My way too early takes on Meta Muse, their foray into the personal AI agent arena
First impressions:
- its very fast and snappy, a nice change from the time you need to spend watching a prompt go through with other products. The product of good model routing? Use a fast/cheap one for the basics, bump up to heavy duty for the computer use?
- not iMessage based (although you can connect WhatsApp), the app is pretty good and offers much more user interaction points than just a long chat. Ideas for prompts, automations, and use cases are prominently displayed, making it easier for a casual user to get started. Ability to have multiple chats is nice. I maintain that iMessage will just be a channel and the long term goal of all personal agents is to have some first class app interface
- payments, security, controls, are front and center. Partnership with Stripe to use Link to check out
- secrets/credentials handling is very smooth, it's nice to just have a place to see everything connected in the app
- payment protections (refunds) for agents is something no one has offered to date, although it's been much debated, will be interesting to see if it drives increased agentic commerce as people trust the agent more
- have yet to do anything heavy duty on it with computer use, so can't say how it compares on that front to Instinct or Grok Bot. As a lightweight assistant, it seems good enough already. Computer use was one of the biggest differentiators for this past wave of assistants so will need to dig in more there. Success rates here will be a big thing to watch
- responses, while very fast, feel a bit more mechanical and not as smart. Tradeoffs! This is probably just the difference between Muse Spark and whatever Instinct runs
- One thing Instinct is particularly good at is being proactive without being pushy - haven't had Muse long enough to know how it does on that front
- Onboarding, not as smooth and premium as Instinct - which makes sense if you think of Instinct as more as premium software and Muse as something for everyone
- Meta having their own models combined with massive amounts of available compute + hardware they own will be a big moat and edge at scale. They can serve results faster, have less downtime, and keep success rates higher
- If they can execute, this can be a very valuable product for Meta and potentially the big consumer AI breakthrough they've been looking for. It may not be perfect for the VC/startup power user that prefers Instinct but it can be good enough for the billions of other people that already use Meta apps. Auto-distribution can make it a Threads level success. Meta ultimately just wants to keep people sticky to their properties and if you end up paying for additional usage, it's a bonus. Just integrating this natively into WhatsApp would be a massive product. This is all a funnel to drive use of Muse Spark
- Pressure may actually be on ChatGPT the most as the preeminent consumer AI product - I get the sense you can get way more out of the free tier of Muse than you can the free levels of ChatGPT. This might be a Trojan Horse for Meta to gain mass adoption for their models. I think Muse is massively subsidized, I have not made a dent on the usage yet. So if you're just a user of ChatGPT, maybe you just come and use Muse instead and get sticky with the personal assistant experience.