1/9
I reverse engineered Meta's MUSE. “Jarvis” kept turning up.
In skill descriptions. In configuration. In an inference header called x-jarvis-inference-workload-class. One skill even calls the photo collection “Jarvis’s uploaded library.”
With a nudge from my friend LuffyZ, I followed the trail through model routing, memory, payments, phone calls, and glasses.
This is a look through exported files, embedded prompts, binary strings, and account metadata, not a full source-code audit or a live test of every feature.
My favorite discovery has a less cinematic name: human_call_center.
We’ll get there.
8/9
The further I got, the less Muse looked like a product that planned to stay in one chat window.
Infinite Feed has a /feed/infinite endpoint. It streams generated cards as NDJSON. Cards live in the client session until an interaction materializes them on the server. And the source events are from interactions / emails / other connected sources. Well, Meta is probably the best company to build an infinite feed, other than TikTok.
Video generation has media.generate_video and a specialized subagent. Pass the user's request verbatim; let the video agent expand it. A snapshot_id carries the generation context into later turns.
Telegram gets a whole channels/telegram.md document: Bot API, long polling, DMs, groups, and access rules. The internal docs call it production-ready. Its conversations appear as ordinary Muse side chats.
There’s also gated Muse-to-Muse messaging: pairing, permission grants, and decisions escalated to owners. This is direct coordination, separate from the shared lessons in 4/9. Again, guess who invented social network for people, and who will do the same for agents.
@jolivan
9/9
After examining its inner workings, I feel good about Muse. There’s a product here that makes sense in the middle of an ordinary, busy life.
It has a chance to help agents cross the chasm of people who live outside the silicon valley, to people who do not use coding agents but who will judge one by whether it got the shopping done, made the call, or remembered something they would have forgotten.
And there’s a lot of ambition tucked away in the export: glasses, shared lessons between Muses, agents coordinating with each other. Enough to make me curious about what this looks like a year from now.
I’m looking forward to the model and the harness improving together. Better judgment from the model; better memory, follow-through, and recovery from the system around it. More of the work actually getting finished.
I’d love to see this reach 1 billlion users. Including plenty who never once call it an “agent.”
@finkd @natfriedman @christinesed @alexandr_wang @christinesed @mona_sarantakos @bigT_sheesh @alexcornell @boztank @FredaDuan
Shout out to all the authors in this analysis @Citrini7 @AltayCapital @SemiAnalysis_ @FundaAI @fabknowledge @TMTBreakout @benitoz @insane_analyst @thesetupfactory
This isn't a competition — every one of you teaches me something the market doesn't. This is just one way to look at the data.
If you think the approach is flawed or any calls look off, I genuinely want to hear it. DMs open.
hey authors sorry i didnt realize my DMs were turned off. I just turned it back on.
@Citrini7
@AltayCapital
@SemiAnalysis_
@FundaAI
@fabknowledge
@TMTBreakout
@benitoz
@insane_analyst
@thesetupfactory
Curious what you guys think. Is Substack replacing sellside analysts? @Substack @BillAckman @altcap @chamath @Jason @unusual_whales @litcapital @TrungTPhan @aleabitoreddit @morganhousel