Draft — This content has not been approved for publication.

Home / Hyperon AI Algorithms / AI-DSL / AI-DSL Full

AI-DSL is a service-composition DSL from SingularityNET with two distinct technical tracks: an Idris dependent-types layer for typed program/service composition, and a MeTTa combinator/ontology layer. It is composition tooling for the SingularityNET marketplace — not a core cognitive algorithm like PLN, ECAN, MOSES, or MeTTa-NARS. Its MeTTa track is what gives it a foothold in the Hyperon runtime; the Idris layer is service-orchestration tooling. In short, it is a "dual citizen" of the marketplace-tooling world and the Hyperon-runtime world.

Last verified: (at repo HEAD aaa702e, 2025-06-25).

On this page

Active maintainer: Nil Geisweiller (ngeiswei).

Stated Goals

The README frames AI-DSL as providing:

  • An expression language for AI service assemblages, with formal mathematical properties (crisp and statistical) — covering algorithmic behaviors, service interactions, and resource usage.
  • A toolset for verifying the validity of assemblages and automatically composing them — given a precise specification, the AI-DSL toolset should fetch and combine appropriate AI services to deliver the requested function.
  • Status (per README): "Up to now the work has been mostly exploratory resulting into prototypes and experimental code covering various aspects of the AI-DSL as opposed to a final product." Two completed phases, with technical reports (May 2021, October 2022).

Implementation: Two Tracks

Idris track (typed composition)

77 Idris source files (*.idr) at HEAD aaa702e. The file tree shows the active substrate is in experimental/ai-algorithms/:

  • experimental/ai-algorithms/NeuralNets/ — Lib.idr, Lib2.idr, Main.idr, test.idr (neural-network primitives in Idris).
  • experimental/ai-algorithms/descent/ — Data/Counter.idr, Data/Matrix.idr, Debug/NoTrace.idr, Examples/LinearRegression.idr, Examples/LogisticLinearRegression.idr, Examples/LogisticRegression.idr (gradient-descent / regression in dependently-typed Idris).

A cross-grep at HEAD (grep -ri 'MeTTa|AtomSpace|PLN|MOSES|Hyperon' --include='*.idr') returns zero hits — the Idris layer does not import or define types over PLN/MeTTa/AtomSpace/MOSES atoms. Its center of gravity is service composition with dependent types.

MeTTa track (combinator + ontology)

30+ .metta files in experimental/ai-service-composition/ across two scenario subdirectories:

  • english-to-chinese-song/: combinator/ontology experiments named etcs-* — etcs-combinator-data-uncurry-xp.metta, etcs-combinator-sd-data-uncurry-xp.metta, etcs-combinator-sd-data-xp.metta, etcs-combinator-sd-xp.metta, etcs-dtl-atw-ontology-syn-sd-xp.metta, etcs-dtl-atw-ontology-syn-xp.metta, etcs-dtl-atw-ontology-xp.metta, etcs-dtl-ontology-xp.metta, etcs-moderate-ontology-xp.metta, etcs-services.metta, etcs-simple-ontology-xp.metta, etcs-xp.metta.
  • recognize-speech-emotion/: combinator experiments named rse-* — rse-combinator-uncurried-xp.metta, rse-combinator-xp.metta, rse-lambda-xp.metta.

This MeTTa track is the AI-DSL slice that connects to the Hyperon runtime — combinator algebra and ontology experiments expressed natively in MeTTa. A recent merge (PR #105 from ngeiswei/combinatory-logic-ai-services, 2025-06-25) confirms ongoing combinator work.

Repository

RepositoryHEADDateCharacter
singnet/ai-dslaaa702e2025-06-25Primary maintained line; an active Idris DSL and a dual citizen (marketplace tooling + Hyperon runtime); Idris 54.4% + Python 33.7% + Shell 9.4% by language; latest merge is the ngeiswei combinatory-logic services PR.

How AI-DSL Relates to Its Neighbors

  • vs MOSES: Adjacent — MOSES evolves programs via genetic search; AI-DSL composes existing services via typed combinators. Different paradigms; no cross-imports at HEAD.
  • vs MeTTa-Motto: Both touch service composition / LLM-agent integration. AI-DSL is Idris-typed plus MeTTa combinator; Motto is a MeTTa LLM-agent library. There may be overlap worth revisiting in a future pass.
  • vs PLN, ECAN, AtomSpace, MOSES, NARS, Pattern Mining: Paradigm-distinct. AI-DSL does not implement reasoning, attention, or learning — it composes services that may include those.
  • SingularityNET marketplace coupling: The marketplace / service-orchestration center of gravity is AI-DSL's main role; the slice relevant to the cognitive-algorithm family is the MeTTa combinator track.

What This Card Is Not

  • Not a core cognitive algorithm. AI-DSL is composition tooling, not a reasoning/attention/learning engine. Position it alongside MeTTa-Motto and the MeTTa runtime, not alongside PLN/ECAN/MOSES/NARS.
  • Not a "neural-symbolic bridge." That framing overstates it — the Idris track is purely typed composition; the MeTTa track is combinator/ontology experiments. Neither bridges neural and symbolic representations.
  • Not located in src/Composition.idr. The repo has no src/ directory. The Idris files live in experimental/ai-algorithms/ as listed above. (The src/Composition.idr citation was a pre-review reviewer error, corrected on inspection.)
  • Not finalized. Per the README "Status" section, the work is exploratory / prototype-level, with completed Phase 1 (May 2021) and Phase 2 (October 2022) technical reports.

Status and Sources

Implementation status: Active, exploratory. Recent commits include a combinatory-logic AI-services PR (2025-06-25), indicating ongoing combinator-level work. Two technical reports completed (2021 + 2022) per docs/technical-reports/.

Primary sources:

  • Geisweiller, N. et al. AI-DSL Technical Report. docs/technical-reports/2021/ai-dsl-techrep-2021-05_may.pdf. Phase 1.
  • Geisweiller, N. et al. (2022). AI-DSL Project Close Report. docs/technical-reports/2022/ai-dsl-project-close-report-2022-oct.md + ai-dsl-techrep-2022-oct.pdf + snet-service-assemblages.pdf. Phase 2.
  • Repository README at singnet/ai-dsl.

Provenance: the verdicts on this card come from a 2026-05-07 source-code review; the extraction archive is at scripts/archive/non_clustered_haa_pilot/source5_NACE_AIDSL_MOSES_MeTTaNARS/.