14:46:52 RRSAgent has joined #webmachinelearning 14:46:56 logging to https://www.w3.org/2026/05/28-webmachinelearning-irc 14:46:56 RRSAgent, make logs Public 14:46:57 please title this meeting ("meeting: ..."), anssik 14:46:58 Meeting: WebML CG Teleconference – 28 May 2026 14:47:07 Chair: Anssi 14:47:15 Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-05-28-cg-agenda.md 14:47:20 Scribe: Anssi 14:47:25 ScribeNick: anssik 14:47:39 Present+ Anssi_Kostiainen 14:47:39 RRSAgent, draft minutes 14:47:41 I have made the request to generate https://www.w3.org/2026/05/28-webmachinelearning-minutes.html anssik 14:58:36 Present+ Victor_Huang 15:00:04 Present+ Mark_Foltz 15:00:19 msw has joined #webmachinelearning 15:00:42 gurusingh has joined #webmachinelearning 15:00:47 Mike_Wyrzykowski has joined #webmachinelearning 15:00:48 Mark_Foltz has joined #webmachinelearning 15:00:54 Present+ 15:01:04 Present+ Ryan_Roemer 15:01:06 domfarolino has joined #webmachinelearning 15:01:22 Present+ Benjamin_VanderSloot 15:01:24 Present+ 15:01:25 Present+ Brandon_Walderman 15:01:31 Present+ Dominic_Farolino 15:01:40 alispivak has joined #webmachinelearning 15:01:48 Present+ Guru_Singh 15:01:49 Present+ Mike_Wasserman 15:01:56 brwalder has joined #webmachinelearning 15:02:04 Present+ Mike_Wasserman 15:02:04 bvandersloot has joined #webmachinelearning 15:02:14 Present+ Thomas_Steiner 15:02:22 Present+ Ben_Greenstein 15:03:05 Present+ Mike_Wyrzykowski 15:03:11 BenGreenstein has joined #webmachinelearning 15:03:31 Anssi: as a reminder, we'll use IRC-based queue management in this meeting: 15:03:35 -> https://www.w3.org/guide/meetings/zakim.html#speakerqueue 15:03:37 Julia has joined #webmachinelearning 15:03:44 Anssi: to suggest agenda topics, use Agenda+ label, e.g.: 15:03:53 -> Agenda+ https://github.com/webmachinelearning/webmcp/labels/Agenda+ 15:04:10 Anssi: please welcome our latest new participant, Tugce Tuncay, joining as an individual contributor 15:04:24 Topic: Prompt API 15:04:29 gb, this is webmachinelearning/prompt-api 15:04:29 anssik, OK. 15:04:33 ryan-roemer has joined #webmachinelearning 15:04:36 Subtopic: Model selection and availability 15:04:44 Anssi: issue #169 15:04:45 https://github.com/webmachinelearning/prompt-api/issues/169 -> Issue 169 [Tag Review] - Model selection and availability (by etiennenoel) [Agenda+] [tag-tracker] 15:05:18 ... TAG review of the Prompt API spec raised a question about how model selection and availability would work in the Prompt API 15:05:23 ... per TAG review comment it would be "acceptable to indicate particular model capabilities instead of 'brand names'" 15:05:34 ... concerns about cost of licensing a model 15:06:01 ... Benjamin shared a reference to Mozilla's standards position 15:06:15 ... two broader concerns "Calcifying around a single model and Lack of model neutrality" 15:06:39 Anssi: AiBrow browser extension for Chrome and Firefox has extended the Prompt API with a model selection mechanism 15:06:44 -> https://docs.aibrow.ai/examples/using-different-models 15:07:16 const session = await AI.AIBrow.LanguageModel.create({ 15:07:16 model: "phi-3-5-mini-instruct-q4-k-m" 15:07:16 }) 15:07:25 Anssi: in the AiBrow implementation, the model is automatically downloaded if not present 15:07:29 ... and the model becomes available to all sites on the machine after that 15:07:38 ... I believe we can see fingerprinting and cross-site tracking issues with this design 15:07:39 A similar model selection mechanism exists in Chrome's extension -> https://chromewebstore.google.com/detail/built-in-ai/aekdeffgkjegkdbdnkllpaeakcapfjdb 15:07:52 johannhof has joined #webmachinelearning 15:07:53 ... however, I think we also see potential solution to mitigate these issues 15:08:37 Anssi: Thomas from Google shared another extension developed by the Chrome team that allows selecting from cloud and local backends 15:09:13 Tom: an experiment, to test polyfills of the built-in AI APIs 15:09:23 ... testing if we can use this extension to use these APIs on other browsers 15:09:35 ... this also works on Edge and mobile 15:09:41 ... Android supported 15:09:46 ... in this extension, cloud backends supported are Google Gemini, OpenAI, Google Firebase 15:09:50 ... and local backend is Transformers.js 15:10:10 ... options allow to override the native implementation 15:10:22 ... an experiment mostly 15:10:23 q+ 15:11:31 Benjamin: picking between models solves calcification to some extend, but makes sites branching their code paths to models and their versions 15:11:56 ... I shared an idea of a possible solution in the GH issue 15:12:07 ... if the browser downloads 1-of-N diverse models, it would make it harder for a site to depend on a particular model behavior 15:12:20 ack reillyg 15:12:51 Reilly: I agree with Ben's idea of downloading 1-of-N diverse models, with fully randomized models, not sure how that'd work 15:13:38 ... the hard line would be, we can't let the site pick which model to choose, model selection API is a non-starter due to fingerprinting aspect, to detect which models are installed 15:13:56 q+ 15:13:58 ... also practical question is it uses a lot of local storage 15:14:23 ... if we allow sites to download multiple models it'd make this storage issue larger than with the current approach 15:14:40 q+ 15:14:55 ... to require that the model B has an open license could be a path to pursue 15:16:00 ack bvandersloot 15:16:37 Benjamin: introduction of openness on the models helps, as well as use of platform model, it is just one of the hard things, without a model picker, assuming language models and interchangeable 15:16:59 ... if that is not the case, confronts the shape of the API 15:17:00 q? 15:17:11 ack msw 15:17:52 MikeWa: appreciate the discusion, I'm more aligned on Ben's comment on make the API threat the models fundamentally interchangeably, it has been out design goals 15:18:17 ... to be able to request specific modalities for example, use that level of abstraction instead or low-level model details 15:18:29 ... that may be more useful that "I want this specific model version" 15:18:57 ... also, we do see Microsoft contributing in Chromium exploring the use of OS-provided models, also explore the use of Apple Intelligence 15:19:21 ... OS-provided models in another good options, also +1 on Reilly to more toward open-weight models and open-source inference stacks, there are on our roadmap 15:20:00 ... slightly different topic, one aspect we hope to explore is to reaffirm the interchangeability of models by common suite of use cases, samples and benchmarks, to test interoperability 15:20:43 ... this is important as Chrome launches new versions of models, to ensure high-value use cases are addressed also when new model versions are being rolled out 15:20:45 q? 15:21:34 Anssi: any preferred next steps in mind? 15:21:50 Benjamin: bringing proposals to the Mozilla's standards position issue might work 15:21:59 ... and continue conversation on mitigation 15:23:04 q+ 15:23:13 Benjamin: I can also help bring Jake into the TAG review and this GH issue here to help drive productive discussion 15:23:39 ack msw 15:24:15 MikeWa: I hope some discussion will continue in the Prompt API repo itself, to help advance the interop project, since Mozilla standards position issue is closed I don't us to engage there at this point 15:24:32 ... appreciate continued engagement from Jake and other Mozillians in this repo and also in the TAG review issue 15:25:26 The latter. 15:25:42 +1 to latter 15:26:48 MikeWa: proposed resolution looks good to me, we also look into interop project with Microsoft to further this aim 15:27:10 RESOLUTION: Explore solutions for the Prompt API to address calcification and model neutrality issues. (issue #169) 15:27:19 Subtopic: Sampling modes 15:27:26 Anssi: issue #203 15:27:27 https://github.com/webmachinelearning/prompt-api/issues/203 -> Issue 203 Proposal: Introduce Categorical Sampling Modes for the Prompt API (by isaacahouma) [enhancement] [interop] [Agenda+] 15:27:32 ... Isaac opened this issue 15:27:44 ... Isaac explains that initially sampling parameters temperature and topK were part of the API 15:27:50 ... based on TAG feedback, there parameters were deprecated in the web pages context, exposed to extension only for now 15:27:58 ... issue was behavior drift across models and versions that breaks API expectations 15:28:13 ... there are valid use cases, so instead of removing entirely, proposal to bucket them into categorical sampling modes, currently: 15:28:21 ... "deterministic", "precise", "balanced", "creative", "imaginative" 15:28:30 Anssi: it is an implementation detail how these modes map to temperature and topK and other sampling parameters 15:29:31 MikeWa: developers want to express more deterministic behavior, and explicit tuning parameters are not interoperable, but the modes we propose, is a better way that can be implemented across model families and version interchangeably 15:29:43 ... currently in Origin Trial in Chrome, getting developer feedback on it now 15:30:05 ... feedback from this group would help determine if there are other options for configuring this aspect in an interoperable fashion 15:30:29 Anssi: the issue has the IDL proposal 15:30:56 ... any comments or feedback on the proposal? 15:30:57 q? 15:31:26 q+ 15:31:32 ack reillyg 15:32:05 Reilly: wanted to clarify that this is intended to resolve TAG feedback 15:32:31 ... hypothesis is this will give developers the level of control over model creativity without developers having to figure specific values 15:32:46 ... we do this as an Origin Trial to explicitly ask developers feedback on this feature 15:33:14 ... by running this experiment we want to get explicit feedback why they are using parameters, so we can build the API in a model agnostic way 15:33:15 q? 15:34:50 We'd love to have the group's blessing to experiment. :) 15:35:06 Victor has joined #webmachinelearning 15:35:22 +! 15:35:23 sgtm 15:35:25 +1 15:35:51 sgtm 15:35:53 RESOLUTION: Experiment with categorical sampling modes to inform the Prompt API design. (issue #203) 15:36:02 Topic: WebMCP 15:36:06 gb, this is webmachinelearning/webmcp 15:36:06 anssik, OK. 15:36:12 Subtopic: WebMCP early wide review update 15:36:26 Anssi: I gave a heads up on our wide review plan to the chairs of the TAG, Security WG, Privacy WG 15:36:45 ... for Architecture review with the TAG, the expectation is to file an incubation review with them, this class of review expects certain materials: 15:36:52 ... - explainer that describes the problem to solve from an end-user's perspective 15:37:14 ... - multi-stakeholder feedback from Chromium, Mozilla, WebKit, and other implementers, web developers, users 15:37:25 ... - any major unresolved issues 15:37:37 Anssi: we have updated the explainer in PR #183, thanks Dominic for leading that rewrite 15:37:38 https://github.com/webmachinelearning/webmcp/pull/183 -> MERGED Pull Request 183 Explainer updates/rewrite (by domfarolino) 15:37:46 ... minor tweak to the flow diagram in review, PR #189, expect this to land soon 15:37:47 https://github.com/webmachinelearning/webmcp/pull/189 -> Pull Request 189 Convert MCP flow diagram to mermaid (by bwalderman) 15:37:52 -> Updated explainer https://github.com/webmachinelearning/webmcp/blob/main/README.md 15:38:02 Anssi: for multi-stakeholder feedback, we have requested standards positions from WebKit and Mozilla: 15:38:06 -> https://github.com/WebKit/standards-positions/issues/670 15:38:07 https://github.com/WebKit/standards-positions/issues/670 -> Issue 670 WebMCP (by domfarolino) [topic: artificial intelligence (AI)] [from: Google] [venue: W3C Web Machine Learning CG] 15:38:17 -> https://github.com/mozilla/standards-positions/issues/1412 15:38:18 https://github.com/mozilla/standards-positions/issues/1412 -> Issue 1412 WebMCP (by domfarolino) 15:38:50 Anssi: when the explainer diagram is updated, we can proceed 15:39:11 ... I propose the editors draft the proposed submission content somewhere in this repo, give the entire group a few days to review before we submitting to the TAG so we have consensus on the content of the submission 15:39:30 ... we can always update the content of the submission later if we need to and as we make progress, but it would be good to have a solid initial submission to not waste busy reviewers' time 15:40:38 ... questions? 15:40:50 brwalder has joined #webmachinelearning 15:41:28 Brandon: the explainer cleanup looks good 15:42:12 Anssi: for W3C's security review with the Security WG, we are expected to write a "Security Considerations" section for the spec, taking into account the Self-Review Questionnaire on Security and Privacy 15:42:33 ... we have combined security and privacy considerations due to the overlap and to avoid duplication of content, this is fine for this spec at this stage 15:42:42 ... Security and Privacy considerations section landed in PR #181 15:42:43 https://github.com/webmachinelearning/webmcp/pull/181 -> MERGED Pull Request 181 Port Security & Privacy considerations from docs/ (by johannhof) 15:44:10 Victor: I looked at the self-review questionnaire while authoring the initial S&P consideration 15:44:18 -> https://webmachinelearning.github.io/webmcp/#security-privacy 15:44:50 Anssi: thank you Johann, Victor and others for your work on this 15:45:39 Dominic: TAG review seem to like to have the questionnaire responses reference, so I propose we document that 15:45:51 Victor: I can open a PR to add that to the repo 15:45:59 Dominic: sounds great, thanks 15:46:32 Johann: I agree with Dominic, let's get the self-review documented too 15:47:06 ... agentic AI risks are covered, traditional security model considerations could be better covered in the content 15:48:10 ... I've proposed we explore the threat model we've discussion in parallel in a non-blocking manner, awaiting confirmation from the Security WG on their expectations 15:48:18 ... comments? 15:48:28 Anssi: for W3C's privacy review, the expectation wrt materials is similar 15:48:45 ... "Privacy Considerations" section, we'll offer the combine S&P considerations for this group as well 15:48:50 -> https://webmachinelearning.github.io/webmcp/#security-privacy 15:48:58 Anssi: I have also informed the Privacy WG chair about our plans, and asked about their expectations for the review 15:49:02 ... will update the group when I have more information 15:50:34 +1 15:51:16 +1 15:51:17 RESOLUTION: The group uses the updated Explainer and the newly authored Privacy and Security Considerations as a reference, and adds a self-review questionnaire responses, then stages review requests for the group to review before submission. 15:51:25 Subtopic: Hint for reversible or consequential actions 15:51:29 Anssi: issue #176 15:51:29 https://github.com/webmachinelearning/webmcp/issues/176 -> Issue 176 Hint for reversible or consequential actions (by johannhof) [Agenda+] 15:51:41 ... the proposed design is that reversible or consequential actions hint is off by default, and the developer needs to explicitly mark the tool as such 15:51:45 ... and last time we resolved to identify use cases to inform this design 15:52:01 ... it looks like Dominic and Johann had a discussion around use cases for tools and concluded that most tools are not destructive or consequential, maybe ~90% fall in this category? 15:52:15 ... Victor took a soft stance to support consequentialHint default=false based on review of existing use cases by MCP tools 15:52:30 ... Johann notes the developer may forget to tag the tools, will use the defaults 15:52:37 ... proposed design assumes tools are "not consequential" by default 15:52:58 q+ 15:53:17 Johann: for the readOnly tooling, agents could enter into e.g. research mode where only read-only tools are allowed 15:53:31 ... explicit read-only tools would be valuable in that case 15:53:49 ... I'm wondering if this is the same situation, do we want to list all consequential tools 15:54:11 ... I tend to agree with others that this should be high user friction point 15:54:26 ... "approve all the tool all the time" should not be the goal, due to prompt fatique style issues 15:54:44 ... I think we have alignment on the proposed design 15:54:53 q+ 15:54:55 q+ 15:55:07 ... all the discussion seem to be supporting the "off-by-default" model hypothesis is worth exploring further, but we don't have enough data to confirm that hypothesis yet 15:55:10 ack domfarolino 15:55:33 Dominic: we're Origin Trialing the feature and will get data from user whether this feature would work per out hypothesis 15:55:59 Johann: we have to make sure to work with partners to understand how to expose consequential tools, what is the right UX treatment for that 15:56:00 q? 15:56:06 ack bvandersloot 15:56:50 Benjamin: want to resurface that if the consequential defaults to false, it might be worth explore hint matching with MCP open world hints 15:57:31 Johann: a bit difficult situation, existing hint naming is not optimal, do we follow the existing naming or better names that convey the meaning better 15:57:32 q? 15:57:36 ack Victor 15:57:55 Victor: I want to respond to Ben and why we should be following MCP inspired hints 15:58:08 ... I think not much consideration was put in the design of MCP hints 15:58:36 ... MCP hints used to just prompt the user somehow 15:59:05 ... copying MCP design may not be the best design for WebMCP hints 15:59:21 ... I prefer the name reflect the intent of the hint 15:59:53 ... in MCP clients, people need to enlist into use of the hints, only then they can be used 16:00:07 ... we vet websites before use, thus defaults are more important than in MCP 16:00:27 ... if people don't label thinks as destructive by default they assume the to be descructive 16:01:09 ... lean on "consequential" hint to be false by default 16:01:22 Dominic: how would this be used other than prompting the user? 16:01:23 q? 16:01:32 Dominic: should document why this is not called? 16:01:37 q+ 16:01:51 Victor: my rationale is, in this situation, we can this to be more of a hint that deterministic contract 16:02:05 ... in the future the model may be so good does not need to prompt the user 16:02:10 ... ack johannhof 16:02:32 Johann: I agree with Victor, people tend to misunderstand, they don't need what needs prompting 16:02:58 q? 16:03:01 ack johannhof 16:03:02 q? 16:04:35 Victor: open to wait a for the Origin Trial data, but also OK to proceed 16:04:39 +1 16:05:16 +1 16:05:22 +1 16:05:22 RESOLUTION: Specify consequential hint as default=false (issue #176) 16:05:31 RRSAgent, draft minutes 16:05:33 I have made the request to generate https://www.w3.org/2026/05/28-webmachinelearning-minutes.html anssik 16:07:38 s/potential solution/potential solutions 16:08:25 s/some extend/some extent 16:09:43 s/and interchangeable/are interchangeable 16:10:35 s/API threat/API treat 16:10:49 s/out design/our design 16:11:07 s/goals/goal 16:11:21 s/instead or/instead of 16:11:35 s/useful that/useful than 16:12:00 s/in another good options/in another good option 16:12:21 s/there are on/these are on 16:13:12 s/us to engage/expect us to engage 16:15:02 s/there parameters/these parameters 16:15:47 s/propose, is/propose is 16:18:09 s/we submitting/we submit 16:24:09 s/from user/from users 16:24:25 s/out hypothesis/our hypothesis 16:25:43 s/should be following/should not be following 16:26:16 s/thinks as/things as 16:26:34 s/the to be/them to be 16:27:12 s/we can this to/this can 16:27:26 s/that deterministic/than deterministic 16:27:53 s/need what/know what 16:28:06 s/a for the/for the 16:28:16 RRSAgent, draft minutes 16:28:18 I have made the request to generate https://www.w3.org/2026/05/28-webmachinelearning-minutes.html anssik 18:15:19 Zakim has left #webmachinelearning 22:18:05 sushraja has joined #webmachinelearning