The Paradigm Shift
The Present: Deterministic Compute
Analogy: The Calculator
Always gives one, precise answer. 5 + 5 = 10.
Core Principle: Fights Randomness
Spends enormous energy (and creates heat) to force transistors into perfect 0s or 1s, suppressing all thermal noise.
The Future: Probabilistic Compute
Analogy: The Sampler
Gives a distribution of answers. "It's 80% likely to be 1, 20% to be 0."
Core Principle: Uses Randomness
Leverages inherent thermal noise. A **p-bit** (probabilistic bit) fluctuates naturally. An input voltage just **biases the probability**, and the output is a **random sample**.
How to "Program" a TSU
You don't write `for` loops. You "compile" your real-world problem into a physical model of energy. This is a three-step process.
The Real-World Problem
"Find the most efficient delivery route for a fleet of 100 trucks."
The "Compiler" (Software)
Translate the problem into an Energy-Based Model (EBM).
Probabilistic Graphical Model (PGM): A high-level "map" of your problem. Nodes = trucks, cities. Edges = routes, costs.
Energy-Based Model (EBM): A *type* of PGM. You define an **Energy Function** where the "best" answer has the "lowest" energy.
Energy = (Total_Fuel_Cost)
+ (Total_Driver_Time)
+ (Penalty_for_Late_Delivery)
+ (Penalty_for_Exceeding_Capacity)
The Hardware (Physics)
The TSU (Thermodynamic Sampling Unit) *physically* settles into its lowest energy state. By feeding it your EBM, the answer isn't "calculated"—it's **sampled** from the hardware's natural equilibrium.
10 Problems for the `thrml` Simulator
You can start today by building these energy functions in software using Extropic's `thrml` simulator. The key is framing the problem as an EBM.
1. Max-Cut Problem
Split a graph's nodes into two groups to maximize the connections *between* the groups.
EBM Frame: Assign `High Energy` (a penalty) to any two connected nodes that are in the *same* group. The TSU will naturally find the state that minimizes this penalty, maximizing the cut.
2. Traveling Salesman Problem
Find the shortest possible route that visits every city exactly once and returns to the start.
EBM Frame: Assign `Energy = Total_Distance`. Add a massive `High Energy` penalty if a city is visited twice or not at all. The TSU will find the route with the lowest total energy (distance).
3. Graph Coloring
Find the minimum number of colors needed for a map so no two adjacent regions share a color.
EBM Frame: Assign a `High Energy` penalty to any two adjacent nodes that have the *same* color. The TSU will find a coloring state that minimizes penalties.
4. Boolean Satisfiability (SAT)
Given a complex logic formula (e.g., `(A or !B) and (B or C)`), find a set of True/False values that makes the whole formula true.
EBM Frame: Assign `Energy = 1` for every logic clause that is *false*. The TSU will seek the lowest energy state, `Energy = 0`, which represents a valid solution.
5. Portfolio Optimization
From thousands of stocks, build a portfolio that maximizes expected return for a given level of risk (variance).
EBM Frame: `Energy = (Target_Risk - Portfolio_Risk)² - (Portfolio_Return)`. The TSU will sample portfolios that minimize this energy, finding states with low risk and high return.
6. Supply Chain Optimization
Find the optimal flow of goods from factories to warehouses to customers to minimize cost.
EBM Frame: `Energy = Total_Shipping_Cost + Total_Storage_Cost`. Add `High Energy` penalties for routes that exceed warehouse capacity or miss delivery deadlines.
7. Image Denoising
Given a grainy or corrupted image, reconstruct the clean, original version.
EBM Frame: Assign `Low Energy` if a pixel's value is similar to its neighbors (smoothness) *and* similar to the original noisy pixel (data fidelity). The TSU "settles" on the clean image.
8. Protein Folding
Predict the 3D structure of a protein from its amino acid sequence.
EBM Frame: The energy function *is* the physics model of atomic bonds. The TSU samples 3D configurations to find the one with the minimum physical (free) energy, which is the stable, folded state.
9. Molecular Binding
Out of millions of potential drug molecules, which one will "fit" best into a target protein's binding site?
EBM Frame: Assign `Low Energy` to a drug molecule's shape and charge that perfectly complements the protein's binding site. The TSU samples molecules to find the lowest-energy (best-fit) candidate.
10. Network Design (5G/Fiber)
Design a telecommunications network that maximizes coverage and bandwidth while minimizing latency and cost.
EBM Frame: `Energy = (Cost_of_Cable + Latency)`. Add `High Energy` penalties for any (user, bandwidth) pair that fails to meet its minimum service-level agreement (SLA).
Start Building the Future.
This hardware is coming. The best way to prepare is to master the software that simulates it.
- 1. Explore the `extropic-ai/thrml` repository on GitHub.
- 2. Pick one of the 10 problems above.
- 3. Start writing the energy function.