Recursive Superintelligence Hires 200 AI Engineers in One Month

Richard Socher's latest venture just pulled off something unprecedented: hiring 200 AI engineers in a single region in 30 days.

Recursive Superintelligence, the San Francisco startup that raised $650 million at a $4.65 billion valuation in May, is building AI systems that improve their own code. The company emerged from stealth with just 20 employees. Now they're at 220.

The Numbers Behind the Sprint

The hiring push concentrated entirely in the Bay Area, according to sources familiar with the matter. Every role required experience with large language models and distributed systems. Base salaries ranged from $265,000 to $450,000, with equity packages that could triple total compensation.

Recursive filled positions across three core teams: infrastructure (80 hires), research (70 hires), and product engineering (50 hires). The company conducted over 4,000 interviews in four weeks.

"We basically turned hiring into a machine learning problem," said one engineering manager who joined from Meta. "Automated screening, parallel interview tracks, instant offers."

Why Speed Matters for Self-Improving AI

The startup's core technology requires a different kind of team. Traditional AI companies hire researchers who train models. Recursive needs engineers who can build systems that modify themselves.

Co-founder Yuandong Tian, formerly director of Meta's FAIR lab, designed the hiring process around a simple test: can this person debug code they didn't write? Every candidate had to fix a broken neural network architecture during their interview.

The company's architecture treats AI improvement as a recursive loop. Models analyze their own performance, generate code patches, test improvements, and deploy updates. This requires engineers comfortable with both ML theory and systems programming.

Building Infrastructure Before Product

Unlike most AI startups, Recursive hired infrastructure engineers first. The initial 80 hires built custom training clusters, deployment pipelines, and monitoring systems specifically designed for self-modifying code.

"You can't use standard MLOps when your models rewrite themselves," explained a senior engineer who joined from Google DeepMind. "We needed version control for neural architectures, rollback mechanisms for model updates, sandboxing for experimental changes."

The company runs its own data centers rather than relying on cloud providers. NVIDIA, a strategic investor, provided early access to their latest chips. GV (formerly Google Ventures) led both funding rounds.

The Talent War Gets Specific

Recursive's hiring sprint reveals how specialized AI talent has become. Generic "ML engineer" roles no longer cut it. The company recruited for ultra-specific combinations: compiler engineers who understand transformers, systems programmers who've built training frameworks, researchers who code in C++.

They poached entire teams from Meta, Google, OpenAI, and Anthropic. One eight-person team working on model optimization at Meta moved together. A distributed systems group from Google DeepMind joined as a unit.

Compensation packages reflected this specificity. Engineers with experience in both CUDA programming and transformer architectures commanded $400,000+ base salaries. Those who'd built training infrastructure at scale got similar offers plus guaranteed liquidity events.

What 200 Engineers Build in Month One

The newly hired teams immediately started on three projects:

1. A custom compiler that optimizes neural networks during training 2. Infrastructure for models to propose and test their own architectural changes 3. Sandboxed environments where AI systems can modify their code without risk

Early results surprised even the founders. Within two weeks, their prototype models were finding optimizations human engineers missed. One model reduced its own inference time by 34% through automated code improvements.

The Recursive Advantage

While competitors focus on scaling compute or gathering data, Recursive bets on a different approach: AI that improves its own efficiency. Their models don't just process information. They rewrite themselves to process it better.

This creates compound advantages. Each improvement makes the next improvement easier to find. The better the system gets at optimizing itself, the faster it improves.

The 200-person team gives Recursive enough engineering firepower to build this vision from scratch. No legacy code. No technical debt. Just purpose-built infrastructure for recursive self-improvement.

The real test comes next. Can a four-month-old company coordinate 200 new engineers effectively? Can self-improving AI actually work at scale? Recursive Superintelligence is betting $650 million that the answer to both is yes.