Muse Spark: 1.0: April 8 1.1: July 9 1.2: August 5 1.3: September 2 1.4: Between now and October 21? Pace was 92 days between versions, then 27, then 28. Average across all is 49 days but recent cadence is ~4 weeks. I’m very interested to see how 1.4 performs, as I think it will say a lot about Muse’s staying power. 1.4 doesn’t need to be the best of all models, given Muse’s exceptional product design. But it can’t be too far behind the frontier either, or Muse will lose users as competitors catch up on design and user experience.

I don’t buy the “nobody uses AI” meme going around. Up to 60% of Google searches now show an AI overview. ChatGPT has 1.2 billion users. Real AI diffusion is going to look boring, like email. It’s happening. With Muse, it’s happening for agentic AI too. The bigger question for me is why there’s already so much diffusion and yet the world feels pretty much the same. What would change that, if anything?

News feed in Muse is underhyped. My Muse curates it, heart taps fine-tune it further, items are digestible and have custom graphics like charts for easier scanning, you can discuss any item with your Muse for follow-up. Reminds me of the old ChatGPT Pulse feature, but with the polish of a Meta product, given their years of experience with feeds.

Update: My Muse (Kit, by the way, in honor of the famous car and Game of Thrones) has decided to get this to compile before my device arrives so we don’t waste time when it gets here.

Just ordered a Waveshare ESP32-S3-Touch-AMOLED-1.75C after discussing it with my Muse. What pushed me over the edge was that it found the gadget SDK has built-in support for dictation and speech. So, voice out of the box? I’m going to later try adding a custom wake word.

99% of my AI interactions now happen in a monothread, either Muse or Dots chat, or a single project thread. (The rest are legacy projects I haven’t yet migrated to a single thread.) Creating a new chat for every request feels archaic.

I wonder if Muse Charm will use Muse Gadgets. That would be cool, showing how even Meta is using this.

Does Meta have a status page for Muse? My Muse has been a bit less reliable today and I want to check a status board. Normally I'd ask Muse, but...

I'm increasingly keeping tasks, subtasks, and context in Todoist so it's easier to switch between AI tools. When I want to try a new one on a task, I just point it at Todoist, rather than having an AGENTS.md or other file on any one computer. Works really well.

Dots biggest value add: proactivity. This is what's impressing everyone using them in our company even though it doesn't translate well into demos on X. Today already, my Dot proactively prepped talking points for a 1:1, suggested times for a proposed meeting someone emailed to request, and warned me about a meeting that got moved. All of this was based on its increasing context about me and activity monitored via connectors.

Meta has enough compute already to give 1 billion people a Muse with its own computer. This isn't GPUs, it's CPUs. Imagine Meta gave away 1 billion free computers. It's essentially that. Strategy? One thought I had: People will ask their agents to do stuff -> agents will use their own powerful computers-> people won't need expensive hardware from companies like Apple -> people will buy lightweight hardware like glasses and Charms from Meta.

Which would you choose to upgrade for an extra $100 a month?

Out: Smart devices, dumb agents. In: Dumb devices, smart agents.

Overly simplistic, but: Consumer: Muse Prosumer: ChatGPT Professional: Claude

The capability of my agents’ computers is becoming increasingly more important than that of my own. Mine are increasingly used to assign agents tasks and review the results. They can be pretty dumb terminals. In this context, Meta’s decision to give every Muse user a good cloud computer makes sense. It reduces the utility of hardware from Apple and other premium providers and redirects those funds to Meta interfaces like glasses and Charm, which don’t have or need much onboard compute because they send stuff to your agents’ cloud computers.

It’s early still, but the anger I feel when my agents get blocked leads me to believe that I’ll choose agent-friendly options whenever they exist and stop using all agent-blockers. Blocking my agents feels like blocking me from getting things done.

With the success of Codex and now Muse, it’s hard to argue against the importance of product experience over raw model capability to a large number of users. A much better product experience is more important than a slightly better model for daily use.

I would love to hear from the Muse team how they’re processing the in-app feedback you can give. I’ve given a ton, and I imagine other people have too. I hope this is being collected, triaged, and turned into requirements and code by agents. Would love to learn about this.