Alexandr Wang on David Senra

Wang traces Muse's roots to a June 2025 memo Zuckerberg published called "Personal Super Intelligence," co-shaped with Wang and Nat Friedman, which gestured at what Muse later became, even though agents barely existed yet. The real trigger, he says, was OpenClaw at the start of 2026: Friedman went all-in on it, running experiments like having it monitor security footage to check he was drinking water, an experience Wang calls somewhere between terrifying and euphoric. Wang treats Friedman's enthusiasm as a credible signal precisely because of his track record as an early Stripe customer and GitHub Copilot contributor. Wang's own first move with his agent was running a Friedman-built "psychoanalysis" prompt that pulled from his email and photos — he describes it as compressing roughly three years of therapy into one session, a factual account of his personal experience rather than a broader claim.

From there, Wang and Friedman each wrote memos to Meta's board in February — Wang's arguing the "super agent" was the final consumer product endpoint, Friedman's titled "the law and trust is a must" — and a prototype was built within one to two weeks and demoed to the board. He states that core elements of the eventual product, including JollyBot and an early Muse Charm, were present in that very first prototype meeting, though not all ideas or tabs existed yet. Zuckerberg, per Wang, used the prototype personally across home life, parenting, and even MMA training, watching footage of himself for feedback — an anecdote Wang offers as evidence of firsthand executive buy-in rather than secondhand reporting.

Wang frames Muse as a deliberate contrarian wager against industry consensus. He says that in early 2026, Anthropic's Opus/Claude models were "way ahead" on coding and the prevailing view was that coding agents were the only technology or capability that mattered for advanced AI, creating pressure to chase Anthropic's lead. Instead, Meta Super Intelligence Labs aimed its model roadmap at personal agents, a choice Wang presents as grounded in conviction about where Meta specifically could win. His argument: Meta's heritage is "connecting people with their friends and family, helping them discover their interests," and its distribution reaches three and a half billion daily users — meaning the natural lane for Meta was consumer AI, not work or coding. He says he told Zuckerberg this thesis even before the Scale deal was finalized, framing AGI's promise as freeing people from obligations so they can spend time on what they want to do — which he ties directly to Meta's existing product surface.

Wang treats the "personal agents came and went" pattern from 2026's open-source agent wave as a solvable reliability problem rather than a fundamental limitation. He says Meta built evaluations around the 100-plus-row behavior spreadsheet, with each row marked either launch-blocking or merely desirable, and tracked checkpoints that were "green on 80 of the rows, but red on 20" until a specially trained MuseSpark 1.3 checkpoint cleared the full set. Readiness, he says, combined this numerical bar with a qualitative "we tried the model, we could tell it was a lot better" judgment. He frames this grinding, iterative sanding process — alongside extensive user testing with fresh downloaders — as the unglamorous labor that separated Muse from OpenClaw-style projects, crediting Peter Steinberger's vision while arguing Meta's edge was execution discipline, not a better initial idea.

Organizationally, Wang says leadership deliberately kept the team small — under 200 across model and product — despite knowing Muse would be one of Meta's most important products of the year, to avoid it becoming what he calls a "Frankenstein" blend of competing product visions. He argues consumer products are felt by users as either having one coherent point of view or fragmenting into a "PM hell" smoothie of mismatched contributions, and credits Friedman's personal taste as the single perspective that shaped Muse, similar to how he frames Steve Jobs' singular imprint on Apple. He also connects this to patience at the top: he says Zuckerberg absorbed seven months of market skepticism — including a Wall Street narrative that Meta was "just burning money" on AI — without rushing the launch, a trait Wang attributes to confidence that Zuckerberg doesn't need outside capital or fear losing control of the company.

Wang describes his viral X activity post-launch as a conscious strategy to cut through what he calls a "torrential stream" of constant AI product releases that numbs most audiences. He says the team held meetings specifically about how to "break through" culturally, since Meta's distribution alone wasn't sufficient, and that he leaned on friends he calls his "schizo friends" — meme-fluent, terminally online people — to understand what cuts through. He admits many tweets were posted on impulse with little forethought and unexpectedly performed best, which he connects to a rediscovered principle that the internet rewards risk and surprise rather than polish. A second explicit goal, in his telling, was make making Meta culturally "cool" again, since he felt Instagram and Facebook's cultural footprint hadn't translated into Meta itself feeling interesting.

He is sharply critical of rival AI advertising, singling out Anthropic's outdoor campaign and OpenAI's Codex ads as failing to show concrete product benefit, calling them confusing relative to simply surfacing user-generated use cases. He ties this to the 1900s-era book Scientific Advertising by Claude Hopkins, arguing its core insight — that audiences are self-interested and don't care about a company's story, only "what this does for you" — still explains why screenshot-driven, benefit-led content outperforms branded campaigns. He also treats the Jolly mascot as an accidental marketing weapon: putting it into memes (including deliberately "racy" ones, which he says he can't recall any product doing before) made screenshots instantly identifiable as Muse-related, which he calls close to the best marketing achievable. He frames user adoption itself as a "trust fall" pattern — small requests succeeding before users escalate to bigger ones — and cites college students fully brain-dumping voice notes to AI as the deepest form of that trust.

Wang's stated personal view is that personal agents are a long-term form factor because it is human nature to have wants and dreams, and that AI like Muse will be "the mechanism and the bridge" enabling people to keep accomplishing bigger versions of them. He explicitly frames this as belief rather than certainty, acknowledging other possible AI relationship "modalities," including fear-driven scenarios where AI operates above humans. His stronger assertion — presented as something close to a mission statement — is that Muse can "give everyone the adulthood they dreamed of in childhood," reversing what he describes as adulthood typically beating ambition out of people as they enter the workforce. He extends this to an image of a more varied, "artistic and cool" future world where billions of people, newly equipped with abundant intelligence, pursue individually distinct creative lives rather than a small number of historical "fanatics" shaping everything alone — a speculative vision he frames in his own words, not as Meta roadmap.

Wang recounts that Zuckerberg first reached out after Llama 4 underperformed internally, in a conversation that began as a request for strategic advice rather than an acquisition pitch, and that the full timeline from that call to the announced deal was five to six weeks. He says he initially judged a deal "doesn't even make that much sense" given Scale's enterprise/government sales business bore little resemblance to Meta's historical acquisitions like Instagram and WhatsApp, and that it became clear Zuckerberg wanted the Llama program's talent, not Scale's business, once the two sides began discussing numbers. The structure that emerged — Meta investing to own 49% of Scale while Wang and several colleagues joined Meta directly — is something he calls "a really fascinating deal construct" and a genuine win for Scale's shareholders, while adding his personal belief that "the best days of scale are ahead of it."

He describes the emotional process of leaving Scale after nine years as difficult, citing Paul Graham's framing that treating a company as your life's work changes how you operate, which he says he had internalized the entire time. What ultimately moved him, in his account, was recognizing Meta's AI organization as "damaged goods" at that moment — which he reframes as an opportunity to rebuild a lab from scratch with chosen principles and talent density, rather than inherit legacy structure. He credits Zuckerberg's decision-making style — rapid, risk-tolerant, willing to pay up for top talent even as OpenAI/Anthropic compensation had also risen sharply — as central to why the rebuild was possible at all.

Wang contrasts his "auteur" style at Scale, where he insisted on understanding every function and became the company's bottleneck by design, with his current role at Meta, which he describes as closer to a coach or steward. He says the research group, called TBD, is built flat by design: researchers report directly to him, which he frames explicitly as "anti-bureaucracy" rather than a literal management claim that he oversees 200-plus people day to day. Technical direction, in his account, is instead handled through a pod structure with designated technical leads who tie-break decisions within their pods, while broader strategic debates happen as group discussions among people he considers all individually accomplished. He likens the organization's payoff structure to "diamond mining" rather than "skyscraper building" — a high-variance bet where most research efforts won't pay off but a few justify the entire investment — and says his central job is setting direction (personal superintelligence, winning consumer AI) and then clearing resource and bureaucratic obstacles so talented people can do their best work without being managed closely.