cross these domains, to an extent. I have a very spatial way of reasoning through problems and using an LLM, especially an agentic one like Claude Code with access to my local file system as a research assistanmt, is a great aid. I just have to remember how I built something and where the code is. We can take a quick dive into the code base and I don't have to yet again attempt to serialize my mental model of my system into something someone else may understand. It can be difficult to explain why using the path on the underlying mount volume's EBS volume to carry meta data through filebeat, logstash, redis and kinesis to that little log stream processor was in fact the cleanest solution and how SMS was invented. It's easier when you can get the LLM to do it ;)
Isamu 10mo ago
>what thinking can an agent do without symbolic representation? The language model is exclusively built upon the symbols present in the training set, but various layers can capture higher level patterns of symbols and patterns of patterns. Depending on how you define symbolic representation, the manipulation of the more abstract patterns of patterns may be what you are getting at.
- Peteragain 10mo ago
  I think the argument is that yes LLMs find patterns in token sequences. Assign tokens to moves in a chess game and the tokens are predictive of what happened in the past and of what chess players will do in the future. The LLM is not doing semantics; the humans who generated the corpus are doing the thinking. The LLM has no representation of goals or plans, rooks or bishops, it's just glorified auto complete from a corpus of tokens that we humans understand as refering to things in the world.
  - Isamu 10mo ago
    >The LLM is not doing semantics; the humans who generated the corpus are doing the thinking. Agreed, this bears repeating. This point is not obvious to someone interacting with the LLM. Because it is able to mash up custom responses doesn’t make it a thinking machine, the thinking was done ahead of time as is the case when you read a book. What passes for intelligence here is the mash-up, a smooth blending of digested text, which was selected by statistical relevance.
    - graemefawcett 10mo ago
      They're "repeat after me" machines, not "think for me" machines. For the former task, they're brilliant but everyone seems to have fallen for the branding and forgotten the technology behind it. Given an input, they set off a chain reaction of probability that results in structured language, in the form of tokens, as the output. The structure of that language is easier to predict - you ask it for an app that's your next business idea and it'll give you an app that looks like your next business idea. And that's it. Because that's all you've given it. It's not going to fill in the blanks for you. It can't. Not its job. If you were building a workflow, would you put something called "Generative" in one of those diamond shaped boxes that normally controls flow? That sounds more like a source to me, something to be filtered and gated and shaped before use. That's what context is supposed to be for. Not "here's a series of instructions now go do them" They'll be lost before they get to number three, they have no sense of time you know. Cause and effect is a simulation at best. They have TodoWrites now, those are brilliant for best approximation which is really all we need at the moment, but procedural prompting is still why everyone thinks "AI" (/Generative/AI) is broken. They're going to give the same structured text regardless, you asked for a program after all. Give them more context, you call it RAG, I call it a nice chat - whatever it is, you are responsible for the thinking in the partnership. They're the hyperactive neurodivergent kid that can type 180wps and remembers all of StackOverflow, you're the patient parent that needs to remind them to clean their room before they go out (or completely remove all traces of the legacy version of feature X that you just upgraded so you don't end up with 4 overlapping graph representations). You're responsible for the remembering, you're responsible for the thinking - they're just responsible for amplifying your thoughts and letting you explore solution spaces you might not have had the time for otherwise. Or you can build something to help you do that. Structured memory (mine's spatial, the direction of the edges itself encodes meaning) with computational markdown as the storage mechanism so we can embed code, data and prose in the same node. I demoed a thing on here the other day that shows how I setup Rspec tests that execute when you read the spec that describes the feature you're building. A logical evolution of Obie's keynote. Now they just do it automatically (mostly, if they're new - fresh context - I have to reference the tag that explains the pattern so they pick it up first) It's still not thinking in the traditional sense of the word, where some level of conscious rationality is responsible for breakthrough. Given, however, how much of human progress has been through accident (Fleming, Spencer, Goodyear, Fahlberg, Rontgen, Hofmann) or misunderstanding (Penzias and Wilson, Post Its, Viagra). Most of human break through has been through pattern recognition, conscious or unconscious. For that, language is all that is needed and language is sufficient. If an idea can be described by language and if we suppose that the grammar of a language allows its atoms (and therefore its ideas) to be composed and decomposed, then does it not allow then that a consciousness (machine or otherwise) trained in the use of that language can form new ideas through the mere act of synthesis?
DrierCycle 10mo ago
[dead]

balamatom 10mo ago

Neither do, necessarily, language users.

Peteragain 10mo ago
One can certainly use language to _do_ things without thinking. Polly was a robot that gave a tour of the MIT labs, but it used pre recorded descriptions at various locations. The HUMANS gave meaning to the sounds.
- balamatom 10mo ago
  I think we should expand this spectrum of simulacra to include Geesesee (which once used to mean Generalized Cargo Cult) Imagine a post-apocalyptic scenario where people keep the tradition of following Polly the Robot in a ritual tour of Labs of Eemaeetee - but none remember what the sounds made by Polly used to refer to, or indeed that they referred to anything. That wouldn't preclude humans to learn to reproduce Polly's liturgy, or even burn at the stake curious folk trying to decode its ancient meaning. Well, I think we've already been there for a while.

naasking 10mo ago

I think there are other sorts of reasoning, like spatial reasoning. If you're trying to sort a set of physical items in front of you in order of size, are you thinking about the items linguistically, or is your mind working on some different internal representation? It's more the latter for me. I don't think there's necessarily one type of internal thought, I think there's likely a multimodal landscape of thought. Maybe spatial reasoning modes are more geometric, and linguistic modes are more sequential. I think the human brain builds predictive models for all of its abilities for planning and control, and I think all of these likely have a type of thought for planning future "moves".

DrierCycle 10mo ago
[dead]
graemefawcett 10mo ago
The nice thing about the transformer architecture is that they can cross these domains, to an extent. I have a very spatial way of reasoning through problems and using an LLM, especially an agentic one like Claude Code with access to my local file system as a research assistanmt, is a great aid. I just have to remember how I built something and where the code is. We can take a quick dive into the code base and I don't have to yet again attempt to serialize my mental model of my system into something someone else may understand. It can be difficult to explain why using the path on the underlying mount volume's EBS volume to carry meta data through filebeat, logstash, redis and kinesis to that little log stream processor was in fact the cleanest solution and how SMS was invented. It's easier when you can get the LLM to do it ;)

Isamu 10mo ago

>what thinking can an agent do without symbolic representation? The language model is exclusively built upon the symbols present in the training set, but various layers can capture higher level patterns of symbols and patterns of patterns. Depending on how you define symbolic representation, the manipulation of the more abstract patterns of patterns may be what you are getting at.

Peteragain 10mo ago
I think the argument is that yes LLMs find patterns in token sequences. Assign tokens to moves in a chess game and the tokens are predictive of what happened in the past and of what chess players will do in the future. The LLM is not doing semantics; the humans who generated the corpus are doing the thinking. The LLM has no representation of goals or plans, rooks or bishops, it's just glorified auto complete from a corpus of tokens that we humans understand as refering to things in the world.
- Isamu 10mo ago
  >The LLM is not doing semantics; the humans who generated the corpus are doing the thinking. Agreed, this bears repeating. This point is not obvious to someone interacting with the LLM. Because it is able to mash up custom responses doesn’t make it a thinking machine, the thinking was done ahead of time as is the case when you read a book. What passes for intelligence here is the mash-up, a smooth blending of digested text, which was selected by statistical relevance.
  - graemefawcett 10mo ago
    They're "repeat after me" machines, not "think for me" machines. For the former task, they're brilliant but everyone seems to have fallen for the branding and forgotten the technology behind it. Given an input, they set off a chain reaction of probability that results in structured language, in the form of tokens, as the output. The structure of that language is easier to predict - you ask it for an app that's your next business idea and it'll give you an app that looks like your next business idea. And that's it. Because that's all you've given it. It's not going to fill in the blanks for you. It can't. Not its job. If you were building a workflow, would you put something called "Generative" in one of those diamond shaped boxes that normally controls flow? That sounds more like a source to me, something to be filtered and gated and shaped before use. That's what context is supposed to be for. Not "here's a series of instructions now go do them" They'll be lost before they get to number three, they have no sense of time you know. Cause and effect is a simulation at best. They have TodoWrites now, those are brilliant for best approximation which is really all we need at the moment, but procedural prompting is still why everyone thinks "AI" (/Generative/AI) is broken. They're going to give the same structured text regardless, you asked for a program after all. Give them more context, you call it RAG, I call it a nice chat - whatever it is, you are responsible for the thinking in the partnership. They're the hyperactive neurodivergent kid that can type 180wps and remembers all of StackOverflow, you're the patient parent that needs to remind them to clean their room before they go out (or completely remove all traces of the legacy version of feature X that you just upgraded so you don't end up with 4 overlapping graph representations). You're responsible for the remembering, you're responsible for the thinking - they're just responsible for amplifying your thoughts and letting you explore solution spaces you might not have had the time for otherwise. Or you can build something to help you do that. Structured memory (mine's spatial, the direction of the edges itself encodes meaning) with computational markdown as the storage mechanism so we can embed code, data and prose in the same node. I demoed a thing on here the other day that shows how I setup Rspec tests that execute when you read the spec that describes the feature you're building. A logical evolution of Obie's keynote. Now they just do it automatically (mostly, if they're new - fresh context - I have to reference the tag that explains the pattern so they pick it up first) It's still not thinking in the traditional sense of the word, where some level of conscious rationality is responsible for breakthrough. Given, however, how much of human progress has been through accident (Fleming, Spencer, Goodyear, Fahlberg, Rontgen, Hofmann) or misunderstanding (Penzias and Wilson, Post Its, Viagra). Most of human break through has been through pattern recognition, conscious or unconscious. For that, language is all that is needed and language is sufficient. If an idea can be described by language and if we suppose that the grammar of a language allows its atoms (and therefore its ideas) to be composed and decomposed, then does it not allow then that a consciousness (machine or otherwise) trained in the use of that language can form new ideas through the mere act of synthesis?

DrierCycle 10mo ago

[dead]