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

What It Covers

Alexandr Wang describes building Muse, Meta's personal AI agent, from a February 2025 prototype to the fastest-growing consumer AI app ever. He covers the seven-month refinement process, the under-200-person team structure, viral marketing tactics, and the Scale AI acquisition deal that brought him to Meta to rebuild its AI lab from scratch.

Key Questions Answered

• •Launch readiness criteria: Build a spreadsheet of 100+ specific model behaviors required for your product, assign a threshold for each row that is "launch blocking," and only ship when every single row turns green. Muse's team ran multiple model checkpoints that were green on 80 rows but red on 20, and held the release until a specially trained checkpoint of Muse Spark 1.3 cleared every threshold simultaneously.
• •Small team discipline for high-stakes products: Muse launched with fewer than 200 people across model and product work combined. Keeping the team deliberately small prevents the "Frankenstein" effect where competing product managers each jam their own vision into the experience. A single cohesive point of view — in Muse's case, Nat Friedman's taste — must drive every design decision to produce a product that feels unified rather than assembled.
• •Agent onboarding via trust-fall progression: New users of AI agents churn when reliability fails early. Structure onboarding so users attempt small tasks first, experience success, then escalate to progressively larger tasks. Each successful completion builds trust for the next request. Products that skip this graduated trust-building phase — like early Open Claw — saw near-universal churn because unreliable experiences at any stage reset user confidence to zero.
• •Consumer AI marketing via use-case screenshots: Effective consumer AI marketing shows what the product does for the user, not what the company built. Muse's growth relied heavily on Wang retweeting user screenshots showing real tasks completed — a friend delegating a lost ID retrieval entirely to Muse being one example. The mascot character appearing in every screenshot created instant brand recognition across group chats and social feeds without paid campaigns.
• •Flat org structure with 200+ direct reports as anti-bureaucracy signal: Wang carries over 200 direct reports at Meta Superintelligence Labs, not as a management practice but as an organizational statement. All researchers in the TBD group report directly to him to eliminate managerial layers. Technical leads within pods handle technical direction, but people management hierarchy is removed entirely so researchers have room to do their best work without approval bottlenecks slowing exploration.
• •Talent acquisition framing for lab rebuilding: Meta attracted top researchers from OpenAI and Anthropic not primarily through compensation — stock appreciation at those labs had made their packages competitive — but through the pitch of building a frontier lab from scratch with a small team, no compute constraints, and a clear north star of personal superintelligence. Founders rebuilding organizations should lead recruiting with creative ownership and mission clarity rather than assuming compensation alone closes top candidates.