Hyperon AI Algorithms

help edit space_dashboard

Core cognitive algorithms expressed in MeTTa that animate static knowledge into active intelligence. These interoperable modules enable cognitive synergy — diverse cognitive skills interacting concurrently on shared memory.

ECAN (Economic Attention Networks)

help edit space_dashboard
Review AI Draft →
Responsible: Matt Ikle, Khellar Crawford

ECAN is the attention-allocation subsystem that regulates which Atoms are actively considered. It uses Short-Term Importance (STI) and Long-Term Importance (LTI) values to implement a dynamic attention economy, preventing exponential blowup in search and inference.

</> Example Implementation
Attention allocation
Shows ECAN's economic attention model with importance values and Hebbian spreading activation.
; ECAN attention dynamics ; STI/LTI economic model !(set-sti (atom "hypothesis-1") 0.9) !(set-sti (atom "hypothesis-2") 0.3) ; Spread attention via Hebbian links !(ecan-step (hebbian-spread 0.1) (rent-collection 0.05) (forgetting-threshold 0.01))

Technical Deep Dive: ECAN Full — STI/LTI dynamics, AttentionBank engineering surrogate, Hebbian spreading, OpenPsi coupling lifecycle, and implementation findings.

MetaMo (Motivational Framework)

help edit space_dashboard
Review AI Draft →
Responsible: Ruiting Lian, Ben Goertzel

MetaMo models motivation in open-ended intelligent agents as a dynamical system coupling appraisal with decision processes. It shapes PLN control dynamics, regulates exploration–exploitation tradeoffs, and embeds safety constraints within motivational dynamics.

</> Example Implementation
Motivational dynamics
Demonstrates MetaMo's motivational framework coupling appraisal with action selection.
; MetaMo goal-motivation system ; Appraisal + decision coupling !(define-goal (goal explore-environment (urgency 0.8) (importance 0.6))) !(metamo-step (appraise current-state goals) (select-action (balance exploration exploitation)))

Technical Deep Dive: MetaMo Full — pseudo-bimonad structure, OpenPsi appraisal comonad + MAGUS decision monad, contractive update dynamics, AGI 2025 papers, and implementation findings.

Semantic Parsing (LLM/NLP)

help edit space_dashboard
Review AI Draft →
Responsible: Leung Man Hin, Roman Treutlein

A neural-symbolic bridge that converts ambiguous human language into executable logic using SENF (Semantic Elegant Normal Form). It collapses idiomatic variations into canonical graph structures, creating grounded atoms for the Hyperon ecosystem.

</> Example Implementation
Natural language to atoms
Shows how the semantic parser converts natural language into canonical graph structures (SENF).
; Semantic Parsing: NL -> SENF -> Atoms ; "The cat sat on the mat" !(parse-to-senf "The cat sat on the mat") ; => (EvaluationLink ; (Predicate "sat-on") ; (ListLink ; (Concept "cat") ; (Concept "mat")))

Technical Deep Dive: Semantic Parsing Full — SENF canonicalization, legacy pipeline, Hyperon-era approaches, and implementation findings.

MeTTa-Motto

help edit space_dashboard
Responsible: Svetlana Besaeva, Alexey Potapov

An interoperability layer embedding LLMs directly into the MeTTa runtime as programmable functions. Supports stateless wrappers, stateful Dialogue Agents, RAG, and functional calling for neural-symbolic integration.

</> Example Implementation
LLM integration
Demonstrates MeTTa-Motto's ability to embed LLM calls directly in MeTTa programs and parse responses as atoms.
; MeTTa-Motto: embed LLMs in MeTTa ; Stateless wrapper example !(llm-query (model "gpt-4") (prompt "Explain quantum entanglement in terms of Atomspace links") (parse-response as-atoms)) ; Response becomes queryable atoms

PLN (Probabilistic Logic Networks)

help edit space_dashboard
Review AI Draft →
Responsible: Patrick Hammer, Peter Isaev, Nil Geisweiller

Hyperon's primary symbolic reasoning system for uncertainty. Represents beliefs with graded confidence, supports deductive/inductive/abductive reasoning, and leverages ECAN for computational tractability over massive Atomspaces.

</> Example Implementation
Probabilistic inference
Shows PLN's probabilistic deduction with strength-truth-values propagating through inheritance chains.
; PLN deductive reasoning with confidence !(add-belief (Inheritance "Tweety" "Bird") (stv 0.95 0.9)) !(add-belief (Inheritance "Bird" "Animal") (stv 0.99 0.99)) ; Deductive inference !(pln-deduce (Inheritance "Tweety" "Animal")) ; => (stv 0.94 0.89)

Technical Deep Dive: PLN Full — STV quantale formalism, backward chaining on MORK, factor-graph belief propagation, historical design decisions, and implementation findings.

MeTTa-NARS (Non-Axiomatic Reasoning System)

help edit space_dashboard
Responsible: Peter Isaev, Patrick Hammer

An open-ended uncertainty reasoning engine built on Non-Axiomatic Logic (NAL). Uses two-dimensional evidence values (frequency and confidence) for continuous learning under the Assumption of Insufficient Knowledge and Resources.

</> Example Implementation
Non-axiomatic reasoning
Demonstrates NARS's two-dimensional evidence tracking with frequency and confidence updating.
; MeTTa-NARS evidential reasoning ; Frequency × Confidence !(nars-judge "bird can fly" (evidence (positive 95) (total 100))) ; => !(nars-revise "bird can fly" (new-evidence (positive 2) (total 10))) ; =>

Technical Deep Dive: MeTTa-NARS Full — NAL inference rules, evidence revision, AIKR principle, NARS-GPT integration, and implementation findings.

NACE (Non-Axiomatic Causal Explorer)

help edit space_dashboard
Responsible: Patrick Hammer, Peter Isaev

An experiential learning agent achieving 1000x data efficiency over deep RL. Constructs logic-based environment models through direct interaction, using curiosity-driven exploration and Non-Axiomatic Logic for evidential tracking.

</> Example Implementation
Curiosity-driven exploration
Shows NACE's curiosity-driven loop where prediction errors drive model refinement and exploration.
; NACE experiential learning ; 1000x data efficiency vs deep RL !(nace-step (observe state-42) (predict (action "move-left") state-43) (actual-outcome state-44) (surprise 0.8) (update-model causal-rule-17)) ; High surprise => explore more

Technical Deep Dive: NACE Full — surprise-driven exploration, NAL-based evidence, environment-model construction, and implementation findings.

AI-DSL

help edit space_dashboard
Responsible: Nil Geisweiller

A type-driven program synthesizer that automatically assembles AI workflows. Uses backward chaining in MeTTa with dependent types and combinatory logic to compose services from the SingularityNET marketplace.

</> Example Implementation
Type-driven composition
Shows AI-DSL's automatic composition of typed services into coherent workflows via backward chaining.
; AI-DSL automatic service composition ; Backward chaining with dependent types !(ai-dsl-compose (goal (-> Image DiagnosisReport)) (available-services (image-classifier (-> Image Labels)) (symptom-mapper (-> Labels Symptoms)) (diagnosis-engine (-> Symptoms DiagnosisReport)))) ; => Composed pipeline: image-classifier ; >> symptom-mapper >> diagnosis-engine

Technical Deep Dive: AI-DSL Full — dependent types, backward-chaining synthesis, SingularityNET marketplace composition, and implementation findings.

AIRIS

help edit space_dashboard
Review AI Draft →
Responsible: Berick Cook

Autonomous Intelligent Reinforcement Inferred Symbolism — a causal ML system that builds deterministic environment models through direct interaction, encoding decision logic in explicit, auditable causal rewrite rules.

</> Example Implementation
Causal model building
Shows AIRIS learning deterministic causal rules through direct environmental interaction.
; AIRIS deterministic causal learning ; Explicit rewrite rules !(airis-observe (state (block-at 3 5)) (action push-right) (result (block-at 4 5))) ; Learned rule: ; (push-right (block-at $x $y)) ; => (block-at (+ $x 1) $y)

Technical Deep Dive: AIRIS Full — causal rewrite rules, surprise-driven exploration, Minecraft demos, Cook & Hammer 2024 paper, and implementation findings.