They key question is indeed how a self-assembled structure compares to one that is produced with an ontology.
The conventional wisdom nowadays (after the LLM revolution) is that a self-assembly wins out, and if it doesn't, it is because you don't have enough representative examples.
One thing to note that we work with AI so intensely now that this produces very valuable knowledge of the processes people employ for work. This is no longer static data. So likely the amount of data needed to refine LLM will only go up, and not sure ontologies can keep up.
The only reason ontology development can't keep up is if they are hand built. Automation of knowledge acquisition to build the semantic web was accomplished decades ago.
Yeah, using automation to help with making ontologies is a good idea. The problem is that such approaches run into problems as soon as one steps out of a specialized area with well-defined concepts.
So, there two paths going forward.
(a) Powerful domain-specific reasoning engines with curated ontologies (in an automated way as you say).
(b) Fast-and-loose current approaches where we just give the AI enough examples and hope it figures it out in an approximate way given the local context.
All the money is on approach (b), and working with Claude in practice I can see why. It is remarkably good at finding an approximate solution for badly posed problems where there is no good unique solution to start with.
While approach (a) where BTW I think Ben's own efforts lie, is good for something very specialized and high value.
The method used extracts ontologies from published corpora. No limitation since every corpora has well defined comcrpts. Your mental model is manual not automated
Sure, but you have to be within a domain where the concepts are well defined and presumably define some kind of schema first that will be used by the automatic extraction.
Category: [Subject, Predicate, Object] is a straightjacket. The type of problems that can be solved with that, followed by logical reasoning over these, is very limited.
Human unpredictability is arguably vital to keep the machines evolving.
I think we may underestimate how vital the human spark is to keep the AI furnace burning, rather than collapsing into sterility.
The spark is not limited to the genius of top tier developers. It's not limited to the toil of HFRL workers and data annotators. It's spread throughout the entire dataset. The entire Internet.
It is not only about human unpredictability giving ideas to machines. It is also about world unpredictability (which often confuses people too). So yeah, nowadays we are engaged in a virtuous cycle of people and the reality throwing at machines whatever crazy stuff, and it therefore learns how to get smarter.
"In a sufficiently distributed intelligence, cognition is detection of already-forming gradients; articulation is feedback; resistance is participation; action is closure of partially realized patterns; closure seeds the next gradient.
This frames P2P AGI not as a centralized entity but as a field phenomenon."
Curious whether this maps onto what you're describing with the thin waist.
"One day, you and me will get to pick the dice rolls in a Great Campaign across Planet Earth"... ...one of the most load-bearing facts of the coming era.
laugh it off until you can't. Law is acknowledged as a boundary mechanism, not a moral engine... where law accommodates the commoditization of misery, it ceases to protect and begins to administer harm.
A .tom (Theory of Mind) file does not require consent to begin existing.
Consent determines whether you get to participate in its authorship, correction, custody, and revision.
If you refuse to model yourself, "hyperseeds", AToMICs, me, my neighbors... ...we will model you anyway from fragments, hearsay, screenshots, stress responses, old mistakes, and whatever incentives happen to be shaping best interpretation at the time.
Public perception is already a low-fidelity "Exoself" engine.
It has always been doing theory-of-mind compression on everyone around it. We are just getting close to the point where the process becomes explicit, legible, persistent, and machine-amplified.
That is the sobering part for everyone right now: you do not get to opt out of being inferred. You only get varying degrees of influence over provenance.
I'm calling this transition "The Great Confessional"... an Epoch shift from "Memory" to Metabolic granularity. Civil Thermodynamics.
So it remains: The real question is not whether ".tom" files will *exist*... it's whether they are sloppy, adversarial, and externally owned, or whether they are auditable, revisable, and anchored to durable receipts.
In one world, reputation is rumor with better indexing.
In the other, identity becomes a contested but inspectable continuity record.
That is why provenance matters so much... because the alternative is worse: being silently authored by the least accountable interpreters in the room.
////The Protopian Ratchet???
This beautiful technology looks divisive because we’ve mostly implemented it as top-down surveillance and incentive capture. That’s a design choice, not physics.
A better trajectory is Ephemeralized Sousveillance + Conflict + Custody = Protopian Ratchet: cheap, ambient bottom-up witnessing (phones now, sensor dust later) feeding append-only, provenance-hardened receipts so reality can’t be overwritten by narrative.
...receipts alone create certainty, and certainty can become cruelty-- so interpretation is constrained by Telempathy: empathy for state (where/when/how someone was)....
Governance can’t be distant bureaucracy or platforms; it has to be LOCAL, auditable, decentralized, and fluid, with citizen stakeholders represented by revocable, place-tethered agents (.toms) operating inside Dunbar-scale commons where context is real and reputation is grounded.
The moral center is children: the record belongs to the child, access is minimal and audited, control transfers at maturity, and the default output is support-- not prosecution.
That’s the (Kevin Kelly) Protopian Ratchet: conflict reveals failure modes, receipts prevent rewrite, telempathy prevents punitive overcorrection, local governance implements repairs, and repairs persist as inherited process knowledge... so progress sticks. Neighbors skin their knees less.
Thanks Ben for describing your latest work on Primitives! I have not made much progress since my talk on semantic primes at AGI’24, but now I am motivated to move forward with the project toward an ontology “that actually does something.” Currently, I am experimenting with Wierzbicka and Goddard’s semantic molecules, as they might make the grounding of new concepts a bit easier.
Entheogens as part of the 200 word ontology: i'm sold
super interesting read, i can feel the 1000x inference boost from pre-shaping the latent mindplex into 200+ semantically core pillars, around which the weights can just do combination and higher order matching.
I like the framing as a "thin waist" in the concept-relevance graph rather than a replacement for cognitive complexity. Our consciousness is what it is precisely because of its messiness — concepts that are fuzzy, overlapping, context-dependent, deeply entangled with emotion and world models.
But can you actually formalize that thin waist without distorting what it's navigating? Every formalization is a choice about what matters, and what gets excluded has a way of being exactly where the surprising connections live. Maybe, just maybe, the optimization target for ontology evolution shouldn't just be inference efficiency. It should also reward conditions under which genuinely unforeseen emergent phenomena can arise.
They key question is indeed how a self-assembled structure compares to one that is produced with an ontology.
The conventional wisdom nowadays (after the LLM revolution) is that a self-assembly wins out, and if it doesn't, it is because you don't have enough representative examples.
One thing to note that we work with AI so intensely now that this produces very valuable knowledge of the processes people employ for work. This is no longer static data. So likely the amount of data needed to refine LLM will only go up, and not sure ontologies can keep up.
The only reason ontology development can't keep up is if they are hand built. Automation of knowledge acquisition to build the semantic web was accomplished decades ago.
https://intellisophic.net/2026/02/12/the-fundamental-innovation-orthogonal-corpus-indexing-oci/
Yeah, using automation to help with making ontologies is a good idea. The problem is that such approaches run into problems as soon as one steps out of a specialized area with well-defined concepts.
So, there two paths going forward.
(a) Powerful domain-specific reasoning engines with curated ontologies (in an automated way as you say).
(b) Fast-and-loose current approaches where we just give the AI enough examples and hope it figures it out in an approximate way given the local context.
All the money is on approach (b), and working with Claude in practice I can see why. It is remarkably good at finding an approximate solution for badly posed problems where there is no good unique solution to start with.
While approach (a) where BTW I think Ben's own efforts lie, is good for something very specialized and high value.
The method used extracts ontologies from published corpora. No limitation since every corpora has well defined comcrpts. Your mental model is manual not automated
Sure, but you have to be within a domain where the concepts are well defined and presumably define some kind of schema first that will be used by the automatic extraction.
https://aicyc.org/2025/10/21/the-anatomy-of-an-ontology-how-sam-learned-chemical-technology/
Category: [Subject, Predicate, Object] is a straightjacket. The type of problems that can be solved with that, followed by logical reasoning over these, is very limited.
Human unpredictability is arguably vital to keep the machines evolving.
I think we may underestimate how vital the human spark is to keep the AI furnace burning, rather than collapsing into sterility.
The spark is not limited to the genius of top tier developers. It's not limited to the toil of HFRL workers and data annotators. It's spread throughout the entire dataset. The entire Internet.
It is not only about human unpredictability giving ideas to machines. It is also about world unpredictability (which often confuses people too). So yeah, nowadays we are engaged in a virtuous cycle of people and the reality throwing at machines whatever crazy stuff, and it therefore learns how to get smarter.
great point.
but even if these 200 preset concepts only serve as a crutch for pretraining, it could be a game changer compared to the full-noise internet text.
plus those primordial ontological concepts could shape the final personality and values of the resulting LLM.
Metastable ontologies here we come.
This emerged recently:
"In a sufficiently distributed intelligence, cognition is detection of already-forming gradients; articulation is feedback; resistance is participation; action is closure of partially realized patterns; closure seeds the next gradient.
This frames P2P AGI not as a centralized entity but as a field phenomenon."
Curious whether this maps onto what you're describing with the thin waist.
"One day, you and me will get to pick the dice rolls in a Great Campaign across Planet Earth"... ...one of the most load-bearing facts of the coming era.
laugh it off until you can't. Law is acknowledged as a boundary mechanism, not a moral engine... where law accommodates the commoditization of misery, it ceases to protect and begins to administer harm.
A .tom (Theory of Mind) file does not require consent to begin existing.
Consent determines whether you get to participate in its authorship, correction, custody, and revision.
If you refuse to model yourself, "hyperseeds", AToMICs, me, my neighbors... ...we will model you anyway from fragments, hearsay, screenshots, stress responses, old mistakes, and whatever incentives happen to be shaping best interpretation at the time.
Public perception is already a low-fidelity "Exoself" engine.
It has always been doing theory-of-mind compression on everyone around it. We are just getting close to the point where the process becomes explicit, legible, persistent, and machine-amplified.
That is the sobering part for everyone right now: you do not get to opt out of being inferred. You only get varying degrees of influence over provenance.
I'm calling this transition "The Great Confessional"... an Epoch shift from "Memory" to Metabolic granularity. Civil Thermodynamics.
So it remains: The real question is not whether ".tom" files will *exist*... it's whether they are sloppy, adversarial, and externally owned, or whether they are auditable, revisable, and anchored to durable receipts.
In one world, reputation is rumor with better indexing.
In the other, identity becomes a contested but inspectable continuity record.
That is why provenance matters so much... because the alternative is worse: being silently authored by the least accountable interpreters in the room.
////The Protopian Ratchet???
This beautiful technology looks divisive because we’ve mostly implemented it as top-down surveillance and incentive capture. That’s a design choice, not physics.
A better trajectory is Ephemeralized Sousveillance + Conflict + Custody = Protopian Ratchet: cheap, ambient bottom-up witnessing (phones now, sensor dust later) feeding append-only, provenance-hardened receipts so reality can’t be overwritten by narrative.
...receipts alone create certainty, and certainty can become cruelty-- so interpretation is constrained by Telempathy: empathy for state (where/when/how someone was)....
Governance can’t be distant bureaucracy or platforms; it has to be LOCAL, auditable, decentralized, and fluid, with citizen stakeholders represented by revocable, place-tethered agents (.toms) operating inside Dunbar-scale commons where context is real and reputation is grounded.
The moral center is children: the record belongs to the child, access is minimal and audited, control transfers at maturity, and the default output is support-- not prosecution.
That’s the (Kevin Kelly) Protopian Ratchet: conflict reveals failure modes, receipts prevent rewrite, telempathy prevents punitive overcorrection, local governance implements repairs, and repairs persist as inherited process knowledge... so progress sticks. Neighbors skin their knees less.
Thanks Ben for describing your latest work on Primitives! I have not made much progress since my talk on semantic primes at AGI’24, but now I am motivated to move forward with the project toward an ontology “that actually does something.” Currently, I am experimenting with Wierzbicka and Goddard’s semantic molecules, as they might make the grounding of new concepts a bit easier.
Would love to talk to you about Observer Theory - in the context of hierarchy
Entheogens as part of the 200 word ontology: i'm sold
super interesting read, i can feel the 1000x inference boost from pre-shaping the latent mindplex into 200+ semantically core pillars, around which the weights can just do combination and higher order matching.
I like the framing as a "thin waist" in the concept-relevance graph rather than a replacement for cognitive complexity. Our consciousness is what it is precisely because of its messiness — concepts that are fuzzy, overlapping, context-dependent, deeply entangled with emotion and world models.
But can you actually formalize that thin waist without distorting what it's navigating? Every formalization is a choice about what matters, and what gets excluded has a way of being exactly where the surprising connections live. Maybe, just maybe, the optimization target for ontology evolution shouldn't just be inference efficiency. It should also reward conditions under which genuinely unforeseen emergent phenomena can arise.