Meta's Chief AI Officer Alexandr Wang Discusses Prototype Refinement, Team Completes Impressive Prototype in One to Two Weeks and Receives Applause from the Board
Meta's Chief AI Officer Alexandr Wang stated that the team completed an impressive prototype in one to two weeks and received applause from the board during the presentation, but it took seven months to refine it for public release.
The prototype worked intermittently, sometimes excellently but often crashed, with core design elements like mascots appearing in the initial version. The goal is to transform it into a highly reliable, finely-tuned consumer product, involving fewer than 200 people.
The team established a spreadsheet containing over 100 specific behaviors and capabilities, setting release blockage thresholds and pass standards for each, reviewing after each model iteration. The first full green light checkpoint is for a specific training version of Muse.
Wang pointed out that stability and reliability in the real consumer market equate to the best product; once it occasionally fails, users consider it garbage. The seven-month journey from the February prototype to product release is seen as long and painful.
Meta previously restructured its AI efforts after Llama 4 underperformed, with Wang joining in 2025 to lead the Superintelligence Lab, focusing on personal AI agents rather than programming competitions. The product emphasizes a single perspective and user experience refinement.
Capital and manpower shifted from rapid prototype demonstrations to reliability engineering, driven by consumer-grade release standards, with reliable products benefiting users in the long term, while intermittent demonstrations and short-term hype face pressure.
Source: Public Information
ABAB AI Insight
After dropping out of MIT in 2016, Alexandr Wang founded Scale AI to provide data labeling and evaluation services. He joined Meta in 2025 as Chief AI Officer after Meta acquired a 49% stake for about $14.3 billion, leading the Superintelligence Lab and overseeing consumer products like Muse.
On the capital front, Meta is investing heavily and allocating internal resources to small, high-density teams (fewer than 200 people) for the seven-month refinement from prototype to reliable product, motivated to compensate for Llama 4's shortcomings and focus on personal AI experiences, shifting resources from model parameter competitions to behavior checklists and release standards.
Similar cases include the early development of products like ChatGPT from demonstration to stable iteration, as well as Apple's hardware refinement path from prototype to consumer reliability. The current AI industry is transitioning from rapid prototype demonstrations to stable consumer-grade products, with most agent products losing users due to unreliability.
Essentially, this reflects the constraints of technological substitution: intermittent capabilities are replaced by reliability thresholds, as the consumer market's zero tolerance for occasional failures makes upward potential dependent on continuous availability rather than a single impressive demonstration, thus requiring strategic restraint and long-term refinement.
ABAB News · Cognitive Laws
Prototype completed in a week, product refined in seven months
Intermittent functionality leads users to consider it garbage
In a fast-paced environment, restraint is the strategy