From 203fc241b01245a5a83d8da3daddd521f0ba6c74 Mon Sep 17 00:00:00 2001 From: mrkvon Date: Wed, 1 Apr 2026 00:07:06 +0200 Subject: [PATCH] Meeting minutes 2026-03-19 Open discussion about Solid and AI --- meetings/2026-03-19.md | 134 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 134 insertions(+) create mode 100644 meetings/2026-03-19.md diff --git a/meetings/2026-03-19.md b/meetings/2026-03-19.md new file mode 100644 index 0000000..edeaeef --- /dev/null +++ b/meetings/2026-03-19.md @@ -0,0 +1,134 @@ +# Solid Practitioners + +## Meeting Date + +- 2026-03-19, 15:00 UTC + +## Details + +- call https://meet.jit.si/solid-practitioners +- chat https://matrix.to/#/#solid-practitioners:matrix.org +- repo https://github.com/solid-contrib/practitioners +- recordings https://spectra.video/c/solid_practitioners/videos + +## Facilitator + +- Jeff Zucker + +## Scribe + +- Michal + +## Participants (please add yourself) + +- [Noel De Martin](https://noeldemartin.com) +- [elf Pavlik](https://elf-pavlik.hackers4peace.net) +- [Timea Turdean](https://timea.solidcommunity.net/profile/card#me) +- [Michal](https://id.mrkvon.org) +- [Luke Dary](https://w3c.social/@lukedary/) +- [James Zaki](https://x.com/james_zaki) +- [Jackson Morgan](https://jackson.solidcommunity.net/profile/card#me) +- Theo [@thhck](https://w3c.social/@thhck) +- Tim Berners-Lee +- Chris Carrier + +## Agenda + +### Administrative + +- Call to add self to hackmd +- Call for Scribe +- Objections to recording the session? + +Jeff: Today's session - there is many people working on Solid and AI - we'll continue with sessions in the future. + +### New Member Check-in + +- David - from Midwest US, interested in Solid many years ago. History of entrepreneurship in internet space, lawyer involved in privacy laws in US, DCC act, heard privacy go off rails with Zuckerberg at Harvard, etc. Interested in promoting Solid and promoting in companies. +- Chris Carrier - new to the community, long time software engineer, experienced disappointment with development in terms of privacy, heard about Solid which fits well into things I want to build. Ideas, building, on top of Solid, experimenting. Happy to find community of other people who are interested. +- looking to commercialize protocol like with http [???] +- James Zaki - company Logos - privacy-first and local-first to some degree, building module that will help conform with Solid protocol. + +### Announcements + +- New alternating Practitioners' meetup schedule starting from April + - North/South meeting schedule + - 1st Thursday 16:00 UTC + - 3rd Thursday 5pm Australian time + - April 16 - University of Australia and Indigenous communities + +- Progressive Stack see https://github.com/solid-contrib/practitioners/discussions/41 + - Take a look, we'll try it out. + - We want this to be an open space welcoming to newcomers and people not already embedded in the community. To be more welcoming to people who might be not heard. + +- ODI Office Hours every second Tuesday - next one at 16 + - 16:00 Europe/London time + +### Solid and AI - Discussion + +[Recording]() + +- Paul, who couldn't attend has been working on Solid/AI, see https://interition.ai/ + +Round, what you're working related to Solid + +- Noel - I'm app developer making Solid apps, like Media Kraken and Umai - I'm building something with AI for Solid - I call it agent, but it will be agent in the future. You log in using Solid, and connect any models you want. Ollama locally, or commercial... There are a lot of AI tools, but I don't like their defaults. Even if ou tell it to use Solid or local-first, it doesn't do it well. I want to provide alternative to AI that doeswe hat I want, all open-source. The app is called Ànima: https://anima.noeldemartin.com +- eP: is it similar to ODI's libre chat? +- Noel - idk +- Luke: First I wanted to have 0-dollar hosted pod server, so I had AI generate https://github.com/daryltd/paa.pub, have personal pod that you can host with CloudFlare. I don't think it has any costs. It also has ActivityPub and ActivityStreams capability. It's hacked, + - also developed RSVP app - share public container similar to SIgnup genius. + - I work at RedHat, internally working on Solid, recently OxiGraph, added SHACL and horizontal scaling, would love people to test that. https://github.com/chapeaux/oxigraph-cloud +- Jackson: I wanted structured RAG - I'd like question in SPARQL query then bring it back. Most databases aren't in RDF, so recently I've focused on tool that automatically scans your databases, converts them to RDF. It does very well with SQL databases. Then it puts it all to Solid, and SPARQL querying it. Automatic integration generation. If anybody has databases. +- Chris: Age verificaion in dystopian way. I'm trying age verification trustless, on top of Solid. Building this with AI tooling and Solid apps. A lot of people doing cool stuff. Also happy to contribute other +- James: I'm interested in how peple are usign zero-proof data. +- Chris: yes, that's the technology I tried to use +- Theo: I experiment similar to ... project - prompting text to React component. Instead of text you give shape, and build React component out of this shape. +- Wilf: Mostly using Vector similarity searches. Interested in light-weight way to give advice. Homeless populations in Portland. Associate vectors with various services. We haven't implemented it yet, but it seems feasible. We're working with BitFocus. We want people to provide with virtual case-worker. They wouldn't have to provide all the data, eventually they could connect pod and fill forms for different social programs (SNAP, ???) +- Jeff: I've been experimenting with it, no specific focus, trying to figure out what it's good for. Opening +- Tim: I wrote Charlie - a chat pre-chatGPT. At the time people were using. Other services work for Amazon etc, Charlie works for you. You can choose any model, then it will add contents of your Pod to the context. Then you can automatically improve responses, when you ask for something you want, you can demonstrate test data, examples, and so on. That's all written up, I'll add link. Inrupt is now working on Charlie - will probably be announced sometime in June. +- Noel: I'm also building something to choose models. For me AI that works for you would be a model trained in a certain way. In the end there's always a system prompt. Whatever you tell them, they still have a prompt that overrules everything you have said. They're not working for you. +- Jackson: For the ETL generator, I've been putting effort for working everything with local models. Open-source models are less powerful, but you can make up for it by providing enough context. +- Noel: They are open-weight in sense you can run them, but they're not open-source in the sense you don't know how they've been trained. In the end that's pretty difficult to know +- Jeff: ODI has [a list of transparency for each model](https://theodi.org/insights/reports/the-ai-data-transparency-index/). Best models only had green light on half of them. Even the best open-source don't tell us what they're doing. +- James: Friend was comparing buying beefy RAM; but the price is much higher than commercial models. +- eP: there some dedicated computers giving 128 unified memory. It's comparably slow, but I think in couple years we might have something that works out. Also difference in how you use it (coding or chatting). +- Noel: I have doubts about practical stuff. Should AI be mounted in a Pod? I have a list of movies, when I talk with AI I want it to recommend movie. If it accesses all mvoies using Solid protocol, it will be slow. I used it with local-first. Does anyone have thought sabout this trade-off? +- Luke: There are ways to do Solid MPC. Not necessarily pure Solid experience, but MPC layer, WebMPC. You could surface a tool for AI which then grabs data. +- Noel: Challenge is: Should the tools rely just on Solid protocol, or something else as well? Maybe AI can use more powerful tools. +- Jackson: Your AI can generate sparql query, then Comunica will communicate with a bunch of pods via different interfaces (SPARQL, TPF, raw). As LWS supports more Sparql, LDF, that's the direction I imagine. +- eP: Does anybody know what's going on in ODI planet? +- Jeff: Samu working on MCP. I invited Samu to come, maybe they'll come. Kwaku has prepared skills for AI. Now there is a [repo with skillsets](https://github.com/solid/solid-llm-skills). +- Jackson: embeddings databases on Solid. There is https://www.semantic-web-journal.net/system/files/swj1738.pdf - generating vectors from RDF graph. We can run RDFtoVec - then we would do RAG in traditional way. +- Noel: AI generating SPARQL and executing it: Are LLM good at generating Sparql? +- Jackson: You need to give it examples and schemata and hodl its hand, then it's fine. +- Luke: it's good at SPARQL, I have MCP tools that write convenience functions. +- eP: do you have links what's useful? I used tools that produced working SPARQL but nothing elaborate. If you have tips. +- Jackson: Luke, do your tools inform AI how to construct SPARQL query? or +- Luke: Lookup entity, add entity, AST tree-walking to look at tool taking code, putting to SPARQL, then queryable. Goal to have code-bases searches. +- Jackson: Do you need to build new tools for this, or are the tools modular enough for reuse? +- Luke: I'd like to have AI tool that generates tools. +- Rahul: I've started writing AI disclosure statement. I have a template, that would be a go +- James: It's a tool, human will be responsible for the result. ... +- Wilf: I was going to say similar: I'm not sure how it matters. When I write code, I'm responsible for the output. +- eP: I don't think it matters. It's important to have test harness. I think it's secondary what's the mixture fo human vs AI +- Rahul: the difference is that AI is probabilistic. Giving sufficiently high probability - did human give enough review? That information is important for me. Right now it's important for me. e.g. food labeling in US is poor, that's a scam. Consumer has right to know. +- Wilf: We make mistakes as humans. I will produce better code with AI. ... +- Rahul: Most code has imcomplete tests. You have people who vibe-code everything +- Wilf: We want to encourage people to use best-practices. Should be suspicious of code without test-suites. But I'm more likely to produce tests with AI. I don't think % of code written by AI matters for your concern. +- Rahul: I recommend Marshall McLuhan: Tool and environment in which it exists matters, and interaction with other tools. Canadian school of media. +- Theo: I agree it's not important the %, but it's nice to see what thoughts have been put in. It's good to know what comes from who. In NLNet best advice is to put prompt in the commit. I think it's an intereting practice to have transpareny over generated code. These things can testify about the quality of efforts with +- Jackson: I can share what we're doing for LDO: One of the problem I see with AI and not disclosing: Disclosing is more useful as social tool for other contributors. We have rule to not use AI for architecture. Even when AI happens to find correct code this time, it may not next time, and it won't help to understand. +- Pavlik: I linked to the policy. https://nlnet.nl/foundation/policies/generativeAI/ I was participating in the original NLNet meeting. I'm happy to share my full OpenCode, but there's a lot of noise. NLNet are also trying to figure out thigs, still. Short leash for AI. +- Th: I agree with noise, there is practice with AI to write plan then review it. I was thinking about linkign this document, and these files are reviewable by human being. +- James: There was a good video from Netflix founder. They'd get AI to have summarize. +- [...] +- Jackson: The biggest concern is that labour will no longer be required to hold the capital class to hold them from destructive tendencies. Will be 5 rulers of AI. Solid could play a role. EU efforts, focus on data ownership. Solid could be part of providing infrastructure to support projects. You can have guardrails on powerful people. +- James: The jobs I started - they're trying to pioneer grassroots campaigns. +- Rahul: There is a modest effort on IDF which is is AIpref. I think rules for bots will have to be enforced by law maybe by EU. Space worth following. +- eP: For me Solid could be more of Social tools rather than technical. I see ValueFlows, Solid, NextGraph help us have more +- Rich: Law aspect: GDPR +- Noel: I agree with Jackson, many people don't pay [...] +- [please watch the recoding ... ] +- Rich: The cost of providing AI is huge. That'll leave out people who can't afford to pay for chatbots. + +[Please watch the recording for the rest.]