Alexander Wang argues conviction, vision and ambition — not model progress — now decide who builds the future.
A year ago Mark wrote a memo on personal superintelligence: all billions of people get a superintelligence tailored to them, knowing their context and expanding their agency. Wang frames the goal as agency expansion, inside a broad ecosystem rather than a totalitarian view of AIs controlling the world.
Wang says updates will continue on the MuseSpark line, with bigger models on the way that he expects to be much more competitive with the very best models available today. A harness for developers and agentic developers launches soon, and Meta is also working on open source models.
Agentic looping. Wang sees astronomical opportunity in systems that spend a thousand or a million times more tokens on a continuous feedback loop. Companies are large feedback loops with humans on every edge; internally at Meta, with the right eval or metric, a swarm of agents beat 100 engineers.
Today the easiest way is to use Open Code, with onboarding on the website, and soon Meta will have a harness of its own. Ultimately Wang wants great models that plug into all available harnesses and enable as much combinatorial innovation across the ecosystem as possible.
Systematic and rigorous thinking stays incredibly important because the abstraction layer keeps changing: from writing code, to orchestrating agents, to getting a million and then a trillion agents working together. Wang calls going all in on word cell a mistake, saying you still need to shape rotate.
At MIT, training a model needed three things: a cloud account like GCP for compute, the code to run, and a dataset. Two of the three you could get by pressing a button online, but there was no effective way to get training data, so that button felt inevitable.
They wanted to build an AI agent to help people get medical care, an idea Wang believes will ultimately exist and is becoming real now, but the timing was wrong. After a month or two Jared Friedman pulled them aside and said he doubted it would go anywhere.
Data stayed unsexy for Scale's first many years. Even with great numbers and revenue, investors at every fundraise were skeptical, asking whether it was a good business, whether it had longevity, whether it was durable. None of them had ever trained a model; many now write think pieces calling data critical.
Develop your own internal compass for how the future will develop and hold deep conviction in it, because you will be inundated with noise and people telling you things, which is hardest when young and inexperienced. Then find the exponential with the steepest, longest curve: once Moore's law, now AI progress.
He grew up in almost New Mexico, now Oppenheimer famous but then the middle of nowhere, doing math and computer science competitions. A friend who interned at Palantir influenced him; he took a gap year at Quora in Silicon Valley, then MIT, where TensorFlow had just launched and Scale's idea came.
“Yeah, I think the key thing is you need to develop conviction in a set of beliefs that nobody else agrees with.”
“Nobody is good at something they've never done before.”
“And so I think for all entrepreneurs, you start out pretty shitty at everything.”
“Like a decade ago, the best AI models could recognize cats in YouTube videos.”
“Top advice, ignore LinkedIn.”
“Yes, which is Meta is proud to offer everyone in this room $1,000 of free credits for the New Spark API.”
“Alexandr Wang: “This is a Once-in-a-Civilization Opportunity”.” https://www.youtube.com/watch?v=sJ4VJWycX9M. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/alexandr-wang-this-is-a-once-in-a-civilization-opportunity.html.