Hi everyone, join us for this week’s call tomorrow to learn the latest about ICP Coder as @Gianm will give us an update on the project! It’s the usual 3pm CET, this is the event in the ICP Discord: ICP
I’m looking forward to it and seeing everyone then ![]()
Hi everyone, thank you for today’s call (2025.10.23). Specials thanks to @Gianm for sharing the latest updates on ICP Coder! This is the generated summary (short version, please find the long version here ): Today’s DeAI Working Group call showcased major updates to ICP Coder, a Motoko-focused AI coding assistant integrated into Cursor and VS Code via the Model Context Protocol (MCP). The team presented a new hybrid Go + Python architecture, with Go handling authentication, user management, and API key issuance, while Python powers the data ingestion and RAG pipeline. They launched a public service with user registration, API keys, and an NPM package for easier IDE setup. The system now ingests Motoko repositories and documentation, embeds them with Sentence-Transformers, and serves context-aware completions through ChromaDB. Deployed on AWS, ICP Coder now supports Linux and macOS and plans to expand beyond Motoko to TypeScript and Rust. Upcoming improvements include advanced retrieval methods, payment integration, and rate-limiting for users, with continued work on performance, model accuracy, and developer usability.
Links shared during the call:
- Simple instructions to give ICP Coder a try: GitHub - Quantum3-Labs/icp-coder: ICP AI assistant. LLM RAG implementation for Motoko Code Generation Enhancement
- Quick demo here: ICP coder Demo: Production | Loom
- Swagger UI: Swagger UI
- Twitter: https://x.com/Q3Labs/status/1975834944363790818
- Rust sdk for MCP servers on ICP: GitHub - ByteSmithLabs/ic-rmcp: Rust SDK for building Model Context Protocol servers on Internet Computer canisters.
Hi everyone, are there any specific items you would like to see on this week’s agenda? Is there anything you’d like to present or discuss? Please let us know and we’ll put them on the agenda ![]()
please note that the call is tied to CET and we just switched from summer to winter time, so the call time might actually have changed for other time zones now. Please double check whether it changed for you: ICP ![]()
Hi everyone, thank you for today’s call (2025.11.06). Specials thanks to @icarus for sharing his expertise! This is the generated summary (short version, please find the long version here ): The WG discussed two strong open small-LLM contenders (IBM Granite 4 & ETH Apertus) and new TEEs running on AMD SEV-SNP. Hardware advances (Zen 5 EPYC Gen 3 nodes) could bring ~50 % faster performance, but protocol-set instruction limits—not memory—remain the bottleneck for running larger models on-chain. Focus remains on hybrid architectures—canister logic + adjacent AI services—for practical DeAI growth.
Links shared during the call:
- IBM Granite 4.0: Hyper-efficient, High Performance Hybrid Models for Enterprise
- IBM Granite 4.0 Tiny Preview: A sneak peek at the next generation of Granite models
- IBM becomes first major open-source AI model developer to earn ISO 42001 certification
- Granite 4.0 Nano: Just how small can you go?
- Apertus: a fully open, transparent, multilingual language model | ETH Zurich
- https://forum.dfinity.org/t/first-sev-snp-enabled-node-deployed/59499?u=icarus
- https://dashboard.internetcomputer.org/network/nodes/hckfw-kshvv-mzhew-h6isd-5q2wd-lde6w-uipmv-7gbuz-53wqz-cz5kx-oqe
- https://dashboard.internetcomputer.org/proposal/136408
- https://www.esafety.gov.au/industry/basic-online-safety-expectations
Hi everyone, hope you’re having a great start to the week! Are there any topics you’d like to see on this week’s agenda, or anything you’d like to present? Please let us know
See you on Thursday
Hi everyone, for our call tomorrow, we’ll have a demo by the LAIN team ![]()
They will present a Mobile Android and IOS / Web Dapp where the user registers with Internet Identity to interact with an AI called LAIN. There will be various channels where you can interact about different topics and the roleplay will be tailored for the channel via a small RAG component. The app is still undergoing development and uses the library ic_llm to generate text, and a database canister to store the conversations of the users as well as a friend system with friend list.
I’m looking forward to the call (at the usual 3pm CET, please see the Discord event here: ICP ) and seeing everyone ![]()
Hi everyone, thank you for today’s call (2025.11.20). Specials thanks to @LainCorp for demoing the LAIN app! This is the generated summary (short version, please find the long version here ): This week’s DeAI Working Group call featured a live demo of LAIN, a decentralized AI-driven social/chat app on ICP, showcasing multi-channel conversations, friend-matching via embeddings, and hybrid use of the IC LLM canister plus DeepSeek. The discussion covered plans for client-side embedding generation, vector-database canisters, Odyssey video integration, and front-end challenges with Three.js on React Native. Participants also explored app distribution hurdles (Android/iOS), constraints of the IC LLM canister, and wishlist items such as an agent-optimized on-chain model and an MCP client library. The session concluded with a broader conversation on distributed vs. decentralized AI, edge hardware experiments, and emerging world-model concepts that point toward increasingly distributed and agentic AI systems.
Links shared during the call:
We released the python cli for Kinic AI memory.
Hi everyone, this week we’ll have an open session. I’ll share a couple of AI projects and resources that I saw recently. Are there any topics you would like to see on the agenda?
Benedict Evans recently gave his semi-annual presentation, "AI Eats the World’. I’d like for there to be discussion on that.
sounds great, cheers! Is it this talk? https://www.youtube.com/watch?v=rTrFb-cSiwE
Hi everyone, thank you for today’s call (2025.11.27). This is the generated summary (short version, please also find it here ): The group opened by reflecting on the macro-tech landscape laid out by Benedict Evans’ recent “AI eats the world” presentation, noting how rapid advances and proliferation of AI reshape not only software but also infrastructure and energy demands. They then discussed ideas from the latest Outlier Ventures’ “Post-Web” report (Chapter 3 update), interpreting its forecast for decentralized, composable Web3 systems and how AI-driven infrastructure and distributed networks could become intertwined. The conversation tied this to parallels in energy and infrastructure planning (including the example of Hydro-Québec and national-level energy strategies like Canada’s SMR Action Plan for small modular reactors) for long-term, systemic planning to meet rising electricity demand. Finally, from a purely technical front, they reviewed the WebNN API — a new web-standard specification enabling efficient on-device neural-network computation — as a component enabling decentralized AI on user devices, without reliance on centralized cloud infrastructure. Taken together, the call underscored that for AI and Web3 to succeed sustainably, it’s not just about software innovation — we must also think about infrastructure, governance, and long-term systemic resilience.
Very interesting and relevant comment by @icarus (to consider for which game DeAI providers are in and how to play it well):
“So I think the game for small players like ICP is some optimal combination of :
a) much lower capital outlay
b) lower hosting & power costs
c) much lower operating overhead
d) find higher value niche inference markets
e) make the inference service tightly integrated into a high value platform of complementary software services”
Links shared during the call:
Hi everyone, this week @icarus will lead an AI4AI session where we’ll focus on hardware and infrastructure to support DeAI on ICP
I’m looking forward to the session! Have a great week and see you on Thursday
Hi everyone, thank you for today’s call (2025.12.04). Special thanks to @icarus for leading it! This is the generated summary (short version, please find the long version here ): This session highlighted ICP’s differentiation in decentralized AI through small, specialized models, ultra-low-overhead DevOps, and strong chain-fusion interoperability. Participants emphasized that ICP’s automated, fast deployment of agents, combined with efficient hardware–model co-design, provides a unique platform advantage. A decentralized “DNS for models and agents,” trust guarantees, and on-chain-adjacent low-latency inference were identified as key future opportunities. Overall, the discussion pointed toward a long-term vision of trusted, composable, decentralized AI agents that can interact natively across Web3 ecosystems.
Links shared during the call:
- https://www.bleepingcomputer.com/news/artificial-intelligence/leak-confirms-openai-is-preparing-ads-on-chatgpt-for-public-roll-out/
- Acumen GSTI – Fetch
- GitHub - icdevs/ICEventsWG: WG for developing an Event System(pub/sub) on the Internet Computer
- Olmo 3: America’s truly open reasoning models
- Introducing Mistral 3 | Mistral AI
- IBM Granite 4.0: Hyper-efficient, High Performance Hybrid Models for Enterprise
- SmolDocling: Vision-Language Model for Document Conversion | Niklas Heidloff
- DeepSeek’s new open-source model could decode large amounts of documents more efficiently | IBM
- Olmo 3 Shows How Far Open-Source Reasoning Can Go
- OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens
- Small Language Models are the Future of Agentic AI: Small Language Models are the Future of Agentic AI
- Small Language Models are the Future of Agentic AI
- How Small Language Models Are Key to Scalable Agentic AI | NVIDIA Technical Blog
Hi everyone, are there any topics you would like to see on this week’s agenda? Or does anyone want to present or share something during the call? Please let us know
have a great week