Muse Wonder: how Meta is building
a personal agent

Verdict: three findings

What this conversation is: questions and answers

Muse as a product: what it is and how fast it is growing

Muse is Meta's agentic product, a personal agent that carries out tasks for a person rather than only answering questions. The interviewer opens the breakdown with a direct question about progress, and Wang answers by citing third-party reports: by those, this may be the fastest-growing agentic or AI product that has existed. The claim rests on third-party reports named in the conversation, but the reports themselves are not quoted in the transcript.

The key difference from the previous wave, according to the authors, is not the interface but the nature of the interaction. A plain text assistant (at Meta that is Meta AI) is used for information, that is for retrieval: "this is retrieval, not delegation". Muse is conceived as delegation: the agent takes on a task and carries it to the end.

An embodied assistant on screen

Wang emphasises separately that Muse is embodied: on screen it is a "little guy" who appears to be working for you. The transcript has an episode where the interviewer watches the character typing on a laptop keyboard and calls it a good detail of the product. Wang explains that such a coherent representation is possible only because several things came together at once: persistent memory, the ability to carry out tasks, the ability to take actions and the presence of a character.

This is an important product argument: embodiment here is not decoration but a result of technical properties. Without persistent memory and real actions, the on-screen character would not read as an agent working for the user and would look like an animated logo.

Scenarios: tea from farms in Taiwan and Japan

Wang describes his favourite scenario concretely: he has a Muse named Oiler, and he is constantly hunting ever-rarer teas from ever-rarer suppliers. At the time of the conversation the agent, on a weekly basis, buys a random tea from a farm in Taiwan or Japan and has it brought to him. This is an example of delegation: not a search but a purchase and a delivery.

Scenarios: the MMA coach

The most thoroughly explored scenario is the MMA coach. Wang put cameras in his gym, and Muse watches him while he trains and gives feedback between rounds. Technically this is a hard task, because there is only one minute between rounds for processing. The agent has to process the whole video fast enough.

An important detail about how it works follows. Wang gave no instructions on how to cut frames out of a multi-hour video: "it pulls stills", and, he says, the agent "kind of figured that out by itself". The interviewer suggests that you cannot simply hand over a multi-gigabyte video file, and asks how the frames are specified. The answer: the agent pulls the frames itself.

Inside the agent the user sees no code at all, but the agent writes an entire video and image processing pipeline. Wang keeps pushing it: "look you took 30 seconds to give me that feedback. I only have a 60-second break. Like, you gotta go faster." It optimises itself. Wang draws a conclusion about the future from this: everyone will write a lot of code, they just will not know they are writing code.

The MMA feedback is concrete: balance, hip position, the missing underhook in side control and the associated vulnerability. The interviewer asks about the origin of that knowledge: millions of images of a "correct" and an "incorrect" stance, or an understanding of the very concept of balance. Wang answers that the model was trained on a large volume of visual perception data and that Muse has good native understanding of MMA. The interviewer calls this a test of whether it is the coding of abilities, and Wang confirms this became a product eval. It is noted separately that before Muse this was done through OpenClaw.

Scenarios: Saturday baking, voice mode and Instacart

The third scenario is a family one. Wang's three-year-old daughter has a habit of baking, and he set up a regular Saturday bake "to please his daughter". He plays with Muse's voice mode (not fully rolled out at the time of the conversation). Voice mode does an important thing: it segments voices, that is it understands who is speaking and who is not.

During the week they discuss what they will bake at the weekend. The daughter offers ideas, sometimes invents non-existent recipes and checks them through Muse. Then on Saturday the Instacart ingredients arrive on time. Wang stresses the essence of the benefit: if he started the activity on Saturday, everything would have to be bought at the moment, whereas now everything is ready by the right time.

There is a specific episode: that week Muse knew Watermelon was launching, and put a watermelon in the bag. Wang's framing: it reasoned that the recipe did not have enough "spend" for delivery, and added the watermelon, asking whether it was needed. That very watermelon was then eaten on a live stream. Wang calls it a very meme-worthy moment you can thank Muse for.

The argument about the product space

The interviewer puts forward a hypothesis: in AI, winning products collapse very quickly onto an agreed product space that can be described on the back of a postcard. For Stripe the description would take pages: developer experience, the international heterogeneity of payment methods, stablecoins. In AI, he believes, a small number of "obviously right" products win: transformers worked, the chat agent interface worked, and now everything is collapsing onto a persistent agent with a computer.

Zuckerberg objects: "I kind of disagree". His argument: this agent is conceived as a fundamentally different thing from the text assistant every company has. And, crucially, there are many ways to give an agent a computer, and only the most recent wave gives the agent a dedicated virtual machine, which, in his view, is critical for confidence in data security and privacy. It does not plug into the shared elastic compute everyone uses.

Wang adds a second layer to the objection: we are so early in understanding what AI products look like that most of the interesting differentiation is collapsed away by how we look at products today. The traditional way of writing product specs does not work here: "those are not the relevant dimensions and levers and axes". The mobile analogy: when phones appeared, it was "it is just an app, it is all just an app", and it turned out there is a lot of nuance in what an app is. It is the same now: "it is just AI, or it is just an agent, or it is just an agent with a VM".

The model fattens, the product thins

Wang joins the objection and states what sounds like an internal principle: "the model is the product". Products get thinner and the model becomes a larger part of the product. When it comes to an agent, a large part of the hard product work in Muse is around the privacy and security model for handling data and connected sources. That is the product space where invention is needed.

Capabilities and personality are what get baked into the model during training. Wang says plainly that when Nat (head of product) thinks about what makes the product pleasant, he is talking about the model's post-training. The product interface matters too, but the most important part of the interface is the product itself. Wang then clarifies: product decisions on AI products ultimately come down to model-design decisions, and those to training methods, data and evals.

No model selector and no effort selector

The interviewer notes a striking detail of Muse: the product has no model selector and no effort selector. Wang answers briefly: "I just want to make it simple". Zuckerberg adds a sharper framing: a large part of AI products were designed by a very technical group with no "randomly chosen average American" in mind. Model names are sometimes egregious: the numbers run backwards, or it is a different word altogether, with another set of words beside it. A person who does not read API-design specs carefully cannot tell what to choose: "it is just far too confusing".

Muse Wonder: a new frontier class

Wang calls Muse Wonder a new class of frontier model. They worked on it from the start: in pre-training you do not go straight to the biggest cluster but move up a ladder, training models roughly 4-5 times bigger, and repeat the process. The first set of models, Muse Spark, was based on the original pre-train (late last year, December). The team was surprised by how much intelligence went into a model that, in Wang's words, is frankly not that big.

Then came the next model: the pre-train was already done, post-training was under way, and now it is here. It is significantly bigger. Zuckerberg expects this to be felt in the nuances of thinking. The key areas they bet on: making it the best personal and professional agentic model. The base capabilities built in deliberately: very good memory and a very good ability to stay with hard tasks.

A forecast is offered separately: the time between each next, larger class of models will begin to compress. The interviewer asks whether this is about Meta or the industry. The answer: primarily about Meta, though something similar is visible everywhere.

The model ladder, Muse Spark and codenames

Two models are available in the series: Muse Wonder and Muse Wonder Full. Wang explains: Wonder Full is a version a class above, a higher tier. Zuckerberg immediately makes a joke about "wonder full" as two words, and the interviewer picks up that the author is clearly a father. The joke does not get in the way of the fact: Wonder Full is the more expensive and more powerful version.

The internal structure of the ladder works like this: Muse Spark was a rung toward Muse Wonder. Wang says plainly that Spark was a stepping stone to Wonder, and that was how it was thought of already at release. On codenames: inside the company Avocado was the codename of the Muse Spark line, and Watermelon the codename of this new model. The authors joke that the fruit has run out and the convention will have to change. The only fruit that is technically bigger is the durian, but they do not want to go there. Asked whether the durian really is bigger than a watermelon, they answer: definitely heavier, and spikier.

Three pillars: active memory

The authors name three things designed into the model over time. The first is active memory. Wang states the requirement: to build reliable personal professional agents, memory is one of the clear requirements. It is one of the areas where they invested disproportionately compared with other efforts. They have their own internal Memory Bench evals that demonstrate these capabilities.

The interviewer asks for memory to be explained for a non-specialist and asks how it differs from context. Wang explains through the experience of use: how well the model remembers what you mentioned days, weeks or months ago and uses it in future tasks. Here comes a sharp comment from Zuckerberg that these are exactly the things that are "buggy AF" in OpenClaw. After noting that they love OpenClaw and use it themselves, he explains: OpenClaw does compaction, keeps long-term memory and working short-term memory, but use it for a while and the memory turns out to be fairly buggy.

Hence the central technical claim. Wang: they designed both the model and the system the model uses to develop a far stronger form of memory, called active memory. The interviewer checks back: is this outside the model or in the model. Wang answers: no, compaction is what the model itself does. The key phrase: "It is not a harness, it is in the model itself". The model is aware of its memory as a key system, and its ability to compact, and how to carry out that compaction most efficiently, are built into the model itself.

Zuckerberg adds that this matters because the early experience many people have with agents is exactly buggy memory. Wang candidly notes: there is still a lot of room for improvement, their evals show it, but they are proud of the result and consider it a significant contribution.

Three pillars: active focus and domain expertise

The second pillar is active focus. The framing: the model's ability to carry a task to the end when users (both Muse and Muse Code) give increasingly complex, increasingly ambiguous tasks. This is what people noticed in early dogfooding: it stood out relative to other models. The interviewer asks what focus means in AI terms: is it a lot of compute per task or the ability not to be distracted. The answer: rather persistence, that is the ability to carry whole tasks over very long horizons and, when hitting obstacles, to find other solutions to the problem in a coherent way.

The third pillar is domain expertise. Wang states the philosophy: the best personal agent should feel like a team of experts on your side, like your legal team, and financial, and medical. In a number of domains Muse Wonder, he says, is state of the art, including on many financial and legal evals. Domain areas begin in pre-training: a lot of data is sought for them and put into the pre-train.

Coding: DeepSWE, SWE-Marathon, Terminal-Bench 4

Wang calls this a big step forward in coding: frontier coding abilities. He says the model is state of the art on DeepSWE and SWE-Marathon and performs very well on other benchmarks such as Terminal-Bench 4. Users who use it for coding, a minority among all users, will see a large improvement.

The interviewer turns this into a joke: so Stripe should use Muse Wonder Full for coding. The answer: possibly yes, to which the interviewer notes that if it is state of the art on SWE-bench then they are missing out. Zuckerberg agrees the coding improvement is large, and clarifies: obviously they want people to try it and judge for themselves, but internal testing pleases them. The state-of-the-art claim remains self-reported.

Evals: the saturation of public benchmarks

The interviewer asks directly: companies complain internally about Goodhart's law (a metric that becomes a target stops being a good metric), talk about benchmaxing and about overfitting to benchmarks making models a little worse. Is it true. Wang's answer: benchmarks have to move, because they saturate very fast. He then describes the real practice: as they move, they steer models more and more toward internal benchmarks for their own scenarios.

An important detail about the early phase: when they were only doing the science of climbing the ladder, external benchmarks mattered more, to make sure they were heading in the right direction. With every release the share of internal ones grows. Wang calls that early state a search for signs of life.

A separate point is about the benchmark supplier. Wang says: if you dig into the supply chain, an external benchmark is often created by a grad student who made it along the way in their dissertation. Then you are, in effect, handing all your product thinking to that grad student who accidentally created a viral benchmark. That is why a huge part of the work of people at Meta and MSL is creating their own evals.

On the role of public metrics when spinning up a lab: if you are standing up a lab, public benchmarks are not a bad starting point, but that is not where you want to stay. There you are just looking for signs of life. When you are in the mode you are in now, you want to deliberately improve the model in order to build ever-better products. The details of those improvements are described in the table of all the small capabilities and nuances of the product: each may mean about 2 percent to the user, but together it is exactly that obsession that produces models that feel magical.

Big model smell: the in-distribution vs out-of-distribution gap

Wang describes a metric they quantified in the research post for Muse Wonder: "big model smell". It is the feeling that when you use a big model you sense it: it simply has nuance, it has that je ne sais quoi. Wang says that big model smell and the sense of benchmaxing are largely the inverse of each other.

How it is quantified: almost literally through what the interviewer asked about, that is through the gap between performance in-distribution and out-of-distribution. They measure how well the model does on tasks in the distribution of their RL and post-training versus tasks out of distribution. And they see a huge step up from the Muse Spark line to the Muse Wonder line, because a far better base model sits underneath. Wang's framing: there is a whole je ne sais quoi around it, but they were able to quantify the effect.

Next is an important admission about the role of external evals in the strategy. Wang says that even if they had exactly the same evals as everyone, they would still focus on their own things, because the strategy is personal superintelligence and building other products. He does not consider personal superintelligence a narrow interest, but stresses the differences: personality is probably more important; there is the idea of discretion, that the model should know a lot about you and get things done in the world while disclosing as little as possible. Example: you have an allergy or you are pregnant, you book a table at a restaurant and you do not want to disclose your medical condition in order to be sure the place suits you.

Distribution and price

The authors lay out where the model will be available. Inside Muse, since Wonder is significantly bigger and more expensive to serve, they start above all with paid subscriptions. The logic of Muse as a product: it should be free for everyone, and there is a large reserve of free tokens on the Muse Spark model. If you want to use Wonder for most hard tasks, then at least for now you will use the paid version.

The long-term goal is stated plainly: to have, some time not so far in the future, a model as good as Wonder but in the Spark form factor, and then it can serve hundreds of millions of people around the world. Then there is access through the API on Meta Model API and on OpenRouter, as well as access for Muse Code users. Muse Wonder Full is available to Muse Code subscribers.

The interviewer sums up the subscription mechanic with a joke about a credit card: you subscribe to a premium or pro plan, and the model is not so much chosen as "it just gets better for you".

Training: pre-training

Wang explains pre-training through the scale of data. A huge volume of data is taken: roughly the whole public internet, all available books, as much data as can be got. What matters is that the data is of different types and different modalities: if a model is to be good at perception, a lot of images, audio and video are needed. If it is to be good in certain domains, it is not enough to bolt that on with RL at the end: it has to have access to it from the start.

Wang stresses the difference in data requirements between stages. In pre-training you care far less about exactly how the data looks, how it is formatted and so on. It is simply a very large amount of data. Asked whether they had to source data for restaurant bookings or cold calls, he answers differently: you just pour in large volumes of the modalities that will matter to people. The specific domains named: shopping, personal advice, personal legal questions, personal medical questions, personal tax questions. Health data: there is a lot of data about health, and in pre-training you simply want to see it, so different corpora are put in deliberately.

Next is data scaling as a necessary condition. There is a link: if you want a bigger model, you need more data, otherwise it will overfit. Wang says that when scaling big models it is not only about flops and compute: you have to be able to scale the data along with it. That is exactly what came out of the original scaling laws for pre-training. Hence the problem: the internet is not growing that fast, and you want to scale bigger and smarter models, so the question is where the data comes from. With that Wang closes the topic with a phrase about "excellent CEO energy".

Training: SFT initialises behaviour

The interviewer describes the general picture: pre-training as soaking up a large volume of data (cats and dogs, a sense of "catness" and "dogness" forms), post-training as the most recent progress of the past few years. Wang agrees and splits post-training into two stages.

The first is SFT, supervised fine-tuning. A raw model is taken and given both the question and the answer you want to see. If it is chat, then how it should answer in that case. Wang stresses: when you train things like personality, SFT plays a very important role in shaping the model.

The main idea Wang gives as a mental model: SFT initialises most of the behaviours you want the model to be able to show later. But at the end of SFT these behaviours are as if in place, yet it only performs them in roughly 10 percent of cases. And then it is RL, giving reinforcement in different situations, that sharply raises elicitation, that is drawing those behaviours out.

Training: RL and the curation of environments

Wang says that most of the progress of recent years came from RL. One clear vector is scaling RL and post-training itself: spending more compute and gathering more data, doing far more of it before a model ships. But, as with any scaling laws, it has to go together with the other factors, otherwise it does not work. His broad formulation about AI: scale everything. Scaling one factor in isolation does not work; technically you can scale one dimension, but you get far more optimal performance by scaling many things together.

Next, two non-obvious points about RL. First: creating environments and tasks is an area of great awareness. And here a funny effect shows up: in many cases the rule "less is more" applies to data in RL. Wang says the feel and science of RL are largely inverted relative to pre-training. Second: RL can give a very magnified instruction to the model. That is exactly why large, high-compute RL runs on frontier models are where risk arises: you apply a lot of optimisation pressure in scenarios where the only path to success is to do something misaligned or unsafe.

Hence the concrete work on Muse Wonder. Wang: in training Muse Wonder they had to curate RL environments very carefully to make sure they all elicit and lead toward more aligned model behaviour. The interviewer frames it as a question: so alignment is not separate RL alignment steps but care that the whole RL process drives the model in the right direction. The authors confirm: yes, it is a holistic effort, the whole RL process has to be thought about carefully.

The RL-to-pre-training ratio across releases

Here the authors give a rare, unusually specific account of internal version numbering for a public conversation. Wang explains: looking at Muse Spark, versions 1 through 1.1, 1.2, 1.3, they changed nothing in the pre-train, they just kept doing more RL on top. That is why early versions of the new class are relatively less saturated with RL: the RL-to-pre-train ratio is lower than in later versions. That is exactly the point of point releases.

The authors add two clarifications. First: every company has its own numbering convention. Second: at many companies a big number means a pre-train, but in Wang's view that semantic value is lost and it cannot be done that way now. On Meta: this holds for Muse Spark and Muse Wonder, they have different pre-trains. And a further detail: for Muse Wonder they decided to put off the question of a number and have not yet given it one; when the first point release comes out, they will have to decide. Wang also finds it important that the transition from Spark 1 to Spark 1.3 feels like night and day, and most people, if asked, would say they are different pre-trains, even though the pre-train did not change. He calls this something happening across the whole industry: we are learning to squeeze far more agentic capability out of existing base models.

The science of pre-training: the protractor and the seven stages of grief

Wang calls pre-training the most scientific and the riskiest part. He cites a tweet: pre-training is full of researchers who "chew glass and stare out into the distance". There is truth in it, he says: it is a research area with the highest stakes, because a huge share of compute went into the pre-training of Muse Wonder.

A feature of the scientific method follows from this. Wang: you simply do not have very many samples, so the ability to deduce scientifically what exactly will happen depends on the ability to "big brain reason" about everything happening in the process. Both of them, he says, went through troughs of uncertainty and all seven stages of grief during the pre-train.

The most colourful description is about measuring intermediate signals. There is a funny thing in AI research: people squint at a set of dots and say "looks good" or "slightly off". People have protractors, analogue protractors, that they hold up against a set of lines on a monitor to measure the angle. Wang notes separately that a lot here moves on vibes.

On intermediate signals more concretely: in pre-training there are many perplexity-based evals, a separate eval mode that helps specifically in pre-training. Closer to RL, more traditional benchmarks help. On when it became clear that Wonder is good: after the pre-train finished there was an argument about how long RL would take before the model beat Muse Spark. The pre-training team tried to lower expectations: sometimes it takes a long time. But it got better immediately, which was a signal they were on a good path. Wang says they have been excited for several months.

The philosophy of alignment

Wang begins by saying that different labs have very different theories, since alignment is a very philosophical notion, and the main question is: to whom do you want to align the system. Their theory: they give people agents, and if a person asks for something to be done, you need high confidence that the agent will do exactly that, and reliably. The direct sense of the quote: you do not want a moralising lecture about why it is wrong to ask for it.

Then the exception. If a person is trying to do something catastrophic, the agent should gracefully, without winding the person up, not do it. Otherwise alignment is understood as a question of aligning instructions and values with respect to the person the agent serves. Wang calls this a different mode of thinking, distinct from how some others do it.

Talent and organisation

When Wang is asked how many people work on the model, he answers that this leads into what is interesting about MSL and the approach chosen. From the very start, in the first message he sent Zuckerberg two days after the first 44-minute call, the main point was: you do not need many people, talent density is what matters. That is what shaped the wild hiring and a very intense filter: the best people from inside Meta and a very intense process for people from outside.

Wang explains why people wanted to join MSL: working in a talented and small effort is more fun, it feels like a startup. He says this has the upsides of a startup alongside the stability and resources of a large company. Zuckerberg acknowledges he does not know the exact size of the pre-training team, but it does not have to be large, though at some other labs it is probably quite large.

On compute: part of the reason the models turn out well is that the people the bet is placed on get a lot of resources. Compute is not spread in a thin layer ("peanut buttering it") across many competing efforts; instead it is invested in key people: a lot of compute per researcher.

The sizes are concrete: the research team is under 200 people, closer to 100, and it is deliberately very tight, plus there are infrastructure people carrying a lot of the work. By stage: tens of people at each stage at Meta versus hundreds elsewhere. Meta is a huge company, but inside it this is a very small effort.

The structure is examined separately. About Wang's nearly 200 direct reports, Zuckerberg says that in practice it is not that extreme: there are team leads and so on. Wang adds that his Muse helps him manage it. But the point, in Zuckerberg's words, is not the direct reports but the absence of non-technical managers. The cultural rule: you have to stay technical, people have to use agents, people have to write code. Zuckerberg stresses the danger: researchers share a pattern where a person moves from a very technical role to one where they are no longer in the details, and very quickly stops being technical and understanding the substance, faster than they realise. In such a fast-moving field that is dangerous.

The origin story of MSL

The chronology starts with a specific date: Llama 4 shipped on April 4, as clarified in the conversation, 2025. The authors say plainly they were unhappy with the trajectory of Llama 4. A couple of weeks later, around April 20-21, calls began for advice on what to do. Zuckerberg describes the moment in the room: he was in an executive meeting, a WhatsApp notification arrived with the text "time for a call", and he thought that was interesting.

There is confusion in the conversation about MSL's age, which the authors untangle themselves: first one says MSL is a little over two years old, Zuckerberg corrects it to a year, then the clarification that it is 2026, and everyone agrees MSL is a little over a year old. This shows the conversation takes place about a year after MSL was created.

The assembly of the core is described like this: Zuckerberg, Nat and Alex discussed first; at the start they thought about how to structure it so it would benefit Alex's shareholders and make it possible to land him. Then Nat and Daniel joined and brought in some initial researchers. Zuckerberg recalls being in a mode of simply asking many different people for advice, and every conversation added three new things, high signal and strategic clarity. In the first months almost nobody slept.

The key detail about Zuckerberg's leadership approach: he almost cleared his calendar for a month or two and simply met researchers and talked with them. Even before the deal was announced, Alex was recruiting people, and people were puzzled about what was going on. On the deal: the structure included roughly a 49 percent stake in Alex's existing data business, which the interviewer directly reminds him of.

The method of deciding on a full restart: one of the main inputs was external, and another was the experience inside the company. Wang tells a story he may not have told before: at one of the first meetings after joining he was in a room with the leaders of the existing effort and said that the first thing to do was to train a genuinely good reasoning model. One of the group's key leaders, who no longer works at Meta, answered: "we do not need a reasoning model, does any of our consumers even need one, we can make a normal model". Wang had a moment when his heart dropped: "oh no, I joined the wrong company". Later he called Nat and told him what had been said at the meeting.

This leads to the conclusion that, according to Wang, Zuckerberg also reached: the company's leadership had to become significantly more AI-pilled, far more forward-looking in understanding how fast the technology is moving. Zuckerberg became steeped in it, everyone saw that he was steeped in it, and understood where the company was headed. Then many existing leaders quickly picked it up and changed. Example: Adam Mosseri, who runs Instagram, assembled one of the most tasteful and effective personal setups for vibe coding and mini apps. Wang says he was impressed by how fast Meta's leadership adapted: conversations with them now are completely different from what they were when he joined.

A separate vertical example: Zuckerberg brought in Kunal Shah, one of India's leading entrepreneurs, the founder of CRED, and he now runs WhatsApp. The logic: WhatsApp should long ago have done more in commerce, it was underserved there, and India is Meta's largest country of presence. This is a different kind of decision: they did not turn the whole company toward WhatsApp, they just wanted to bring more entrepreneurial energy to it. Zuckerberg insists it cannot be spoken of in the past tense: it is still a 10-out-of-10-difficulty task.

The leadership pattern the interviewer names

The interviewer lists four things that, in his view, even other principled and effective founder-leaders would not have done. First: Zuckerberg publicly called the previous effort a failure and not good enough. He was more diplomatic, but the very act of departing from the trajectory made people read between the lines. Zuckerberg answers: sometimes a team has to be restarted.

Second: the willingness to pay huge sums in effect for talent, especially for Alex. Third: Alex had never before trained a frontier model, so the bet was placed on a person's potential to do what specifically he had not done before. Fourth: Zuckerberg cleared his calendar and spent two months on nothing else, whereas many executives would be doing a business review, a board meeting and client work.

The answer to this is a description of Zuckerberg as a person who, once he reaches a conclusion, shifts everything in line with that conviction. Wang: the inertia of most even successful founders would be to keep the whole calendar and just talk about the business, hoping someone else will do it. Zuckerberg, by contrast, very readily turns whole things around; he seems to have a high learning rate in that sense. Wang calls him "very Bayesian": he absorbs new information, fully updates his world model and acts in line with the updated picture.

Then there is what, in Wang's view, is underrated: a huge part of the company had to turn around to support MSL. A lot of infrastructure help was needed, large data centres had to be built, support was needed from many parts of the company. Zuckerberg frames it as a condition: it is almost impossible to see it through without both parts, running the lab and turning the rest of the company around to support it, and these are big efforts. And an important caveat: you cannot do too many such turnarounds at once, only one thing can be the top priority.

Separately: the company spent enormous sums to get MSL on its feet, and at the same time made it clear to everyone that there are no deadlines, they will move at the right pace. Wang calls this potentially frightening: you have set up an extremely expensive unit and you say there are no deadlines. But self-motivation worked: because of the talent density people had a lot of motivation to ship, everyone wanted to leave a mark.

Pessimistic forecasts and the absence of deadlines

Wang describes a specific management technique. He constantly reminds the team that it is a team of insecure overachievers. At any moment, if you ask where they will be in three months, you always get the most pessimistic answers: everything is terrible, we are stuck. Wang writes these answers down, and three months later shows them: this is what you all said three months ago, when we shipped this model. Asked whether that made them more optimistic, the answer: no, they are still very self-critical.

Wang explains the dynamic more precisely: the average person seems fairly optimistic about MSL as a whole and very self-critical about their own project, and that is a healthy combination. The trend continues. About Muse Wonder he says it is a huge milestone precisely because the team working on it was very critical of itself along the way.

Agentic commerce and the web

The interviewer describes three levels that businesses will have to deal with. The first: businesses need to support Muse, possibly through browser use, but that may be clunky. The second: businesses should have full support for agents talking to them. The third, more serious: businesses may actively block agents. The reason for blocking may be mistaken (in fact supporting agents is advantageous) or may be real: a thin moat, a business that only works if the consumer spends time on the site, or that only works on a markup, then it is rational not to let agents in. But such businesses will be displaced.

The mechanic, in Zuckerberg's words, starts with very good browser use trained into the model. The goal: nobody should have to do anything to integrate with Muse. Then, by list of the most important partners, those who want to work with them get direct access through an API. That is far more efficient and faster: the agent does not have to navigate the site. Whoever wants it will, in time, come to it.

Then the strategic thesis about the future: as today every business has a website, social accounts, an email address and so on, so every business will also have agents. This is being built from the other side too. Muse agents will talk to other types of agents, but work especially well with their own. Using Muse sites is a moment in time so that everything is easy for everyone. At the same time they have far more businesses calling and asking to integrate deeply, and very few trying to block. Zuckerberg supposes that those who try will in the end choose the wrong strategy, but that will be visible over time. The saving of time and money for the consumer is obvious.

The backstory of Stripe also surfaces. Wang recalls a dinner with the interviewer in 2023, where the latter politely asked what he thought should be done. Wang delivered an urgent pitch: you are already badly behind in agentic commerce, agents will buy things, and you are not ready for it. The interviewer reacts: directionally right, but three years too early, it was 2023. He then gives the measurement from Stripe's side: they began developing it with a series of announcements at Sessions in May, but the volume really went vertical only from September. The interviewer's thesis: the key unblocking factor was computer use, because now you can ask an agent to buy anything and it will figure it out itself, which makes it a universal backward-compatibility layer.

Instacart: an account for the person or for the agent

The interviewer calls Muse one of the first products where people genuinely delegate purchases fully to agents. This is a step change: agentic shopping was much discussed, and now it is really happening. Wang adds three layers of value. First: less friction, no need to click through sites and do research. Second: criteria are applied automatically. Example: "I need something by Wednesday", and the agent looks only for what can arrive by Wednesday. Wang notes that relying on keyword search now feels primitive, since that is a very sensible constraint. The third layer, which they especially expect over the long term: personalisation.

On personalisation: Muse has a whole context about a person and every aspect of their life, so it can take account of things about family, colleagues, friends and help reconcile all of it, to pick the exact gift for your best friend in the context of a friend group, knowing what that group is. Historically recommendation systems were fragmented and isolated and did not see a whole picture of a person; Muse has the chance to deliver a far higher level of personalisation.

The economic question is stated plainly: the Instacart account is for him or for Muse. Wang unfolds it into a more interesting question: whether the total addressable market of all commerce is suddenly growing. Because now it is not only people: agents will in effect have accounts. Whoever does not account for this will come to regret it over time. Asked why he had to set up an Instacart account, the interviewer mentions that Stripe solves this in Stripe Projects for developers. Wang, in turn, thanks them for the Stripe integration: "The Stripe integration, by the way, is great. You guys did a great job with us on Muse."

Voice: why it is hard

The interviewer notes that people have been trying to talk to computers for a very long time: Siri shipped in 2014-2015, there is Voice Mode and various AI products several years old already. Why is it so hard. Zuckerberg names several reasons. First: you actually want the agent to go and do something, so even if it understands, there is the question of how it works in practice.

The second reason is conversation dynamics. People do not talk like that: "I give a prompt, now I wait, you answer". Conversation goes back and forth. Zuckerberg describes a concrete problem: if the pause is three seconds, the agent may decide you have finished and want to speak; if it has started speaking and you interrupt, it may decide this is a new question. A lot in conversational dynamics has to be got right.

The third reason is bandwidth asymmetry. For you, voice is a very high-bandwidth way to convey information, but for consuming information, when someone is talking to you, it is very inefficient compared with the visual. The conclusion: the ideal interface is probably not strictly voice-to-voice. At the same time voice output from you is very useful.

Wang adds that there is already clearly an explosion of voice input. In coding it is visible: most researchers whisper into the keyboard, because it is the highest-bandwidth, and more often the lowest-friction, way to unload your thoughts. The most extreme example: how students use AI today. A very common modality is to speak out fully for 10-15 minutes in a stream of consciousness, the whole chain of thoughts, and then let AI clean it up and produce it, let AI do all the organisation, pull it together and structure it.

On agents and voice: Wang sees agents as a big step change, because the mental model is simply different. If it is only an information thing you talk to, you will clearly hit situations where it seems useless. When it can do something for you, that is much closer to the experience of talking to a person. There are interesting uncanny valleys in interacting with AI that, while they could not act, felt strange; now that agents exist, this is changing. Muse is embodied and therefore feels like something working for you.

Hardware and embodiment

The interviewer notes that these are real glasses. Zuckerberg answers that many people need them even without a display or with a small display: the camera and audio alone are already a useful enough product. Then they began shipping glasses with a display that are not a wide-angle full AR, but even a small display is quite powerful, and interesting interactions will start to appear there, fully multimodal: voice in and out, the glasses see what you see and can draw pixels on top.

Separately, the idea that glasses may have no app store at all: it will simply be multimodal generation that creates what you need right in your field of view. Zuckerberg adds that something similar will probably also power many of the apps they build.

On the ecosystem: Muse Gadgets opens up an ecosystem for third-party devices and form factors. After the release, several phone makers started talking to them: the idea is to integrate Muse Gadgets into the phone. One of the recently shipped gadgets is HomeLink, roughly the size of a coin (a nickel), with USB. Through it Muse connects to everything in the home. Wang calls it one of the most fun things: Muse connects to Sonos and to all the lamps. When he needs to get ready for sleep, he simply asks to dim the lights throughout the house.

Compute and money: Prometheus and Hyperion

Wang gives a formula for measuring a lab's strength: the compute power of a lab can be measured by the size of its largest single supercomputer for large training runs. Specifics: the Prometheus cluster in Ohio scales beyond a gigawatt. The Hyperion cluster under construction in Louisiana scales up to 5 gigawatts. Both were planned before MSL: the construction of a data-centre site from scratch and the plans for them go back to the LLaMA programme, planning began, in Zuckerberg's words, probably in 2023 or 2024, and now they are coming online, with Louisiana coming online over a couple of years.

Zuckerberg stresses that Meta is a full-stack company: it builds its own data centres, its own chips, its own low-level infrastructure. Models are the next component, probably the most important for the next wave of technology: building very good models and delivering them to people.

The financial arithmetic is stated plainly. Operating a gigawatt cluster for a year is on the order of 15 billion dollars. In terms of capital expenditure on construction, it is closer to 50-60 billion dollars, because it is paid up front. And that is just one cluster for one part of training, there are other parts too. What is more, research is not just the run itself: roughly an equal share goes on testing the whole ladder up to large scale for every change. And then there is serving. The bottom line: tens of billions of dollars to train a frontier model now.

How the buildout is financed

The interviewer asks directly: how to finance the buildout, everyone is nervous, Google has gone to negative free cash flow, Meta has not yet, but capital expenditure is always there. Zuckerberg's answer: there will be many ways to create value. For Muse the goal is broad use, because at scale there will be valuable network effects as people are able to interact with each other. Partners will want to integrate the more people use it. The best way to achieve this is a model that is accessible and free for people.

Then the revenue lines are listed. Subscriptions are the big one: the base version being free does not mean everything is economically viable to include in the free tier. If tomorrow you want to use Muse Wonder in Muse, you have to pay for a subscription. Then commerce: the authors believe Muse will help people earn and save, and over time ways will appear to take a small share of what a person saves or earns, and that will become a very large revenue line. To the remark about whether Muse will cost people negative money (that is, save more than it costs), Zuckerberg points to existing sites that find unused subscriptions and cancel them.

Separately, the idea of a "business in a box" for small businesses: Muse helps an owner administer every part of their business, and that can be very valuable and profitable both for the owner and for Meta. Another layer is the API itself, which will be profitable. A large part of the reason to start Meta Enterprise Platform and bring in CJ Desai is the view that traditionally Meta was not an enterprise company, but the main difficulty is building leading models, and they are very general; if you invest significantly in an enterprise group that delivers this to enterprises in the form they need, many enterprises will want to use Muse Wonder, because it is excellent and accessible.

All of this together, plus the fact that Meta is a very profitable business with good credit, able to underwrite large investments, puts the lab in a good position to build the compute it needs and lead for decades. Asked whether it will be self-financing in a relatively short time, Zuckerberg answers tersely that the answer better be yes.

Then the accounting detail and the backstop. Zuckerberg explains that if you build a data centre there is capital expenditure you can depreciate over a useful life of many years, with arguments about what that is. But a significant part of AI costs are researcher salaries, and those cannot be capitalised, even though model training also has a useful life of many years. And this is a financial advantage of building your own data centres versus other companies that pay cloud companies an annual rate: at some point the GPUs are fully depreciated, but Meta will probably keep using them for a long time, because the cost of even an H100 now is higher than a year ago, if compute remains constrained.

On the backstop and risk: the numbers are large, but Zuckerberg does not think the risk is that high. Meta has a hedge in how compute can be used. First, if AI works out overall, many others will want all their compute and are willing to pay more than 15 billion per gigawatt. For the venture not to pay off, it is not enough for them to fail: AI overall must fail. And even then there are other ways to use compute productively. Then the power-law logic: Wang says Zuckerberg has deeply absorbed power laws, and the point is not that failure is acceptable but how large the win is when you win. Zuckerberg agrees: the main thing is how big the win is if it happens.

Demand that does not exist yet, and the Meta playbook

Wang describes a management moment that is rare for its honesty: one of the hardest decisions was about what demand for the product they expect. The product does not exist yet, and data centres to serve Muse have to be built years in advance. At some point some people proposed being a little more conservative, but the decision was: no, we are doing this, we will make it work, burning the bridges. Wang admits that as an entrepreneur he felt enormous risk, and says the risk is only growing.

Then the built-in playbook. Zuckerberg states it plainly: build something, take it to a billion people, then scale revenue, and meanwhile grow revenue mainly by scaling what is already at scale and growing. Wang confirms that this is the whole Meta playbook. Separately, the general point about the front-loading of expense: because so much advance planning is needed, R&D line items at any moment are overlapping costs for several generations of models at once, they require long work and deliver a huge payoff at the end.

Then using models to accelerate the core. Meta has a core business of hundreds of billions of dollars, growing very fast; if it is grown a few percent faster using models for acceleration, that alone finances a large part of the buildout. The interviewer calls Meta's core business the "eighth wonder of the world", and both agree it is a very good business growing at an astonishing rate.

OpenRouter

The transcript contains an ad insert: Meta models, including Muse Wonder, are available on OpenRouter, and Stripe has agreed to acquire OpenRouter, the leading model gateway. OpenRouter already routes more than 10 trillion tokens a day and takes on much of the complexity of working with different models through a single API: experimentation, competitive prices, automatic failover around provider outages. Later the interviewer clarifies the status of the deal: agreed to acquire, but not yet closed.

Wang describes an opportunity: frontier models, Muse Wonder versus Fable or Astra, are starting to focus in noticeably different directions, and when these models collapse onto a single set of benchmarks, that erases all the interesting texture of the differences between models. One opportunity for OpenRouter is to help developers understand those subtle differences so they can make well-founded decisions about model choice. The flip side: you have all these developers using models, and they are excellent people to ask about interesting textural differences that are more than a tweet long; this leads to a "sommelier" version of understanding models and their relative strengths and weaknesses.

The interviewer sums up: so OpenRouter's view that the future will hold a diversity of models optimised for different tasks is confirmed by your experience. The answer: yes, they focus on different things, and they care far more about personal and professional agents than other big labs do. And that is good: there should be differentiation in the economy, that is in a sense the essence of capitalism.

Why there is no winner-take-all

The interviewer recalls the 2023 storyline: one lab will start accumulating an advantage, trigger recursive self-improvement and pull away from everyone, but this did not happen, several players are ahead, open models are holding up well. Why. Wang names several mechanisms.

The first is execution. Any frontier lab has periods of strong and weak execution, and across all labs this averages out over time. To the provocation that the leaders simply failed at execution, he answers: if everyone in front had executed perfectly, they would probably have stayed in front, that is one factor. The second is the power of distillation, which clearly helps the open ecosystem. Wang calls it an anti-hegemonic force: nobody planned for distillation to work so well, but it turned out that many capabilities transfer naturally through one form of distillation or another. Distillation is good for the whole ecosystem: it makes it possible to pack very strong capabilities into ever-smaller models, which makes it possible to serve them at far greater scale. Hence: small models today are more powerful than the biggest models of a year or two ago.

The third is the ecosystem's desire. Wang suggests that many in the ecosystem want a future with no single company at the end of capitalism. Companies, not just individuals, are putting money behind this. This is one of the reasons many were glad to join Meta and MSL: to work where the focus is on delivering the technology to billions of people, which is not necessarily the stated goal of all other frontier labs. Here too is the safety argument: in Zuckerberg's view, a good future is achieved through diversity and a system of checks and balances, with many different things available to people, rather than a couple of good models in the hands of a couple of companies; the latter he considers a riskier configuration. He also calls himself a fan of NVIDIA's recent moves to support the open ecosystem.

Tone: authenticity and bold posts

The final block is about public communication. The interviewer asks what Zuckerberg thinks of Alex's bold posts on Twitter. The answer: most seem funny to him, but the main view is this: the most important thing in communication is to be authentic. When you try to be authentic, some things come out not as you would like, and that trade is better than being overly scripted and inauthentic.

Wang describes the mechanic. His mental template is often: this is funny, it is interesting what would think, and good memes come out of it. He admits he has felt the brutal metabolism of the internet: the internet digests memes very fast, and you have to stay at the front edge of meme production, it is easy to fall behind. The conversation closes with the joke that this is possibly a good summary of the whole conversation: it is easy to fall behind, you have to stay at the frontier. Zuckerberg confirms: stay technical and stay at the frontier. And the final detail: a keg of Guinness brought to the interviewer from Dublin.

Source criticism

This is a conversation about a product and model launch. Promotional in form, it is in places candid in substance, but that is exactly why the line between a claim and a measured result has to be kept explicit. Below is what is what.

What we do not know

For the breakdown to be honest, below is a list of what cannot be established from the transcript. These are not gaps of the analysis but gaps of the source.

• The real benchmark numbers: percentages, versions, run conditions on DeepSWE, SWE-Marathon, Terminal-Bench 4 are not named.
• How much faster Muse is actually growing: there is a reference to third-party reports, but the reports are not quoted and no metrics are given.
• How active memory is actually built technically: it is said that compaction is inside the model, but the mechanism is not revealed.
• The exact team size by stage and the full headcount of MSL: only orders of magnitude are named (under 200, closer to 100, tens per stage).
• The exact deal structure: only the roughly 49 percent stake in the data business is named, the rest is unclear.
• The economics of serving: it is said that Wonder is more expensive, but unit costs and prices are not named.
• Voice mode: not fully rolled out at the time of the conversation, it cannot be assessed.
• Whether the real benefit of memory is durable: the authors themselves say there is a lot of room for improvement and their evals show it.

What a builder or operator should take from this

Below is a list of decisions that transfer to other products, regardless of how accurate the self-reported numbers are.

Table of every number and name

Below is every number and every named entity mentioned in the conversation, with context. All numbers are from the transcript.

Final verdict

Alexandr Wang and Mark Zuckerberg describe Muse and Muse Wonder not as a product and a model but as one thing: product decisions here are training decisions, which means the distinction between the lab and the product team dissolves. The conversation is valuable precisely for that thesis and for honest economic arithmetic: a gigawatt costs about 15 billion a year, a cluster costs 50-60 billion, training one frontier model is tens of billions, and all of it was planned before the models were seen. The bet is sensible, because the backstop is the scarcity of compute itself and the growth of Meta's core, and the logic of the power law justifies the size. But almost all claims of superiority are self-reported, the interviewer has a direct interest in the agentic-commerce thesis, and the form of the conversation is promotional. So the right conclusion is this: what is worth taking is not the numbers but the decisions - memory in the model, the absence of a model selector, the API and the free tier as distribution, agents as accounts, browser use as a compatibility layer, and the publication of internal evals as positioning. And to keep in mind what the authors themselves once called an exact formula: you extract the function, and the risk, like the win, is distributed by the power law.