14:47:31 <RRSAgent> RRSAgent has joined #ai-model
14:47:35 <RRSAgent> logging to https://www.w3.org/2024/09/25-ai-model-irc
14:47:35 <tpac-breakout-bot> RRSAgent, do not leave
14:47:36 <tpac-breakout-bot> RRSAgent, make logs public
14:47:37 <tpac-breakout-bot> Meeting: AI Model Management
14:47:37 <tpac-breakout-bot> Chair: Michael McCool
14:47:37 <tpac-breakout-bot> Agenda: https://github.com/w3c/tpac2024-breakouts/issues/15
14:47:37 <Zakim> Zakim has joined #ai-model
14:47:38 <tpac-breakout-bot> Zakim, clear agenda
14:47:38 <Zakim> agenda cleared
14:47:39 <tpac-breakout-bot> Zakim, agenda+ Pick a scribe
14:47:41 <Zakim> agendum 1 added
14:47:41 <tpac-breakout-bot> Zakim, agenda+ Reminders: code of conduct, health policies, recorded session policy
14:47:41 <Zakim> agendum 2 added
14:47:41 <tpac-breakout-bot> Zakim, agenda+ Goal of this session
14:47:42 <Zakim> agendum 3 added
14:47:42 <tpac-breakout-bot> Zakim, agenda+ Discussion
14:47:42 <Zakim> agendum 4 added
14:47:42 <tpac-breakout-bot> Zakim, agenda+ Next steps / where discussion continues
14:47:43 <Zakim> agendum 5 added
14:47:43 <tpac-breakout-bot> tpac-breakout-bot has left #ai-model
19:02:17 <alcooper> alcooper has joined #ai-model
20:03:27 <lambda> lambda has joined #ai-model
20:09:42 <sebastian> sebastian has joined #ai-model
20:12:32 <sebastian> scribe sebastian
20:12:45 <Max> Max has joined #ai-model
20:13:20 <McCool> McCool has joined #ai-model
20:13:43 <gkellogg> gkellogg has joined #ai-model
20:14:25 <HXavier> HXavier has joined #ai-model
20:16:01 <McCool> slides: https://github.com/webmachinelearning/hybrid-ai/blob/main/presentations/WebML%20Discussion%20-%20Hybrid%20AI%20for%20the%20Web%20-%20AI%20Model%20Management%20TPAC%202024%20Breakout.pdf
20:17:05 <gkellogg> present+ Gregg_Kellogg
20:17:06 <anatoly-scherbakov> anatoly-scherbakov has joined #ai-model
20:17:11 <anatoly-scherbakov> present+
20:17:12 <Evan> Evan has joined #ai-model
20:17:38 <asully> asully has joined #ai-model
20:17:57 <sebastian> <McCool showing agenda>
20:18:16 <ora> ora has joined #ai-model
20:18:23 <ora> present+
20:18:26 <ayu> ayu has joined #ai-model
20:18:49 <sebastian> Problem Statement ... Issues and Constaints ... Discussions ... Next steps
20:19:01 <sebastian> topic: Problem statements
20:19:16 <chrisp> chrisp has joined #ai-model
20:19:20 <sebastian> McCool: AI models ...
20:19:26 <sebastian> ... can be very large
20:19:35 <sebastian> ... will often be shared
20:19:43 <sebastian> ... are often updated
20:19:57 <kenji_baheux> kenji_baheux has joined #ai-model
20:20:30 <reillyg> reillyg has joined #ai-model
20:20:33 <reillyg> present+
20:20:44 <sql> sql has joined #ai-model
20:20:46 <sebastian> McCool: There is also the current same-origin storage portioning policy
20:21:02 <sebastian> ... this preserves privacy issues
20:21:22 <zolkis> zolkis has joined #ai-model
20:21:56 <dezell> dezell has joined #ai-model
20:22:04 <dezell> present+
20:22:10 <kenji_baheux> present+
20:22:33 <sebastian> ... this works for images (not shared) and for software libraries (often shared)
20:23:12 <tomayac> tomayac has joined #ai-model
20:23:33 <sebastian> topic: Use Cases for Large Models
20:23:56 <sebastian> McCool: Language translation, ...
20:24:07 <sebastian> ... meeting captions
20:24:20 <sebastian> ... background removal
20:24:33 <sebastian> ... video creation and editing
20:24:49 <sebastian> ... written language recognition
20:24:55 <sebastian> ... personal assistant
20:26:05 <kenji_baheux> who is speaking?
20:27:28 <sebastian> XXX: Shall we specify a shared repository which only has my personal information?
20:28:25 <sebastian> @@@
20:28:39 <sebastian> topic: Why run AI?
20:29:14 <sebastian> McCool: There are pros such as latency and being used offline.
20:29:39 <sebastian> ... cons are size limitations, download time, and storage costs
20:29:58 <sebastian> topic: Model Size vs Download Time
20:30:40 <sebastian> McCool: Average home networks speeds are between 45-216 Mbps
20:32:01 <sebastian> ... downloading Phi-3-mini will result,e.g, to 22mins for baseline
20:32:12 <sebastian> topic: Existing APIs and Experiments
20:33:08 <sebastian> McCool: For the same-origin there exists HTTP Cache, ...
20:33:13 <sebastian> ... Cache API
20:33:20 <sebastian> IndexedDB API,
20:33:44 <sebastian> s/IndexedDB API,/... IndexedDB API,
20:33:56 <sebastian> ... Origin Private File System API
20:34:26 <sebastian> McCool: For the cross-origin there exists the File System Access API
20:34:48 <sebastian> topic: Caching Desired Properties
20:35:34 <sebastian> McCool: we need to reduce latency, ...
20:35:41 <sebastian> ... Bandwidth
20:35:46 <sebastian> ... Storage
20:36:00 <sebastian> ... and preserve privacy
20:36:11 <sebastian> topic: Security and Privacy Considerations
20:36:14 <Domenic> Domenic has joined #ai-model
20:36:55 <sebastian> McCool: browsers implement only per-origin local caches
20:37:34 <sebastian> ... the cross-site privacy risk based on cache timing analysis
20:37:42 <ErikAnderson> ErikAnderson has joined #ai-model
20:37:48 <dsinger> dsinger has joined #ai-model
20:38:01 <kirkwood> kirkwood has joined #ai-model
20:38:24 <sebastian> ... the per-origin caches tolerable for “typical” (non-AI) web resources. But AI Models are large and potentially shared
20:38:36 <sebastian> topic: Issue Starter Pack
20:39:13 <RafaelCintron> RafaelCintron has joined #ai-model
20:39:16 <kadirtopal> kadirtopal has joined #ai-model
20:39:31 <sebastian> -> https://github.com/w3c/tpac2024-breakouts/issues/15
20:40:32 <sebastian> McCool: Here are first issues for discussions:
20:40:53 <sebastian> ... background model download and compilation,
20:41:03 <sebastian> ... model naming and versioning
20:41:19 <sebastian> ... allowing for model substitution when useful
20:41:48 <sebastian> ... common interface for downloadable and “platform” models
20:41:58 <sebastian> ... storage deduplication
20:42:10 <sebastian> ... model representation independence
20:42:17 <sebastian> ... API independence
20:42:27 <sebastian> ... browser independence
20:42:42 <sebastian> ... offline usage, including interaction with PWAs
20:42:50 <sebastian> ... cache transparency
20:43:13 <sebastian> s/discussions/discussion
20:44:58 <dezell> q?
20:45:03 <dezell> q+
20:45:10 <kenji_baheux> q+
20:46:43 <sebastian> ack de
20:46:44 <anatoly-scherbakov> anatoly-scherbakov has left #ai-model
20:46:51 <dsinger> q+ to mention similar problems
20:47:39 <sebastian> David: Seems a big issue. Are interested are you working on specific APIs?
20:48:04 <reillyg> q+
20:48:27 <reillyg> ack kenji_baheux
20:49:51 <sebastian> ack ds
20:49:51 <Zakim> dsinger, you wanted to mention similar problems
20:50:40 <tomayac> q+
20:50:51 <ErikAnderson> q+
20:51:15 <kenji_baheux> ack reillyg
20:52:15 <estade> estade has joined #ai-model
20:52:33 <andrewnolan> andrewnolan has joined #ai-model
20:52:37 <sebastian> topic: Alternatives
20:53:56 <sebastian> McCool: One option can be that we define model-aware caches
20:54:39 <sebastian> ... use 'fake misses' to avoid redundant downloads
20:55:02 <Domenic> q+
20:56:28 <sebastian> ... identify cache items by content-dependent hashes
20:56:32 <kenji_baheux> q+
20:57:58 <sebastian> McCool: the idea is that model caches would behave as if they were per-origin caches
20:59:07 <sebastian> McCool: there is another alternative, the auto-expedite common models
21:00:06 <sebastian> ... the more common a model is, the less of a tracking risk it is
21:01:12 <sebastian> q?
21:01:18 <reillyg> ack tomayac
21:02:06 <reillyg> ack ErikAnderson
21:03:56 <estade> q+
21:04:02 <reillyg> ack Domenic
21:07:31 <reillyg> ack kenji_baheux
21:10:32 <reillyg> ack estade
21:11:32 <ErikAnderson> q+ arthur_coleman
21:12:08 <kenji_baheux> echoing Erik's point on finding if there is an actual problem; want to share what I've heard so far from partners. Enterprise / Edu customers would like to use a custom LLM that speaks their users' language / lingo across different origins (typically internal websites and/or popular 3p solutions to Enterprise/Edu needs). Other extreme, some
21:12:08 <kenji_baheux> partners want a custom LLM that speaks to their community but they wouldn't share it with other sites. Likely it would be running on the server as most of their users wouldn't necessarily want the big download.
21:13:33 <ErikAnderson> It's a proposal supported only by Chrome and Edge, but Related Website Sets is an example of a technology that may be applicable here _if_ we hear from customers that it's a common use case to share a company-specific model across multiple top-level eTLD+1s.
21:14:08 <ErikAnderson> ack arthur_coleman
21:17:24 <sebastian> adjourn
21:32:43 <gkellogg> gkellogg has joined #ai-model
21:41:15 <gkellogg> gkellogg has joined #ai-model
21:41:49 <gkellogg> gkellogg has left #ai-model
21:46:49 <reillyg> reillyg has left #ai-model
21:51:13 <dsinger> dsinger has joined #ai-model
21:53:20 <dsinger> dsinger has left #ai-model
22:32:08 <alcooper> alcooper has joined #ai-model
22:33:05 <alcooper> alcooper has left #ai-model
23:19:40 <kirkwood> kirkwood has joined #ai-model
23:51:44 <Lei_Zhao> Lei_Zhao has joined #ai-model
23:57:01 <kirkwood> kirkwood has joined #ai-model