Technical Working Group DeAI

Developers
304

Hi @stdevelopr , thanks for reporting. Which error did you get? Are you a member of the ICP Discord?

305

simply cannot join the server…
“Unable to accept invite”

306

does https://discord.internetcomputer.org work for you?

307

No. It is the same link i got from the calendar.

308

For this week’s call we’ll have a workshop on Accelerated Infrastructure for AI on ICP where we pick up last week’s discussion on which AI models, GPUs and related aspects will be worth pursuing for ICP. @icarus will lead the session and it’ll be at 3pm CET again (so like last week, 3 hours earlier than usual). Please also take a look at last week’s notes and resources if you’re interested :thumbsup: have a great week and see you then

309

Hi everyone, thank you for today’s call (2025.05.29). Special thanks to @icarus for leading the call. This is the generated summary (short version, please find the long version here ): The DeAI Working Group discussed and advanced a hardware-runtime comparison framework for AI accelerators. Key updates included runtime benchmarking (TensorRT, vLLM, Modular), open model selection (LLaMA, Mistral), quantization pipelines via GGUF, and tools for format conversion and model interoperability. Real-world experiments with 3B–7B quantized models showed promise for on-device inference, supporting IC-compatible AI agent deployment across browser, edge, and mobile environments.

Links shared during the call:

310

:spiral_calendar: Upcoming Calls & Topics

Date Topic Presenter
June 05 World Computer Summit review
Future Session on Efficient LLMs and DeepSeek
future Session on instruction limit
Future Expert session on Active Inference implementations
311

Hi everyone!

I’m happy to share one early stage of ic-rmcp crate

Currently this SDK is supported MCP tools as example we share in last WG.

Current feature

  • Support MCP tools
312

That’s incredible, looking forward to the progress, I’m currently in the process of figuring out different ways to architect agents on ICP so this is very helpful. Ideally MCP integration should be also added to the ic_llm crate and the ICPAnda tooling so that it’s more or less standardized and people won’t have to rewrite the same functions over and over again.

313

Btw we’re looking for some collaboration.

If you’re project owner of some AI agent want to support MCP. Please DM me

314

Hey friends, I have a question. Is it possible to test the ic_llm crate (I mean like a backend canister I wrote which uses the crate) locally? I wrote some code and tried deploying it locally and it couldn’t do anything cause the llama model is hosted on the IC mainnet. Any way I can work around that or do I just have to deploy to mainnet?

315

that’s great @baolongt , thanks for sharing!

If you want and can join today’s call, we can talk about it during the session. Maybe otherwise next week or so :+1:

316

Hey Typonomy, yes you can. There’s a guide here. So basically, you can simply add the LLM canister as a dependency into your dfx.json like this:

 "llm": {
      "type": "pull",
      "id": "w36hm-eqaaa-aaaal-qr76a-cai"
    },

Run dfx deps deploy (this deploys the llm canister locally) and then run ollama locally as it’s described in the guide. Let me know if you run into any issues.

317

Hi everyone, thank you for today’s call (2025.06.05). This is the generated summary (short version, please find the long version here ): The meeting introduced new projects, including an AI-driven wallet and a spam prevention solution leveraging ICP security features. Highlights from the World Computer Summit emphasized Caffeine AI’s transformative potential, with discussions covering infrastructure costs, decentralized AI integration, and security enhancements through AI-driven auditing. The group recognized the need for open-sourcing ICP-specific training datasets, improving hardware optimization through SIMD and WASI, and collecting structured developer feedback to accurately gauge demand and capabilities, proposing clear next steps for collaborative research and community-driven initiatives.

Links shared during the call:

318

Hi everyone, during tomorrow’s session @tiago89 will talk about vibe coding on ICP, including a demo :muscle:
If you like, you can already take a look at these resources upfront:
IC Vibe Code session (https://www.youtube.com/watch?v=ZuNUy13wmlI&t=5753s)
and template (GitHub - pt-icp-hub/IC-Vibe-Coding-Template: Full Stack + LLM Canister template, ready with tests and AI workflows that accelerate your work on the IC.).

I’m looking forward to the session and seeing everyone then! And if you have any other items you would like to share or see on the agenda, please let me know :+1:

319

and please note that the call was moved to 3h earlier (3pm CET), this is the event in the ICP Discord (as usual it’ll take place in the voice channel): ICP

320

Hi everyone, thank you for today’s call (2025.06.12). Special thanks to @tiago89 and @aespieux for the demo! This is the generated summary (short version, please find the long version here ): In today’s DeAI Working Group call, the main focus was on Tiago’s demo, which showcased Vibe Coding with VS Code and a template he and Antione created to support live feature implementation, test generation, and AI pair programming. The tool, already used in hackathons, impressed the group with its potential for speeding up development. It could in the future also integrate with the MCP server to support agent workflows on ICP. Beyond the demo, the group discussed ICP’s strategic strengths for AI development and Caffeine AI, debated fine-tuning versus RAG approaches, explored WASM-level optimizations, and aligned on three priority tracks: infrastructure enhancements, knowledge sharing, and cross-platform AI interoperability.

Links shared during the call:

321

Thanks again Patrick for having us on today’s call, it was great to share what we’ve been building. A Motoko version of the template is on its way soon :eyes:

322

Thanks for presenting! Awesome, that sounds great, please share it once available :muscle:

323

Hi everyone, this week @icarus will lead our next AI4AI call where we focus on hardware and infrastructure for AI on ICP :thumbsup: In particular, we will continue our initiative to analyze options for GPU hardware, AI frameworks/runtime and LLMs to run as collected in this spreadsheet: https://docs.google.com/spreadsheets/d/1a-NzZeh6D9smxovW7ByYOx7Ls1tRqBJyaS9gXVzRy-Y/edit?gid=712966265#gid=712966265 If you like, take a look at the spreadsheet and add any options or findings you have. I’m looking forward to our call!