Top
Best
New

Posted by yks 19 hours ago

Muse – Meta’s personal AI agent(ai.meta.com)
568 points | 625 commentspage 8
wxw 12 hours ago|
Playing around with it tonight and I feel like a killer feature is its speed. Really fast response times.
Centigonal 19 hours ago||
Exciting to finally see glimmers of Meta's vision for its massive LLM investment.

It's not a bad product idea, but adoption might be rough because of all the negative goodwill that Meta has accumulated over the years. Also, I wouldn't be surprised if someone (maybe one of the big AI labs) ran Meta's own playbook against them here by creating an imitator and aggressively funding adoption to try and win the market as a fast follower.

vinni2 18 hours ago||
I don’t understand why Apple can’t pull something off like this.
dxxmxnd 17 hours ago|
I think they will eventually and I will just simply keep doing things manually until then. I will not use any of Meta’s AI products for personal use.
mike_tyson 6 hours ago||
this is like if Satan released an AI app that ran on your phone 24/7
ernsheong 13 hours ago||
"We couldn't continue with that account. Try again."

Is it geo-restricted to the US for now?

llIIllIIllIIl 16 hours ago||
So this is like installing codex and then using remote on chatgpt app? Niche AF. But you rent the computah from meta?
nickysielicki 12 hours ago||
People are unfortunately just ranting about not trusting meta, which… okay, fine, but having played around with this for the past hour or so it’s probably my favorite mobile LLM experience so far across Claude and ChatGPT and Gemini. It’s very good and in its own class.
russellbeattie 11 hours ago|
Agreed. I'm not a fan of Meta in any way, but I've had fun messing with the app this afternoon - both on my phone and the web version at muse.ai - to see what it could do.

Like the announcement said, it's running in its own VM, but it's an impressively transparent system: You have access to its System Files right in the GUI, which surprised me. You can look at the AGENTS.md and other harness files it uses as well as some shell scripts, bash files, cronjob folder etc. It seems a lot more obvious to me how Muse works than say Claude or ChatGPT, which don't make it nearly as clear where stuff lives and how long its there for. I mean, I get they use VMs as well, but there's definitely no "System Files" menu option in those phone apps at least. The Muse bot seems much more "permanent" in my mind as a result.

The AI is quite helpful as well. It let me download its 350Mb "hatch" binary which runs the bot/harness/agent (written in Rust), plus it zipped up the library of skills it uses and all the system files in my VM for me to look at on my machine. Nothing extra special, but it was nice to browse through all the files and the binary strings to see how it ticked. It also explained the limits of the VM to me - for example, that I can't SSH into the VM as it runs through a proxy and can't be reached directly, but was willing to help me add itself to my Tailscale network.

It's basically a nice looking UI on top of a typical agent harness using standard Markdown files, all running in a little Linux cloud VM spun up just for you. And Muse Spark 1.3 is finally good enough that it was actually useful (and fast), so I can see why they launched the app/service now.

From a non-geek perspective, I asked it about selling my old Nintendo Switch on Facebook Marketplace - something I've been meaning to get around to for about a year or so now - and it said I can upload a picture of all the stuff I have and it'll check prices and create multiple listings for me. That alone might make me hate Meta a little less after this if it works as advertised.

reinhash 9 hours ago||
yeah nice try meta, if there is anyone I would not allow access to all my private data, its you!
modeless 14 hours ago|
It's hilarious that Meta made a good AI integration for Gmail before Google did.

For a while now I've had a startup idea for an agent that would monitor your inbox for content-free notification emails that just say "you have a new message, click here to read it" or "package delivered" with no details about the package contents. Then it would login to the website, do the 2FA dance, click through the UI, extract the actual important information, and put it back in my email inbox where it should have been in the first place.

I haven't worked on this idea because I think general purpose AI agents are likely to be able to do it very soon. And it seems like Muse is almost there already. It almost worked with MyChart, but hit a snag navigating the website after logging in. I expect this will be fully working in a matter of months.

I honestly can't think of any software ideas that won't be subsumed by general purpose agents or world models or some other form of AI in the very near future. Time to pivot to hardware ideas!

More comments...