The other day I signed up for Muse (through a VPN) just to see what all the fuss was about and I have to say it’s a really nice implementation of a personal assistant. It genuinely feels great and makes me want to use it!
It has a simple surface but you can dig deep, from customizing personality, rules, communication style, avatar, etc. You can even install Tailscale onto its VM to SSH into it, but sadly once the machine shuts down, the overlay resets and that’s gone. Couldn’t figure a way to persist it out yet. Still very cool to dig around and explore all the CLIs Meta created for Muse to interact with stuff (you can scp them off the VM)
If I had to rank them I would say: Grok Bot -> Muse -> Hermes -> dot -> OpenClaw
I would love to use Muse more but same as everyone else, there’s no way I’m giving Meta more access to my life. So for now, my Muse is jailed to only Meta owned services and general research, and for that it’s great.
The privacy issue blocks me from adopting this or other personal assistants: do I want to give muse my identity data and access to my mail, calendar, and phone so it can do even basic tasks like make appointments or pay bills?
It will be interesting to see if muse's privacy features can get users over this.
A lot of use cases for agents require long running sessions. And you're likely to have many agent sessions running in parallel. These aren't going to be running on your personal devices
I agree. I currently do run a bunch of dev containers on my local Docker but that is not good enough, they neeed to be able to work while I am on the run and close my laptop. So we are trying having personal containers on a managed Kibernetes cluster for every developer. We’re still working on making this work but it seems promising, the hardest thing being just permissions to let the agent use MCP servers, http APIs etc. If anyone has done this, how did you manage that?? I know there are products like OpenCLI that help with the problem but we want something we can build ourselves since AI makes it so much easier to build things!
Why not? Why wouldn't your phone be able to do this?
The other advantage long running sessions have is that very low tok/s are perfectly acceptable. In fact, hardly noticeable.
(aside, of course, from Google and Apple both having incentives to prevent this from happening, as both depend on selling centralized services to you and advertisers, and the local model breaks both)
I’m sure there is probably a skunkworks like team secretly working at Apple trying to make the hardware needed to accelerate local models enough to handle personal data processing needs without decimating the battery. I’d guess another team is probably tasked at researching a E2E encrypted remote inference mechanism where the personal data that does get out of the iPhone/ipad/Mac is encrypted with something Secure Enclave on processor, and “even Apple can’t read it”..
Apple is my long term bet for personal AI computing.
No way in hell am I letting FaceBook/OpenAI/Anthropic/Deepseek/anybody have complete access to my personal digital life. That’s just dumb.
Just look at OpenAI digging through ChatGPT logs to get the idea to solve the math problems etc. I’m sure all the subscription plans and subsidies are aimed at gathering enough breadth of real people comms to get the models to improve on wider signals…
Worst case, I don’t want to be voluntarily giving up my digital footprint to get profiled a certain way and be fed what AI overlords (or the companies that wield them really) think I should get as an answer/response.
My thoughts and activity stays private as much as practical and feasible.
I agree, most of the proactive agent assistance use cases requires me to give out data I’m hesitant to give like email, financial data, health data…
I’m hoping someone makes a lower maintenance version of open source Muse or Grokbot or Dots (more user friendly than OpenClaw or Hermes) that I can easily deploy to my own AWS, Cloudflare, Vercel, or Fly account and use some ZDR friendly inference provider like Fireworks.
I'm a designer, but I admit it's really hard to predict what products and UX are going to work in the AI age.
Chatbots have been tried for a decade and were a total failure. With AI, it has become the fastest-growing way to interact with technology, ever.
GPT-5 tried to have an auto router and create abstraction for the user; people hated it. Instinct and Muse are doing the same with a different framing; it's going wild.
There are tons of these little examples where a small difference in framing, a small threshold in tech, and product assumptions completely flip on their heads.
It's amazing to follow but requires you to be way more flexible in your assumptions.
I think devil’s in the details on all of your examples, although I very much agree with your overall point - it’s so nuanced in terms of timing, current capabilities, etc.
Re: chatbots were tried but it’s a wrong model for a deterministic system. It’s the best one for a probabilistic one.
Re: chatGPT auto-router - 1) I suspect explicit model selection matters less in agentic contexts (this thing is coordinating a bunch of other agents for me) vs QA context where you want to make sure you’re getting the best answer. 2) I suspect auto-routing matters less now that models are way better
But I have to admit that in the design field we have a strong tendency to over-rely on "proper rules" and to overgeneralize learnings. (My pet peeve is when B2C patterns are transferred as-is in B2B or prosumers)
AI makes some old UX viable but can also unlock tons of value once you break your assumptions and reluctance. An example I see a lot here and on socials is how much of the code developers actually read nowadays. In the case of Instict, it is how many people actually share their credit card with an unknown startup's IA to buy stuff. I just heard on a podcast that 40% of them actually used their AI to buy something online (I didn't verify the %)
I'm the kind of guy who didn't understand why people use phones so much when a desktop is way more comfortable, so I have to keep track of the crazy ones!
Honestly I just think the fact that I don’t get spawned into a new chat every time I open it is an amazing. Simple and clean, AI seriously needs its Steve Jobs
"clunkiness of ChatGPT is stark. The whole desktop ChatGPT vs Codex switch is dumb. The chat vs work switch is dumb. ChatGPT projects vs Codex local vs Codex cloud projects is mega dumb. I need to make 10 choices each time I want to start or continue a chat. Just give me one chat that orchestrates all of it in the background." This times a 100x. Every step forward seems 2 steps back. The sidebar is jungle, and trying to find chats is tough, trying to remember if its codex or chat, a local one or remote, and that chat ones can't be put in a codex projects and countless other issues that make it a nightmare. While I can they are working on improving the mess to at least presumably get rid of the Chat vs Work toggle, I haven't heard anything that eventually it will just be one, where we can just toggle how it responds it, vs being limited on it can be organized and searched.
Seems to me like they just copied Claude's distinction of Cowork because it sounded better to enterprises but it should never have been pushed to the userbase.
And you accidentally end up asking Astra Max what time the game is tonight or something stupid like that and it still has to finish and burn credits up
Meta stupidly chose to open Muse only in the US.
If anyone wants to take the world, just copy it, you have a month. (Before someone do it. Not Meta. They stick to dumb decisions)
I'm not sure to what extent they chose that, versus their privacy woes in the EU and elsewhere chose it for them. Turns out meta's reputation among US consumers is sparkling compared to their overseas reputation...
FB has a huge advantage launching any consumer app over any start up. Even ones that have higher valuations. If you think the market position isn't important, imagine what would happen if Apple changes their TOS such that Muse stops working as well on their devices. Or take it off the store because it's "dangerous".
I don't understand Muse, couldn't agents already do all this? I have ChatGPT for example and I can have it book tickets and other things for me, it even pulls up a cloud browser if it can't access APIs. It can check for details from my email and use those for the ticket. Now OpenAI has Dots too which is a 24/7 version of the same thing.
Sure, it's nothing super unique but the way that Meta has put all the pieces together including security works really well in practice. Muse seems to be the best of the AI personal assistants for regular people right now.
Does anyone believe the security works though? This is the company that brought us facebook, pervert glasses, and got sued by its E2EE content monitoring team (which apparently exists!) for emotional trauma because they had to read so much stuff from private WhatsApp messages.
I believe that Muse security is about as good as any other large cloud environment. In the long run maybe all of them will eventually be hacked but so far my other Meta accounts have been 100% secure. So I'll take my chances with Muse because it saves me time and allows me to get more done.
Those other issues, while perhaps examples of bad corporate behavior, aren't really security lapses per se.
It does not at all. ChatGPT is trying to be an everything app and it’s extremely convoluted by this point and hard for most people to make sense of, especially if you’re not in tech.
The article emphasises exactly what you said: technically it has no new features, it's just better UX than anything else, and more friendly to non-techies.
When it comes to being mainstream, an approachable UI sells
No, I'm asking why ChatGPT isn't sufficient because they literally do the same thing unlike Dropbox vs rsync. In other words ChatGPT has the same UI UX but people flock to Muse for some reason.
That’s pretty much the question the article tries to answer! TL;DR - good product design, personality mechanics, and ad-funded tokens is what makes it different.
It’s the most consistent and user friendly agent with the ability to also be a harness for multiple custom agents. I literally had people running around the office amazed by muse last week. It has brought agents to the masses in a way a lay person can understand and use.
“Interestingly, what’s making Muse a success also explains why Google has mostly failed to capture the zeitgeist so far. They have the compute and the ad money to fund something like Muse. But they have no product chops to build something simple and understandable in consumer AI.”
Umm what? Having billions of hourly active users across YouTube, Google, Google Maps, and Chrome still doesn’t mean they have the product chops to build another consumer facing app/service? I think it’s inevitable that Muse will get crushed.
You listed products that have been around for ~20 years. Keeping up already entrenched products running doesn't correlate with being able to launch new products.
Those are all quite mature products which they have maintained and grew super well. They haven’t had a good track record in consumer AI, there’s no coherent product strategy or offering. They had some localized hits like NotebookLM and nano banana but it hasn’t translated into organic growth and usage for Gemini
Google Search and Google Maps are utilities, not consumer apps that vie for leisure time, so I don't think they are really comparable. The next frontier for AI apps like Muse and Dots is competing with the social media giants.
We should look to see how Shorts vs IG Reels have adapted to the short form video landscape for product chops in this area. IMO IG has adapted far better to the new era, and I very often get people sending me Reels (they used to send TikToks). I very rarely see anyone send Shorts but it still happens. I expect this trend to continue and IG to extend their lead here. And Muse certainly has first mover advantage in personalized AI apps, who knows when Google will finally release one.
Google has huge amounts of training data but very little of it is relevant to creating an AI personal assistant with good UX for common tasks. Like Gemini might be really good at answering specific questions or generating synthetic content but that doesn't mean it can securely, reliably carry out workflows in a way that feels natural to regular users. At least for now, that part still requires some human vision and product management.
I like that you can get root access to the VM your instance runs in, connect it to your tailnet, automate connection approvals, and operate a torrent seedbox on Meta’s infra.
Have you figured out a reliable way to persist stuff? Like a launch script or similar that doesn’t reset when the VM shuts down and can be used to restore tailscale
bashrc, bash profile, cron, systemd, etc all get wiped clean the moment the VM resets
If you can design something that a Baby Boomer can understand, then you have a mass market product at full scale. Otherwise your hopefully building something that a niche group finds attractive and that grows from there, outwards to scale.
61 comments:
The other day I signed up for Muse (through a VPN) just to see what all the fuss was about and I have to say it’s a really nice implementation of a personal assistant. It genuinely feels great and makes me want to use it!
It has a simple surface but you can dig deep, from customizing personality, rules, communication style, avatar, etc. You can even install Tailscale onto its VM to SSH into it, but sadly once the machine shuts down, the overlay resets and that’s gone. Couldn’t figure a way to persist it out yet. Still very cool to dig around and explore all the CLIs Meta created for Muse to interact with stuff (you can scp them off the VM)
If I had to rank them I would say: Grok Bot -> Muse -> Hermes -> dot -> OpenClaw
I would love to use Muse more but same as everyone else, there’s no way I’m giving Meta more access to my life. So for now, my Muse is jailed to only Meta owned services and general research, and for that it’s great.
How'd you do it with a vpn? Doesn't it still need your facebook account to be from the US?
The privacy issue blocks me from adopting this or other personal assistants: do I want to give muse my identity data and access to my mail, calendar, and phone so it can do even basic tasks like make appointments or pay bills?
It will be interesting to see if muse's privacy features can get users over this.
There will come a point where the local models are good enough (and iPhones powerful enough) for this to be possible without privacy downsides.
A lot of use cases for agents require long running sessions. And you're likely to have many agent sessions running in parallel. These aren't going to be running on your personal devices
I agree. I currently do run a bunch of dev containers on my local Docker but that is not good enough, they neeed to be able to work while I am on the run and close my laptop. So we are trying having personal containers on a managed Kibernetes cluster for every developer. We’re still working on making this work but it seems promising, the hardest thing being just permissions to let the agent use MCP servers, http APIs etc. If anyone has done this, how did you manage that?? I know there are products like OpenCLI that help with the problem but we want something we can build ourselves since AI makes it so much easier to build things!
Why not? Why wouldn't your phone be able to do this?
The other advantage long running sessions have is that very low tok/s are perfectly acceptable. In fact, hardly noticeable.
(aside, of course, from Google and Apple both having incentives to prevent this from happening, as both depend on selling centralized services to you and advertisers, and the local model breaks both)
Ok but look at the list... Google likely already has all that. It's not really something a normie cares about self hosting.
I’m sure there is probably a skunkworks like team secretly working at Apple trying to make the hardware needed to accelerate local models enough to handle personal data processing needs without decimating the battery. I’d guess another team is probably tasked at researching a E2E encrypted remote inference mechanism where the personal data that does get out of the iPhone/ipad/Mac is encrypted with something Secure Enclave on processor, and “even Apple can’t read it”..
Apple is my long term bet for personal AI computing.
No way in hell am I letting FaceBook/OpenAI/Anthropic/Deepseek/anybody have complete access to my personal digital life. That’s just dumb.
Just look at OpenAI digging through ChatGPT logs to get the idea to solve the math problems etc. I’m sure all the subscription plans and subsidies are aimed at gathering enough breadth of real people comms to get the models to improve on wider signals…
Worst case, I don’t want to be voluntarily giving up my digital footprint to get profiled a certain way and be fed what AI overlords (or the companies that wield them really) think I should get as an answer/response.
My thoughts and activity stays private as much as practical and feasible.
I agree, most of the proactive agent assistance use cases requires me to give out data I’m hesitant to give like email, financial data, health data…
I’m hoping someone makes a lower maintenance version of open source Muse or Grokbot or Dots (more user friendly than OpenClaw or Hermes) that I can easily deploy to my own AWS, Cloudflare, Vercel, or Fly account and use some ZDR friendly inference provider like Fireworks.
I'm a designer, but I admit it's really hard to predict what products and UX are going to work in the AI age.
Chatbots have been tried for a decade and were a total failure. With AI, it has become the fastest-growing way to interact with technology, ever.
GPT-5 tried to have an auto router and create abstraction for the user; people hated it. Instinct and Muse are doing the same with a different framing; it's going wild.
There are tons of these little examples where a small difference in framing, a small threshold in tech, and product assumptions completely flip on their heads.
It's amazing to follow but requires you to be way more flexible in your assumptions.
I think devil’s in the details on all of your examples, although I very much agree with your overall point - it’s so nuanced in terms of timing, current capabilities, etc.
Re: chatbots were tried but it’s a wrong model for a deterministic system. It’s the best one for a probabilistic one.
Re: chatGPT auto-router - 1) I suspect explicit model selection matters less in agentic contexts (this thing is coordinating a bunch of other agents for me) vs QA context where you want to make sure you’re getting the best answer. 2) I suspect auto-routing matters less now that models are way better
Yes definitely.
But I have to admit that in the design field we have a strong tendency to over-rely on "proper rules" and to overgeneralize learnings. (My pet peeve is when B2C patterns are transferred as-is in B2B or prosumers)
AI makes some old UX viable but can also unlock tons of value once you break your assumptions and reluctance. An example I see a lot here and on socials is how much of the code developers actually read nowadays. In the case of Instict, it is how many people actually share their credit card with an unknown startup's IA to buy stuff. I just heard on a podcast that 40% of them actually used their AI to buy something online (I didn't verify the %)
I'm the kind of guy who didn't understand why people use phones so much when a desktop is way more comfortable, so I have to keep track of the crazy ones!
Honestly I just think the fact that I don’t get spawned into a new chat every time I open it is an amazing. Simple and clean, AI seriously needs its Steve Jobs
"clunkiness of ChatGPT is stark. The whole desktop ChatGPT vs Codex switch is dumb. The chat vs work switch is dumb. ChatGPT projects vs Codex local vs Codex cloud projects is mega dumb. I need to make 10 choices each time I want to start or continue a chat. Just give me one chat that orchestrates all of it in the background." This times a 100x. Every step forward seems 2 steps back. The sidebar is jungle, and trying to find chats is tough, trying to remember if its codex or chat, a local one or remote, and that chat ones can't be put in a codex projects and countless other issues that make it a nightmare. While I can they are working on improving the mess to at least presumably get rid of the Chat vs Work toggle, I haven't heard anything that eventually it will just be one, where we can just toggle how it responds it, vs being limited on it can be organized and searched.
Then Tibo is on X asking things like “If we were to add one more sidebar, what should be in it”
Seems to me like they just copied Claude's distinction of Cowork because it sounded better to enterprises but it should never have been pushed to the userbase.
And you accidentally end up asking Astra Max what time the game is tonight or something stupid like that and it still has to finish and burn credits up
Meta stupidly chose to open Muse only in the US. If anyone wants to take the world, just copy it, you have a month. (Before someone do it. Not Meta. They stick to dumb decisions)
> Meta stupidly chose to open Muse only in the US
I'm not sure to what extent they chose that, versus their privacy woes in the EU and elsewhere chose it for them. Turns out meta's reputation among US consumers is sparkling compared to their overseas reputation...
FB has a huge advantage launching any consumer app over any start up. Even ones that have higher valuations. If you think the market position isn't important, imagine what would happen if Apple changes their TOS such that Muse stops working as well on their devices. Or take it off the store because it's "dangerous".
Didn’t help Google, Apple, or Amazon to do the same even though they have market position and pre-existing AI products.
I don't understand Muse, couldn't agents already do all this? I have ChatGPT for example and I can have it book tickets and other things for me, it even pulls up a cloud browser if it can't access APIs. It can check for details from my email and use those for the ticket. Now OpenAI has Dots too which is a 24/7 version of the same thing.
Sure, it's nothing super unique but the way that Meta has put all the pieces together including security works really well in practice. Muse seems to be the best of the AI personal assistants for regular people right now.
Does anyone believe the security works though? This is the company that brought us facebook, pervert glasses, and got sued by its E2EE content monitoring team (which apparently exists!) for emotional trauma because they had to read so much stuff from private WhatsApp messages.
I believe that Muse security is about as good as any other large cloud environment. In the long run maybe all of them will eventually be hacked but so far my other Meta accounts have been 100% secure. So I'll take my chances with Muse because it saves me time and allows me to get more done.
Those other issues, while perhaps examples of bad corporate behavior, aren't really security lapses per se.
Meta has a long history of treating their user’s data recklessly. Cambridge Analytica being a prominent example.
They’ve also had a long history of using shady/deceptive practices to gather data, such as Onavo.
Is their internal security up to par with other cloud providers? Maybe. Security from them? Definitely not to be trusted.
(Former employee)
Grok as well seems to have a hit a sweet spot.
UI/UX
Many times, UI/UX makes all the difference in user adoption & use.
But ChatGPT already has the same UI UX.
It isn’t though.
Muse is:
- a single thread
- has avatars that you can name
- has approval cards
Now dots is similar in concept but that only launched a few days ago and Muse launched a month ago.
It does not at all. ChatGPT is trying to be an everything app and it’s extremely convoluted by this point and hard for most people to make sense of, especially if you’re not in tech.
That's strange because they still have a billion users. So they're all using it somehow even without being in tech. The core UX is still a chatbot.
OpenAI does now with Dots but it’s still desktop only so they flubbed the ease of use that has made Muse and Grok such growth engines.
It’s also not free.
The article emphasises exactly what you said: technically it has no new features, it's just better UX than anything else, and more friendly to non-techies.
When it comes to being mainstream, an approachable UI sells
I don't understand C, couldn't assembly already do all this?
But chatgpt is still user friendly or claude I think also do these kinda of things
And dropbox is just rsync and some UI
Difference being they're already largely similar so the analogy would be between C# and Java.
This is the Dropbox comment of the AI assistant era. At least it's not quite as out of touch.
No, I'm asking why ChatGPT isn't sufficient because they literally do the same thing unlike Dropbox vs rsync. In other words ChatGPT has the same UI UX but people flock to Muse for some reason.
Facebook's install base is 3x that of cgpt
That’s pretty much the question the article tries to answer! TL;DR - good product design, personality mechanics, and ad-funded tokens is what makes it different.
Think of it as a Meta-hosted OpenClaw VM for each Facebook user.
Meta is reducing the friction for your parents to run OpenClaw without any assistance.
It’s the most consistent and user friendly agent with the ability to also be a harness for multiple custom agents. I literally had people running around the office amazed by muse last week. It has brought agents to the masses in a way a lay person can understand and use.
“Interestingly, what’s making Muse a success also explains why Google has mostly failed to capture the zeitgeist so far. They have the compute and the ad money to fund something like Muse. But they have no product chops to build something simple and understandable in consumer AI.”
Umm what? Having billions of hourly active users across YouTube, Google, Google Maps, and Chrome still doesn’t mean they have the product chops to build another consumer facing app/service? I think it’s inevitable that Muse will get crushed.
> YouTube, Google, Google Maps, and Chrome
You listed products that have been around for ~20 years. Keeping up already entrenched products running doesn't correlate with being able to launch new products.
Those are all quite mature products which they have maintained and grew super well. They haven’t had a good track record in consumer AI, there’s no coherent product strategy or offering. They had some localized hits like NotebookLM and nano banana but it hasn’t translated into organic growth and usage for Gemini
Google Search and Google Maps are utilities, not consumer apps that vie for leisure time, so I don't think they are really comparable. The next frontier for AI apps like Muse and Dots is competing with the social media giants.
We should look to see how Shorts vs IG Reels have adapted to the short form video landscape for product chops in this area. IMO IG has adapted far better to the new era, and I very often get people sending me Reels (they used to send TikToks). I very rarely see anyone send Shorts but it still happens. I expect this trend to continue and IG to extend their lead here. And Muse certainly has first mover advantage in personalized AI apps, who knows when Google will finally release one.
Google has huge amounts of training data but very little of it is relevant to creating an AI personal assistant with good UX for common tasks. Like Gemini might be really good at answering specific questions or generating synthetic content but that doesn't mean it can securely, reliably carry out workflows in a way that feels natural to regular users. At least for now, that part still requires some human vision and product management.
Youtube and google maps are used because of content they have. As software they are awful.
I like that you can get root access to the VM your instance runs in, connect it to your tailnet, automate connection approvals, and operate a torrent seedbox on Meta’s infra.
Have you figured out a reliable way to persist stuff? Like a launch script or similar that doesn’t reset when the VM shuts down and can be used to restore tailscale
bashrc, bash profile, cron, systemd, etc all get wiped clean the moment the VM resets
Totally theoretical of course
right, it could be made up
https://neuromatch.social/@jonny/117329100531227330
I appreciate the edit
perhaps its for plausible deniability next time they torrent training data /s
If you can design something that a Baby Boomer can understand, then you have a mass market product at full scale. Otherwise your hopefully building something that a niche group finds attractive and that grows from there, outwards to scale.
It’s boomers
I don't get it. What muse does it better than others? it's just chat app connected to AI service.
Did you read the article? Some suggestions there.