← Back to WILLIAM
Responsible: Ben Goertzel, Arthur Franz (original WILLIAM concept)
Papers: Goertzel (2025), Hyperon for AGI⇒ASI Whitepaper, §5.12, §7.6; Franz, A., WILLIAM: Adaptive Compression for AGI
Status: Experimental. Core compression-based pattern detection is under development. MORK trie instrumentation is being implemented. Full integration with PLN scheduling, ECAN attention, and neural acceleration remains a research goal.
This card provides technical depth beyond the concise WILLIAM index card. WILLIAM is an adaptive-compression-based approach to cognitive pattern discovery, now being integrated into MORK's trie infrastructure to serve as a real-time guide for both symbolic reasoning and neural processing.
WILLIAM embodies a fundamental insight: the patterns worth remembering are those that compress experience most effectively. It acts as a cognitive feature detector that continuously asks "what's the simplest explanation that captures what I'm seeing?" The weakness prior provides the theoretical foundation — simpler patterns generalize better, and compression naturally identifies what matters without hard-coded heuristics.
Each pattern application is scored by the description-length reduction it achieves:
Formal definition:
\[\text{gain}(r, S) = L(S) - L(S') - C(r)\]When data follows a hierarchical generative process with bounded-size reusable templates, reuse probability \(\rho > 0\), and heavy-hitter separability margin \(\gamma\):
\[K(x) = \sum_i \ell(f_i) + K(r_s) + O\bigl((\log \ell(x))^2\bigr)\]The key implementation step is adding instrumentation directly to MORK's trie nodes. Each node carries:
| Field | Purpose |
|---|---|
| Local occurrence counts | How often this exact node is accessed |
| Subtree totals | Aggregate weight counters across all descendants |
| Compression-gain sums | Cumulative compression benefit when this pattern is applied |
| Top-\(k\) children rankings | Ranked lists of most important children by various metrics |
This instrumentation enables weighted iterators that return heavy subpatterns directly from any point in the graph — no global scans required.
The API remains minimal:
iter_prefix_topk(prefix, k) — returns the \(k\) most important patterns under a given prefixiter_any_topk(k) — finds globally significant patterns across the entire graphThe implementation handles concurrent access through per-core write buffers for counts, with wait-free readers using snapshot/RCU (Read-Copy-Update) techniques. This ensures that pattern discovery does not block the main reasoning pipeline.
WILLIAM's iterators serve multiple cognitive processes simultaneously:
| Consumer | How It Uses WILLIAM |
|---|---|
| PLN | Prioritizes inference on high-value subgraphs; follows "heavy edges" during backward chaining |
| Schedulers | Allocates resources based on compression-adjusted priorities |
| ECAN | Receives compression-driven importance signals for attention allocation |
| Pattern mining | Uses heavy subpatterns as seeds for deeper structural discovery |
| Symbolic Heads | Template library creation: mines frequent subgraphs from training text for key-value template stores |
WILLIAM's integration with neural networks — particularly transformers and predictive coding networks — uses compression metrics to guide computation allocation. Applied to neural internals (especially networks using local learning, which have greater propensity toward compositional representations), WILLIAM finds "heavy-hitter features" to:
The weakness prior provides the theoretical foundation for all three operations — the same quantale-based framework that guides symbolic pattern selection also guides neural sparsification, ensuring a uniform simplicity bias across both domains.
In the whitepaper's §5.13 integration picture, WILLIAM occupies a specific position:
Patterns discovered by WILLIAM feed directly into ActPC-Chem as chemical rules, and WILLIAM-discovered structures can be transferred to new domains via TransWeave.
Related cards: PRIMUS Full · MORK Full · ECAN Full