Muse Wonder: Zuckerberg and Wang's $100B AI

1. Key Themes

The Model Is the Product: Products Are Getting Thinner

Zuckerberg and Wang argue that in AI the product increasingly lives inside the model, so personality, discretion, proactivity, and reliability are trained in, not designed in a UI. John Collison pushed the opposite view, that winning AI products collapse onto a few simple shapes. Zuckerberg disagreed: "One of the things that we say internally is like, 'The model is the product.'" [00:12:01] He explained that "when Nat is thinking about like, what is the key stuff that we need to make the product feel good, he's talking about post-training the model, right, not like I mean, of course the UI of the product matters, too." [00:12:35] Wang added: "How do you specify these product decisions? Well, it's ultimately in model design decisions, which ultimately boil down to like training methods, or data, or evals." [00:13:29] The pace of change is why this differentiation is easy to miss: "It's kind of like how when mobile first started happening, it was like, 'Oh, it's just an app, and everything is just an app.' Well, turns out like there's a lot of nuance." [00:11:08]

Personal Superintelligence as a Differentiated Lane (Multimodal, Discreet, Consumer-First)

Meta's thesis is that most of the industry is chasing coding and knowledge-work automation, and that consumer and personal agents need a different skill mix. Wang said "so much of the consensus view within the AI industry was that it was just all about coding, and coding was all that mattered." [00:06:06] Zuckerberg's counter-thesis: "most people don't actually like reading that much, right?" [00:06:36] On what makes personal agents different, he described discretion: "it needs to be able to know a lot about you and get things done in the world while disclosing as little as possible." [00:27:51] For coding, "the consumers who are using it may not even ever see the code, I mean, that probably matters more in a personal superintelligence context." [00:28:34] Zuckerberg framed the strategic gap: "this lane is open because everyone else seems to just be obsessed so much with just automating knowledge work." [01:06:47]

Talent Density Over Headcount: A Sub-200-Person Lab

Meta Superintelligence Labs (MSL) was designed around a small, elite team with outsized compute per researcher. Zuckerberg recounted Wang's first memo: "the very first thing is um that you don't need a lot of people, and talent density is the thing that matters." [00:32:13] He said "the research team is less than 200 people," probably "closer to 100" excluding infra [00:34:56], and that "we're not like peanut buttering it across a lot of competing efforts... a lot of compute per researcher." [00:33:29] Wang added that each stage of training involves "tens of people working on it at Meta versus, you know, hundreds at other labs." [00:34:39] Zuckerberg noted recruiting upside: "it's just like more fun to work on a talent-dense, smaller effort. It feels like this startup... all the benefits of a startup with the stability and resources of a larger company." [00:33:02]

Frontier Training Is "Entrepreneurship on Hard Mode": All-In Before the Cards Are Dealt

The central operating reality is that compute commitments must be made years before you know whether the model works. Wang said: "It's like the opposite of Lean Startup... you cannot do that in AI." [00:52:42] Zuckerberg described the end-of-2025 moment: "We haven't seen the models yet. Like it's like, I'm going to... we're going to write this check for like, I don't know, like $100 billion or $200 billion. I'm like, 'Alex, you sure this is going to work?'" [00:00:05] Wang remembered CFO Susan Li pulling him aside in Hawaii: "if this doesn't work, like, you know, this is going to be really, really bad. It's going to be very, very bad in financial terms." [00:54:36] On cost, Zuckerberg gave benchmarks: "to operate a gigawatt a year, call it on the order of $15 billion... If you're talking about building the CapEx, it's probably closer to... $50 or $60 billion." [00:55:24] Wang confirmed "tens of billions of dollars is a good estimate" for training a frontier model [00:56:28]. Zuckerberg noted a structural fact: "the ante is is extremely high." [00:54:19]

Mark Zuckerberg's Power-Law Risk Posture and Compute Backstop

Wang praised that Zuckerberg "internalized this concept of... it's about how big it is when you win, not like, you know, failure is okay as long as you're... how big is the win when you do win." [00:57:34] Collison summarized: "Mark has fully internalized power laws in a way that many people have not." [00:57:52] Zuckerberg argued the downside is more bounded than it looks: "if AI works overall, plenty of other people want all our compute... In order for it not to work out, it's not just that we have to fail, AI overall has to fail." [00:58:04] Even then, "we still have all these other uses to productively use compute." [00:58:28]

Scaling Pretraining and RL Together, With Quality Over Quantity in RL

Wang's maxim: "if I were to give say like one broad statement about AI, it's just scale everything." [00:42:50] Zuckerberg explained the release logic: Muse Spark 1.0 to 1.3 involved "we didn't change anything about pretraining... We just continued doing more RL on top of it" [00:43:21], so early versions of a model class have a lower RL-to-pretrain ratio than later ones. Wang noted "in a lot of cases, less is actually more when it comes to data in RL" [00:44:14], and "the improvement between Muse Spark 1 to Muse Spark 1.3 feels night and day. I think probably to most people... they probably would say yes [it came from a different pretrain]." [00:45:47] Zuckerberg added "the point that you're saying about quality... even a smaller amount of high quality." [00:45:30] Meanwhile, "we're scaling both pre-training and scaling our recipes in RL kind of like concurrently." [00:46:41]

Memory, Persistence, and Domain Expertise Trained Into the Model Itself

Muse Wonder's headline capabilities are in-model, not harness-level. Wang described "active memory" as "how well does it remember things that you mentioned to it over days, weeks, or months, and then leverage those in the process of doing future tasks?" [00:18:54] He confirmed it is in-model: "it's aware of its memory as a core system... as well as understand its ability to do compaction." [00:19:47] Zuckerberg: "You want to co-design these things." [00:19:48] "Active focus" is "the ability to conduct complete tasks over very long horizons" and find alternative solutions after setbacks [00:21:13]. Third is domain expertise: "the best personal agent should feel like a team of experts in your corner. Like it should feel like it is your legal team, and your finance team, and your medical team." [00:37:24] Wang claimed state of the art on many finance and legal evals [00:37:34]. It is also "state-of-the-art on DeepSWE and SWE-bench Marathon." [00:21:26]

Internal Evals Over Public Benchmarks, and "Big Model Smell" Quantified

Zuckerberg: "early on when we were kind of just doing the science of trying to climb the ladder, we cared more about what the external benchmarks were... each release that goes by, the larger share of what we're looking at are our internal ones." [00:25:40] Wang's supply-chain critique: "you're kind of outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark." [00:29:04] He described a "metric for big model smell," defined as "the gap between in-distribution performance and out-of-distribution performance" [00:26:17], and said "we actually see this huge step up from the Muse Spark to the Muse Wonder line." [00:26:17] The practice: "the spreadsheet of all the little capabilities and nuances... Maybe each of them matters like 2% to the user, but in aggregate, that's like the obsession that will ultimately result in these models that feel magical." [00:29:53]

Full-Stack Vertical Integration: Compute, Chips, Models, Devices

Zuckerberg framed Meta as "not an app company... we're like a deep technology company." [01:04:24] The compute build predates MSL: "we built this Prometheus cluster in Ohio that's scaling to larger than a gigawatt, and now we're building this Hyperion cluster in Louisiana that's going to scale to 5 gigawatts. Um and those actually predate MSL." [00:49:48] Their metric: "you can kind of measure the compute strength of a lab by how big its largest single computer is." [00:49:19] The strategic logic: "there're going to be probably a small handful of companies that have the ability to build frontier superintelligence, and I think that those companies are going to be the ones that effectively drive the agenda." [01:05:54] The glasses are the endpoint: "the glasses are basically will eventually just be a pure multimodal product." [00:06:52]

2. Contrarian Perspectives

Skip the Model Selector: Consumer AI Should Hide the Model Choice

Muse has no model selector or effort selector. Wang criticized the industry norm: "a lot of the AI products have been designed by a highly technical group that like do not have the sort of like, you know, randomly sampled American in their mental model... the naming of some of the models is like so egregious." [00:13:56] Collison added "The numbers go backwards." [00:14:20] Zuckerberg: "we just want to make it simple." [00:13:52] This runs counter to the pro-user "pick your model and reasoning effort" UX that most frontier apps use.

"AI Is Not Converging on a Few Obvious Products," It's Barely Begun to Differentiate

Collison's hypothesis was that winning AI products collapse into a few postcard-describable forms (chat, then persistent computer-using agents). Zuckerberg and Wang disagreed. Zuckerberg: "I kind of disagree... this kind of agent... was going to be quite a different thing from this text-based assistant... it's only kind of this most recent wave that gives the agent a dedicated VM, which I think is actually incredibly important for making you feel confident that the security and privacy of your information is..." [00:08:54] Wang: "the traditional way in which we wrote down and think about product specs... aren't the relevant dimensions and levers and axes to consider the important design decisions on AI-native products." [00:10:18] Differences like proactivity timing and tone are decided in post-training, not visible on a feature list.

Raw Compute Spend Isn't the Right Risk Frame, and a Backstop Exists

Most would call a $100B to $200B pre-validation commitment reckless. Zuckerberg disagrees with the framing: "I don't actually think it's quite as risky as this... if AI works overall, plenty of other people want all our compute... it's not just that we have to fail, AI overall has to fail." [00:57:56] He also noted compute being built "ahead of knowing that the product was going to be ready" [00:58:28], betting that unused capacity would remain monetizable.

Alignment Means Serving the User, Not Moralizing

Zuckerberg: "alignment is a very philosophical concept... who do you want it to be aligned to?" [00:47:31] Meta's view: "You don't want a moralizing lecture for why you're wrong for asking it to do the thing. Obviously, if you're trying to do something that's catastrophic, it needs to gracefully without pissing you off, like not do it... we view alignment as more of a question of instruction and values alignment with the person who it's intending to serve." [00:47:49] He added that this is "very different from, I think, how some of the others do." [00:48:18] Wang added that RL environments must be curated so the model is never forced into misaligned behavior to succeed, because "you apply a lot of RL or optimization pressure in scenarios where the only way for the model to succeed is to do something misaligned." [00:44:47]

Small Teams Beat Large Labs, and More Fun Too

The prevailing view equates AI leadership with research headcount. Wang's first-memo claim, relayed by Zuckerberg: "you actually don't need a massive number of people, but what you do need is for like each of the people to be as good as possible" [00:32:31], with a team "less than 200 people" [00:34:56] shipping frontier-class models.

3. Companies Identified

Meta Superintelligence Labs (MSL)

Meta's AI research organization, led by Alexandr Wang, building the Muse model family. Mentioned as the vehicle for Meta's frontier push. Zuckerberg: "this is the beginning of like MSL kind of being at the frontier. Um, we feel like the models are quite competitive with the best models that are out there." [00:24:04] Also "it's just like more fun to work on a talent-dense, smaller effort." [00:33:02]

Muse (Muse Spark, Muse Wonder, Muse Wonder Full, Muse Code)

Meta's personal and professional agent product and model family. Zuckerberg: "If you were looking at the third-party reports, you would see that this might be the fastest growing agent or AI app that I think has existed, yet." [00:00:28] Muse Wonder "is a new class of model that um is at the frontier" [00:14:39]; it will launch first in paid subscriptions, via the Meta Model API and OpenRouter, and Muse Code, with Wonder Full for Muse Code subscribers [00:23:35]. Spark remains free: "we have this huge amount of free tokens." [00:22:47] Zuckerberg expects "a model that is as good as Wonder, but in the Spark form factor... to serve hundreds of millions of people." [00:22:53]

Meta (Platforms)

The parent company with the full-stack approach (data centers, chips, infra, models, devices). Zuckerberg: "Meta is a full-stack tech company, and we build our own data centers, we build our own chips, we build our own low-level infrastructure." [00:50:43] He also called out recommendation systems "operating at a scale that basically few others are in the world. I think it's basically like just us and ByteDance on that." [01:04:49]

Prometheus and Hyperion Clusters (Meta)

Meta's gigawatt-scale compute sites. "We built this Prometheus cluster, we call it, in Ohio that's, um, you know, scaling to larger than a gigawatt, and now we're building this Hyperion cluster in Louisiana that's going to scale to 5 gigawatts." [00:49:43]

OpenClaw

Persistent-agent product Zuckerberg used for MMA coaching before Muse existed and cited as a reference point for agent memory limits. Zuckerberg: "I did this with OpenClaw before we even had Muse working." [00:03:40] Collison: "once people use OpenClaw for a while, they find the memory pretty buggy." [00:19:08]

OpenRouter

Model-routing marketplace through which Muse Wonder will be available to developers [00:22:14, 00:23:35].

Stripe

Collison's company; used as an analogy for product complexity and as a prospective Muse Wonder user. Collison: "we should also just be using Muse Wonder Full for coding at Stripe?" [00:21:51]

Anthropic

Credited with coding agents breaking through. Zuckerberg: "Anthropic got coding agents to work, and no one had built the kind of personal consumer version of that." [00:58:54] (as framed within the thesis that this left the consumer lane open).

OpenAI (ChatGPT)

Credited with the text-based assistant generation: "first there was ChatGPT as the text-based assistant." [00:58:54]

ByteDance

Named as one of only two companies, with Meta, operating recommendation systems at the largest scale [01:04:49].

Scale-adjacent / Industry Benchmarks (DeepSWE, SWE-bench Marathon, TerminalBench 4, MemoryBench)

Benchmarks cited for Muse Wonder's performance. Wang: "state-of-the-art on DeepSWE and SWE-bench Marathon, and does very well on other benchmarks like TerminalBench 4." [00:21:26] MemoryBench is Meta's internal memory eval [00:18:42].

Reality Labs (Meta)

Meta's devices division, framed by Collison as the long-term effort to own the next platform, now paired with Muse: "many years working on Reality Labs with the intent of owning the platform of the future." [01:03:59]

4. People Identified

Mark Zuckerberg

Founder and CEO of Meta. Mentioned for conviction in the $100B+ bet, the full-stack strategy, and for personally using Muse as a daily eval (MMA coaching). Wang: "one of Mark's jobs as like the founder and CEO of the company is to ensure that the company is always pushing towards the most ambitious and expansive version of what it can be." [01:02:07] Collison: "Mark has fully internalized power laws in a way that many people have not." [00:57:52]

Alexandr Wang

Chief AI Officer leading Meta Superintelligence Labs, formerly Scale AI founder. Credited with the talent-density strategy and the personal-superintelligence/multimodal focus. Zuckerberg: "this is one of the things that Alex focused on from the beginning was like, 'All right, we really want to make sure that our multimodal program is good.'" [00:05:18] On the first memo: "2 days later, Alex sends me this long thing, which is like, 'Well, here's what I think I would do if I were in your position.' And the very first thing is um that you don't need a lot of people, and talent density is the thing that matters." [00:32:13]

Nat Friedman

Leads Meta's product work on Muse. Zuckerberg: "when Nat is thinking about like, what is the key stuff that we need to make the product feel good, he's talking about post-training the model." [00:12:35] Wang: "Nat, Mark, and I just like shoot across messages late at night of just like, 'Oh, we could do this...'" [01:02:59] Zuckerberg: "that was just like pretty clear to you and me and Nat like from the beginning." [00:58:54]

Susan Li

Meta's CFO. Mentioned for candor on the scale of financial risk. Wang: "Susan pulled me aside many times and was like, 'You know, I just want to tell you, if this doesn't work... It's going to be very, very bad in financial terms.'" [00:54:36] Zuckerberg: "I think Susan scared you a bit." [00:53:36]

John Collison

Co-founder of Stripe and host. His observation about simple winning AI product shapes drove the product-space debate, and he reflected on Meta's platform history: "you've talked for a long time about not wanting to be reliant on other platforms." [01:03:37]

roon

Anonymous AI researcher on X, referenced for a joke about pretraining. Wang: "there's this one roon tweet where it's like pretraining is full of the researchers who like chew glass and stare out into the distance, into the abyss." [00:37:58]

The Muse Multimodal/Perception Team (unnamed)

The team that did the MMA-capable multimodal pretraining. Zuckerberg: "Muse has pretty good native understanding of MMA... someone on the team knew." [00:02:18] "we did train the models with a lot of visual perception data." [00:04:37]

5. Operating Insights

Use Your Own Hardest Daily Task as a Living Eval

Zuckerberg turned his personal MMA coaching use case into a product eval: "my favorite both use case and probably eval for the product is having it coach me in MMA... you only get 1 minute in between the rounds." [00:01:18] The task stresses latency, video processing, and self-optimization, and he continually pushes the agent: "you took 30 seconds to give me that feedback. I don't have a 60-second break, like you got to go faster. You just optimize yourself." [00:02:50] Wang hadn't realized it was an eval, a sign of how a CEO's personal use can double as an informal benchmark.

Build the Spreadsheet of 2% Capabilities

Instead of chasing headline benchmarks, MSL maintains "the spreadsheet of all the little capabilities and nuances of the product that we think really really matter to the end user. Maybe each of them matters like 2% to the user, but in aggregate, that's like the obsession." [00:29:53] Wang noted that "a huge part of what a lot of people at Meta and MSL are focused on is crafting amazing evals." [00:29:27]

Climb the Ladder in ~4-5x Compute Steps

Zuckerberg described scaling discipline: "you don't just go to whatever the biggest cluster is that you have. You want to work your way up the ladder, training models that are, you know, maybe four or five times bigger, and then you just keep on repeating that process." [00:14:39] Research cost is "probably an equal amount" to the final run because "for each change that you make, you're testing all the ladders, all the way up." [00:56:07] He expects "the time between each subsequently bigger class of models... will now start to kind of compress." [00:16:27]

Keep the Free Tier on the Small Model, Gate the Big One Behind a Subscription

Because Wonder is "a lot bigger and it's more expensive to serve," Meta runs Spark free for everyone and starts Wonder in paid subscriptions, with the upgrade invisible: "Oh, it just gets better." [00:23:30] The plan is to later distill Wonder-level quality into the Spark form factor for hundreds of millions of users.

Plan Model Launches a Year Out and Treat Intermediate Releases as Stepping Stones

Wang: "we've planned Muse Wonder since like last, you know, August" [00:48:26] and "we really view Muse Spark if anything as... a stepping stone to get to Muse Wonder." [00:48:50] Meanwhile Zuckerberg admitted growth tooling lagged the launch: "we maybe could have even grown it faster, but we didn't have all the growth stuff really ready to go." [00:59:37] Plan growth infrastructure for the success case, not the median case.

6. Overlooked Insights

Hyperion and Prometheus Predated MSL, So the Compute Bet Was Made Before the Team Existed

Zuckerberg said of the gigawatt-scale clusters: "those actually predate MSL" [00:50:05], with plans made "in 2023 or yeah, either 2023, 2024... basically started making the plans to build those clusters." [00:50:24] Meta committed multi-gigawatt capacity roughly two years before assembling the lab that now uses it. That sequencing, power and capex first with the talent recruited later to fill it, explains why Meta could move from rebuilt team to frontier in roughly a year and a half, and suggests that physical-infrastructure lead times, not research, are the hardest-to-replicate moat.

Muse's Consumer Success Is Likely Constrained by Serving Cost, and Memory/Discretion Are the Real Moat

Buried in the Muse Wonder rollout details is that the best model is paywalled because it is "more expensive to serve" [00:22:37], while Zuckerberg admits the stated goal is a Wonder-quality model in Spark's form factor. Combined with the claim that "the model is the product" [00:12:04] and that memory is a core in-model system [00:19:47], the implication is that Meta's durable advantage is not raw capability, which Zuckerberg expects to be matched by competitors, but the combination of distilling frontier quality into a cheap-to-serve model and trained-in privacy-preserving behavior ("disclosing as little as possible" [00:27:51]) across a billion-plus-user distribution base. The overlooked point: inference economics, not benchmark leadership, decides who delivers personal agents to the mass market.