Meta Adopts External Platforms for AI Speed

In the high-stakes arena of artificial intelligence, where billions of dollars and the future of tech dominance are on the line, even giants stumble.

Meta, under the audacious gaze of CEO Mark Zuckerberg, has poured immense capital and ambition into the pursuit of “superintelligence”.

The company has assembled elite teams and lured top researchers with nine-figure pay packages.

Yet, beneath the gleaming facade of this grand vision, an uncomfortable truth has emerged: the very infrastructure meant to power Meta’s AI revolution is, well, slow.

Internal memos, obtained by Business Insider, reveal a surprising strategic pivot within Meta Superintelligence Labs (MSL).

Far from relying solely on its famed in-house engineering prowess, the company is now actively pushing its AI teams to ditch its cumbersome internal systems in favor of external, off-the-shelf tools like Vercel and GitHub.

It’s a candid admission of technical debt and a stark illustration of the urgency driving the AI arms race, forcing even a tech titan to look outside its walls for agility.

The scale of Meta’s AI bet is undeniable.

Over the past year, Zuckerberg has reorganized the company’s AI efforts under MSL, a dedicated unit tasked with spearheading breakthroughs.

This isn’t just about incremental improvements; it’s about a generational leap, a race against formidable rivals like OpenAI and Google.

But this race, it turns out, isn’t just won by brilliant minds and vast computing power.

It also hinges on the mundane, yet critical, efficiency of developer tools.

Aparna Ramani, head of infrastructure for MSL, didn’t mince words in a late September memo to staff.

Meta’s existing systems, built for an era of billions of users and gargantuan engineering teams, simply aren’t fit for the nimble, fast-moving AI units within MSL’s Product and Applied Research (PAR) group, led by former GitHub CEO Nat Friedman.

“Deploying changes takes too long (hours vs. minutes), and the overall tech stack is not conducive to vibe coding,” Ramani wrote.

The phrase “vibe coding” itself is a fascinating window into the evolving world of software development.

It describes the practice of leveraging AI tools to write code, a rapidly spreading trend that demands an infrastructure capable of rapid iteration and experimentation.

When a system takes hours to deploy a simple change, it stifles that creative, AI-assisted flow, turning what should be a sprint into a frustrating crawl.

This inefficiency, Ramani noted, meant engineering tasks that would be quick elsewhere became a manual, time-consuming slog at Meta.

The proposed solution is two-pronged.

First, an immediate embrace of external platforms.

Vercel, a rapidly growing player in web and AI app building, has emerged as a key partner.

The company, which recently raised $300 million at a $9.3 billion valuation, boasts clients like Netflix and Adobe.

Paired with Microsoft-owned GitHub, a ubiquitous code-hosting and collaboration platform, Vercel promises to slash deployment times from a staggering 99 minutes to a mere two minutes or less.

Meta is already running 10 projects on this Vercel-GitHub tandem, with more in the pipeline.

It’s a pragmatic move, prioritizing speed over the traditional “not invented here” ethos that often characterizes large tech companies.

Adding another layer of intrigue, both Nat Friedman, who leads the PAR group, and Alexandr Wang, Meta’s new chief AI officer, are investors in Vercel.

While this alignment might raise eyebrows in some contexts, it also highlights a deep understanding of the tool’s capabilities and a vested interest in its successful integration within Meta.

It suggests a strategic alignment that goes beyond mere vendor selection.

The second part of Meta’s strategy involves building an internal alternative called “Nest.”

This platform will host apps built with TypeScript, a popular industry programming language, on Meta’s own systems.

While Nest is envisioned as the “de facto” long-term option, the memos explicitly state that Vercel will remain available, serving as an “escape valve where there are capability gaps.”

This hybrid approach underscores the immediate need for speed coupled with a desire for eventual internal control and deeper integration with Meta’s vast data ecosystem.

This isn’t the first time Meta has turned to outside technology when its internal offerings fell short.

Earlier this year, the company reportedly used rival AI models like Anthropic’s Claude for its internal coding assistant, Devmate, after its own Code Llama struggled with complex tasks.

Similarly, its “Vibes” feature incorporated Midjourney, an external AI image generator, rather than Meta’s own technology – another instance where Friedman’s advisory role to Midjourney is noteworthy.

The shift at MSL is more than just an operational change; it’s a profound commentary on the nature of innovation in the age of AI.

It reveals that even with seemingly boundless resources, the agility required to compete at the bleeding edge sometimes means shedding pride and embracing the best tools available, regardless of their origin.

Meta’s desperate scramble for speed highlights the immense pressure and existential stakes involved in the race for artificial superintelligence, a race where even a few minutes saved in deployment time could make all the difference.