Ah great, delighted that you find this interesting (because I do!). And for the push back, thank you.

from below: “i really cannot overstate how much i appreciate your illumination of present approaches and terms, so that i might synchronise. so helpful, thank you”


https://tech.lgbt/@ngaylinn/111153861857195130

nate: “I agree, de-duplication is fundamental to intelligence. In fact, I like the suggestion that metaphor came before cognition. This fits really well with the story of evolution, since it’s much easier to see interacting molecules as ‘metaphors’ than to see them as ’thought’”

Yes. Though perhaps i’d say metaphor relates to specific related behavioural characteristics of a phenomena rather than the phenomena entirely – though in common with molecules, all behaviour is circumstantially evaluated: biology increases conditionality, as does interpretation (of map and territory), and cognition (which includes simulation). And i’ll touch on generalisation as distinct from metaphor later also.

“important distinction: system signal distinction” – (or “consciousness is distinct from its contents”)

“present approaches to ai ignore the system of cognition (the gap between brain and behaviour) – but there is a system. consider that present approaches mistake (memristor like) state persistence, for the system (’the files are not the app’). constraints exist at the context level (map scopes), for reasons. so while map scopes can be flattened, this ignores system architecture, and evaluable constraint (at each step)

.

nate: “I’d push back a little on the assertion that AI does not de-duplicate. I see what you mean, in that there’s no ‘sleep phase,’ where the system goes offline and its representations are actively rebuilt. However, the training process we use for those AIs is a bit like that, and totally different from how humans learn” “I think AI does de-duplicate, but it doesn’t take the important next step: building symbolic representations that it can manipulate and reason with. In fact, it doesn’t even have a workspace or a thinking agent to do that.”

Yes, on a technicality; and although you later revisit this statement, you are right in both cases (i ought to update my language to reflect this point)

So yes, there is some de-duplication, though not equal nor equivalent to the significance of what i’m referring. And yes, for each scope, although this process can be described in domain specific terms, by de-duplication I refer to the general mechanism (pertinent note: de-duplication also accounts for generalisation! pages have been updated)

“aside: i’m not totally wedded to ‘de-duplication’ as a term, but it seems to fit from first representations up through to paradigm level (i have some math for this), and is accessible”

“further, it aligns with a similar primitive pattern of territory. that is, a pattern which describes how map aligns with territory (i believe). consider that it is the fabric of map which aligns with territory, not the maps themselves, per se. as such, mathematics charts the territory of maps. the precision available by some mathematical accounts is a result of proximity ’to the metal’ of the substrate. i’ll add diagrams and list observable instances”

“important notes!”

“i think i use the term symbol differently, to refer to primitive representation. see: map scopes

“and, i think what you call stereotypes falls under what i call evaluable constraints – or an associated or tangential step, perhaps (i have an account of that mechanism too, so comparing notes will be interesting!)


https://tech.lgbt/@ngaylinn/111153817475158463

nate:

  1. “Important caveat: there’s no one algorithm that is ‘AI,’ there’s a huge variety of them. People have experimented with AI that ‘sleeps,’ AI that ’thinks,’ all sorts of weird variations.

  2. They’re just not in the mainstream yet, mostly because researchers haven’t gotten those ideas to pay off in a big way”

1 - On sleep (summarising). The important points are: 1) offline operations and edits; 2) with low likelihood of interruption 3) the cognitive goal is optimal alignment – coherence, throughout and across contexts, and the risks are high. The biological goal is physical, chemical logistics. Together, this is the silent high stakes, for cognition based survival.

“for biology, sleep is opportunistic. for ai, offline operations and edits are somewhat schedulable; and we can design the system to avoid pitfalls of biological survival, principally by hindsight and informed design, but also because ‘substrate and survival constraints are different’”

2 - Great! – I’d really love to know who I might be able to get in touch with in this space!

“a related question – given your experience with big tech, who would you give ‘real agi’ to (at least first, to avoid heavily resourced predators getting too much of a head start).

i’ll expand this if necessary, consider that i’m not talking about a descendent of llm etc”

musing: “the map is not the territory, the word is not the thing” — the word is not even the map – but an inherently ambiguous serialisable pointer

“for language based ai – without a specific system for (arbitrarily deep) contextual interpretation (and validation) – all derived form is mimicry, akin to stumbling upon a face on a wing”


https://tech.lgbt/@ngaylinn/111154278143277378

nate:

  1. Thinking about this more, there’s a subtle distinction that’s important.
  2. I think of the mind as generally operating with two basic primitives: stereotypes and symbols. For any concept in our ontology, we have both.
  3. A stereotype is just a sort of average over all impressions of a thing. This is what current AI specializes in.
  4. A symbol is a bit more like an entry in a database. It’s a logical object with properties and relationships. You can reason with them.
  5. A really good stereotype can approximate a symbol, but not exactly. This is why LLMs can answer many questions correctly, but make ridiculous assertions at other times. This is why they struggle with ’not’ and questions about categories and inference on topics they don’t have extensive training data for.
  6. I think when you say ‘de-duplication,’ maybe you mean both of these things? I consider them separately, mostly because I think AI does one and not the other.
  7. Folks like @garymarcus specialize in neurosymbolic AI to fill this gap.

1 - I really cannot overstate how much I appreciate your illumination of present approaches and terms, so that I might synchronise. So helpful, thank you.

2 - Term collision (well, partial collision actually): symbol. I initially used symbol as the fundamental of pattern, trend, etc; and not for “associated graph nodes” (by whatever term). I’ll need to review all notes/ communication with this in mind…

3 – Stereotype, this sounds a bit like a pass constraints. Constraint evaluation legal form? Or a definition of a “finite space of permissible”? (Permissible, possible, plausible; threshold/ tolerance/ conditionality?) (regardless, I think we see this mechanism from different sides. Interesting)

“musing: ok. so for this take, constraints are fundamental, though scoped. deep, aligned, coherent – perhaps present ai (that i am aware of) does so ‘shallowly’?”

So some additional distinctions:

  1. System signal
  2. Recall derivation
  3. Alignment distinction (this manifests in several ways)

On alignment (commonality) vs distinction.

“we identify by minimal-viable distinction. we reason by commonality (try to reason about two things which you have not yet established commonality)

(note: commonality fades from perception unless maintained, for reasons). so much to say on this

4 – Yes, for me, commonality defines graph relations (like a relational database, and your symbols), and we must walk nodes to reason deeply; and each node implicates and frames unique context, by association. This space leads to full asynchronous simulation of arbitrary (though constrained) geometric worlds. (I think the missing piece is my perspective of constraint mechanism, your stereotypes. Curious to see what you do presently)

Operationally, synchronisation is involved, often asynchronous to awareness of immediate surroundings. We do not do this unless safe, and able to “sit back” as a cognitive posture (though under pressure we can leverage previously isolated insights from the graph).

I can speak much more about this all, in terms of survival (cognitive operational latency); cognitive mode/ persistence formats (recall v derivation); operational logistics (consequences of simulation, gamma, batch edits, etc), and more.

5 – Nodes must comply with constraints. Once integrated, constraint evaluation is unconscious, and asynchronous to thought, with results communicated asynchronously as visceral sensation, so as not to obscure representational perception (think code smell)

ftr, when i talk about a specific feature with certainty, it is always open to pushback or feedback

“i perceive this geometrically, as a isolated special context and navigable landscape, and i have refined special constraints to evaluate from pre-first principles…”

most features comply with constraints across several domains of concern, including fundamental and biological constraints, fractional accounting of priors (constituents), concordance with causal peers, and instructive to observable (testable) downstream experience, behaviour, society, etc.

but! – while i’m “absolutely certain of the geometry of it” – i may easily miscommunicate!

please do question or push back as often as necessary to check what i mean. i’m going to begin introducing diagrams to help with this kind of thing

6 – Yes that’s right. I’m pointing to the general form. Though good point, I haven’t touched on mechanism of constraints yet… But your “stereotype” might be sufficiently equivalent (if we’re pointing at the same thing, the generalisation between implementations is likely sufficiently equal)

Fun, i’ll take a look.

7 – Ah, yes, i’ve previously followed gary on twitter (alt account). I’ll revisit now I better understand the term collisions you’ve helped point out. Thanks again.


beginning to introduce diagrams

edit: beginning here the pattern