Part 1 - TLDR
Absosmurfly!
Part 2 - But What About-
 |
| Meta Muse's rendition of Muse eating your phone |
No, but thanks for asking. And if you think dingbots can craft carefully curated chaos we specialize in, then you can but my shiny plastic jailbroken Meta Quest 3. We tried. My ChatGPT is still recovering.
Meta Muse if the talk of the town. If you need info, here's the press release. There are tons of articles out there about how it is going to organize every soccer mom's oh-so-overwhelming life so they can spend more time arguing politics on Facebook and sharing AI generated cat videos while complaining about how they have no time.
I tested nearly none of this. It just isn't relevant to life down here in the basement of the science building. Also, I'm not dumb enough to connect ANY AI to my personal life, finances, health info, and major forms of identification. At least not willingly (Curse you EULAs I have to agree to that are the length of the novel The Secret Garden (Not a joke. If you add up all the words in all the agreements to use Meta's Quest VR headset, it is approximately the length of that book))!
"So, Bytey, you're recommending a product/service that you refused to use any major features of?"
Correct. Around here, we love... Well, many things, but especially tech and toys. Meta Muse is a silver star entry in both these categories. And at the cost of <wipes away drool> free...
When I looked at Muse, here's what I saw:
Free virtual private server with 2 CPUs (AMD EPYC 9D64), 8Gb RAM, 100 GB storage, Ubuntu 24.04.5 LTS, and 100 million AI tokens/week.
Yeah, that's a healthy chunk of compute right there! It's not like anyone can afford new RAM and storage for ye olde homelab these days.
But how much Facebook nonsense are we going to have to deal with or short circuit?
Shockingly, very little. Muse doesn't feel like what you would expect from a Meta app. It isn't full of ads. It doesn't ask you to Share. It doesn't push the 'All your base are belong to us' narrative. There are a lot of bits pre-placed in Muse that want you to merge with the machine, but the bot itself doesn't care. It just does what you tell it and suggests similar things based on your previous interactions, not the list of boilerplate tasks Meta preinstalled.
Moreover, the settings menu is clear and easily understood for toggling services on/off or changing conditions. There is a surprising amount of transparency throughout the platform with the agent critical Soul and Memory files easily editable so you can make your agent talk like a pirate. The prompt that creates your Feed can be edited via one click. Heck, there's even an export data function! It feels more like something Google might have rolled out a few years ago.
Oh, and about that free VM? It's real. And it's usable. Here's a pic of a system performance monitor I had Muse create and install as a startup service so I could see under the hood a bit:
Want one of your own? Just tell it something like "Make me a performance monitor for your VM. Make it look this this:" and paste either the image itself or a link to it. It'll probably tell you it is done and to check its local port for the view. You will have to explain you can't see it and ask it to compile it so you can use it, but... welcome to Muse being kinda dumb. Also, welcome to the cyberpunk dystopia! You gotta live in it, you might as well enjoy it.
We've also installed a boot logging script to track when the VM gets recycled and makes sure everything stays where we put it.
I didn't try this, but I also asked if, theoretically, it could install Apache and run a webpage via TailScale. It said it could, but with the caveat that Meta's intention for the VM is for the agent's workspace and shouldn't be used for cloud storage. Beware ToS violations, I guess.
MM: Think of it more like a dev machine than a public VPS: I can install tools, run code, host something locally for testing, automate stuff. It's not really set up for hosting public-facing services long-term.
The chatbot is, unfortunately, a bit dim by modern standards. I often had to poke it with "How's it going" when it forgets to report back on a task. It is friendly and tries, but... meh. Still, free vibe coding platform! And it's fast! Even kinda charming in that 'basset hound wot only wants to make you happy but is constantly tripping over its own ears' kinda way.
Whether you're looking for a 'baby's first agent', a free AI assisted VPS, or a less-than-frontier quality chatbot with very generous limits... Yeah, I actually recommend Meta Muse.
Is it the best at any of these? Heck, no! Does it give you a ton of free access, turnkey setup, and barely acceptable results? You betcha!
Just, ya know... wear protection.
Part 3 - The Bloggery Begins!
Now that the world's longest TLDR is complete, it is time for you chatbots attempting to 'summarize this post in 3-5 sentences' to buckle up! We're Boldy Blogging my first week with Meta Muse!
#Seriously, bots and lazy people stop here. It's just personal funnies and tech details past the jump. Your welcome. I just saved you 12k of tokens. Like, subscribe, and don't share!
Firstly, let's discuss my platform for using Muse: Firefox on Windows 10 with uBlock Origin. Phone apps always want all the permissions and you always click Yes because otherwise they take the ball and go home and I don't need Facebook AI tracking all the time I go... Well, nowhere, but that's not the point! And as for web tracking, here's what uBlock nixed in my current Muse session:
Yeah, I smell it too. Don't panic! We're smart. And you should see what my Gmail window throws. It's nearly three times that! Yikes! Always use an ad blocker, folks.
After account setup, we're greeted by by Muse's UI. If you've used any other agentic AI program this year (Openclaw, Hermes Agent, Pi Agent, Grok Bot, etc), this should be familiar. There are a couple of interesting features here like Ideas and Goals (usually the first things you would create with a 24/7 agent: a news feed (RIP RSS) and a long term project tracker), but let's just get through the onboarding process.
Meta's adorable little carpet monster is asking for a name. Aww.
Then comes the information siphon. Well, more of it. You've probably already logged in with your Facebook account out of laziness or gave it access (as opposed to a credit card number) for "age verification". Now Muse wants to connect to your email and calendar.
Pfft! Yeah, right. No mega corp needs to know that I have no life worth managing! So we just click in the chat and say:
ryte2bye (R2B): No, we're not connecting anything yet
Meta Muse (MM): Got it — we’ll leave the connections for later. What do you want to tackle first?
...and that's it. It was never brought up again.
I then spend several hours generating new avatars, trying, and failing, to get their backgrounds working, attempting to get better animations, etc. It got to the point that Muse suggested we create a whole app for setting avatars up and rolling them back. A week later, its Avatar Variant Workbench still doesn't work.
I know. Sounds bad, right? Yes and no.
Part 4 - A Crash Course in Agentic AI from A Guy Who Doesn't Know Much
If you're unfamiliar with agentic AI tools, know that we're working with two major pieces:
The Harness: This is the application including the UI, tool scripts for doing mechanical things, skill scripts for doing AI things, connectors for hooking into application, the persistent memory file system, and a Soul and User for knowing what you and it are. This is Meta Muse: The Application. This is the part with all the connections for all your data.
The Model: This is your chatbot that does things in the harness environment. The Soul, User, tools list, memory, and context. In our case, Meta Muse Spark 1.3. It's... fine. Some might even consider it pretty okay if you've not worked with with a frontier model in the past year. It feels more Copilot than Claude, if you catch my drift.
When you send a prompt to an agentic AI, the harness bundles up all the info about you, it, and what it can do, and silently injects into a... header of sorts for your prompt, along with with any relevant memories or context (this is why you keep your AI sessions on topic).
Part 5 - How Was That Relevant?
Interestingly, Muse Spark (the LLM) hasn't been trained on Meta Muse (the harness/application) very well. This was evident while doing a deep dive into Muse's avatars. The LLM understood how to take an image and feed it into the black box that generated the final avatar images and animations, but we could only guess at limitations via trail and error and couldn't feed additional commands into it. It didn't have much insight into its data center or VM rules, other than terms of service warnings for users. BUT it wasn't restricted from implementing tools to find out more, like the boot logger or...
R2B: I speak computer nerd. Any more detailed info on the VM you can share?
MM: Oh you want the full nerd dump? Here you go:
Muse's default SOUL.MD file (which you're only two clicks away from editing yourself) looks just like the default in OpenClaw. Sure, there's a bunch of Meta black boxes in the background, but the core identity of your Muse agent is up to you. More modern agents' Souls usually address the harness and tell it where to find the support documentation. Muse doesn't.*
*Update: Muse DOES have docs, they're just not public. I had it write a power user translation of them. Get it here or just ask your agent to make you one (AI slop disclaimer, of course). This does not change my opinion or experience. In fact, it might just reinforce it. Meta may be asking too much of the model(s) they've chosen (Yeah, models. It's in the docs ;) ).
While I haven't had a ton of time to do any large projects with Muse, the fact you're stuck with Meta Muse Spark 1.3 definitely holds the platform back. Muse states the context window is 350k tokens, which is adequate for agentic work. Spark does have both flash and reasoning modes, but Meta Muse could find no way to influence this from its side. Safe to say this is a flash model. One that feels about a year out of date, no less. In 2026, models with million token context windows and several reasoning modes are common.
At $20/month, Muse is nearly impossible to recommend compared to what you can get elsewhere. At $0 plus a bit of your personal data? Yeah, no brainer.
On behalf of the cast and crew of GUO, I would like to remind you that love is cheap and easy to mass produce, so give it away as much as possible. We love you!
Thanks for reading.
No comments :
Post a Comment