Below you will find the internal Tech – Team reviews of the proposals made for the Singularitynet-RFPs.
For more information and an explanation of each of these RFPs, the available rewards, etc. please read our previous blog on the topic: https://deepfunding.ai/a-new-deep-funding-pool-snet-rfps/
For the peer reviews of all other proposals, please view: https://deepfunding.ai/dfr3-proposals-peer-review-recommendations/
All in all, we are extremely happy with the quality and quantity of the proposals submitted!
We have a few that are specially recommended, but also the ones that do not have this label are basically good candidates for implementation. Therefore we will give an overview of all RFP-submissions, but with special acknowledgment of the ones that are recommended. We are looking forward to seeing the voting results, and hope we will be able to bring in some great value-adding projects.
Proposals that have received the qualification ‘Recommended by SingularityNET’ are:
- RFP 1 – Community Engagement Scores by Photrek
- RFP 1 – Cross-Platform Reputation Scores by Togethercrew
- RFP 2 – DeFiGraph – Knowledge Graph (KG) for DeFi by L3A / MIT / Neo4J
- RFP 2 – HiveMind – augmented Q&A bot and P2P knowledge base by Togethercrew
- RFP 3 – Memory-augmented LLMs: Retrieving Information using a Gated Encoder LLM (RIGEL) by MLabs
- RFP 4 – Architectural Design for Uncertain Knowledge Graphs by Photrek
Read along for some more details on these AND the other RFP submissions
RFP 1
For RFP 1 we decided to compare the 2 proposals side by side since only one can be selected but both are recommended by SNET
Team 1: Photrek, has developed the first iteration (PoC) of the tool and has now submitted their plan for an MVP version
Team 2; Togethercrew, has developed a tool that measures the health of an organization’s community in Twitter or Discord, and would like to expand their existing tool with the new features defined in the RFP.
There are advantages to both proposals. In an ideal situation, the teams will combine their strengths and collaborate on the project. We did explore the possibilities for that but did not yet reach a solution that was satisfactory to both teams. Therefore it’s up to you, the community to decide who should be awarded.
Recommended: Community Engagement Scores by Photrek
https://proposals.deepfunding.ai/collaborate/ed600af3-885c-45bc-a874-56d2dde371ce
Funding Requested: $80,000
Advantages:
The team has an excellent track record of delivery, being the first DF team ever, to onboard a service onto the platform. They have also proven with the PoC that they are capable of delivering this engagement score solution. They are intimately familiar with Deep Funding and the SNET mission and vision. I fully expect them to deliver on their promises and be entirely focused on the needs of Deep Funding.
Recommended: Cross-Platform Reputation Scores by Togethercrew
https://proposals.deepfunding.ai/collaborate/abb80969-ec08-437c-a8f7-bd0289fbf33e
Funding Requested: $80,000
Advantages of awarding Togethercrew.
Their proposal shows that they have a deep understanding of engagement scores including concepts such as liquid weighted ranking, and the positive/negative role of experts in the system. They have an existing product with features (integrating behavior on Twitter or Discord, with Telegram and other features in the pipeline) that could potentially be leveraged for Deep Funding. They also have an existing network of other platforms that are already using their tool which could help adoption of our approach across a wider ecosystem.
Because we wanted to ensure that we could build on their existing fundament we (SNET) did a code review of their existing solution. Without going into details, we were very happy with our impression. Their MicroService Architecture offers very good potential for component-based expansion.
So what to choose?
Photrek deserves credit for its accomplishments and can be trusted to be focused on further development that is targeted at DF.
Togethercrew might need to juggle multiple project goals and will need to refactor some of the existing code, but they have a significantly larger development team and have committed to the full scope extended by their current feature set.
As SingularityNET we decided not to give one recommendation over the other. Both have their advantages, both are decidedly recommended, and we will leave the final decision in the hands of our community.
RFP 2,3,4 recommended proposals
Recommended: DeFiGraph – Knowledge Graph (KG) for DeFi
https://proposals.deepfunding.ai/collaborate/91872492-7fd9-49e4-9792-dc39cece2143
RFP2 Unique Knowledge Graphs Or Knowledge Bases Coupled With Clear User Stories
Funding Requested: 40,000 USD
They propose to develop a knowledge graph(KG) with rich DEX-related data and integrate it with a chatbot. Users of the service will be able to get price-related information about particular projects as well as recent stats like which pools on a decentralized exchange have the most trading volume this week.
Pros:
- They have an excellent team with participation from MIT and Neo4J.
- The team has a working software that aggregates different stats across exchanges.
- The domain of KG is relatively narrow, which makes it more likely to be completed.
- here are a large number of tools that can be built on top of such a service.
Cons:
- –
Recommended: HiveMind – augmented Q&A bot and P2P knowledge base
https://proposals.deepfunding.ai/graduated/under-review/7f1791e9-9b18-44f4-bf07-f7c99ca2cc67
RFP3 Memory-Augmented LLMs
Funding Requested: 72,000 USD
The proposal is about building a question-answering chatbot, which will aggregate information from different platforms, e.g. Discord and Github.
Pros:
- There are somewhat similar Q&A chatbots, which are limited to a single platform, they are very useful since people tend to ask similar questions multiple times.
- The team has large experience analysing online communities
- The budget is moderate given how much potential such a product has.
Cons:
- —
Recommended: MLabs – Memory-augmented LLMs: Retrieving Information using a Gated Encoder LLM (RIGEL)
https://proposals.deepfunding.ai/graduated/under-review/e5b9301b-f1e3-4198-b6f3-66ee3484b311
RFP3 Memory-Augmented LLMs
Funding Requested: 140,000 USD
Large language model augmented with embedding-based memory. The main idea is using a large database of texts and corresponding embeddings. Whenever a user asks something relevant embeddings are retrieved and fed into the LLM. Thus it will be possible to link an answer to the source of information. A number of experiments are to be conducted to determine certain architectural details.
Pros:
- Having a llm on the marketplace is very important. It can serve as a base for other projects.
- LLM augmented with memory can learn new facts without retraining.
- Overall it is a sound research and software engineering proposal.
Cons:
Recommended: Architectural Design for Uncertain Knowledge Graphs
https://proposals.deepfunding.ai/graduated/under-review/4c89ca73-386e-4846-9e21-6340dbe78512
RFP4 – Tools For Knowledge Graphs And LLMs Integration
Funding Requested: 40,000 USD
The proposal is about the research and development of UKG (Uncertain Knowledge Graph). UKG is a knowledge graph (KG) where each link has an associated confidence estimate. The prototype will integrate KN with UKG with a large language model, where UKN will play the role of temporary memory. The prototype will be used to score for comments on the Deep Funding portal.
Pros:
- Estimating the quality of comments on the Deep Funding portal is a very useful feature.
- There are a number of related publications about uncertain knowledge graphs
Cons:
- We are not aware of huge success stories in UKG applications, so it might be hard to develop algorithms that do something useful with UKGs. On the other hand, the prototype will be dealing with a very narrow domain, which makes the problem much easier.
RFP 2,3,4 Other proposals
Smart Crypto Investment Recommendation System
https://proposals.deepfunding.ai/graduated/under-review/121d8cee-12cc-493c-84a3-3549017293b5
RFP2 Unique Knowledge Graphs Or Knowledge Bases Coupled With Clear User Stories
Funding Requested: 40,000
The proposal is about development of a knowledge graph for assessment of crypto projects.
The system will aggregate different sources of information and give a user recommendation i.e. how risky it is to invest into that particular project. Estimates that they want to provide:
- Token Risk/Return Profile
- Project Underlying source of return
- Token Liquidity
- Fundamental KPIs (Revenue & Earning Multiples, Monthly Active Users, Token holders, and so on)
Pros:
- It sounds like a useful project. More sources of information about crypto the better.
Cons:
- I am not sure it is even possible to give good recommendations, let alone automatically. On the other hand, basic functionality will be useful anyway.
Blasphemy! The game.
https://proposals.deepfunding.ai/graduated/under-review/2bbe3766-dfa5-4c44-ae20-2bbca6d69a50
RFP2 Unique Knowledge Graphs Or Knowledge Bases Coupled With Clear User Stories
Funding Requested: 40,000
A game on a blockchain where users can create and delete (even not owned) NFTs. An NFT is linked to some kind of medium, e.g. text or video etc. Creating or deleting an NFT costs the user a fee, and creating costs more than deleting. Creators of long-living NFTs are given a share of fees.
Pros:
- Can drive AGIX token utilisation if the game becomes popular.
- The content that will remain in the KG might have some social reference such as being inoffensive or generally supported
Cons:
- The gameplay is not that articulated. There is a risk that the game won’t be that popular
- It requires KYC which decreases the number of players
Unique Knowledge Graph for Source and Content Reliability
https://proposals.deepfunding.ai/graduated/under-review/0e86d730-40ce-4190-8b0b-8edc00d76bb4
RFP2: Unique knowledge graphs or knowledge bases coupled with clear user stories
Funding Requested: 40,000
The proposal is about building software for analysis of news articles and linkage of these articles with knowledge graph(KG). The KG is computed from a large corpus of such articles. KG will contain entities that are mentioned in articles, alongside with text sentiment, crosslinks etc. With such a knowledge base it will be possible to estimate the truthfulness score of an article, to estimate biases of a particular news source etc.
Pros:
- Overall design is sound
- Some use cases such as validation of product reviews look particularly useful.
Cons:
- No real objections, but we preferred ‘TrustLevel Integration into SNET/AI Marketplace’ in ‘New projects’ over this one.
Developing Memory-Augmented Language Models for Enhanced Information Retrieval
https://proposals.deepfunding.ai/graduated/under-review/9f6984cb-f2a2-4a26-8966-129db0555e86
RFP3 Memory-Augmented LLMs
Funding Requested: 110,000 USD
The proposal is too generic, it lacks necessary details. It’s not clear what kind of external memory will be used, so I can’t really evaluate this proposal.
SnetDoc: A SingularityNET Expert at Your Fingertips
https://proposals.deepfunding.ai/graduated/under-review/0d8e88c1-5dcf-49ab-93e3-aff8ea131423
RFP3 Memory-Augmented LLMs
Funding Requested: $55000
The proposal is about building helpfull chat bot that is familiar with documentation of snet projects. Chat bot is based on opensource Llama model augmented with embedding-based memory.
Pros:
- It’s a useful thing to have.
- Architecture is sound, it builds upon open source llama model and a vector database.
- The budget is sound.
Cons:
- It would makes sense to add such a bot to discord and telegram groups, but this proposal features a standalone interface.
- Documentation is usually missing something that community members know. This information can be obtained from conversations, but this proposal lacks this feature. My recommendation is to update the proposal to include this feature. The architecture can be relatively easily updated to include this additional source of information.
Self-Supervised Learning for Memory Optimization
https://proposals.deepfunding.ai/graduated/under-review/43a4ead0-94ac-4f36-a23a-88fdb590d253
RFP3 Memory-Augmented LLMs
Funding Requested: 100,000 USD
The proposal is about augmenting a large language model(LLM) with an external memory. The LLM will summarise given text pieces and these summaries will be stored in the FIFO queue. This queue constitutes the external memory.
The memory retrieval part is a bit vague. Lang-chain library has summary buffer memory which works in a similar fashion. The difference is that the queue contains unmodified text, but when the queue is overflown the last element not just gets deleted, but is used to update conversation summary. https://python.langchain.com/docs/modules/memory/types/summary_buffer
Another thing is that modern LLMs are quite good at summarising out of the box, while a significant part of the budget goes to summarisation training. I recommend to test for example llama-2-70b together with lang-chain summary buffer to see if additional training is necessary, or perhaps the goal can be achieved by simple modifications of the summary buffer.
Phase 1 of Huggingface datasets library to SingularityNET Pipeline
https://proposals.deepfunding.ai/graduated/under-review/b6bac171-634d-44c6-bf18-15884cc04e98
RFP4 – Tools For Knowledge Graphs And LLMs Integration
Funding Requested: 19,500 USD
Huggingface has a datasets Python library. This library can be used to access knowledge graph question-answering (KGQA) datasets. There are many of them https://github.com/KGQA/KGQA-datasets
The idea of the project is to publish an snet service that provides an access point for these datasets.
Pros:
- Can be used by other snet projects in case KGQA is needed.
Cons:
- It is useful only for a limited number of research projects, which can probably use KGQA offline, so the number of API calls is likely to be low.
Bringing Network Pharmacology to OpenCog Hyperon
https://proposals.deepfunding.ai/collaborate/51348a9c-c5d3-40d3-bee3-6ad7ff240990
RFP4 – Tools For Knowledge Graphs And LLMs Integration
Funding Requested: 40,000 USD
The project’s goal is to import several biological knowledge graphs into opencog hyperon.
Pros:
- It’s a good test case for hyperon. Hyperon should be fast enough and convenient to work with on such datasets.
- Robert Haas is well known in the community, We have full confidence that he can finish this project and provide valuable feedback to hyperon developers.
Cons:
- Just importing knowledge graphs is only a first step. Biological knowledge graphs are used in pharmacology to gain different insights into biological processes. TBD in collaboration with our internal teams and external users how the KG will be put to use.
This concludes our feedback.
As mentioned above, both the recommended ones and the proposals that do not have this label are basically good candidates for implementation. Therefore we invite you to not only rely on this review, but also use your own best judgement when voting. We are looking forward to your choices in the voting event!
