01 · Substrate
Computer
Voltage transitions across transistors, organized into Boolean gates, composed into arithmetic units, memory hierarchy, and I/O. Does math. Only does math.
silicon gates deterministic
AI is software.
A file and a runtime performing Digital Information Synthesis. Every synthesizer has a coordinate. Every coordinate has an energy cost.
Scroll
1970s
The chemistry, changed.
Pre-1974
CFCs as engineering triumph
Non-toxic, non-flammable, low-cost. CFCs looked like pure success — refrigerators, AC, hairspray.
1974 — 2010
Ozone hole · Montreal Protocol
The cost was hidden in a layer no one was looking at. Once measured, the world moved fast — without ever stopping refrigeration.
Today
Refrigerators still refrigerate
We changed the chemistry. We didn't stop the appliance.
2020s
The arms race, named.
2024
Every query through a GPU
Route everything through 700 W silicon. The results are impressive. The infrastructure spend is $2T and climbing.
2026 — 2030
945 TWh · 1010× above the floor
General-purpose silicon runs roughly ten billion times above the thermodynamic minimum. Most everyday queries have closed-form answers and never need the generative path.
The Move
Compute differently.
The work is real. The compute path it runs on is the choice the field can change.
The map of the moves is the Periodic Stack.
Foundation
Three terms. Precisely.
Name what is being mapped. Precise definitions give every subsequent claim a fixed referent.
02 · Orchestration
Software
Orchestration of the computer's math toward a designer's goal. A specification of which operations happen, in which order, with which data. A file plus a runtime.
file runtime impedance mismatch
03 · AI = DIS
Digital Information Synthesis
AI is software. Not a separate category, not magic. A subset of software characterized by synthesizing information in digital form, constrained by the universal cost of bridging algorithm to physical computation.
synthesis ⊂ software
AI ⊆ Software ⊆ Algorithm × Hardware
The two-bound universe · everything else is between
The Three Zones
Where information lives.
Every problem a synthesizer addresses lives in one of three zones — named by the relationship between math and the words describing it.
Z₁ math = words
Bounded determinism
Closed-form problems. Arithmetic, formal parsing, proof checking, deterministic control. Outputs exactly derivable from inputs. Verification is a comparison, not a measurement.
L₀–L₁ full verify
do not compute → stored answers should not be recomputed. If a closed-form exists, evaluate it directly.
Z₂ math ≈ words
Bounded inference
Authoritative structure constrains a space of valid answers without fully determining them. Law, regulated medicine, tax, contract interpretation, credit. The bridge zone.
L₁–L₂ citation
do not compute → Z₂ inference where a Z₁ substrate would do. Retrieve, don't regenerate.
Z₃ math ≠ words
Unbounded generation
No authoritative ground truth. Open-ended creative work, aesthetic generation, emergent dynamics. Only what tends to happen, with statistical regularities over human-produced artifacts.
L₂–L₂ₘₐₓ statistical
do not compute → Z₃ generation without named grounding. Cost scales with distance to the grounding point.
The transitions — Z₁↔Z₂ and Z₂↔Z₃ — are where most real synthesizers actually live.
The Physics
Shannon, Landauer, and the thermodynamic floor.
Information is physical. Erasing it costs energy. The gap between where we operate and where physics says we could is vast — and recoverable.
1948
Shannon
Information is physical. Every channel has a maximum rate at which it can be transmitted reliably, bounded by noise, bandwidth, and power.
1961
Landauer
Erasing one bit of information dissipates a minimum quantity of heat. Not an engineering estimate. A consequence of the second law.
2026
OpenIE
Measure every operation. Route to the physics-permitted optimal path. You cannot optimize what you cannot see.
Emin = kBT · ln 2 ≈ 2.87 × 10⁻²¹ J / bit
Landauer's limit at room temperature (300 K)
General-purpose silicon runs roughly 10 billion times above the thermodynamic minimum. Generative inference per output bit runs even higher — in the 1018–1020 range above the floor. Every step down that curve is recoverable energy.
E(x) = ∑p ∈ P(c(x)) θ(p) · μ(p, H)
Energy per input · Landauer floor θ(p) × impedance-mismatch factor μ on hardware H
Two levers reduce per-input energy: change which primitives are in the composition, or change the mismatch on the hardware.
The field is converging
Three research traditions are arriving at the same impedance-mismatch argument from different angles. Energy-based models score whole answers by how well they satisfy constraints — reasoning as optimization, not next-token prediction. Joint-embedding predictive architectures learn world representations directly, skipping pixel-level reconstruction. Formal verifiers like Lean check whether a proof compiles, not whether it sounds right. The convergence isn't coincidence — it's the same answer to "what's the right mechanism for the work" arriving from three rooms at once. Language for communication. Energy for constraints. Formal systems for verification. The Periodic Stack is the map; the layered reasoning stack is what gets built on it.
Every Optimization
Every optimization is a coordinate move.
Once a synthesizer is located on the stack, the next move becomes a structural question with measurable consequences.
P × T
Mixture of Experts
P: dense → sparse · T: L₂ → lower L₂
Sparsity along the composition axis activates fewer ops per token. Strong when the effective composition is genuinely sparse.
T
Quantization · Distillation
T: L₂ → L₁
Pure thermodynamic-axis move. Lower precision, lower erasure cost. Precision loss can shift verifiability.
R × V
RAG · Tool Use
R: facts → navigation · V: none → citation
Shifts what the weights encode and the verification the output can carry. High-leverage for known-knowledge domains.
E
Agentic Frameworks
E: reactive → reactive (chained)
Extends reactive substance through chained calls. True active behavior needs continuity + internal initiation — beyond chaining.
P × T
Long Context · KV Cache
P: same ops, larger inputs · T: L₂ → L₂ₘₐₓ
Same primitives, larger inputs. Cost scales quadratically. Effective when long-range context is genuinely required.
Hardware
New Silicon · NPU
P: same · T: L₂ → lower-cost L₂
Shifts the hardware boundary itself. The mismatch factor μ(p, H) is exactly what silicon co-design changes — specific primitives become lower-cost to execute on silicon designed for them.
Silicon co-design
Each silicon class shifts a different region of the cost surface downward. LUT silicon makes Z₁ closed-form lookups effectively free. Retrieval-accelerated silicon shrinks Z₂'s pattern-bound costs by collapsing memory access patterns into native ops. Verifier silicon hardens the Zc constraint-bound mechanism — SAT solvers, theorem provers, formal checkers as ISA-level primitives. The frontier model still runs at L₂; the rest of the workload moves down the surface where the silicon's μ is smallest. → tesilicon.thermoedge.ai for ThermoEdge TESilicon — a heterogeneous system on module designed for exactly this surface.
Where It Matters
Not a meter. A force multiplier.
When you can measure, budget, and route every joule per decision, intelligence becomes feasible where it was previously impossible.
+5 hrs
Mission endurance. A soldier carries 250 Wh of battery. A 5 W AI processor cuts a 20-hour mission to 14. An architecture at 1 W average recovers 5 hours. That's not efficiency. It's capability that didn't exist.
+300 min
Persistent coverage. A 20-drone swarm on 100 W power budgets. 20–80% compute savings means 200–300 extra minutes of ISR coverage per day across the swarm.
7.1 W
Onboard intelligence. A satellite's full solar budget. The difference between a 10 W AI chip and a 2 W adaptive runtime is having onboard intelligence — or not.
Browser
Last-mile intelligence. Intelligence that runs in the browser. No data center, no cloud, no subscription. A health worker in rural Malawi with a $100 phone covers what used to require three.
The Loop
The machine that bends the curve.
The thesis names the structure. The flywheel names the verb. Four steps in the inner cycle, two outputs.
The receipt makes the efficient coordinate visible. The cost function lets you choose it. The cache makes it compound. Energy-floor pricing makes choosing it the dominant strategy. The empty cells on the Stack become the markets that can't be re-entered from a V₄ baseline. Each part is load-bearing; remove one and the loop stops turning.
Where this breaks
The thesis has limits. Name them.
Naming the failure modes is what makes a framework measurable. Four places where the Periodic Stack is incomplete on purpose.
Limit 01 · Shape ambiguity
The router can pick the wrong mechanism.
At the Z₂↔Z₃ seam, shape detection is bounded statistical inference — not deterministic classification. Some inputs genuinely sit on the boundary. The fix is a human override path, not a better classifier. Routing is a parser, not an oracle.
Limit 02 · Cold-start
First-time tasks have low-confidence fit estimates.
The cost function's confidence output requires history. The first time a candidate sees a task class, the fit estimate is a guess. The honest default is conservative: bias toward the higher-V mechanism until enough receipts accumulate to calibrate.
Limit 03 · Receipt overhead
The receipt itself costs energy.
Decomposing a decision, attributing θ and μ per primitive, signing the trace — non-trivial. The receipt is worth its cost on expensive decisions. On nanojoule decisions, it can exceed the decision's cost. The cost function applies recursively: budget the receipt against the value of the receipt.
Limit 04 · Frontier necessary
Sometimes the expensive coordinate is the right one.
Novel design, open-ended synthesis, true Z₃ work — the frontier model is the right mechanism. The thesis isn't don't use the frontier. It's use it where it fits. When it engages, the energy is well spent.
None of these breaks the framework. They sharpen it. The Stack is a planning surface, not a closed-form oracle — what makes it useful is that the limits are nameable too, and the cost of the limits can be budgeted against the value of the framework.
Continue
Read the map.
Locate any synthesizer on seven orthogonal axes. Watch the table light up. Read the artifact that makes claims falsifiable.
Companion stacks
One method, applied across fields.
The same axes recur across the family. The thermodynamic class (T) inherits from the compute stack. Verifiability (V) and encoding regime (R) are shared with the data and state stacks. A primitive carries consistent coordinates throughout.
Compute
258 primitives, 33 families.
compute.openie.dev →
Data
The ten-axis data-system stack.
data.openie.dev →
State
The nine-axis state stack.
state.openie.dev →
Communication
The eleven-axis channel stack.
communication.openie.dev →
Memory
The eleven-axis retention stack.
memory.openie.dev →
Knowledge
Seven axes · the object the others handle.
knowledge.openie.dev →
Agent
Four axes · the actor and selection.
agent.openie.dev →
Stack
The overview · three names, one substrate.
stack.openie.dev →
Verify
Seven axes · joules per verified output.
proof.openie.dev →
OpenIE
The parent thesis.
openie.dev →