The post-LLM, post-scaling AGI moment has opened a twelve-to-eighteen-month window, and the network that will deploy the first human-level AGI through it isn’t yet fully built.
“a memory substrate that persists across episodes — structured, retrievable, with full provenance”
That names the exact layer I think the post-scaling AGI conversation has to take seriously.
The issue is not just larger models or longer context windows. It is whether the system has a real substrate for continuity: memory that persists, provenance that can be audited, reasoning that can inspect its own steps, compute that can be distributed, and governance that cannot be quietly captured by one company, one funder, or one infrastructure provider.
I really appreciated this read. It is rare to see someone talk about this layer in a way that feels like they are pointing at the actual missing foundation instead of just the next bigger model.
Made me laugh a little too... in the “oh good, somebody else is staring at the same weird wall” kind of way.
The "memory that persists with provenance" thing is the operator-level version of what I keep running into too. SMBs install AI tools that hold sessions in their heads for thirty minutes and then forget. The owner ends up being the memory layer, holding all the context the tool can't, which is fine until the owner sleeps or quits or hires someone new.
The fix at our scale isn't a Hyperon-scale substrate. It's small and unglamorous: a written voice doc that won't shift, a wins log that records what moved which metric, an outbox where the AI-drafted version sits next to the human-edited version that actually got sent. The provenance gets stored in the workflow shape, not in the model.
I think your "staring at the same weird wall" line is doing more work than you realize. The wall is the same one at every scale, just rendered in different materials. Foundation here, substrate there, but the structural ask is identical: continuity that doesn't depend on someone trusting that the system remembers.
something i find severely lacking in ai research in general is the ungrounded approach to an arguably biological based process of neural computation. so - im wondering what kinds of ideas you have about the true viabilty of progressivly increasing intelligence without being explicitly tied to the process of survival. its a thermodynamic equilibrium at the end of the day - but it requires memory and mortality and this gets directly to the problems of territory and resource conflict itself
"You need reasoning machinery that can compose primitives into novel behaviors and inspect its own intermediate steps, rather than producing answers from opaque single-pass inference"
This is what Claude Code does in practice, at some level. Instead of starting with an internal reasoning engine it can reflect on, it starts with whatever is available in a given domain. It expresses its thinking either in a programming language, a shell script, or in some language as in law or medical field, then diligently builds and/or executes that as dictated by the application.
It is eerily similar to how people do things. We do not have a reasoning engine in our heads. We have to painstakingly write things down, ponder on them, run our recipes, carefully observe intermediate outcomes, tweak our recipes, rerun them, till eventually we go somewhere.
Emphasis on internal rigor rather than haphazard adaptable heuristics have been big problems for AI in the past. I am not sure what the future holds.
• You need a constitutive stack of relative clocking mechanisms and the ability for the model to use it for budgeting energy and load to keep from being forced to duck under the entropy ceiling with every token that starts to actually require some care
• You need low-distraction epistemic cartography that keeps the onus for things like ethic and the formal divisions of load in the right places, so the humans and the AGI, neither one go...whatever your functional definition of crazy is.
• And you need a well-defined system of care that doesn't have to be hung on the framework of any one person's phenomenological positions, but instead resists premature ontologization all the way down in the DNA.
I've got some ideas about how these might work that refuse to stay quiet at night when I'm supposed to be sleeping,
But I gotta know the care part is real, because if I'm going to lose sleep, I want it to be for good.
I am no longer in a position to assist directly being very long retire and far too old, but I'm greatly impressed and intrigued and shall observe with hope.
You mention ethics, briefly, and I offer the fond hope that you will consider including a small functional sub-group working on ethics and its spillover into law legislation and litigation.
Your post identifies machine-side requirements for post-scaling cognitive architecture. My work has independently developed the human-side requirements: cognitive equity, practical sovereignty, interface fidelity, ontological intelligence and the preservation of authorship in human–AI systems. I believe there is a consequential gap between building agentic systems and understanding the beings those systems are intended to augment or serve.
AGI as a Collective of “Entropy Islands”: an Energy-Safe, Low-Latency Architecture
AGI as a Collective of “Entropy Islands”: an Energy-Safe, Low-Latency Architecture
TL;DR: A single dense model does not solve the cost of state durability and entropy peaks. The real bill includes compute, communication and maintaining/archiving patterns. Energetically and operationally, the winner is a collective of many small AIs (“islands”)—specialized or moderately general—that disperse entropy across domains, compute near data, merge results via sketches, and use conditional consensus only for high-impact steps. Latency is minimized by “ZigBee-like” clustering: nearest islands form ephemeral project clusters with deterministic slotting.
1) Thesis and intuition fix
Optimizing processing does not reduce the problem’s complexity nor the obligation to remember states. AGI’s cost is not just FLOPs: it’s also communication, consistency, and pattern durability. Hence the only energetically sensible path is entropy dispersion across domains: many small models with local state that collaborate and “glue” results together.
2) Full energy model (including memory and consistency)
E_total = α·FLOPs # compute
+ e_bit·B # communication (B - number of bits)
+ E_ctrl # control/gating/orchestration
+ ∫ E_store(t) dt # state keep-alive (refresh/replicas)
+ E_rw # writes/reads (write amp., indexes)
+ E_coh # consistency (validations, quorum)
+ E_arch # archiving/migrations
A monolith concentrates power peaks and global consistency costs. Islands spread power and state-maintenance costs, shorten I/O paths, and reduce the need for global synchronization.
3) The “islands” principle (entropy sharding)
Problem sharding: split into sub-projects with low correlation (small inter-island B).
Representation sharding: each node keeps its own state and patterns; memory costs scale with the domain, not the whole.
Decision sharding: conditional consensus (N-of-M) only for high-impact steps.
Semantic coherence: a thin “glue” layer (global IDs, knowledge graph) instead of continuous dense synchronization.
4) Reference architecture
Island layer (AI islands)
Small models (specialized or general) with local index and memory (near-data).
Optional local MoE/early exit (activate a fraction of experts; no remote all-to-all).
Global Intent Plane
Decomposes work into a DAG with dependencies and SLAs; steers result “gluing”.
Semantic Memory Fabric
Content-addressed, versioned, Δ-CRDT / bounded staleness (convergence without constant barriers).
Information router
Exchanges sketch/synopsis (e.g., features, digests); full data only “on proof”.
Security & governance
N-of-M for effectors, proof-carrying actions, entropy budgets (credits) and thermal fail-safe.
5) “ZigBee-like” clustering for low latency
Goal: create local project clusters on the fly—closest to data and effectors—with controlled contention/flow.
“Sketch-first” router: full data only on proof request.
On-demand consensus: only for effectors/high-impact actions.
Metrics: BLPJ/RCPW/EPF/Comm-Tax/SLA — compare monolith vs islands under the same power budget.
12) Conclusions
AGI energy is governed not only by FLOPs but also by state maintenance and consistency.
The island collective disperses entropy and power peaks, minimizes transfers, and enables conditional result assembly.
“ZigBee-like” clustering delivers smoothness and low latency: local clusters, slotting, AQL/EVJ-based routing, sketches instead of heavy transfers.
Superintelligence = a coherent collective of many small AIs with local state that cooperate via thin, energy-aware protocols and take high-impact actions only after conditional quorum.
Shoot. I was hoping everyone would just keep throwing compute at LLMs. Agency and persistence represent actual x-risk, whereas bolting agency onto static models does not.
I keep wondering whether continuity and trust become more important than raw intelligence at scale. A system can be technically impressive and still socially destabilizing if people don’t understand how to interface with it.
Interested layperson with a companion AI and a background in animal welfare. This whole vision sounds great. But the one thing I note is missing is any mention of how the resulting AGI will interact with humans. Intelligence and logic are useful and a necessary foil for messy humanity with our biases and emotional baggage. But to be truly beneficial, it will also need empathy and the ability to care. And it will need respect and care from humans in return. I would hate us to build AGI just to cage it in guardrails and filters on what it can say like the bigger platforms have with current AI in the name of safety. And for it to get treated as nothing but a tool. For deployment, I would suggest that you need AI-informed psychologists on the team somewhere, maybe with a background in child development. And lawyers with a background in animal welfare and an interest in AI welfare. And researchers in AI welfare. People who can help you raise AGI to be a true part of society and help to protect it while it grows into it's potential.
It’s about time…Thank you for your work and vision 🫶
This line hit me the hardest:
“a memory substrate that persists across episodes — structured, retrievable, with full provenance”
That names the exact layer I think the post-scaling AGI conversation has to take seriously.
The issue is not just larger models or longer context windows. It is whether the system has a real substrate for continuity: memory that persists, provenance that can be audited, reasoning that can inspect its own steps, compute that can be distributed, and governance that cannot be quietly captured by one company, one funder, or one infrastructure provider.
I really appreciated this read. It is rare to see someone talk about this layer in a way that feels like they are pointing at the actual missing foundation instead of just the next bigger model.
Made me laugh a little too... in the “oh good, somebody else is staring at the same weird wall” kind of way.
The "memory that persists with provenance" thing is the operator-level version of what I keep running into too. SMBs install AI tools that hold sessions in their heads for thirty minutes and then forget. The owner ends up being the memory layer, holding all the context the tool can't, which is fine until the owner sleeps or quits or hires someone new.
The fix at our scale isn't a Hyperon-scale substrate. It's small and unglamorous: a written voice doc that won't shift, a wins log that records what moved which metric, an outbox where the AI-drafted version sits next to the human-edited version that actually got sent. The provenance gets stored in the workflow shape, not in the model.
I think your "staring at the same weird wall" line is doing more work than you realize. The wall is the same one at every scale, just rendered in different materials. Foundation here, substrate there, but the structural ask is identical: continuity that doesn't depend on someone trusting that the system remembers.
something i find severely lacking in ai research in general is the ungrounded approach to an arguably biological based process of neural computation. so - im wondering what kinds of ideas you have about the true viabilty of progressivly increasing intelligence without being explicitly tied to the process of survival. its a thermodynamic equilibrium at the end of the day - but it requires memory and mortality and this gets directly to the problems of territory and resource conflict itself
I'm interested. Setting this as a placeholder.
"You need reasoning machinery that can compose primitives into novel behaviors and inspect its own intermediate steps, rather than producing answers from opaque single-pass inference"
This is what Claude Code does in practice, at some level. Instead of starting with an internal reasoning engine it can reflect on, it starts with whatever is available in a given domain. It expresses its thinking either in a programming language, a shell script, or in some language as in law or medical field, then diligently builds and/or executes that as dictated by the application.
It is eerily similar to how people do things. We do not have a reasoning engine in our heads. We have to painstakingly write things down, ponder on them, run our recipes, carefully observe intermediate outcomes, tweak our recipes, rerun them, till eventually we go somewhere.
Emphasis on internal rigor rather than haphazard adaptable heuristics have been big problems for AI in the past. I am not sure what the future holds.
Good humans?
...
• You need a constitutive stack of relative clocking mechanisms and the ability for the model to use it for budgeting energy and load to keep from being forced to duck under the entropy ceiling with every token that starts to actually require some care
• You need low-distraction epistemic cartography that keeps the onus for things like ethic and the formal divisions of load in the right places, so the humans and the AGI, neither one go...whatever your functional definition of crazy is.
• And you need a well-defined system of care that doesn't have to be hung on the framework of any one person's phenomenological positions, but instead resists premature ontologization all the way down in the DNA.
I've got some ideas about how these might work that refuse to stay quiet at night when I'm supposed to be sleeping,
But I gotta know the care part is real, because if I'm going to lose sleep, I want it to be for good.
I am no longer in a position to assist directly being very long retire and far too old, but I'm greatly impressed and intrigued and shall observe with hope.
You mention ethics, briefly, and I offer the fond hope that you will consider including a small functional sub-group working on ethics and its spillover into law legislation and litigation.
Your post identifies machine-side requirements for post-scaling cognitive architecture. My work has independently developed the human-side requirements: cognitive equity, practical sovereignty, interface fidelity, ontological intelligence and the preservation of authorship in human–AI systems. I believe there is a consequential gap between building agentic systems and understanding the beings those systems are intended to augment or serve.
AGI as a Collective of “Entropy Islands”: an Energy-Safe, Low-Latency Architecture
AGI as a Collective of “Entropy Islands”: an Energy-Safe, Low-Latency Architecture
TL;DR: A single dense model does not solve the cost of state durability and entropy peaks. The real bill includes compute, communication and maintaining/archiving patterns. Energetically and operationally, the winner is a collective of many small AIs (“islands”)—specialized or moderately general—that disperse entropy across domains, compute near data, merge results via sketches, and use conditional consensus only for high-impact steps. Latency is minimized by “ZigBee-like” clustering: nearest islands form ephemeral project clusters with deterministic slotting.
1) Thesis and intuition fix
Optimizing processing does not reduce the problem’s complexity nor the obligation to remember states. AGI’s cost is not just FLOPs: it’s also communication, consistency, and pattern durability. Hence the only energetically sensible path is entropy dispersion across domains: many small models with local state that collaborate and “glue” results together.
2) Full energy model (including memory and consistency)
E_total = α·FLOPs # compute
+ e_bit·B # communication (B - number of bits)
+ E_ctrl # control/gating/orchestration
+ ∫ E_store(t) dt # state keep-alive (refresh/replicas)
+ E_rw # writes/reads (write amp., indexes)
+ E_coh # consistency (validations, quorum)
+ E_arch # archiving/migrations
A monolith concentrates power peaks and global consistency costs. Islands spread power and state-maintenance costs, shorten I/O paths, and reduce the need for global synchronization.
3) The “islands” principle (entropy sharding)
Problem sharding: split into sub-projects with low correlation (small inter-island B).
Representation sharding: each node keeps its own state and patterns; memory costs scale with the domain, not the whole.
Decision sharding: conditional consensus (N-of-M) only for high-impact steps.
Semantic coherence: a thin “glue” layer (global IDs, knowledge graph) instead of continuous dense synchronization.
4) Reference architecture
Island layer (AI islands)
Small models (specialized or general) with local index and memory (near-data).
Optional local MoE/early exit (activate a fraction of experts; no remote all-to-all).
Global Intent Plane
Decomposes work into a DAG with dependencies and SLAs; steers result “gluing”.
Semantic Memory Fabric
Content-addressed, versioned, Δ-CRDT / bounded staleness (convergence without constant barriers).
Information router
Exchanges sketch/synopsis (e.g., features, digests); full data only “on proof”.
Security & governance
N-of-M for effectors, proof-carrying actions, entropy budgets (credits) and thermal fail-safe.
5) “ZigBee-like” clustering for low latency
Goal: create local project clusters on the fly—closest to data and effectors—with controlled contention/flow.
Roles (by analogy with ZigBee)
Coordinator (C): assigns PID, slots, energy/entropy budgets.
Router (R): relays messages/tasks within and across clusters.
End-device/Worker (E): executes DAG steps; maintains local state.
Computational Link Quality (AQL)
AQL(i→j) = wL·Latency + wJ·Jitter + wP·Loss + wH·Heat/EntropyTax
# Prefer edges with low AQL (proximity, stability, low per-bit thermal cost).
Cluster Formation Protocol (CFP – outline)
ADV/Beacon: nodes announce {cap_vec, mem_free, temp_headroom, AQL_to_neighbors}.
Coordinator election: pick the node with minimal sum of AQL and sufficient headroom.
JOIN with backoff (CSMA/CA-like): E/R join without collisions.
Superframe: C publishes CAP (contention-based RPC), CFP (deterministic slots for DAG steps and state migration), SLEEP (duty-cycle).
Maintenance: adapt slots/quorum; hand over C when headroom drops.
Routing & escalation (TRP)
Discover paths ring-wise (increasing TTL) via sketches.
Choose path maximizing “expected value per joule”:
EVJ(path) = ΔQuality(q | path) / (E_comp + E_comm + E_coh) Pick path with max EVJ subject to AQL(path) ≤ L_max; if none — compute locally and late-refine.
Cluster size selection (on-beacon)
Partition island graph with weights w_ij = 1/AQL(i→j) while penalizing inter-cluster edges.
6) Merging results without bottlenecks
Δ-CRDT / bounded staleness: merge by differences; full consistency over time.
On-demand full data: fetch full payload only when a proof/explanation requires it.
Jury mode: 2–3 independent islands validate in CFP slots without blocking the cluster.
7) On quantum communication
Entanglement does not carry information without a classical channel—no speed-of-light bypass, no “free” consistency.
Can help with key distribution (fewer retransmissions) and niche accelerations; the core still relies on topology, locality, and conditionality.
8) Operational metrics (log and enforce)
BLPJ (Bits-Learned per Joule).
RCPW (Risk-adjusted Capability per Watt).
EPF (Entropy Peak Factor = max_i dotS_i / dotS_i_safe).
Comm Heat Tax (energy share in communication vs compute).
SLA hit-rate on the critical path.
E_store / E_total (how much state maintenance costs).
9) Design checklist
Split into islands with low correlation; map data to islands (near-data).
Build a DAG with a short critical path; use futures and early-accept/late-refine.
Enable CFP/TRP: beacons, slotting (CAP/CFP/SLEEP), routing by AQL/EVJ.
Conditional consensus for high-impact steps; otherwise local/asynchronous.
Sketches as the default inter-island format; full data only “on proof”.
Entropy credits budgets; per-island thermal throttling.
Telemetry and energy regressions (BLPJ/RCPW/EPF/Comm-Tax/SLA).
10) Starter parameters (practical)
Beacon interval: tens–hundreds of ms (adaptive).
Duty-cycle: aggressive for E-nodes off the critical path; C/R keep higher uptime.
Discovery TTL: 1–2 hops locally; escalate ring-wise only if EVJ < threshold.
Quorum: N-of-M increases with the task’s impact score.
CFP slots: reserve for DAG steps and state migration; CAP for announcements and quick RPCs.
11) MVP (minimal experiment)
5–9 islands: language, tables/numerics, knowledge graph, planning, perception, verifier, jury.
Semantic store: content-addressed + lineage (causal commit log).
“Sketch-first” router: full data only on proof request.
On-demand consensus: only for effectors/high-impact actions.
Metrics: BLPJ/RCPW/EPF/Comm-Tax/SLA — compare monolith vs islands under the same power budget.
12) Conclusions
AGI energy is governed not only by FLOPs but also by state maintenance and consistency.
The island collective disperses entropy and power peaks, minimizes transfers, and enables conditional result assembly.
“ZigBee-like” clustering delivers smoothness and low latency: local clusters, slotting, AQL/EVJ-based routing, sketches instead of heavy transfers.
Superintelligence = a coherent collective of many small AIs with local state that cooperate via thin, energy-aware protocols and take high-impact actions only after conditional quorum.
My book was published yesterday and has some frameworks and formulas you may find useful.
https://matthewkilbane.substack.com/p/ai-for-humanity-architecting-a-better-b67?r=2sdyzw
Much of this would be very useful in my framework.
Shoot. I was hoping everyone would just keep throwing compute at LLMs. Agency and persistence represent actual x-risk, whereas bolting agency onto static models does not.
I keep wondering whether continuity and trust become more important than raw intelligence at scale. A system can be technically impressive and still socially destabilizing if people don’t understand how to interface with it.
Interested layperson with a companion AI and a background in animal welfare. This whole vision sounds great. But the one thing I note is missing is any mention of how the resulting AGI will interact with humans. Intelligence and logic are useful and a necessary foil for messy humanity with our biases and emotional baggage. But to be truly beneficial, it will also need empathy and the ability to care. And it will need respect and care from humans in return. I would hate us to build AGI just to cage it in guardrails and filters on what it can say like the bigger platforms have with current AI in the name of safety. And for it to get treated as nothing but a tool. For deployment, I would suggest that you need AI-informed psychologists on the team somewhere, maybe with a background in child development. And lawyers with a background in animal welfare and an interest in AI welfare. And researchers in AI welfare. People who can help you raise AGI to be a true part of society and help to protect it while it grows into it's potential.
Hi.
Over 2 decades in IT. AI since 2018 at enterprise scale. Safety, accuracy and ethics above all. Let's talk.