A Request for Comment: A Vision for a Strategy-Native Neural System What I Mean by NLP in This Context
When I say (NLP) neuro-linguistic programming here, I’m not speaking of machine NLP, but of the older, more psychological frame that modeled how humans think and act. Out of that tradition, I take a few clean and useful ideas. Strategies: Human beings run internal programs for tasks. We switch into a math strategy when solving equations, a persuasion strategy when making an argument, a motivation strategy when driving ourselves forward. Each is a flow of steps triggered by context. Modalities: Strategies draw on representational channels — visual, auditory, kinesthetic, and language. In machines, this translates less literally, but the principle holds: different channels or flows combine to shape different behaviors. TOTE (Test → Operate → Test → Exit): This is the backbone of strategy. We test our current state, operate to move closer to a goal, test again, and either exit (done) or loop back for another attempt. It is feedback incarnate. Intensity/Desire: Not all goals burn equally. Some pull with urgency, others linger in the background. Intensity rises and falls with context and progress, shaping which strategies are chosen and when. This is the essence of NLP that I want to carry forward: strategies, feedback, and desire. Executive Summary I propose a strategy-native neural architecture. At its center is a controller transformer orchestrating a library of expert transformers, each one embodying a strategy. Every strategy is structured as a TOTE loop — it tests, it operates, it tests again, and it exits or adjusts. The Goal Setter is itself a strategy. It tests for needs like survival assurance, operates by creating new goals and behaviors, assigns an intensity (a strength of desire), and passes them to the controller. The controller then selects or creates the implementing strategies to pursue those goals. This whole system rests on a concept network: the token embeddings and attention flows of a pretrained transformer. With adapters, controller tags, gating, and concept annotations, this substrate becomes partitionable and reusable — a unified field through which strategies carve their paths. The system is extended with tools for action and RAG memory for freshness. It grows by scheduled fine-tuning, consolidating daily experience into long-term weights. I offer this vision as a Request for Comment — a design to be discussed, critiqued, and evolved. The Strategy System Controller and Expert Strategies The controller transformer is the orchestrator. It looks at goals and context and decides which strategies to activate. The expert transformers — the strategy library — are adapters or fine-tuned specialists: math, planning, persuasion, motivation, survival, creativity. Each is structured as a TOTE loop: Test: measure current state. Operate: call sub-strategies, tools, memory. Test again: check progress. Exit or adjust: finish or refine. Strategies are not just black boxes; they are living feedback cycles, managed and sequenced by the controller. Goal Generation with Desire and TOTE The Goal Setter is a special strategy. Its test looks for overarching needs. Its operate step generates candidate goals with behaviors attached. Its test again evaluates them against constraints and context. Its exit or adjust finalizes goals and assigns intensity — the desire to act. These goals are passed into a Goal Queue, where the controller schedules them based on intensity, value, urgency, and safety. This is how the system sets its own direction, not just waiting passively for prompts. Tools and RAG The strategies reach outward through tools: calculators, code execution, simulators, APIs, even robotics. They also reach into retrieval-augmented generation (RAG): an external vector memory holding documents, experiences, and notes. Tools are the system’s hands. RAG is its short-term recall. Together, they keep the strategies connected to the world. Daily Consolidation At the end of each day, the system consolidates. It takes the most important RAG material, the traces of successful strategies, and runs scheduled fine-tuning on the relevant experts. This is long-term memory: the system learns from its own actions. RAG covers freshness, fine-tuning covers consolidation. The strategies sharpen day by day. The Substrate: A Concept Network of Tokens A pretrained transformer is already a concept network: Tokens are mapped to vectors in a meaning space. Attention layers connect tokens, forming weighted edges that shift with context. By the later layers, tokens are transformed into contextualized vectors, embodying concepts shaped by their neighbors. This is a unified substrate, but raw it doesn’t separate strategies. To make it strategy-native, I propose: Adapters: LoRA or prefix modules that bias the substrate toward particular strategy flows. Controller Tags: prompt tokens like [MATH] or [PLANNING] to activate the right flows. Gating and Attention Masks: to route or separate flows, allowing strategies to partition without isolating. Concept Annotations: clusters and labels over embeddings, marking areas as “narrative,” “mathematical,” “social,” so strategies can claim, reuse, and combine them. This makes the transformer not just a black box but a living concept network with pathways carved by strategies. Safety and Reflection Every strategy’s TOTE includes policy tests. Unsafe plans are stopped or restructured. Uncertainty checks trigger escalation or deferral. Logs are signed and auditable, so the system’s actions can be replayed and verified. Meta-strategies monitor performance, spawn new strategies when failures cluster, and adjust intensity rules when needed. This keeps the growth of the system accountable. Conclusion: A Call for Comment This is my vision: a strategy-native neural system that does not merely respond but calls strategies like a mind does. Every strategy is a TOTE loop, not just the Goal Setter. Goals carry intensity, giving the system direction and drive. The controller orchestrates expert strategies, tools, and memory. A concept network underlies it all — a transformer substrate refined with adapters, tags, gating, and annotations. RAG and tools extend its reach. Scheduled fine-tuning ensures it grows daily from its own experience. I put this forward as a Request for Comment. What breaks here? What’s missing? How do we measure intensity best? Which strategies deserve to be trained first? Where are the risks in daily consolidation? How should gating be engineered for efficiency? This is not just an assistant design. It is a sketch of a mind: one that sets goals, desires outcomes, tests and operates with feedback, reaches outward for tools and memory, and grows stronger with each cycle. I welcome input, critique, and imagination. Together we can refine it — a mind of strategies carved into a unified network of concepts, guided by goals that pull with desire. How can this be interfaced with, joined with Opencog? -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/opencog/b41659bc-7f3a-4383-83d2-fffd35bea4abn%40googlegroups.com.
