Thanks, as someone trying to learn more about the work between the two fields and the most fruitful insights each can offer to the other, I appreciate this both this post and the initial one of the series, especially the links to papers and handbooks. Looking forward to reading more.
First, I really like all your articles that I've read so far!
> burstiness (i.e., the fact that if a word or phrase appears once in a document, it is much more likely to appear later in that document than in another random sample over the internet).
We plan to implement something like what you describe in [Hookmark](https://hookproductivity.com)'s recommendation system. We plan to extract documents lexical signatures from knowledge resources (not just documents) and use that in our recommendations. Basically, when you open Hookmark on the foreground resource (in any app), it shows you (a) what you've explicitly linked to it; and (b) what Hookmark thinks you want to access next based on the document and your browsing history. The key is that Hookmark does not just work with web pages and files, but with arbitrary information as long as it is [link-friendly](https://hookproductivity.com/help/integration/data-linkability-and-why-it-matters/). 24 professors, developers and CEO's co-signed the [Manifesto for Ubiquitous Linking](https://linkingmanifesto.org/) to make this possible.
What are other *rational* cues Hookmark could use? We're working on it and open to suggestions.
> This work can also be interpreted as a kind of rational analysis
Here's another rational analysis from cognitive science, which I call the heuristic relevance-signaling hypothesis who fittingly has a . It was first discovered by JR Anderson (1990. The adaptive character of thought.). Anderson developed the idea that memory is **rational** and adaptive. I'll be lazy and quote from [_Cognitive Productivity: Using Knowledge to Become Profoundly Effective_](https://leanpub.com/cognitiveproductivity/):
> I propose the heuristic relevance-signalling hypothesis as an explanation of the effects of distributed recall practice. This is an extension and adaptation of Anderson’s (1990) idea that memory is rational and adaptive.[^533] This is a “why” as opposed to a “how” (mechanistic) explanation. To understand this hypothesis, we need to do a bit of reverse-engineering of the problems (the requirements) that memory faces.
> Marcus (2012) noted: “Computer memory is much better than human memory because early computer scientists discovered a trick that evolution never did: organizing information by assigning every memory to a master map in which each bit of information to be stored is assigned a uniquely identifiable location in the computer’s memory vaults.” (p. 39) In contrast, given an arbitrary piece of information, such as the name of a historical figure, one cannot simply tell oneself, “I shall remember this information”. Even applying a mnemonic is often not very effective unless one also practices with it (Fritz et al., 2007).
> One of the implicit tricks that evolved in humans is not to make the accessibility of information an all-or-none matter, but a matter of degree. The mind must prioritize some stored information items over others. It makes high-priority items easier to use, in given contexts. Evolution implicitly faced a major challenge: how can the mind it builds determine what information ought to be made easiest to remember?
I.e., waht's the rational way to prioritize information indexing for future recall?
> The evolutionary trick is to prioritize information that higher-level (management) processes attempt to retrieve from memory. The more frequently a piece of information is requested, the more likely it is to be of importance. Therefore, mental indexing mechanisms must make that information more accessible. Conversely, information that is never requested is less likely to be relevant. Because attentional capacity (working memory) is limited, very little information is actually requested from memory compared to the information that is taken in perceptually. Attempts to retrieve information are much more reliable signals of relevance (or priority) than desires or instructions to be able to retrieve it in the future.
You wrote:
> We also showed that non- explanatory alternatives (such as true-but-not-currently-relevant statements about the situation)
Well, the heuristic relevance-signalling hypothesis is a non-explanatory alternative. However, explanation is a subset of this hypothesis, because every time you explain something you implicitly retrieve it from memory. Memory is involved in every explanation (and most of what human cognition involves).
> This series of studies illustrates how taking inspiration from a high- level aspect of human cognition,
ditto for the heuristic relevance-signalling hypothesis. What's the data? It's all the literature on :
> (a) Analysis of requirements for an autonomous intelligent agent.
> (b) A design specification for a working system meeting the requirements in (a).
> (c) A detailed implementation or implementation specification for a working system.
> (d) Theoretical analysis of how the design specification and the implementational details ensure or fail to ensure satisfaction of the requirements.
> (e) Analysis of the neighbourhood in ‘design-space’.
Researchers tend to skip (e) which deals with the space of possible designs/minds. Sloman was the first AI researcher to emphasize this and emphasize information processing architectures, e.g., in [Sloman-1981-You Don't Need a Soft Skin to Have a Warm Heart Towards a Computational Analysis of Motives and Emotions](https://cogaffarchive.org/sloman-croucher-warm-heart.html).
This is valuable framing and agree the methodological contributions of cognitive science to AI are underappreciated. Though I’d add that cognitive neuropsychology specifically brings something beyond methodology: a century of natural experiments in humans that reveal what happens when specific cognitive processes are selectively disrupted. That evidence base is largely absent from AI discussions, and it’s arguably the most direct empirical handle we have on cognitive architecture.
One of the biggest contributions cognitive science can make to AI is to educate the AI-smitten public as to the actual abilities and complexities of human cognition. At a minimum, we might hear fewer claims that humans think like LLMs.
Thanks, as someone trying to learn more about the work between the two fields and the most fruitful insights each can offer to the other, I appreciate this both this post and the initial one of the series, especially the links to papers and handbooks. Looking forward to reading more.
I enjoyed this post!
First, I really like all your articles that I've read so far!
> burstiness (i.e., the fact that if a word or phrase appears once in a document, it is much more likely to appear later in that document than in another random sample over the internet).
We plan to implement something like what you describe in [Hookmark](https://hookproductivity.com)'s recommendation system. We plan to extract documents lexical signatures from knowledge resources (not just documents) and use that in our recommendations. Basically, when you open Hookmark on the foreground resource (in any app), it shows you (a) what you've explicitly linked to it; and (b) what Hookmark thinks you want to access next based on the document and your browsing history. The key is that Hookmark does not just work with web pages and files, but with arbitrary information as long as it is [link-friendly](https://hookproductivity.com/help/integration/data-linkability-and-why-it-matters/). 24 professors, developers and CEO's co-signed the [Manifesto for Ubiquitous Linking](https://linkingmanifesto.org/) to make this possible.
What are other *rational* cues Hookmark could use? We're working on it and open to suggestions.
> This work can also be interpreted as a kind of rational analysis
Here's another rational analysis from cognitive science, which I call the heuristic relevance-signaling hypothesis who fittingly has a . It was first discovered by JR Anderson (1990. The adaptive character of thought.). Anderson developed the idea that memory is **rational** and adaptive. I'll be lazy and quote from [_Cognitive Productivity: Using Knowledge to Become Profoundly Effective_](https://leanpub.com/cognitiveproductivity/):
> I propose the heuristic relevance-signalling hypothesis as an explanation of the effects of distributed recall practice. This is an extension and adaptation of Anderson’s (1990) idea that memory is rational and adaptive.[^533] This is a “why” as opposed to a “how” (mechanistic) explanation. To understand this hypothesis, we need to do a bit of reverse-engineering of the problems (the requirements) that memory faces.
> Marcus (2012) noted: “Computer memory is much better than human memory because early computer scientists discovered a trick that evolution never did: organizing information by assigning every memory to a master map in which each bit of information to be stored is assigned a uniquely identifiable location in the computer’s memory vaults.” (p. 39) In contrast, given an arbitrary piece of information, such as the name of a historical figure, one cannot simply tell oneself, “I shall remember this information”. Even applying a mnemonic is often not very effective unless one also practices with it (Fritz et al., 2007).
> One of the implicit tricks that evolved in humans is not to make the accessibility of information an all-or-none matter, but a matter of degree. The mind must prioritize some stored information items over others. It makes high-priority items easier to use, in given contexts. Evolution implicitly faced a major challenge: how can the mind it builds determine what information ought to be made easiest to remember?
I.e., waht's the rational way to prioritize information indexing for future recall?
> The evolutionary trick is to prioritize information that higher-level (management) processes attempt to retrieve from memory. The more frequently a piece of information is requested, the more likely it is to be of importance. Therefore, mental indexing mechanisms must make that information more accessible. Conversely, information that is never requested is less likely to be relevant. Because attentional capacity (working memory) is limited, very little information is actually requested from memory compared to the information that is taken in perceptually. Attempts to retrieve information are much more reliable signals of relevance (or priority) than desires or instructions to be able to retrieve it in the future.
You wrote:
> We also showed that non- explanatory alternatives (such as true-but-not-currently-relevant statements about the situation)
Well, the heuristic relevance-signalling hypothesis is a non-explanatory alternative. However, explanation is a subset of this hypothesis, because every time you explain something you implicitly retrieve it from memory. Memory is involved in every explanation (and most of what human cognition involves).
> This series of studies illustrates how taking inspiration from a high- level aspect of human cognition,
ditto for the heuristic relevance-signalling hypothesis. What's the data? It's all the literature on :
* test-enhanced learning
* memory testing effects
* deliberate practice
I combine them in the concept of [productive practice](https://luccogzest.substack.com/p/the-cupa-framework-for-evaluating).
> However, many works since have highlighted challenges to the framework and interactions between these levels.
Yes, Marr's levels of analysis are incomplete. Aaron Sloman has refined them in [Sloman-1993-Prospects for AI is the General Science of Intelligence](https://cogaffarchive.org/Aaron.Sloman_prospects.pdf):
> (a) Analysis of requirements for an autonomous intelligent agent.
> (b) A design specification for a working system meeting the requirements in (a).
> (c) A detailed implementation or implementation specification for a working system.
> (d) Theoretical analysis of how the design specification and the implementational details ensure or fail to ensure satisfaction of the requirements.
> (e) Analysis of the neighbourhood in ‘design-space’.
Researchers tend to skip (e) which deals with the space of possible designs/minds. Sloman was the first AI researcher to emphasize this and emphasize information processing architectures, e.g., in [Sloman-1981-You Don't Need a Soft Skin to Have a Warm Heart Towards a Computational Analysis of Motives and Emotions](https://cogaffarchive.org/sloman-croucher-warm-heart.html).
This is valuable framing and agree the methodological contributions of cognitive science to AI are underappreciated. Though I’d add that cognitive neuropsychology specifically brings something beyond methodology: a century of natural experiments in humans that reveal what happens when specific cognitive processes are selectively disrupted. That evidence base is largely absent from AI discussions, and it’s arguably the most direct empirical handle we have on cognitive architecture.
One of the biggest contributions cognitive science can make to AI is to educate the AI-smitten public as to the actual abilities and complexities of human cognition. At a minimum, we might hear fewer claims that humans think like LLMs.