Alexandr Wang

John Collison names four moves he thinks even an opinionated founder would not make. Mark Zuckerberg treated the recent prior AI effort as not good enough. Standing up Meta Superintelligence Labs is the reboot, and people can read that between the lines. He was willing to pay a lot of money to bring Alexandr Wang over, including 49 percent of what Collison calls a very good grindy business. Wang had worked in AI and had never trained a frontier model, so the hire was a bet on potential rather than on having done that exact job. Zuckerberg also cleared his calendar and did only this for two months.

Wang says that once Zuckerberg decides something is right, he shifts the work to match that be

Mark Zuckerberg says he and Alexandr Wang had talked about once every few months for years, and he was always impressed by how Wang thought about strategy. Wang was running what Zuckerberg calls a very operational data business. Zuckerberg still thought he was the sharpest entrepreneur of his generation among the people he had met.

When the current program's trajectory was not good enough, people asked how the social media guy and the data guy were going to build a frontier lab. After talking to the leaders of the labs and the heads of research, Zuckerberg's conviction was that Wang was the person, and that they could recruit the team they needed. He wanted Wang's strategic clarity, the ext

Alexandr Wang says a reason people were excited to join Meta and Meta Superintelligence Labs is that they wanted to work somewhere focused on giving this technology to billions of people. He says that is not the stated goal of every other frontier lab. People are choosing the world they want, in where they work and who they do business with, and those choices push away from a single winner.

The lab's standing start is recent. Llama 4 launched on April 4, and they date that to 2025. Around April 20 or 21, the calls for advice on what to do began. John Collison first says the lab is a little over two years old. The room corrects him: it is a little over one year old.

John Collison recalls the 2023 story that one AI lab would get ahead, start recursive self-improvement, and pull away from everyone else. He says that has not happened. Several labs are still at the frontier, Muse Wonder is arriving as Meta's frontier-class model, and open-source models have kept up.

Alexandr Wang gives three reasons. Execution moves around. Every frontier lab has strong stretches and weak ones, and those average out. If a lab in the lead had executed perfectly the whole time, he thinks it would probably have stayed ahead. Distillation is the second reason. He calls it an anti-hegemonic force in the technology, not a plan anyone chose. It packs strong capability into smalle

Alexandr Wang says frontier AI is the opposite of a lean startup. A lean startup gets something working and only then doubles down. You cannot do that here. The bets have to be made before there is proof the thing works, or you end up radically behind. He calls it entrepreneurship on hard mode, and he gives Mark Zuckerberg credit for taking those decisions.

Zuckerberg retells the end of last year. They were excited about the progress and had not seen the models yet. The check they were going to write was, as he put it, like $100 billion or $200 billion, and he asked Alex if he was sure it was going to work. John Collison describes the same constraint as pushing the chips all in before the c

Alexandr Wang's broad rule for AI is to scale everything, and to scale it together. Scaling one thing in isolation does not work. Pushing a single dimension gets much less than scaling many things at once. Scaling laws only hold when the other factors move with them.

Muse Spark's point releases show the split. From version 1 to 1.1, 1.2, and 1.3, Meta did not change the pre-train. It kept doing more reinforcement learning on the same base. Early versions of a new class have a lower ratio of reinforcement learning to pre-training, and later versions raise that ratio. Muse Spark and Muse Wonder are different pre-trains. In reinforcement learning, less data is sometimes more, which is the oppo

Alexandr Wang says Muse Wonder is state of the art, by Meta's own measurement, on a slate of personal and professional agent tasks. Active memory is one place the lab invested disproportionately. It is not the context window, and it is not how much the model can hold while it works on a single task. It is how well the model remembers things a person mentioned over days, weeks, or months, and then uses those memories on later tasks.

Wang says the team is extremely fond of OpenClaw, and that OpenClaw's memory gets buggy after a while because it compacts a long-term memory against a short-term working memory. Meta designed both the model and a system for the model to use, and calls the result

Alexandr Wang says Muse Wonder is a new class of model at the frontier, and the lab has been working toward it since the start. A pre-train does not jump to the biggest cluster. The team climbs a ladder, training models about four or five times bigger, then repeats that. Muse Spark came from the first pre-train, done in December at the end of last year. That model is not that big, and the team still built a lot of capability into it. Wonder is the next class and significantly bigger. The target is the best personal and professional agentic model, with very good memory and the ability to persist through complicated tasks.

The time between each bigger class should now start to compress, at Me

Alexandr Wang heard a key lead on Meta's existing effort reject a reasoning model because the lead did not think consumers needed one and said a normal model would do. In one of Wang's first meetings after he joined, he told the leads that the first milestone was to train a really good reasoning model by some timeframe, since that was clearly the most important step. The answer hit as a heart-sinking moment. He felt he had joined the wrong company and would have to take 10 steps back.

The decision to do a full reboot had more than one input. A main input was external, and the meeting was another. Wang took the comment as proof that Meta needed leadership much more committed to how fast the

Alexandr Wang says an agent is a big step change because the mental model is different from a chatbot that only shares information. When the system can only talk, the exchange gets into situations that feel useless. When it can do things, the experience moves much closer to talking to a person. Talking to AI sat in an uncanny valley until the systems could take actions. Agents are what make that strangeness recede.

Muse can show up as a small embodied character that feels like it is working for someone, and that picture only holds because the pieces arrive together. It has persistent memory, it can do things and take actions, and it has a bit of personality. Wang says that mix is what makes

Alexandr Wang says Muse Wonder is the state of the art model on a lot of the finance evaluations and a lot of the legal evaluations, and the domain skill starts in pre-training, where the team looks for a lot of domain data to include.

Mark Zuckerberg named what people care about: shopping, personal advice, personal legal questions, personal health questions, and personal tax questions. Muse Wonder is very good on personal advice questions, and no score was given. The same skills can be reinforced to some degree, but pre-training still needs a large amount of data. Health has a lot of data out there, and pre-training is where the model is meant to see it.

The philosophy is that the best pe

Mark Zuckerberg says he does not know how big a pre-training team is, and it does not need to be that big. Other labs are probably quite big. The models are doing well because the people Meta is betting on have a lot of resources. Compute stays with the key people instead of being spread across a lot of competing efforts, with a lot of compute per researcher so they can do the work.

Wonder was the next pre-train and climbed a ladder through a tiering system, and then came Spark. Avocado was the internal codename for the Muse Spark line, and Watermelon was the codename after it. The fruit names are used up. Durian is the only fruit that is technically larger, and it is heavier and spikier.

He is the founder of Instant Pot. More than 5.5 million people follow Instant Pot related accounts on Facebook alone. People even make decals and knit sweaters for their Instant Pot.

He thought the problem wasn’t that it was going to be popular on the market, but that stores had limited coverage. The best way to let people know of his product was through online businesses.