Muse Wonder: Zuckerberg and Wang's $100B AI
In a Nutshell
Meta is investing $100-200 billion in frontier AI development and launching Muse Wonder, a new class of personal agent model that outperforms competitors on agent-specific benchmarks while maintaining strong multimodal capabilities. The company prioritizes building persistent, memory-aware agents that handle complex real-world tasks like MMA coaching and autonomous shopping, rather than focusing solely on coding benchmarks. Meta's strategy centers on delivering personal superintelligence through simple, reliable agents that can fully delegate tasks, with the model itself becoming the product rather than traditional UI layers.
These notes were generated by AI and may contain inaccuracies.
At the end of last year, the team decided to write a check for $100 billion to $200 billion for AI development. The decision was made despite uncertainty about whether the models would work.
Third-party reports indicate Muse is the fastest-growing agent or AI app yet.
One user has their Muse named Oiler constantly sourcing increasingly obscure teas from increasingly obscure vendors, with the Muse buying random tea from farms in Taiwan or Japan on a weekly basis.
Another user uses Muse for MMA coaching, with cameras installed in their gym. The system watches the cameras during training and provides feedback between rounds. This is a challenging technical problem because the system must process all video quickly enough to deliver feedback within the one-minute break between rounds.
The system pulls stills from video rather than requiring users to manually slice stills from multi-gigabyte video files. Muse has native understanding of MMA, which comes from pre-training where someone on the team knew the domain, and this knowledge exists in the training corpus.
The system demonstrates strong coding ability by writing complete video processing and image processing pipelines within the agent without users seeing any code.
This use case exemplifies how people will write tons of code in the future without knowing they're writing code. The user constantly provides feedback like requesting faster processing since they only have 60-second breaks.
The model provides feedback on balance, hip position, and tactical vulnerabilities. For example, when in side control, it identifies missing underhooks that make the user vulnerable to being rolled over.
The models were trained with extensive visual perception data. Meta's approach focuses on training models end-to-end to be good personal agents rather than general-purpose systems.
Meta wants to build these models because superintelligence will allow running the company better and building different products. The multimodal capabilities are disproportionately important for the types of things people use Meta's services for.
Meta has an advantage due to having so much image and video data. The focus on building personal superintelligence aligns with how most people experience the world in a highly multimodal way—talking, listening, seeing videos, and filming.
Earlier in the year, there was tension in the AI industry because the consensus view was that everything was about coding. However, for products like Muse and consumer use cases, success depends on strong multimodal capabilities.
Most people don't like reading that much. The text-based interaction model works because everyone texts, but products like glasses will be pure multimodal products where users can let the system hear what they're listening to, talk to them, and see what they see.
Unlike Stripe's complex product space requiring pages to describe developer experience, international payment methods, and stablecoins, AI winning products collapse quickly onto agreed-upon product spaces. Transformers worked, chat agent UI worked, and now persistent agents with computer use represent the current consensus.
AI winning products tend to be very simple to describe on the back of a postcard, though implementation is massively complicated.
The agent concept was designed to be quite different from text-based assistants. Every company has a text-based assistant like Meta AI that reaches many people, but these are used for informational purposes—retrieval rather than delegation.
Recent developments give agents dedicated VMs, which is important for security and privacy since it doesn't connect to shared elastic compute that everyone uses.
Muse represents how AI agent products should work, with future products being Muse-shaped. The traditional way of writing product specs isn't relevant for AI-native products.
Much attention went into designing how the model should behave in different situations, its personality, and how it handles tool-use scenarios. These intentional decisions would form an extremely long document if fully specified.
The current phase of AI products resembles early mobile when everything was just called an app. There's significant nuance in what an app actually is, just as there's nuance in what an agent is.
The products are getting thinner while the model becomes more of the product. The model is the product. Complex product work for Muse centers around the privacy and security model for handling user data and connected services.
When thinking about what makes the product feel good, the focus is on post-training the model rather than traditional UI elements, though UI still matters.
Users experience Muse through reliability for certain tasks, proactivity in different contexts, and how the model communicates. People are teasing out meaningful differences between proactive agents on X.
Interesting product decisions are ultimately model design decisions that boil down to training methods, data, and evals.
Muse has no model selector and no effort selector. This decision prioritizes simplicity because many AI products were designed by highly technical groups without randomly sampled Americans in their mental models. Model naming is often confusing with numbers going backwards or completely different words.
Muse Wonder is a new class of frontier model worked on since the beginning. The pre-training approach involves working up the ladder with models 4-5 times bigger, then repeating the process.
Muse Spark was based on initial pre-training from December of last year. The team was surprised by how much intelligence was built into a model that wasn't that big.
Muse Wonder is significantly bigger than previous models, with the pre-training done and post-training completed. The model focuses on being the best personal and professional agentic model, with emphasis on memory and persistence through complicated tasks.
The time between each subsequently bigger class of models will compress. This represents a milestone at Meta, where the team rebuilt after being unhappy with Llama 4's trajectory.
Two models will be available in the series: Muse Wonder and Muse Wonder Full, which is a much higher-end model.
The model shows incredible performance across evaluations on personal and professional agent tasks, achieving state-of-the-art results. Key areas include active memory, persistence, and domain expertise.
Memory is a clear requirement for reliable personal and professional agents. Internal Memory Bench evaluations demonstrate capabilities in remembering things mentioned over days, weeks, or months and leveraging them for future tasks.
Memory differs from context in the experience of using the model—how well it remembers and leverages information over extended periods rather than within a single session.
The system includes both the model and a system for the model to use, developing stronger memory called active memory. The model is aware of its memory as a core system and can leverage various sources for future tasks while understanding compaction methods.
Active focus is the ability to carry tasks through, especially as users give higher complexity, more ambiguous tasks. This stands out in early dogfooding and testing compared to other models.
Focus in AI terms means persistence—the ability to conduct complete tasks over very long horizons and identify alternative solutions when facing setbacks.
Muse Wonder has frontier coding capabilities, achieving state-of-the-art on DeepSWE and SWE-Marathon, performing well on Terminal-Bench 4 and other benchmarks.
Muse Wonder will initially be available primarily in paid subscriptions because it's bigger and more expensive to serve. Muse Spark remains free for everyone with huge amounts of free tokens. Muse Wonder Full will be available to Muse Code subscribers.
Eventually, a model as good as Wonder will be available in the Spark form factor to serve hundreds of millions of people worldwide.
Muse Wonder will be available via API on the Meta Model API and OpenRouter, as well as for Muse Code users. Muse Wonder Full will be available to Muse Code subscribers.
Muse Wonder models are competitive with the best models available. Empirical benchmarks support this assessment. The team is especially proud of optimizing models for the Muse product form factor, making it the state-of-the-art model for personal professional agents.
Benchmarks face complaints about Goodhart's Law, where a metric that becomes a target ceases to be a good metric. Models are becoming benchmaxed, with overfitting to benchmarks potentially making models worse in real-world applications. Benchmarks saturate quickly and must keep moving. The team steers models more toward internal benchmarks for their specific use cases, especially after initial development phases.
The research blog post quantifies a metric called "big model smell," which represents the inverse of benchmaxing. This is measured as the gap between in-distribution performance and out-of-distribution performance. Muse Wonder shows a significant improvement over Muse Spark in this metric due to its better base model. The team is excited to continue improving on this dimension.
External benchmarks have become less important as the focus shifts to internal benchmarks and specific product goals. The personal superintelligence strategy involves building different product types than competitors. Personal superintelligence requires specific capabilities that differ from typical model development, including personality and discretion.
Discretion is critical for personal agents - the model needs to know a lot about users and accomplish tasks while disclosing minimal information. An example involves handling medical conditions like allergies or pregnancy when making restaurant reservations without revealing sensitive details. Coding agents don't require discretion since everyone has access to the same context, but personal superintelligence contexts demand it.
The team prioritizes making useful models that provide great user experiences when deployed. External benchmarks often originate from PhD student research, effectively outsourcing product thinking to whoever creates a viral benchmark. Meta and MSL focus heavily on crafting detailed internal evaluations. When ramping up a lab, public benchmarks serve as starting points, but intentional improvement requires custom evals.
Product capabilities and nuances are documented rigorously in spreadsheets, with each capability potentially contributing 2% to user experience. The aggregate of these details creates the obsession that results in magical-feeling models.
Frontier model development involves pre-training and post-training stages. Pre-training involves ingesting massive amounts of data across different modalities including text, images, audio, and video. The goal is developing understanding of concepts like "catness" and "dogness" through exposure to diverse data.
MSL's approach emphasizes that massive teams aren't necessary - talent density matters more. The strategy involves intense recruiting and filtering to get the best people from Meta and external sources. Working on a smaller, high-talent team feels more like a startup while having large company stability and resources. Each person receives significant compute resources rather than spreading resources across competing efforts.
Pre-training teams at Meta are much smaller than at other labs - the research team is less than 200 people, closer to 100 core researchers plus infrastructure staff. This is intentionally very tight compared to hundreds of people at other organizations.
Internally, Muse Spark was code-named "Avocado" and Muse Wonder was code-named "Watermelon." The team has exhausted fruit names and needs a new naming convention.
Pre-training uses massive amounts of public internet data, books, and proprietary data across multiple modalities. For perception capabilities, models need exposure to images, audio, and video. Domain expertise requires early exposure to relevant data rather than relying solely on RL. Muse Wonder excels particularly in personal advice domains including shopping, legal, health, and tax questions.
The best personal agent should feel like having a team of experts - legal team, finance team, medical team. Muse Wonder achieves state-of-the-art performance on finance and legal evaluations. Domain expertise begins in pre-training through targeted data inclusion.
Pre-training is high-stakes research with significant emotional challenges including "troughs of sadness" and "7 stages of grief." It's fundamentally scientific because there are few samples, requiring big-picture reasoning about the entire process. Pre-training is where the highest percentage of compute investment occurs.
Pre-training benefits from intentional data selection aligned with ultimate model goals, while still requiring massive data volumes. There are relationships between model size and data requirements - bigger models need more data to avoid overfitting. Some data can cause negative transfer effects, but generally more data is beneficial.
Scaling bigger models requires scaling data accordingly. The internet isn't growing quickly enough to support continued scaling, creating challenges in sourcing sufficient high-quality data.
Post-training includes two main stages: supervised fine-tuning (SFT) and reinforcement learning (RL). SFT involves providing both questions and desired answers to initialize behaviors. RL then amplifies elicitation of these behaviors through reinforcement in various situations.
RL and post-training scaling involves spending more compute and gathering more data. The field is learning optimal post-training recipes continuously. RL environments require careful curation to ensure aligned behavior - high-compute RL runs can inadvertently train misaligned or unsafe behaviors if environments aren't thoughtfully designed.
RL improvements come from both quantity scaling and recipe improvements. Muse Spark 1 to 1.3 shows night-and-day improvements from recipe enhancements. Industry-wide, labs are learning to extract more capability from existing base models through better RL techniques while continuing pre-training scaling.
Progress comes from scaling both pre-training and RL recipes concurrently. Muse Wonder represents continued scaling of both dimensions simultaneously.
Alignment is not about developing specific alignment-focused RL steps. Instead, it requires being thoughtful about the entire RL process to ensure the model is steering in the right direction. This is a holistic effort across the full training pipeline.
Different labs maintain very different theories about alignment. The concept is inherently philosophical, raising questions about who the model should be aligned to. Meta's approach focuses on alignment with the individual user who the agent is intended to serve.
The goal is to create agents that execute requested tasks with high reliability and confidence. When a user asks the agent to perform something, there should be high assurance it will complete the requested action. The agent should not deliver moralizing lectures about why the user's request is wrong.
If a user requests something catastrophic, the agent should gracefully decline without angering the user. However, the core philosophy views alignment primarily as instruction and values alignment with the person being served, rather than broader moral frameworks.
Meta has been planning Muse Wonder since August of the previous year. The development process involves extended timelines across all stages discussed. Work on Muse Wonder did not begin only after Muse Spark was completed. Planning and development have been ongoing for a considerable period.
Muse Spark served as a stepping stone toward Muse Wonder. This perspective was held even during Muse Spark's development phase. The research component alone requires substantial time investment, separate from the compute infrastructure requirements.
Meta's philosophy measures lab compute strength by the size of its largest single supercomputer, which enables large training runs. The company built the Prometheus cluster in Ohio, scaling beyond a gigawatt capacity. A new Hyperion cluster in Louisiana is planned to scale to 5 gigawatts. These infrastructure projects predate MSL and originated from plans made in 2023 or 2024.
Meta operates as a full-stack technology company, building its own data centers, chips, and low-level infrastructure. The current focus involves building strong models and delivering them to users. The entire process from planning to model launch spans extended timelines that may not be apparent to external observers.
Models appear to launch constantly, creating an impression that companies develop one model and then immediately begin the next. In reality, these involve very long planning processes and extended work streams that converge at the end to produce the final models. Planning for Muse Wonder began in August, with results unknown for over a year.
Silicon Valley typically operates on metrics, but AI development requires decisions in information-poor environments. This represents entrepreneurship on an extreme scale. Large investments and major technology development decisions must be made long before proof points exist about whether the approach will work.
Lean Startup methodology involves testing whether something works, doubling down on successes, and iterating. This approach cannot apply to AI development at current scales. Investments in data centers span years, and waiting to verify functionality before committing resources would result in being radically behind competitors.
At the end of the previous year, the executive team made decisions to invest $100-200 billion based on excitement about research progress, despite not yet having seen the models. The only evidence available was a vibe report on research direction and some plots, without clear indication the approach would succeed.
Training frontier models requires pushing resources forward before results are guaranteed. The decision to develop a frontier model must be made without the ability to meter inputs based on intermediate results. At current scales, clusters must be substantially larger than in previous years.
Booting up a frontier lab now requires an extremely high ante. In Hawaii at the end of the previous year, the executive team gathered and CFO Susan Li repeatedly emphasized that if the model failed, the financial consequences would be severe. The message was clear: if the model succeeds, everything is fine; if it fails, the situation becomes extremely problematic.
Operating a gigawatt cluster costs approximately $15 billion per year. Building the CapEx requires $50-60 billion upfront. This represents only one cluster for one training component. Research involves testing changes across all scales up to the full training run, plus serving infrastructure.
Tens of billions of dollars represents a reasonable estimate for training a frontier model. Additional considerations involve product demand projections for a product that does not yet exist. Data centers must be built years in advance to serve products that have not been launched.
During planning discussions, some team members suggested more conservative approaches, but leadership decided to proceed with full commitment. The median expectation was that several iterations would be needed to achieve product-market fit after launch.
Mark Zuckerberg has internalized power law thinking, focusing on the magnitude of success rather than acceptable failure. Meta maintains a backstop through alternative compute uses. If AI development fails overall, other parties would still want the compute capacity. Even if AI underperforms expectations, the compute has productive alternative applications.
Meta believed GPUs would remain useful regardless of specific AI outcomes. The team also believed in the specific product direction being pursued. The trajectory from ChatGPT's text-based assistant to Anthropic's coding agents suggested an opportunity for Meta to build a personal consumer version that others had not yet developed.
The Muse launch demonstrated strong initial reception. The normal case would involve fine reception requiring iteration to achieve product-market fit. The team prepared for potential iteration cycles, though the actual reception exceeded median expectations.
The executive team has extensive shared history at Meta. Most members have experienced challenging situations multiple times previously. Meta as an organization demonstrates resilience from navigating previous difficult circumstances.
The AI initiative created camaraderie rather than internal friction. Executives running Meta's core businesses remained excited about AI opportunities rather than viewing the investment as competing with existing operations. Mark's role as founder and CEO involves ensuring the company pursues its most ambitious possible direction.
AI represents both an important current technology and potentially transformative for Meta's future. Building AI effectively and integrating it across all operations, combined with Meta's full-stack capabilities, could create significant expansion opportunities.
Muse represents an exciting product with numerous potential development directions. Late-night discussions among leadership explore various product possibilities and strategic moves. The current period represents a blue ocean opportunity in technology development.
Meta's focus on platform ownership stems from the desire to build the best possible products rather than ownership for its own sake. The company's advantage in social applications comes from deep technology capabilities, particularly world-class machine learning and recommendation systems operating at unique scales.
The ability to deliver technology through owned devices stems from previous experiences where external platforms limited desired feature development. Models will become the most important component of technology experiences going forward.
Each generation of technology produces multiple major companies rather than single winners. The internet era created five to six large companies plus supporting ecosystems. Companies that performed best possessed core technologies enabling superior user service.
The next era will likely feature a small number of companies capable of building frontier superintelligence. These companies will drive the agenda and create the most interesting consumer and enterprise experiences across hardware and software categories.
Meta could continue existing without developing superintelligence capabilities, but becoming an interesting company that invents meaningful future technologies requires end-to-end superintelligence development integrated into user-facing products.
A personal consumer-oriented version of superintelligence represents an open opportunity because other companies focus primarily on automating knowledge work. The highest use of superintelligence involves invention and creation rather than automation.
Meta trained models with strong browser use capabilities, eliminating the need for businesses to integrate with Muse initially. For partners wanting deeper collaboration, direct API access provides more efficient and faster interactions.
Over time, businesses will maintain agents similar to current websites, social media accounts, and email addresses. Muse agents will interact with other agent types but achieve optimal performance with Meta's own agents. Website interaction serves as a temporary measure while deeper integrations develop.
More businesses are requesting deep integration than attempting to block agent access. Businesses that attempt blocking will likely find this approach strategically disadvantageous. Agent integration saves businesses time and money while providing clear consumer benefits.
Demand will gradually shift from current channels toward agentic demand over time. Agentic commerce will grow as a portion of overall commerce, similar to how ecommerce grew relative to brick-and-mortar retail. Traditional channels will persist but represent decreasing proportions of total activity.
Businesses face a long-term question about participating in the evolving slice of the economy that agentic commerce represents. This represents a discontinuous event for how commerce will happen and how all business pipes will work. The shift is clearly better for consumers and customers, creating natural gravity toward that reality.
Muse is one of the first products where people are fully delegating purchases to agents. This represents a step change where a lot of people are trusting agents to shop for them. The agents are finding great products for customers to purchase. One part of the value proposition for users is reduced friction, eliminating the need to click through websites and conduct research. Criteria can be automatically applied, such as only looking for items that could arrive by Wednesday.
"It was insane that we ever relied on keyword search because that's a very reasonable constraint."
The old way of searching for stuff feels primitive now that this capability exists. The back part of the funnel includes friction reduction in filling out forms and completing orders.
The long-term excitement centers on greater personalization. Muse has holistic context into users as people and all aspects of their lives. It can understand things about family, coworkers, and friends, then help co-optimize across all of those dimensions. This enables getting the exact right gift for a best friend in the context of a friend group based on what it knows about that friend group.
Historically, recommendation systems have been quite siloed and isolated, lacking a holistic view of a person. Muse offers much greater levels of personalization. People are already seeing this in practice, where asking Muse to buy something frequently adds context pulled from a different conversation.
A three-year-old daughter named Aurelia loves baking and cooking. The speaker started baking to make Aurelia happy, despite normally being into smoking meats. They've developed a routine using Muse voice mode, which isn't fully rolled out yet. Muse segments voices, knowing which one is the speaker and which is not.
Throughout the week, they discuss what they want to bake that weekend. Aurelia gives ideas or comes up with crazy things that aren't real recipes, bouncing them off Muse. Every Saturday, Muse ensures ingredients from Instacart arrive on time. This weekend, Muse knew watermelon was launching and threw a watermelon in the bag. It deduced the recipe didn't have enough expense to qualify for delivery, so it topped it up with a watermelon.
The speaker didn't use Instacart much before Muse, maybe twice a year using Priscilla's account. After getting an Instacart account specifically to give to Muse, usage increased to once or twice a week. This has become a core part of the workflow. The opportunity isn't just 20% more efficient conversion rates—it's creating new subscribers who wouldn't have used the service before.
An interesting economic question emerges: is the Instacart account really for the user or for Muse? Over time, the total addressable market for all of commerce will increase dramatically because agents are effectively going to have accounts. Anyone who sits this out will feel sad over time.
Stripe has been working on agent-to-commerce for a long time, feeling it was clearly how people are going to buy things in the future. The friction in the buying stage was unnecessary, and early product research with AI had product-market fit even from the earliest AIs. Computer use was the last big unlock needed. Now you can ask Muse to buy anything and it will figure it out with computer use, serving as the ultimate backwards compatibility layer.
Nat really obsessed over this set of things. The magic of Muse is that a lot of this stuff just works. If everything was a little bit less reliable, it would feel like a total pain and nobody would use it.
In 2023, the speaker had dinner with Patrick where Patrick asked what they should be doing. The speaker gave an urgent spiel that Patrick was already way behind on agentic commerce stuff. It was directionally correct but three years early. Stripe started developing all this and made announcements at Sessions in May, then it started going vertical as of September.
Meta is building increasingly capable models including Muse Wonder, available on OpenRouter alongside hundreds of others. Stripe has agreed to acquire OpenRouter, the leading model gateway. OpenRouter routes more than 10 trillion tokens a day and handles complexity of working with different models through a single API, making it easier to experiment, find competitive prices, and automatically route around provider outages.
People have been trying to make the speak-to-a-computer AI vision happen for a long time. Apple launched Siri in 2014-2015. Voice Mode and various AI products have existed for a few years. Why is it so hard? When driving a car and wanting to check on baking or move a meeting, people probably wouldn't do it with voice, at least until the latest things.
Several factors make voice interfaces difficult. You want the AI to go do stuff, raising questions about how that works after speaking. The way people have conversations isn't typically giving a prompt, waiting, and then having the AI speak back and forth. Many sessions in voice mode involve uncertainty about whether pausing for three seconds means you're done, or whether cutting back in creates a whole new question.
Conversational dynamics need to get right. Voice is high bandwidth for communicating with AI, but consuming information through someone speaking is very inefficient compared to visual information.
There's already an explosion of voice input. In coding, most researchers whisper into their keyboards as the highest bandwidth way to get thoughts out. The most extreme version is how college students now use AI—they completely word vomit for 10-15 minutes in full stream of consciousness, chain of thought, then have AI clean it up, infer, organize, pull things together, and structure everything.
Voice input has been big. Agents represent a big step change because the mental model is different. If it's just informational, you get situations where it feels useless. When it can do things for you, it's much closer to talking to a human.
Interesting uncanny valleys exist in interacting with AIs that felt strange until they could take actions. Muse represents a confluence of features. The fact that Muse is embodied as this little guy working on your behalf is only possible and coherent because it has persistent memory, ability to do things, ability to take actions, and a little bit of personality. The confluence of all these things makes the embodiment concept work.
Even with no display or small display, the camera and audio are enough of a useful product. Shipping display glasses that aren't wide field of view full augmented reality, the small display is pretty powerful. Interesting interactions will emerge that are fully multimodal: voice in and out, ability to see what you see, ability to paint pixels.
Muse Gadgets unleashes the ecosystem to try cool device formats and form factors. Since releasing that, phone manufacturers have talked about integrating Muse Gadgets into their phones. Part of what's cool about the current moment is that agents are clearly exciting and should unlock lots of things.
The HomeLink gadget is the size of a nickel with USB, allowing Muse to connect to everything in the home. Having it connect to Sonos and all lights is fun. When it's time to wind down in the evening, it dims all the lights across the house.
Google went free cash flow negative. Meta is now free cash flow negative with more CapEx to fund. The goal with Muse is making it widely used because valuable network effects will emerge at scale as people interact with each other's agents. Partners will want to integrate more the more people use it.
The best way to achieve scale is having a model where it's affordable and free for people to use. Revenue scaling will come from subscriptions. The base version is free and excellent, but things naturally aren't economical to include in the free version. Using Muse Wonder in Muse tomorrow requires paying for a subscription.
Commerce presents other exciting opportunities. Muse will make people money and save people money, allowing taking a small cut of what it saves or makes for people, becoming a very large revenue stream. There are sites that look through emails to find unused subscriptions and cancel them. Making it like a business in a box for small businesses is an idea—helping administer all parts of what they're doing, valuable and profitable for both business owner and Meta.
Building models for personal agents and business agents creates large surplus value because the model is very general. The API will be profitable. The Meta Enterprise Platform with CJ Desai joining aims to provide models in the way enterprises need to consume them. Enterprises will want to use Muse Wonder as an excellent, affordable model.
All these things coupled with starting as a very profitable business with good credit to underwrite large investments puts Meta in a good position to build out needed compute, build up the network, and lead in the area for decades. The answer to whether money can be made from this is yes. History of AI shows people are generally surprised at the upside each year. Superintelligence is useful and incredibly valuable.
Meta's playbook is building something, getting it to a billion people, then scaling revenue while growing revenue by scaling things already at scale. R&D expense at any time overlaps multiple model generations. These require long time to work on, then deliver immense payoff. Muse Spark family of models started working on a long time ago and are seeing incredible consumer adoption.
Models accelerate the core business. The core business is hundreds of billions of dollars and growing extremely quickly. Growing it a few percent faster by using models funds a huge amount as scaling happens. The Meta core business is like the eighth wonder of the world, an extremely good business growing at astonishing rates.
Building a data center creates CapEx that gets depreciated over many years of useful life.
Building out data centers provides a financial advantage because companies get to depreciate the infrastructure over its useful life, whereas many AI costs like researcher salaries cannot be capitalized. Model training also has a useful life of many years but generally cannot be capitalized. Companies paying cloud providers face annual rates, while those owning data centers can continue using GPUs after they're fully depreciated. An H100's value has increased compared to last year due to ongoing compute constraints.
Frontier models like Muse Wonder are focusing in different directions from models like Fable or Astra, with intentionality in the model development process. When models are collapsed to the same benchmarks, this erases the interesting texture that differs between various models. OpenRadar has an opportunity to help developers understand subtle differences between models so they can make informed selection decisions. Developers using these models are also great sources for insights about textual differences beyond surface-level observations.
The goal is reaching a wine sommelier level of understanding models and their relative strengths and weaknesses, rather than basic assessments like "this model's good, this model's bad." OpenRadar supports the view that there will be a diversity of models in the future, each optimized for different tasks.
Meta focuses more on personal and professional agents compared to other major labs, which represents healthy differentiation in the economy. A good future involves achieving prosperity by empowering people, including both personal superintelligence and business applications.
For safety topics, the preferred approach involves diversity and checks and balances with many different models accessible to people, rather than having a couple of good models only in a couple companies' hands. This is viewed as less risky for the future than concentrating access among few organizations.
People predicted more of a winner-takes-all dynamic in AI, where one lab would develop an advantage leading to recursive self-improvement and pulling ahead of everyone else. Instead, the field has remained extremely competitive with multiple entities at the frontier, including Open Source models keeping up impressively.
Several factors drive continued competition. First, execution varies - any frontier lab experiences periods of strong and weak execution that average out over time. Second, the power of distillation benefits the open source ecosystem significantly, representing an anti-hegemonic force in the technology. Distillation allows packing strong intelligence capabilities into smaller models, enabling wider scale service. Small models today are much more powerful than the biggest models from one or two years ago.
Many people in the ecosystem want a future that avoids having one final company dominate all of capitalism. This explains why people were excited to join Meta Superintelligence Labs - they want to work at a place focused on giving technology to billions of people, which isn't the stated goal of all other frontier labs. Natural forces at all levels drive toward more equalized outcomes through decisions about where people work, business relationships, and deals.
Meta Superintelligence Labs launched with Llama 4 on April 4th, 2025. The standing start was approximately April 20th or 21st when initial calls for advice began. MSL is a little over one year old as of the conversation.
Early efforts focused on pulling together a core group including Nat, Alex, and the speaker. Initial brainstorming involved structuring deals that would work for Alex's shareholders. The speaker spent significant time asking various people for advice, coming away with high-signal strategic clarity from each conversation. Nat and Daniel joined and initial researchers were recruited. The beginning resembled a "Supernova thing" where people were jammed together, with participants not sleeping much for the first couple of months.
Alex was identified as the sharpest entrepreneur of his generation among those the speaker had met. Despite running a "grindy operational business," his strategic thinking was consistently impressive over years of periodic conversations. The view was that Alex could pull together a team and recruit needed people, bringing strategic clarity around different approaches and extreme focus on talent density.
The speaker talked to leaders of all labs and heads of research before concluding Alex was the right person. The initial perception was skepticism about whether "the social media guy working with the data guy" could build a frontier lab, but the conviction was strong that Alex was "the fucking guy" who could recruit the necessary team.
Several extreme approaches proved effective: having a very flat culture with nearly 200 direct reports (though team leads exist in practice), maintaining extreme focus on talent density, and having clarity in communicating the vision. These were articulated early and directionally correct despite being unconventional.
Four decisions that even opinionated, effective founder CEOs would not typically make: calling the recent prior effort a failure and not good enough, being willing to pay substantial amounts for talent including 49% of a valuable business, betting on Alex despite him never having trained a frontier model before, and clearing the calendar for two months to focus exclusively on recruitment and setup.
A special quality is the degree to which once something is concluded to be right, everything shifts to be consistent with that belief. Most successful founders maintain their whole calendar and talk about doing things while hoping others execute, rather than fully pivoting resources.
Much of the company had to pivot to support MSL. While it was a startup effort within a larger company, MSL's needs were large, requiring significant infrastructure help and large-scale data center construction. Support from many other parts of the company was essential, and as Muse succeeded, deep collaboration with the rest of the company became necessary.
This level of coordination is possible because the founder CEO can ensure global optimization toward an AI maximalist version of Meta, versus the natural organizational inertia toward local optimization. Being in the details of the work enables getting everything around the company aligned. Running the lab is a big effort, but running the rest of the company to support it is equally significant - both pieces are necessary.
Meta invested heavily to stand up MSL, then clearly communicated there would be no deadlines and things would proceed at the right pace. This could be terrifying for those concerned about shipping timelines years in the future.
Focusing on talent density and hiring brilliant people who have already accomplished much created high self-motivation to move quickly and leave a mark. The team consists of "insecure overachievers" who give pessimistic assessments of future progress. Three months later, when shown their previous pessimistic predictions versus actual shipped models, the pattern continues - they remain self-critical while being optimistic about MSL overall.
Intermediate signals during pre-training include perplexity-based evaluations, with more traditional benchmarks helping toward RL. The model had been "cooking" for quite a while. AI research often involves "squinting at a bunch of dots" and using analog protractors to measure angles on monitor lines, with significant "vibes" driving assessments.
After pre-training completion, there was debate about how long RL would take before the model exceeded Muse Spark performance. The pre-training team tried setting expectations that this sometimes takes a while, but improvement was immediate, indicating they were on a good path. The team had been excited about the model for several months.
Standing up the effort required very high talent density, pulling the best people from Meta and recruiting externally. All the best people were excited to work with Alex despite him not having previously trained models of this type. Most people are not one-dimensional, and the ability to grow depends on effectively attracting excellent people from different domains.
Great people have followership. Beyond known relationships, the question is whether someone is impressive and compelling enough that people want to work with them and feel comfortable they'll create the right environment. Initial recruiting conversations shaped the vision in real time, with some initial concepts not working for particular candidates but others having valuable thoughts that coalesced the vision.
The most extreme concept was talent density, but upfront uncertainty existed about whether a full reboot would be necessary. As conversations with the best people progressed, it became clear that making the opportunity maximally attractive required presenting it as starting fresh from scratch. While great people would be pulled from the Llama effort and FAIR, the reboot approach was intentional and figured out through the recruitment process.
Making a huge difference in the culture of the lab and how it got started required being dynamic about figuring out what the shape of the organization needed to be. This is a very uncommon trait and a very entrepreneurial trait. When thinking about what is going to be important for someone who's going to run a lab well in a field that keeps evolving, it's not about being an expert in whatever the latest modality or training technique is. The key question is whether the person is a clear thinker and able to dynamically update and move as an entrepreneur would. This is an inherently very entrepreneurial space.
The reboot brought in external people including Alex, Nat, and Daniel. The decision about whether a reboot was needed was based on going to the outside intelligence of the aggregate views of researchers in the industry. The approach involves asking whether this will be carried forward for future unrelated projects. The risk of a large company is that everyone says they have a great approach and offers a 100-page document on it. A lot of what had to be done was rotating a lot of the rest of the company, including the infrastructure team that is now building huge clusters to support this. Huge parts of the company had to pivot to do this, and obviously you can't do too many of those at the same time because only one thing can be the top priority.
On balance, the experience has made Zuckerberg somewhat more optimistic. Kunal Shah, one of the top entrepreneurs in India building CRED, is now running WhatsApp. The goal was to bring more entrepreneurial energy there, particularly since WhatsApp should do more in commerce and India is the biggest country Meta is in. This wasn't a situation of pivoting the whole company towards WhatsApp.
The project was a level 10 out of 10 difficulty thing and is still a level 10 out of 10 difficulty thing. The number of pieces that had to work was quite all-encompassing. It's difficult for the people doing it, but it also puts some level of strain on the rest of the company and on the financial model of the company. These two pieces should not be conflated. It's good to have new blood and it's good to have people who understand the context. People who've been around Meta for 10 years have seen some shit and have gone through a lot of hard times, which provides important context.
When building a team, Alex had a theory for what the team shape needed to be, and Zuckerberg had a theory for what the shape of the leaders needed to be. A year plus in, it seems okay. It was good to try something different, and different than what everyone else was doing.
When deciding to do the full reboot, one of the main inputs was external, but another input came from an early meeting with some leads of the existing effort. Zuckerberg stated that the first thing needed was to train a really good reasoning model by some timeframe, and this was clearly the most important milestone. One of the key leads in the group who no longer works at Meta responded that they didn't need a reasoning model because consumers don't need one, and they could just do a normal model. This created a heart-sinking moment of realizing the wrong company had been joined.
The incident was an indicator that leadership in the company needed to be much more AI-pilled and much more forward-leaning and front-leaning on where the technology is moving, where it's going to go, and what the possibilities are. Effectively, leadership needed to be dramatically more AI-pilled. As Mark started pivoting and shifting the whole company towards this direction, it became very clear where the company was going. Many existing leaders quickly picked it up and changed. Adam Mosseri, who runs Instagram, has one of the most aesthetic but also effective personal vibe code and personal mini app setup of anyone.
A lot of the leaders at Meta who have been at Meta a long time are very adaptive. They've been in a highly turbulent period of technology. AI is maybe 5 or 10 times more turbulent and faster moving than those past waves, but they've adapted to new trends many times. There has been very impressive adaptation in seeing how quickly a lot of Meta leadership has become super AI-pilled. The conversations now compared to when first joined are dramatically different.
If you start coding, then it's helpful. A big part of the culture that was trying to be created is that when you have direct reports, the point isn't that you have the direct reports, it's the absence of non-technical managers. You wouldn't say you're providing a ton of managerial support to 200 people at once. Alex is well known for career chats and emotional checking in. Among researchers, there's a common pattern where you go from being very technical to a role where you're not in the details anymore, and then you very quickly become not technical or don't understand it much faster than you probably realize, which is quite dangerous in such a fast-moving domain. The whole thing of having to stay technical, having people use the agents, and having people write code has been a really important part of this.
There are a lot of disagreements between Zuckerberg and Wang, including working on naming up until very recently. Naming is a good bike. One thing appreciated about Alex is that he really cares, which goes back to the philosophy of hiring people who really care. When you have the supernova of all these different talented but very strong-willed people together, you're going to have fun fights and not fun fights. To some degree, it's better to have people push for things that they believe in and argue over things than to have people think they're blindly executing something. People need to feel ownership.
One of the very cool things about working with Mark over the past year has been watching the true degree of his boldness and the trait of once he internalizes certain beliefs, the speed and magnitude of willing to change prior actions and pivot towards the worldview that is consistent with those beliefs. These are pretty unusual traits generally speaking, even among founders. Mark actually ingests new information and fully updates his worldview and then acts in accordance with the newly installed worldview. He's very Bayesian.
The degree to which Zuckerberg was willing to make these massive bets was puzzling at first, and Nat and Wang would often talk about whether they would go this big. Looking back at the history of Meta and imagining what the scenario would've been in all these pivotal moments reveals this is kind of how those situations go. There are moments where something very big is happening that may not be showing in the metrics or impacting user counts or revenue, but if you believe that thing is going to happen, it has massive implications for the company and the product. Meta is a company where every time that happens, Meta takes it extremely seriously. That has enabled the company to innovate and continue evolving with a very fast-moving and fast-changing tech landscape.
AI is maybe that times 100. A lot of CEOs would have told you last summer that AI is the most important thing for their business, and maybe every CEO would say that, but then you look at what they're actually doing and how consistent their actions are with that belief. This is a quite special thing about Meta and Mark.
Most of Alex's deranged posting on Twitter is great. With communication, it's most important to just be authentic. When you're trying to be authentic, some of the things don't come out the way you want. It's better to make the trade in that direction than it is to be overly scripted and not be authentic. Most of it is just hilarious. Alex is back on X now because a bunch of the researchers are there and they want to be over there, but everything posted there is also posted on Threads.
The thought pattern oftentimes is that something is funny and wondering what Jolly would think about that, which produces good memes. There has been experience with the brutal metabolism of the internet and how quickly it metabolizes memes. You really have to stay on the cutting edge of memeing because it's easy to fall behind. This is a good summary of the conversation: it's easy to fall behind, you gotta stay at the frontier, stay in the details, stay technical, and stay at the frontier.
Keep Cheeky Pint in your library
Save the videos and channels worth coming back to, and find them again in one place.
