Meta’s Superintelligence Labs is trying to speed up AI development by steering its teams toward new tools that allow developers to build and test products more quickly, according to internal memos obtained by Business Insider.
The shift highlights the stakes of Meta CEO Mark Zuckerberg’s massive bet on AI. Over the past year, Meta has poured hundreds of millions of dollars into luring top researchers with nine-figure pay packages and reorganized its AI work under the new Meta Superintelligence Labs (MSL). Now, its new leaders are revamping the internal engineering tools that those teams use — a shift that underscores the urgency with which Meta is moving as it races to catch up with rivals like OpenAI and Google.
Internal documents show that MSL’s Product and Applied Research (PAR) group, led by former GitHub CEO Nat Friedman, is pushing staff away from Meta’s slower, home-grown infrastructure toward mainstream developer platforms like Vercel.
One memo from late September notes that Meta’s existing systems, designed for billions of users and giant engineering teams, take “too long” to deploy changes and are “not conducive to vibe coding,” making it harder for small, fast-moving AI teams to experiment. The practice of vibe coding, or having AI tools write code, is becoming more widespread among engineers in the tech industry. Meanwhile, Meta is also building an internal alternative called “Nest” that will integrate more easily with Meta’s data for quick experimentation.
Another internal post shows that Meta has paired Vercel with GitHub, a widely used code-hosting and collaboration platform owned by Microsoft, to give its PAR teams a faster way to work.
Friedman and Alexandr Wang, Meta’s new chief AI officer, are investors in Vercel. Vercel has become one of the biggest names in the market for building web and AI apps. In September, it raised $300 million at a $9.3 billion valuation from backers including Accel, Singapore’s GIC, BlackRock, Khosla Ventures, and General Catalyst. The company’s customers include giants like Netflix, Adobe, and Stripe.
The memos offer a rare glimpse into how Meta is trying to accelerate its AI ambitions by relying on outside development and engineering tools, and sidestepping the limitations of its own systems that were built for a different era.
Meta declined to comment. Vercel and GitHub did not respond to requests for comment from Business Insider.
Meta systems are ‘not conducive to vibe coding’
In one of the memos sent in late September, Aparna Ramani, head of infrastructure for MSL, told staff that the company’s internal systems, which were built to support billions of users and massive teams of developers, don’t work well for the small, fast-moving engineering teams inside PAR.
“Deploying changes takes too long (hours vs. minutes), and the overall tech stack is not conducive to vibe coding,” Ramani wrote.
In other words, engineering tasks that would be quick and easy to deploy elsewhere often require more manual work at Meta, slowing teams down.
Ramani’s memo outlines a plan with two approaches to quickly build app prototypes and reduce deployment times from 99 minutes to two minutes or less.
The first approach uses Vercel, which, Ramani wrote, is “widely used and can unblock MSL immediately.” She wrote that Vercel is willing to work with Meta to mitigate some data and hosting restrictions. Ramani also wrote that Meta already has one tool running on Vercel, with multiple other projects in development.
The second approach is to build an internal platform called “Nest” for hosting apps built with TypeScript — a popular Microsoft-created programming language widely used across the industry — on Meta’s systems. Ramani wrote that a working prototype was expected within two weeks of the memo.
Nest will eventually become the “de facto” option, and keeping Vercel available provides “an escape valve where there are capability gaps,” the memo said.
Meta is increasingly using Vercel and GitHub
A second internal memo, published on September 17, shows how Meta used Vercel with GitHub to provide its PAR teams with a faster way to build and prototype apps. Meta is already running 10 projects on Vercel and GitHub, which could now be updated “in minutes,” the post said.
Meta has often turned to outside technology when its own tools fall short. Earlier this year, Business Insider reported that Meta rolled out an internal coding assistant called Devmate that uses AI models from rivals, including Anthropic’s Claude, after employees said Meta’s own Llama AI model struggled with complex tasks.
Meta has also recently launched a feature called Vibes, which uses Midjourney, an external AI image generator, rather than its own image-generation technology. Friedman is an advisor to Midjourney.
Have a tip? Contact Pranav Dixit via email at pranavdixit@protonmail.com or Signal at 1-408-905-9124. Use a personal email address, a nonwork WiFi network, and a nonwork device; here’s our guide to sharing information securely.
Read the full article here
