There is a company that built an AI in two weeks but didn’t release it for seven months. It’s Meta.
Alexander Wang, who leads Meta’s AI organization, spent 70 minutes on David Senra’s podcast released on the 11th of this month describing how he created the personal AI agent “Muse.” The direction was set in February, a prototype was built in one to two weeks, and it was demoed to the board. Many of the features that ended up in the final product were already present in that first demo. Yet the launch didn’t happen until September. What did they do during the remaining seven months?
1. The 80‑point checkpoint was a disqualification
Wang explains it like this: he listed more than 100 behaviors and capabilities a personal agent must have in a spreadsheet. For each row he drew a “cannot launch below this line” threshold and a “nice to have above this line” line. As the model was trained, checkpoints appeared where 80 rows were green and 20 were red—those were not launch‑ready. Only when a checkpoint showed every row turned green did they finally launch.
I lingered on this point. When I look at AI project reports from domestic companies, they usually end with a single number: “Accuracy 85%.” The average looks respectable, but the report never shows where the remaining 15 % fails. Sending a customer’s data to the wrong person and a simple typo are both counted as one point. So the PoC passes, but the business says it can’t use it.
I’ll call this the “launch‑blocking line.” It’s not an average score; it’s a behavior‑by‑behavior threshold. If you can’t cross that line, you don’t ship, no matter how high other scores are. Meta’s seven months were spent not on adding features but on erasing the red cells.
2. A “smoothie” product is known first by the user
The second thing that stood out was team size. The combined model‑and‑product effort was under 200 people. Even though it was the company’s most important product, they deliberately kept the team small.
Wang likened a product that emerges from a large organization to a “smoothie.” When each PM pushes in the feature they want, the result is a blended mixture of many visions. Users sense that muddle. He said clearly that Muse was shaped entirely by the taste and sense of one person—Nate Friedman.
Does this scene feel familiar in your own organization? When an enterprise AI task force is formed, sales pushes sales‑related requirements, finance pushes finance‑related ones, audit pushes audit‑related ones. Everyone wants their items in the meeting, and the resulting chatbot reflects every department’s demands—yet no department actually enjoys using it. In my view, the first thing you need to decide in an AI project isn’t the model; it’s the single “final decision‑maker.”
3. Trust is a launch condition, not a feature
Before the prototype was built, one of Friedman’s memo titles was “trust is a must.” Wang said that sentence defined Muse’s core architecture. Muse is an agent that can handle payments and shopping on behalf of the user, so if trust collapses, the product itself fails.
To be fair, we still don’t know whether Meta passed this test. After launch, U.S. media criticized how much information Muse accumulates about a user’s conversation partner and whether deletion requests are actually honored. A market study showed that the day‑after‑install revisit rate was almost twice that of ChatGPT, but after about ten days the trend reversed. Meta has the distribution power of an app used by 3.5 billion people daily; most of us lack that kind of weapon.
That makes the issue even more critical for us. If your company can’t match Meta’s distribution or capital, losing trust makes it extremely hard to regain a foothold. What data an internal agent sees, what it remembers, and who can erase those records must be part of the launch‑blocking criteria—not an after‑the‑fact add‑on.
4. Demo in 2 weeks, launch in 7 months
The moderator likened seven months in the AI industry to ten years in other industries. Wang credited that endurance to Mark Zuckerberg’s patience. Executives can’t read that number backward: it’s not “We got a demo in two weeks, so a month will do,” but rather “If a demo took two weeks, a usable product will take at least ten times longer.”
So, what if we re‑evaluate our own AI initiatives against this yardstick?
Do you have a behavior list and a launch‑blocking line instead of just an average score?
Is there a single person designated as the final decision‑maker?
Are data‑access and record‑deletion requirements baked into that launch‑blocking line?
If any of those three items are missing, what you need now isn’t a better model—it’s a stricter checklist.
If you’d like to collaborate on defining the launch criteria for internal AI projects, feel free to get in touch. 노우버스 AX 컨설팅