author here. I kept digging through the Muse filesystem after my first post hit the front page here this week.
Going through my logs I found a background agent that used a model called azure/muse-special while building my website. I found this interesting and dug a little deeper.
The transcript and daemon binary point to an OpenAI model running on Azure. Still unclear which one or why it was selected.
The runtime also ships with an Anthropic client and a catalogue listing Claude, GPT and Kimi models. I didn’t observe Claude being used in my own sessions, but this is my Part 2 of exploring the Muse file system.
The Muse public APIs seem to be heavily inspired by OpenAI's, they even support the richer "responses" API. They have a similarly shaped compaction API, they have the same "encrypted" reasoning. Even the Muse harness seems heavily inspired by (open source) Codex. Considering Azure's historic involvement w/ OpenAI, it seems more plausible that Azure is used as spill-over capacity, and the same Azure infra used to encrypt GPT models encrypts Muse models.
My guess that the generally visible Anthropic/OpenAI details are leftovers from the whole meta "move fast" behavior. There are a few rough edges on the product where it leaks internal codenames (eg. signing in w/ WhatsApp required consenting to using "hatch" on one screen), so I'd believe that this was a leftover on the VM from the prototyping stage, before the Muse models were ready, or because the dev's got to benchmark against different models.
I'm not no. Someone from meta would have to chime in... but its very openai shaped and served differently from all the other models listed in the daemon, under a mysterious name.
a fine-tuned OpenAI model on Azure for the purpose of compaction or something could make sense I guess but that would still be OpenAI's weights with meta ontop, and they already have a compaction model served under azure/avocado-compaction-v1
If it's using the same weird (and awful) "backend prompt encryption" pattern and mechanism that Codex also uses (https://news.ycombinator.com/item?id=48905028), then I'd also say it points to OpenAI being involved. Hopefully that shit isn't becoming a more popular pattern, absolutely awful for troubleshooting stuff.
Meta already serves its own models on Azure under their real names. azure/avocado-compaction-v1 and azure/avocado-memory-flush-v1 are in the same catalog. Those sessions do not come back as gpt_responses_v1 items with rs_ ids and encrypted reasoning.
Depends on the size of the deal. OpenAI would have to disclose it in their S-1 if it crossed the threshold of being material information for investors.
I had a totally benign chat with OpenAI and it titled it as “amateur porn” in Chinese characters, it was very alarmed when I pointed the conversation name out to it. It almost never misses these days, but when it does the failure modes are very strange.
I remember getting a bunch of „Thanks for watching! Subscribe and smash that like button“ in the middle of chat sessions a few times (like, more than 3 times over the past 3y)
It kind of makes sense when you think about it. If the training data is YouTube transcripts, people often abruptly switch from content to asking for a subscription.
I'm not sure I understand why they would route it to other models. It can't be that they don't have enough compute. Maybe worried that the answer from their own models would be bad? Doesn't really make sense, but I could be missing something.
i agree, and it's what made me spend time exploring today.
fair q. call_ ids and gAAAAA blobs arent damning but the rs_ reasoning ids embed a unix timestamp that matches the session to the second, then OpenAI's 819x marker. plus the summary is in OpenAI's summarizer voice.
author here. I kept digging through the Muse filesystem after my first post hit the front page here this week.
Going through my logs I found a background agent that used a model called azure/muse-special while building my website. I found this interesting and dug a little deeper.
The transcript and daemon binary point to an OpenAI model running on Azure. Still unclear which one or why it was selected.
The runtime also ships with an Anthropic client and a catalogue listing Claude, GPT and Kimi models. I didn’t observe Claude being used in my own sessions, but this is my Part 2 of exploring the Muse file system.
original post: https://x.com/heypeterjames/status/2103545183400800746
pete at mouse dot dev
Nice work!
Are you sure this isn't just a custom model that is API-compatible with OpenAI?
This feels like the far more likely option.
The Muse public APIs seem to be heavily inspired by OpenAI's, they even support the richer "responses" API. They have a similarly shaped compaction API, they have the same "encrypted" reasoning. Even the Muse harness seems heavily inspired by (open source) Codex. Considering Azure's historic involvement w/ OpenAI, it seems more plausible that Azure is used as spill-over capacity, and the same Azure infra used to encrypt GPT models encrypts Muse models.
My guess that the generally visible Anthropic/OpenAI details are leftovers from the whole meta "move fast" behavior. There are a few rough edges on the product where it leaks internal codenames (eg. signing in w/ WhatsApp required consenting to using "hatch" on one screen), so I'd believe that this was a leftover on the VM from the prototyping stage, before the Muse models were ready, or because the dev's got to benchmark against different models.
I'm not no. Someone from meta would have to chime in... but its very openai shaped and served differently from all the other models listed in the daemon, under a mysterious name.
a fine-tuned OpenAI model on Azure for the purpose of compaction or something could make sense I guess but that would still be OpenAI's weights with meta ontop, and they already have a compaction model served under azure/avocado-compaction-v1
It uses a similar API to openAI. Source: https://dev.meta.ai/docs/overview
Source 2: I work on AI at Meta.
why not the same ones, is there a unique need for a different surface?
Why host it on Azure though?
Microsoft owns 27% of OpenAI.
If it's using the same weird (and awful) "backend prompt encryption" pattern and mechanism that Codex also uses (https://news.ycombinator.com/item?id=48905028), then I'd also say it points to OpenAI being involved. Hopefully that shit isn't becoming a more popular pattern, absolutely awful for troubleshooting stuff.
Anthropic does the same thing with reasoning signatures. It has already become the standard pattern.
I asked muse. It’s for codex-cli, and it routes through “OpenAI inference proxy”
What endpoint would Claude be calling?
I don't really see evidence that it's an "OpenAI" model, the title is IMO very misleading.
Meta already serves its own models on Azure under their real names. azure/avocado-compaction-v1 and azure/avocado-memory-flush-v1 are in the same catalog. Those sessions do not come back as gpt_responses_v1 items with rs_ ids and encrypted reasoning.
Avocado and Muse are different models. Muse uses the openAI API.
https://dev.meta.ai/docs/overview
Source: I work on AI at Meta.
router makes sense. fill a gap with someone else’s model wearing a mask, swap it out with your model not wearing a mask when (or if) you reach that.
definitely tracks with alexandr’s hand-wavey influencer turn.
I thought it was 1.3 spark. No?
It'd be hilarious that meta chose to pay openai. If there was such a deal, would it come to light in any public/official filings?
Depends on the size of the deal. OpenAI would have to disclose it in their S-1 if it crossed the threshold of being material information for investors.
No reason for it to. Software companies use products from other software companeis all the time.
meta already pays openai and anthropic for employee use despite having their own models available
Muse 1.3 spark occasionally spits out Chinese character responses to me like internal instructions. “Go fast” or “Get help”…
I’ve had Claude opus do that too.
I had a totally benign chat with OpenAI and it titled it as “amateur porn” in Chinese characters, it was very alarmed when I pointed the conversation name out to it. It almost never misses these days, but when it does the failure modes are very strange.
If you ever look at what's in the Common Crawl the amount of Chinese porn crap is jarring.
I remember getting a bunch of „Thanks for watching! Subscribe and smash that like button“ in the middle of chat sessions a few times (like, more than 3 times over the past 3y)
It kind of makes sense when you think about it. If the training data is YouTube transcripts, people often abruptly switch from content to asking for a subscription.
If AI is typeahead on crack, then this tracks.
Got Russian from it once. As a child of the '80s I jumped right on it, but the explanation seemed believable.
Still sleeping with one eye open.
A Hindi word in the middle of a normal reply for me (but once I translated the word it was right in context heh)
This used to happen a lot with GPT 5.4. It would start outputting entire sentences in Korean for no apparent reason.
People that know multiple languages sometimes code switch too funnily enough
zuck wouldn't do this, he is too proud imho
I'm not sure I understand why they would route it to other models. It can't be that they don't have enough compute. Maybe worried that the answer from their own models would be bad? Doesn't really make sense, but I could be missing something.
i agree, and it's what made me spend time exploring today.
fair q. call_ ids and gAAAAA blobs arent damning but the rs_ reasoning ids embed a unix timestamp that matches the session to the second, then OpenAI's 819x marker. plus the summary is in OpenAI's summarizer voice.
its just a best guess.