Technical Working Group DeAI

Sorry I’ve missed these meetings, I’d love to get these into my Calendar automatically somehow, is that possible?

2 Likes

Hi @lastmjs , does this event link work? https://calendar.google.com/calendar/event?action=TEMPLATE&tmeid=NHM4bTkxZGthaGdlY3I1cm5hYWFpajJha2dfMjAyMzEyMDdUMTcwMDAwWiBwYXRyaWNrLmZyaWVkcmljaDkzQGdvb2dsZW1haWwuY29t&tmsrc=patrick.friedrich93%40googlemail.com&scp=ALL

Otherwise, you can also send me your email address and I’ll send you a direct invite :slight_smile: Looking forward to having you!

Is there a recording of this meeting?

What was the answer to this?

It says “could not find the event”: lastmjs@demergentlabs.org

1 Like

I had planned to record it but unfortunately Google Meet didn’t let me. We might have to switch the video call platform to be able to record the meetings going forward.

Concerning Singularity.Net, it was an open question for us how they are achieving decentralization and we didn’t have an answer yet. I’m planning on looking into that a bit and will post what I find. If anyone knows more about them and their marketplace though, please share :slight_smile:

ok, thanks, I just sent you an invite. Please let me know in case you didn’t receive it.

If anyone else wants a direct invite, please share your email address (or DM me with it) and I’m happy to invite you too

as far as I can tell from the their website (e.g. https://dev.singularitynet.io/docs/ai-developers/marketplace/) and whitepaper (https://public.singularitynet.io/whitepaper.pdf), SingularityNet has an on-chain (Ethereum) registry of AI services and the app then combines that with off-chain metadata to display the different services available on the marketplace. The services in the registry are vetted by the SingularityNet foundation and seem to have to adhere to certain (interoperability) standards. Not sure they have to be decentralized though. So in general, it seems to me like the decentralization aspect is mostly the on-chain registry. Haven’t seen anything too impressive yet, happy to learn more though if anyone has more insights.

2 Likes

Hi everyone, looking forward to our call tomorrow, Thursday 12/07 at 5pm UTC. This is the link to it: https://meet.google.com/ddh-jnpq-dji

Agenda-wise, I think we can take a few minutes in the beginning to share any cool (De)AI-related news we’ve come across. We can then discuss some of the action items; wish list of needed features/components on the IC, showcasing and promoting IC DeAI projects and developing the GPU subnet as a community effort. As it came up a few times now and there seems to be an interest from several of us, talking about implementation efforts and ideas for vector databases/embeddings within canisters would be another topic. Let’s see how much we can cover :slight_smile:

Would you like to see anything else on the agenda? Do you have any priorities in terms of topics to cover?

For any of the action items, it’d be great to decide if and how we’d like to tackle them as a group as well as see if anyone would like to take the lead on any of them. We could get more into “action mode” that way and push these items forward. What do you think?

See you all on the call,
Pat

1 Like

Sorry I missed the 2min warning for the end of the call. Thank you for the call, I think it was a great discussion and I’m looking forward to continuing it!

As a next step, I’m proposing that we organize ourselves such that we can push the different action items we’ve been discussing further. From the discussions so far, they seem to be

  • Showcasing and communicating DeAI projects and innovations on the IC (within the community and beyond)
  • Working out unique value propositions for consumers and users of DeAI
  • Educational materials and documentation
  • Wish list of features and components that’d be beneficial for development (incl. GPU subnet as a community effort)
  • Technical initiatives like on-chain vector databases

What do you think about these? Are any other items missing?

Would you like to work on any of these items or even take the lead?

And I unfortunately just noticed that the screen recording software I used didn’t capture the audio correctly, so the recording is not very helpful. Any good solutions for the recordings anyone can propose?
These are the few notes I took:

Please add to these notes with your own.

3 Likes
  • migrating libraries: size (below 2MB) → zipping, no multi-threading

The wasm size limit will be removed soon. I heard in the Tooling WG today that chunked wasm upload will be available early next year.

So, I think it is worth the effort to look into porting some the following C++ libraries:

  • OpenBLAS
  • FAISS (depends on a BLAS library)
6 Likes

@ildefons
Regarding collaboration on porting AI related C++ libraries, would you be OK if we discuss that in the OpenChat C++ community?

We have pretty lively discussions around issues and solutions for gaming related C++ libraries that will be relevant for the AI libraries as well.

2 Likes

100%. I will join the chat

1 Like

The story of deploying the knowledge graph on ICP that was mentioned in the SingularityNet collaboration, even if technically feasible, seems challenging. SingularityNet’s sister team, openCog, will launch an alpha version of Atomspace in late Q1 2024, will ICP be involved in any way?

4 Likes

regarding wasmtime benchmarking idea, I found some interesting snippets, not sure what to make of them yet:

3 Likes

Hi everyone, we’ve got a dedicated channel in the ICP Developer Community Discord now: Discord
Thank you @domwoe for creating it!
I hope we can use the Discord channel for all the interesting discussions, ideas and news we cannot fit into our meetings :slight_smile:

Hi everyone, we’ll have our next call tomorrow, Thursday Dec 14 at 5pm UTC. This is the link to the video call: https://meet.google.com/ddh-jnpq-dji
As a preparation, @icpp and I talked about the different action items and how we could tackle them as a group. Posting our ideas here and please add yours. We can also discuss these further tomorrow.

  • Showcasing and communicating DeAI projects and innovations on the IC (within the community and beyond)
    Let’s all get our projects on Awesome IC (GitHub - dfinity/awesome-internet-computer: A curated list of awesome projects and resources relating to the Internet Computer Protocol)
    How about a GitHub repo for the DeAI WG where we can showcase the projects in more detail, we can link it back to Awesome IC

  • Working out unique value propositions for consumers and users of DeAI
    Async brainstorming via a Miro board where we can all add ideas, additional discussions via Discord
    Dedicated meeting (one of our calls on Thursday where we focus on this topic)

  • Educational materials and documentation
    We could have a dedicated section on our GitHub repo (to give an overview and starting point, e.g. programming languages, examples, concepts, libraries to use)
    Document learning journey (e.g. by someone getting started with DeAI or a DeAI project on the IC)

  • Wish list of features and components that’d be beneficial for development (incl. GPU subnet as a community effort)
    We could have a Wiki in our GitHub repo (so it can be easily edited in browser) for all of us to add to; how would we like to structure this list/the Wiki?
    We’ve got some ideas from our discussions so far to start off with

  • Technical initiatives like on-chain vector databases
    We could have Wikis in our GitHub repo dedicated to each technical topic/initiative, also as a way of knowledge transfer between us, how would we like to structure this?

Looking forward to seeing everyone tomorrow!

2 Likes

During tomorrow’s call, we’ll also have a 10min presentation by @ildefons about his work on MotokoLearn and on-chain Machine Learning training plus inference (Introduction of MotokoLearn V0.1: on-chain training and inference of machine learning models). That’ll be interesting to learn more about!

2 Likes

Absolutely on board! ELNA.ai was made for this.

2 Likes

Thank you everyone for today’s call!
Thank you @ildefons for walking us through MotokoLearn (GitHub - ildefons/motokolearn). I think it’s worth sharing your work also beyond the IC community as this must be one of the first, if not the first, on-chain training and inference libraries for some common Machine Learning techniques.

One point mentioned which I thought was interesting were cost comparisons and statistics on how expensive it is to run AI applications on-chain vs off-chain. That could also be helpful then to understand which use cases to focus on and plan the underlying hardware infrastructure around it.

As discussed, we’ll move the calls to the Discord voice channel (Discord) going forward to have it close to our DeAI channel there and not being cut off after 1h :slight_smile: thank you @mnl for proposing this idea.

It’d be great to have discussions around the different action items in the Discord and plan some next steps there as well. Any ideas or feedback you have, please also share them there or here (or also send me a direct message) :slight_smile: Looking forward to continuing the conversation

3 Likes