We Need To Talk About Meta Muse

Meta launched Muse on Sept. 8, an agent that actually executes on your behalf, and I bought Meta. I sold Palantir to do it.

I wasn't expecting Meta to launch something like this at this point in time. The fact that they did really signals a huge thing. This came out of a conversation on Amit Kukreja's stream, and I want to lay the whole thesis out here.

I currently own Meta stock.

Three generations of agents

In a way, this all started with Claude Code. Anthropic built an agentic harness for Claude and figured out how to make the model come alive. Before that it was question and answer. Harness these models, give them tools, and as they get more intelligent they can take actions on our behalf.

Then OpenClaw. Completely open source, a fun hobby that took the world by storm. It was very clunky and always kind of didn't work, but it did enough that you'd go, holy crap, this is mind-blowing. It was basically the beta version of what Muse and Grok Bot are now: a do-anything AI that can do anything on a computer a human could.

Hermes took that to a completely different level, but it was still hard to set up. You had to put in the work, set up a little server or run it on your local machine. It wasn't for the normie.

Muse and Grok Bot are the third generation. You sign up, you start talking to it, and it starts doing things for you. That's the one that's really going to take hold. Grok Bot has the first-mover advantage, for sure.

And the models have hit a threshold. They're good enough now to do most things, especially work that's administrative, routine, repetitious, part of a process, or simple.

Meta has three things going for it

This is very similar to my SpaceX thesis. For this agent process, though, Meta is even more so. Three reasons.

One, compute. Meta is investing very heavily in it. They're going to own the layer underneath, the data centers that power these agentic systems.

Two, distribution. Meta has the mother of all distribution. Facebook, WhatsApp, Instagram. More than 3.5 billion people use Meta's apps every day.

Three, the harness. Meta decided to build an agentic harness and an agent that will actually execute on people's behalf.

Here's why I care. Put those together and you have a recipe where, overnight, Meta can become the world's go-to personal assistant for first-tier work. Go through my email and tell me the most important thing. Organize this. Buy something for me from the store. Do this recipe for me. That's a wide open gate for them.

Tokens at a discount, plugged into everything

Because Meta owns the compute, let's say they start hosting all these different models. Then they can serve those tokens at a discount. As those tokens get more intelligent, they take more actions. As the models get better, Muse gets better and can do more complex things.

I'm very impressed with Zuck on this next part. Meta partnered with plug-in companies like Plaid, so Muse can hook directly into your financial accounts. As far as I know, Grok Bot doesn't have a built-in Plaid plugin yet. Meta has sent a very clear signal. They're willing to take the risk to plug into basically anything a human would want plugged in, so the agent can run their lives.

Agents can basically do anything in the digital world, especially as the models get better. I don't think we have fully wrapped our heads around what that means long term.

Not financial advice. But fast forward three to five years and, to me, Meta is a very different company. It will have captured where we're going from a technology perspective, with a huge advantage in distribution, compute, and the model. They'll hyper-optimize their models to run on their own agent harness. Obviously.

Why didn't Google or Apple do this?

This is also where Elon's X purchase from, what, four years ago now starts to make a lot more sense. X is distribution for these agentic harnesses.

In theory, any large distribution company should be doing this. So why hasn't Google done it? Why doesn't every Android device have something like this? Why doesn't Apple?

Culture. It has to be culture. It has to be leadership.

Look at what Zuck and Elon have in common. They both have that control where it's, we're doing what I want, and guess what, you're not getting me out of this.

Give power to the psychos

I think we live in a time where you have to let the psychos run. You have to give power to the psychos and empower them. Make sure they're not killing anybody, and that's probably going to be hard.

I bet this is why these AI companies talk about a 10% chance of AI killing everybody. They're handing the keys to the kingdom to these psychos to run with it.

But you almost have to. If you don't, somebody else will, and whoever does it and executes at a high level wins. So it's a race between all these companies to see who can most responsibly give the keys of the kingdom to the psychos to build things with artificial intelligence. Because AI empowers the most creative people.

Why I bought Meta

I'm investing in leadership. I'm investing in risk-taking. I'm investing in culture. And I'm investing in vertical integration.

Meta controls four things: compute, distribution, the model, and the harness. Control all four and you control your whole destiny.

SpaceX has one more, which is power. Through Tesla they'll make the batteries. Through Starship they'll put data centers in space and harness the sun. SpaceX has taken it one level deeper, and it has objectively the greatest entrepreneur of our time leading that effort.

But Zuckerberg? This is legit.

For this agent process specifically, though, Meta is the one I think is incredibly well positioned.

This is how I typically invest. I try to look 5 to 10 years into the future and position myself for that future. The bet is that the culture, the leadership, and what they're working on add up to a moat big enough that it's going to be very difficult to overcome. Then I let those babies run.

Yes, I sold Palantir to buy Meta. I still love Karp. I just love the combination of playing in the physical world and the digital world. Data centers paired with the models, the agent harness, and the distribution. To me, that's the moat of the future. Without the atoms you don't have the model. You cannot serve an extremely powerful model to a lot of people without chips and data centers and ultimately energy. If you own that entire stack, let's go.

Glasses and word vomit

Amit pushed back on whether Muse has staying power, and on how Meta actually makes money on it. That's a very fair point. Here's my answer.

Where Meta wins is that so many people already live in Instagram, WhatsApp, and Facebook.

Then there's the glasses use case, which is fascinating. Picture an always-on agent on your face, in glasses rather than a VR headset. If they get to where the glasses can show stuff clearly in your field of view, with a microphone and all-day battery, I could constantly interface with an agent. Maybe not out in public in front of everyone.

The future of this technology is spoken word. You could still type if you want. Number one, talking is way faster. Number two, and this is something I've found in my own use, the AIs have gotten so good that you don't even have to be that coherent or clear about what you're trying to describe. The AI will just figure it out. You just word vomit.

Fast forward that technology a year down the road and leverage that distribution. Do anything for anyone is unbelievably powerful when you can get it in front of that many people.

The fallback: rent the tokens

Now say I'm wrong about Muse. Another model comes out, your ChatGPTs, your Anthropics, and people just don't use Muse that much.

Then Meta does what SpaceX does. Rent the compute out. Make that money on those tokens, baby. Those are going at what, 80, 85, 90% margins.

Tesla: undervalued on what time scale?

Amit asked me if Tesla is undervalued. My first question back was, on what time scale?

Over the next 12 months, who knows. Over five years, I'd be lying if I said I didn't think it would be multiple times higher. That's why I'm in the name and always have been.

Robotaxi and Optimus, those two verticals, are going to be unbelievably valuable. And I think people are massively underestimating the margins Robotaxi is going to generate.

Look at Austin right now. They're charging roughly Uber-level prices. There's no driver in there, and the thing drives nice.

The key variable is the ratio of supervisors to Robotaxis. Based on my model, if they can get it down to 3:1, three Robotaxis per supervisor, they're at sub-dollar per mile cost. If they charge above roughly $2.80 to $3 a mile, they're making about two bucks a mile in gross profit.

So why is the stock stuck? I have no idea. It's an expensive stock, and it always has been. But you don't buy Tesla for next year. You buy it for the long, long term, because you think Elon is going to change things from an industrial perspective and build moats that are nearly impossible to overcome.

Uber's way out

I just don't see how Uber competes.

Jason Calacanis and I got into it on X over this. He called me later and was super gracious. He apologized for how he came across.

His thesis is that there will be multiple winners, and Uber is one of them, because not every automaker is going to want to own its own distribution. If that's your thesis, that's fair.

I just don't see why a company with very large scale wouldn't run its own network. Rev it up and undercut Uber on price as you add cars. Now Uber is in a pickle. It has to pay drivers less to compete, especially without enough driverless vehicles on the fleet. And the more driverless vehicles Uber adds, the more likely the fleet provider says, why do I keep giving you a cut of the margin?

I'll grant the other side. There's regulation. There's hesitancy.

The more I think about it, though, the more I think Uber might be very well positioned if it gets out of what it does today and goes straight into the human-to-human gig economy. Anything with a machine involved is going to be fully automated. That won't stop humans from wanting to do things with other humans. The Fiverr of the physical world. Why doesn't Uber do that?

SpaceX buys Tesla: the Meta thesis on steroids

I think SpaceX is buying Tesla this year. If not this year, next year for sure.

The timing isn't that important, because they're already doing so many things together. You already have Grok Bot inside Tesla's cars. You've got Tesla Megapacks helping power Colossus. They're already co-mingling. The question is when they make it official. What that does to the stock, I have no idea.

The way I see it, the merger is the Meta thesis on steroids. The entire vertical of digital AI on top of physical AI. Do-anything software, plus the supply chain for do-anything hardware: self-driving vehicles for transportation, humanoids for labor.

That feels way, way bigger than a $3 trillion market cap. At that point it's literally the foundation of human civilization.

What my agents are doing right now

For the last few weeks, my agents have been making my Shorts. They go into YouTube analytics for live streams, find the best moments, and make the Shorts using a DaVinci MCP on a machine, with Claude Opus doing the edit work. Then they upload to YouTube Shorts, TikTok, Instagram, and X at the right times. I know for a fact Grok Bot does this. I don't know if Muse does. Experiment with both.

Hermes wasn't for the normie. Muse and Grok Bot are. You sign up and start talking, and Meta has the distribution to put that in front of more than 3.5 billion people a day.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=EQYj6ALsFR8.