Public transcript

Creating Muse: The Fastest-Growing AI Product Since ChatGPT | Alexandr Wang, Meta

By: David Senra
Length: 1 h 12 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Alexandr Wang, founder of Scale AI, talks with David Senra about the creation of Meta's Muse personal AI agent, from the founding vision of Meta Super Intelligence Labs through the OpenClaw-inspired prototype, the seven-month refinement process, and the product's viral launch. He also recounts how Mark Zuckerberg recruited him after Llama 4, the unusual deal where Meta took 49% of Scale while he joined Meta, and how he manages a flat, 200-plus-direct-report research organization. He argues personal agents are the long-term consumer AI form factor and that Meta's heritage and distribution make it uniquely positioned to win there.

Key points

Meta Super Intelligence Labs was founded around Mark Zuckerberg's June 2025 memo on Personal Super Intelligence, co-written with Alexandr Wang and Nat Friedman.
Wang and Friedman's February 2026 experiences with OpenClaw convinced them personal agents were special, leading to a board demo prototype built in one to two weeks.
Muse launched seven months later after meticulously sanding a spreadsheet of over a hundred required behaviors, with a specially trained Musepark 1.3 as the first all-green checkpoint.
The team stayed under 200 people to keep a single point of view; Wang credits Nat Friedman's taste as shaping the product.
Wang's unhinged meme-heavy posting on X was a deliberate strategy to break through and make Meta cool again.
Meta has said Muse has millions of users, and Wang retweeted graphs implying it is the fastest-growing consumer AI app of all time.
After Llama 4 disappointed, Mark Zuckerberg recruited Wang in five to six weeks with a deal where Meta owns 49% of Scale and Wang joined Meta.
Wang has over 200 direct reports at MSL as an anti-bureaucracy design choice, acting as a coach rather than an auteur.

Questions it answers

00:00Origins of Muse

Where did the idea for Muse come from?

Meta Super Intelligence Labs was founded around Mark Zuckerberg's June 2025 memo on Personal Super Intelligence. In February 2026, after Wang and Nat Friedman had intense experiences with OpenClaw, they concluded personal agents were something special, wrote memos to the board, and built a working prototype in one to two weeks that included the Jollybot and an early Muse Charm.

Mark Zuckerberg published the Personal Super Intelligence memo in June 2025.
Wang wrote a memo calling the super agent the final consumer product.
Nat Friedman's memo lines: the clause, the law, and trust is a must.

09:23Why Meta and personal agents

Why did Meta build a personal agent instead of competing on coding?

Wang believed Meta's heritage of connecting people with friends, family and interests, plus distribution to three and a half billion people, made consumer AI its natural winning zone. While the industry rushed to compete with Anthropic on coding agents, MSL aimed its model roadmap at personal super intelligence.

Meta has distribution to three and a half billion daily users.
The industry crowded into coding agents; Meta chose consumer AI.
Wang told Mark early that Meta's apps represent what people want to do.

12:46Seven months of refinement

How did Muse go from a fragile prototype to a reliable product?

The team spent seven months meticulously sanding out details, tracking a spreadsheet of more than a hundred required behaviors with launch-blocking thresholds for each. The first model checkpoint green across the board was a specially trained version of Musepark 1.3, and the team stayed under 200 people to keep a single point of view.

Over a hundred specific behaviors were tracked with launch-blocking thresholds.
A specially trained Musepark 1.3 was the first all-green checkpoint.
Wang credits Mark's patience and restraint through the doubting period.

26:00Launch and marketing

How did Muse break through culturally after launch?

Wang posted heavily and 'unhinged' on X, leaning on memes and the Jolly mascot to make Meta cool again and cut through the torrential stream of AI releases. He argues modern products must 'work on the screenshot,' and the mascot appearing in screenshots became the best marketing.

Wang writes every tweet himself; the internet rewards risk and surprise.
Ads should show the product's benefit, not the company.
The mascot in screenshots instantly communicates a Muse screenshot.

35:01Retention and philosophy

How does Muse keep users engaged, and what does Wang believe it means?

Successful users go through a trust fall cycle, giving the agent slightly bigger tasks and having it work each time. Philosophically, Wang believes personal agents let people offload what they have to do and keep dreaming bigger, giving everyone the adulthood they dreamed of as kids.

Users give the agent progressively bigger tasks in a trust fall.
Wang: Muse can give everyone the adulthood they dreamed of in childhood.
The world has been shaped by a small number of fanatics; that could be everyone.

45:31The Scale AI deal

How did Mark Zuckerberg recruit Wang from Scale AI?

After Llama 4 disappointed, Mark reached out asking what Meta should do; within five to six weeks they struck a deal where Meta invested and owns 49% of Scale, Scale continues independently, and Wang and a few others joined Meta. Wang says it was a win for Scale shareholders and let him refound Meta's AI lab from scratch.

First contact to announced deal took five to six weeks.
Meta invested and owns 49% of Scale; Scale continues independently.
question_start_marker_removed

59:00Managing MSL

How does Wang run Meta Super Intelligence Labs differently from Scale?

At Scale he was the auteur whose point of view shaped everything, which made him a bottleneck. At Meta he hires brilliant people and acts as a coach, setting the north star and giving them room, with over 200 direct reports in a deliberately flat, anti-bureaucracy structure.

TBD researchers report directly to Wang as an anti-bureaucracy statement.
He acts more like a coach or steward of environment than an auteur.
AI research is diamond mining; data businesses are skyscraper building.

Notes

Origins of Muse

Meta Super Intelligence Labs (MSL) was founded around Mark Zuckerberg's June 2025 memo on Personal Super Intelligence, co-written with Alexandr Wang and Nat Friedman.
In February 2026, after deep experiences with OpenClaw, Wang and Friedman concluded personal agents were special. Wang wrote a memo calling the super agent "the final consumer product"; Friedman wrote memos with the lines "the clause, the law, and trust is a must."
A prototype was built in one to two weeks and demoed to the board; the Jollybot and an early Muse Charm were already present.

Why Meta bet on personal agents

While the industry crowded into coding agents after Anthropic's lead, MSL aimed its model roadmap at personal super intelligence.
Wang argued Meta's heritage (connecting people with friends and interests) and distribution to three and a half billion daily users made consumer AI its natural winning zone.

The seven-month refinement

The team tracked a spreadsheet of over a hundred required behaviors, built evals for each, and set launch-blocking thresholds.
The first all-green model checkpoint was a specially trained version of Musepark 1.3.
The team stayed under 200 people to preserve a single point of view — Wang calls Muse "a single work of art" and credits Nat Friedman's taste, comparing it to Steve Jobs.

Launch and marketing

Wang posted heavily on X, leaning into memes and the Jolly mascot to "make Meta cool again" and break through the torrential stream of AI releases.
He criticizes Anthropic and OpenAI ads for being about the companies rather than showing product benefits, citing the book *Scientific Advertising* by Hopkins.
Modern products must "work on the screenshot"; the mascot in screenshots instantly signals a Muse screenshot.
Meta has said Muse has "millions" of users; Wang retweeted graphs implying it is the fastest-growing consumer AI app of all time.

The Scale AI deal

After Llama 4 disappointed, Mark reached out; first contact to announced deal took five to six weeks.
Meta invested and owns 49% of Scale; Scale continues independently; Wang and a few others joined Meta.
Wang calls it a win-win-win and a chance to refound Meta's AI lab with a small, flat, highly technical team.

Managing MSL

At Scale, Wang was the auteur involved in everything, which made him a bottleneck. At Meta he acts as a coach: set the north star, spot talent, provide environment.
He has over 200 direct reports — a deliberate anti-bureaucracy statement; researchers in the TBD group report directly to him.
He frames AI research as "diamond mining" (power-law payoffs) versus "building skyscrapers" (linear, operational businesses like data annotation).

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.