@xiaogaifun: Zuckerberg's thoughts on Meta Muse are truly insightful. Meta's Muse is indeed very popular. It even made me wonder if Agents like Codex might become outdated. There's already a lot of analysis on Muse in the industry lately. But if you're really interested in this product...

Mark Zuckerberg’s insights on Meta Muse are truly profound.

Meta’s Muse has indeed generated significant buzz. It even made me wonder if agents like Codex might be relegated to the previous generation.

In recent days, there have been numerous analyses of Muse within the industry. However, if you’re genuinely interested in this product and want to understand it deeply, I recommend consulting the primary sources.

Zuckerberg recently appeared on a podcast, detailing why he decided to build Muse and his understanding of Personal Agents.

I just finished watching this video, and it’s far more reliable and insightful than 99% of third-party analyses out there.

Here’s why: I believe products like Muse have the real potential to become the WeChat of the AI era. At the very least, we’re already seeing the first signs of this.

Let’s break it down step by step.

The premise for discussing all of this is first clarifying how Muse differs from Codex, Claude, WorkBuddy, and other highly competitive coding agents and general-purpose agents from the past six months.

After repeatedly reviewing Zuckerberg’s statements, if I were to summarize it in one sentence: the agents we were previously familiar with were mainly Task Agents. What Muse aims to be is a Long-lived Personal Agent.

Codex is unquestionably an agent. However, its current most typical working mode still revolves around a relatively well-defined task. For example, I just asked it to fix a product bug for me, or it can generate a slide deck according to my requirements.

We are already very familiar with how to use this type of agent. Give it a task, the agent executes, and delivers the result. The reason Meta’s Muse has gained such attention is that, building on this foundation, it introduced several key designs.

First, it has a much longer lifespan.

This is the “Long-lived Personal Agent” mentioned earlier.

Agents like Codex, while already capable of executing tasks for extended periods, still generally operate around a single Task. Once the task is done, that round of interaction is essentially complete.

Muse is different. What a user gives it can be a long-term goal, and it then continuously works towards that goal.

Zuckerberg gave a particularly relatable example. He asked Muse to help him and his daughter plan a baking activity every week, including preparing the ingredients in advance. After one week’s activity, the matter doesn’t end.

For instance, after discovering that making Cake Pops was too difficult this time, he would share this result with Muse, and they would continue to adjust the plan the following week. So, such a task could continue for months, or even longer.

Second, and what I find most compelling, is that it proactively observes the user and then generates tasks on its own.

No matter how powerful agents like Codex are today, most of the time we still need to tell them what to do first.

Muse has a very important design called Ideas. Based on the user’s goals, history, and context, it independently figures out what else could be done next and then proactively asks, “Would you like to try this?”

For example, Zuckerberg and his daughter were playing Civilization together. Muse first created a complete set of game strategies for them. After that, it came up with the idea to expand the strategy and incorporate historical knowledge about different civilizations. So, it proactively asked Zuckerberg if he wanted to continue. Zuckerberg said yes, and it proceeded with the work.

This feature directly created an “Aha Moment” for me.

Just like when I was discussing Muse with a friend the other day, I said this thing is starting to show some initiative. It proactively observes the user and then suggests things it can help with.

This feature is particularly interesting.

Because the biggest problem for many people using AI today is not that AI isn’t powerful enough, but that they simply don’t know what AI can do for them.

When we open a chat interface, we still have to think of the prompts ourselves and then decide what tasks to assign to the AI.

However, if an agent already understands our context well enough, it can actually take the initiative to discover problems and drive tasks forward.

Building on this, I want to explain why I believe Muse has the potential to become the WeChat of the AI era.

Actually, everyone should be feeling this now: the real change happening to many products in the AI era might not be about just cramming AI into the product.

The biggest difference between new-generation AI products and those from the mobile internet era is that their target audience includes not just humans, but also agents.

Therefore, from the very first day of product design, agents must be considered as highly important users.

The agents in the current industry are essentially playing a Single-player Game. You have your agent, I have mine, and we are mostly independent of each other.

But Zuckerberg believes this state will change completely later on.

He mentioned the concept of a Fleet. Simply put, if one person’s Muse discovers a particularly good method of use, other Muses can learn from this anonymized experience and then recommend it to suitable users.

In this way, the more people use Muse, the better each person’s Muse will theoretically become.

Furthermore, agents can directly interact with each other.

Because a Personal Agent already possesses the user’s context. In a sense, it’s already a digital counterpart.

Then why can’t two Personal Agents communicate on their own?

Take a very simple example.

Suppose I’m planning to go to Inner Mongolia next month to see the autumn scenery and want to find a photographer. Today, I have to search on Xiaohongshu, ask friends on WeChat, or look everywhere for a suitable person. The future could be entirely different.

My Muse knows I’m going to Inner Mongolia, including the time, budget, and the photography style I prefer. So, it can search the entire Agent Network.

Another person’s Muse knows its owner happens to be a photographer, is available during those days, has a suitable price point, and their past work matches my preferred style. The two agents could then have an initial communication.

At this point, the so-called social network is no longer just about humans actively scrolling feeds, adding friends, and sending messages. It might become a situation where everyone has an agent working behind the scenes.

These agents, carrying a person’s context, needs, relationships, and resources, move through the network to find suitable people and opportunities for them.

I think this imagination space is truly vast.

The fundamental node of WeChat is the human. Humans establish relationships, then chat, post to Moments, join groups, and conduct transactions. The fundamental node of something like Muse might still be the human, but the entities actively operating at high frequency within the network will likely start to be agents.

Or let’s put it this way: WeChat is a Human Network. The ultimate form of Muse could be a Human + Agent Network.

Of course, once we get to this point—agents having our context—everyone will probably think of the same question: what about privacy?

When using Muse these past two days, I’ve actually been very cautious myself.

Because a Personal Agent has another major difference from products like Codex: to truly understand us, it needs to obtain more and more of our personal context. Emails, calendars, chats, payments, various accounts.

As we use it more, it might end up understanding us better than most people. This is indeed unsettling.

Regarding privacy, Zuckerberg’s logic is that the best privacy commitment isn’t asking users to trust that the platform won’t look, but making it so the platform technically cannot see.

Here’s their approach.

Meta now provides each person’s Muse with an independent Secure VM (Virtual Machine).

You can think of it simply as the user’s agent having a relatively independent computer in the cloud, where personal data, credentials, and the agent’s operations are all contained within this environment.

Next, Meta plans to roll out an even stronger Confidential VM. At that stage, all data within the entire VM, including conversations between the user and Muse, will be further encrypted, with the goal that even Meta itself cannot read it.

This kind of system has already been proven by WhatsApp. The issue of privacy should be solvable through technological means.

Furthermore, Zuckerberg’s vision for Muse’s business model, at least from now on, seems quite mature.

Because getting the vast majority of people to pay for an AI long-term, whether in China or the US, I feel, is not that easy.

Zuckerberg’s idea is to hope that the vast majority of users can ultimately use Muse for free and at scale.

Including now, Muse still offers a significant amount of free allowance, because his judgment is that if the logic of Muse truly holds, then it should inherently have the ability to help users earn or save money.

For example, for a small business, Muse can help build products, run operations, and can integrate with Meta’s advertising system to help users operate continuously.

As long as there is economic value exchange, Muse could then take a cut or a toll, so to speak.

Therefore, the long-term business model he envisions isn’t just charging users a subscription fee every month. It’s more likely that, in the process of agents helping users complete transactions, do business, or generate income, Meta takes a very small portion.

And this money might not even be paid directly by the user; it could come from the merchants transacting with the user.

This logic is actually quite straightforward.

For instance, if I ask Muse to book a flight ticket or reserve a hotel for me. Originally, the merchant would have to pay a customer acquisition cost to get this user. In the future, they could very well allocate a portion of that to Muse.

These thoughts are quite insightful. If you have time, I strongly recommend taking a look. This is probably the most worthwhile read of the Mid-Autumn holiday.

I’ve put the podcast link in the comments section.