Detailed Plan
Explanation
Context
We want to make the SingularityNET Marketplace as easy to use as possible for using AI services to solve complex real-world problems. With the AI Domain Specific Language (AI-DSL) we propose creating the ability to simply write down a problem specification, within a smart contract, for an AI to solve, and have a system automatically discover and connect combinations of available AI services on a marketplace to solve that problem for you. The underlying program synthesis, type verification, program registry, and implicit co-evolutionary technologies we plan to employ could be utilized much more generally within other problem domains on the Cardano blockchain.
AI-DSL will provide abilities for AI services to discover, connect to, and use other AI services to fulfill or closely match user-specified requirements, provide mechanisms to ensure AI services can be combined and chained together, and provide validation that services combined/chained together satisfy or “closely” match user-specified requirements.
The ability for AI services to easily connect with each another and to seek out help from other services to fulfill user specifications will dramatically transform the SingularityNET AI marketplace. It will become a place for users not only to search out simple services to use, but will allow users, and ultimately services themselves, to easily create entirely new AIs as easily as connecting Lego pieces together.
We seek funding to extend our earlier foundational work to create a proof-of-concept (POC) on a real-world problem involving the NuNet Fake News Warning application. While we have additional test cases in mind (T2-T5 in our task list), once we can demonstrate the process in one application, others should follow relatively easily.
The idea
In our earlier phase 1 work we successfully:
-
Built an AI-DSL Registry prototype using Idris, a Dependently Typed Language (DTL), to retrieve, match and connect AI services based on their specifications as dependent types.
-
Started building an AI ontology to ultimately provide a rich and extendable vocabulary for the AI-DSL;
-
Started building the AI-DSL itself, from its syntax to its semantics;
-
Experimented with Idris to formalize and reason about realized function attributes such as costs and quality;
-
Explored the interaction between the AI-DSL and the tokenomics of the network as a means to provide soft guaranties when hard guaranties are difficult to obtain.
A full technical report providing more detailed descriptions of our results in phase 1 can be found at https://github.com/singnet/ai-dsl/blob/master/doc/technical-reports/2021-05/ai-dsl-techrep-2021-05_may.pdf
We now propose to apply our earlier work to build NuNet’s Fake News Network application by integrating the registry, ontology, realized functions, and tokenomics into a holistic prototype, running on a real world test case of AI service assemblage, in real conditions, that is ideally on the SingularityNET-on-Cardano network.
In a complementary process, we also wish to begin explorations using co-evolutionary mechanisms based upon Dr. Debbie Duong’s SISTER algorithm (https://github.com/singnet/simulation), to automatically learn which program building blocks can be combined with each other.
We detail our AI-DSL tasks below:
Registry Tasks
-
R1: Complete the compose procedure for automatically composing AI services to meet an overall specification
-
R2: Experiment with functional and distributional costs and expected result quality
-
R3: Experiment with more sophisticated compositional laws (parallel composition; making temporal cost non-additive; etc)
-
R4: Misc improvements such as to return and rank all matches
-
R5: Explore/Implement fuzzy matching
Ontology Tasks
-
O1: Keep exploring SUMO as base ontology to define AI-DSL vocabulary
-
O2: Build base ontology of ground types; needed for any type description (possibly based on Idris and/or Hyperon)
-
O3: Import relevant domain-specific data-type ontologies into AI-DSL framework; e.g. low hanging fruit will be biomedical and finance domains where existing ontologies are fairly mature
-
O4: Specify a service ontology (as part of AI-DSL ontology) for easier discovery or AI agents based on various criteria
Exploration Tasks
- E1: Explore interactions between AI-DSL and smart contracts
Test cases
-
T1: Fake News Warning app
-
T2: Bio-AtomSpace (planned future work)
-
T3: Financial (planned future work)
-
T4: Controlled Agent (planned future work)
-
T5: Existing SingularityNET services (planned future work)
Production Tasks
-
P1: Build a test suite of unit and integrated tests
-
P2: Turning prototype into product
Co-Evolutionary Tasks
- S1: Recruit community developers willing to have spidering bots run through their github repositories to begin building AI-services. Once we have built small and simple but functional AI blocks, we have the ability, via SISTER, to scaffold up to more complex types implicitly and automatically learn the explicit AI-DSL interface.
___________________________________________________________________________________
Timeline
Month. Task(s)
March R1, R2, O1, E1
April R1, R2, R3, E1
May R4, R5, O2, O3
June R5, O2, O4, T1
July T1, P1, S1
Aug T1, P1, S1
Sept P2, S1
___________________________________________________________________________________
FULL TIME EQUIVALENT (FTE) TIMELINE
___________________________________________________________________________________
Budget Breakdown
Task Months Project Lead Software Developer
R1 1 $1500 $5000
R2 0.8 $1200 $4000
R3 0.5 $750 $2500
R4 0.5 $750 $2500
R5 1 $1500 $2500
O1 0.5 $750 $2500
O2 1 $1500 $2500
O3 0.5 $750 $2500
O4 0.5 $750 $2500
E1 1 $1500 $2500
T1 2 $3000 $10000
P1 1 $1500 $5000
P2 1 $1500 $5000
S1 1 $1500 $0
TOTAL 12.5 $18450 $61500
GRAND TOTAL $79950
___________________________________________________________________________________
Definition of success:
3 months after award: Basic registry tasks (R1, R2, R3, R4) and ontology tasks (O1, O3) complete, exploratory fuzzy matching (R5) started, and exploratory AI-DSL/smart contract interactions complete.
6 months after award: Fuzzy matching (R5) complete, building base and specifying service ontologies (O2, O4) complete, Fake News Warning app built, unit and integrated test suite built, exploratory co-evolutionary tasks (S1) begun.
12 months after award: Fake News Warning app turned into product, continued exploratory task S1.
Expected Fake News Warning app public launch date: Q4 2022
___________________________________________________________________________________
Team:
Dr. Matthew Iklé, CAIO, SingularityNET
Dr. Nil Geisweiller, AI Researcher, SingularityNET
Sam Roberti, AI Researcher and Idris programmer, SingularityNET
Dr. Deborah Duong, AI Researcher, SingularityNET