Alexandr Wang: From Los Alamos to Superintelligence
In a Nutshell
Alexandr Wang founded Scale AI after spotting that data was the missing piece for training models, then joined Meta to build a frontier lab around talent density and agentic systems. He argues the biggest bottleneck now is spreading AI through the economy, not model breakthroughs, and that vision plus conviction will matter more than intelligence as AI becomes abundant. The practical takeaway is to identify the steepest, longest exponential curves, develop an internal compass, and build feedback loops that let swarms of agents outperform large teams.
These notes were generated by AI and may contain inaccuracies.
Alex Wang grew up in Los Alamos, New Mexico, which he describes as the middle of nowhere. He participated in numerous math competitions and computer science competitions during his youth. Even at a young age, he had ambitions to do really big things, though the exact paths to achieve that were not clear to him initially.
A friend who was really into programming became a significant influence, securing an internship at Palantir. This inspired Wang to pursue a similar path. After finishing high school, he took a gap year to work at Quora in Silicon Valley at age 19, then attended MIT at age 18. He was 19 when he started Scale.
Wang describes the period from ages 17 to 19 as one of constant change and rapid learning, where he was "drinking from the fire hose." He emphasizes two critical experiences during this time:
- Working at a company provided invaluable insight into how companies actually function, how to build products, how to iterate, and how groups make decisions
- Attending MIT offered the opportunity to explore interesting areas and experiment
At MIT, Wang trained his first models and experimented with TensorFlow shortly after its release. This experimentation ultimately led to the idea for Scale.
After one year at MIT, Wang applied to Y Combinator and was accepted. He describes YC as critical to his entrepreneurial journey, noting that it provides a unique combination of being supportive while also giving honest feedback when founders are making mistakes.
Wang and his co-founder initially pitched a different idea to YC: building an AI agent to help people access medical care. They worked on this concept for one to two months before Jared Friedman advised them that the timing was wrong, despite the idea being fundamentally sound. This feedback prompted them to return to the drawing board.
Wang's experience training models at MIT revealed a fundamental problem: to train a model, you needed three things—a cloud computing account, code, and data. While the first two could be obtained instantly online, there was no effective way to acquire training data. This insight became the foundation for Scale.
Wang explains that Scale was founded through first principles thinking rather than following trends. At the time, large language models had not yet emerged, but self-driving cars were becoming prominent, making computer vision the initial market focus. Wang notes that data was considered "unsexy" for many years, with investors expressing skepticism about the business model's durability despite strong revenue numbers.
Wang emphasizes that successful companies are built on convictions that nobody else shares initially. He states that the most successful companies toil in obscurity for years before their ideas become mainstream. He argues that entrepreneurs must develop their own compass for what the future will look like rather than following the herd, which leads to confusion and poor decisions.
Wang acknowledges that nobody is good at starting a company when they begin—the key is developing the ability to continuously improve and learn quickly. He notes that investors who passed on early Scale funding now write think pieces about data's critical importance in AI.
Wang believes the current bottleneck in AI is not model progress but diffusion throughout the world. Even if models remained at current capabilities, decades of economic and operational upheaval would follow. He describes this as a once-in-a-civilization opportunity for dreamers with vision and ambition.
Wang contrasts the startup landscape of a decade ago—where founders had to be clever and find market angles against resource-rich incumbents—with today, where AI agents make it possible for startups to compete more directly with larger companies.
Wang describes Meta's vision of personal superintelligence, where billions of people will have superintelligent systems tailored to their individual contexts and goals. The core concept is "agency expansion"—helping people accomplish things they could not have previously imagined.
Meta envisions an ecosystem approach rather than centralized control, with billions of personal superintelligences working alongside an explosion of entrepreneurship. Currently, 200 million businesses operate on Meta platforms, and the company believes this number should grow to billions through AI-enabled creativity.
Wang joined Meta approximately one year before this conversation. He describes implementing a zero-based rebuild of how to construct a frontier AI lab. Within nine months, the team launched Muse Spark 1, followed by Muse Spark 1.1 two months later.
Key insights from building the lab:
- Talent density was the core bet and compounds naturally as more talented people attract additional talent
- Frontier AI work is fundamentally scientific research requiring different mindsets than traditional internet companies
- The lab must function as an organism capable of compounding with exponential growth across capabilities, compute, and adoption
Wang mentions several upcoming releases:
- Continued updates to the Muse Spark line
- Bigger models competitive with the best available today
- A harness to empower developers and agentic developers
- Open source models to decentralize AI capability development
Wang emphasizes making models accessible rather than rationed for wealthy developers and companies. He notes that Muse Spark 1.1 performs comparably to Opus-level models for agentic workflows while being significantly cheaper—approximately 8x less expensive.
Wang describes AI development as occurring in waves, each approximately 10 times larger than the previous:
- First wave: self-driving cars
- Second wave: large language models and chatbots
- Third wave: coding agents
He predicts continued emergence of new modalities and form factors, each dramatically larger than predecessors.
Wang identifies speed and reliability as critical factors for agent harnesses. The goal is extreme reliability and extensibility to support complex multi-agent setups. He believes current models are already powerful enough to fuel significant GDP growth expansion.
Looking back on the decade, Wang predicts people will recognize that debates about exact timelines for superintelligence were shortsighted. He notes that a decade ago, the best AI models could recognize cats in YouTube videos, while today systems demonstrate capabilities that would have seemed impossible.
Wang argues that intelligence and agency will become abundant, shifting the scarce resource to vision and ambition. The question becomes whether individuals have clear views of desired future states and the drive to realize them.
Wang acknowledges builders have responsibilities to prepare the world for AI technology, including helping enterprises and governments adapt, securing against biosecurity and cybersecurity risks, and managing associated dangers. However, he emphasizes this is also a time of unprecedented opportunity for developing new sciences, solving health and biology problems, building previously unimaginable businesses, and creating novel creative opportunities.
Systematic and rigorous thinking remain critically important as abstraction layers continue changing. Previously, companies were built by writing code and organizing human teams, but the current focus is orchestrating agents together, then developing organizations of agents, scaling from a million agents working together to potentially a trillion agents. The need persists to structure workflows at whatever abstraction layer is being operated at. Systems thinking never goes out of style.
Going all-in on "Word Cell" is a mistake - shape rotation is necessary. What's increasingly necessary is having a deeper compass and philosophical view on how the world should develop, drawing lessons from human history about how civilization moves through periods of change.
Humanity will change more in the next decade than it has in the last 100 years. The imperative to have positive visions for that change and coherent articulations of how it should develop is more important than ever.
There is astronomical opportunity in agentic looping and developing systems that enable spending 1,000x more or 1 millionx more on tokens to drive outcomes in continuous feedback loops. Companies function as large-scale feedback loops where humans operate each edge - getting customers, making them happier, increasing spending, hiring more people, and repeating the cycle. Micro feedback loops exist within these systems.
Developing agentic systems that operate and optimize these feedback loops offers huge amounts of alpha. Internally at Meta, cases exist where the right agentic loop with the right eval or metric can have a swarm of agents accomplish more than a team of 100 engineers, very easily.
Figuring out what the world looks like with lots of agentic coordination problems is one of the most interesting problems today. Mechanically, this involves markdown files, cron jobs, and pointing agents at enough data so they can figure out things that aren't in distribution. The process comes down to figuring out the metric, then skills, markdown files, cron jobs, and goals.
Everything is more mundane once you dig into it. Ignore LinkedIn threads about magic stuff - LinkedIn is where you get customers.
Develop your own internal compass for how you think the future will develop and have strong conviction in it. You will get inundated with noise and people telling you what to do, and it will be very confusing and hard, especially when you're young. Having true conviction in what you believe and what you want to do is the most important thing.
It took deep conviction in what was being built to weather the storms of many years of chaos in the market, industry, and among people around us.
Try to identify what is the exponential in the world that has both the steepest curve and will go the longest. Many decades ago, this curve was Moore's Law, which was clearly the right thing to invest in at the time. Currently, it's AI progress, but there will be more of these very steep curves in the future.
It's fine if these curves start at a very boring point or don't even seem interesting. When Scale started, there were cat detectors in YouTube videos, which was hard to explain as the most important technology of our time, but it was on an unbelievable exponential.
Meta is offering everyone in the room $1,000 of free credits for the new Spark API. Meta will keep making the models better. Spark is currently 8x cheaper than Opus, which converts to significantly more value in Opus dollars. Details on how to get these credits will be provided.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
