Meta AI Chief Threatened to Quit Days After Joining

💡 TL;DR - The 30 Seconds Version

⚡ Shengjia Zhao, ChatGPT co-creator, threatened to quit Meta days after joining in August, forcing the company to hastily name him Chief Scientist.

🚪 At least three other AI researchers have already left Meta's superintelligence lab after brief tenures, with two returning to OpenAI.

🔄 Meta has reorganized its AI division four times in six months and shelved its flagship Llama Behemoth model after poor performance.

💰 Despite offering nine-figure compensation packages, Meta announced a hiring freeze across AI teams in late August to "plan 2026 headcount."

🏗️ The exodus reveals that elite AI researchers value autonomy and stable direction alongside massive paychecks and computing resources.

🎯 Meta's next major model training runs will test whether financial firepower alone can win the AI talent wars.

Three early departures, a hiring pause, and a Behemoth rethink raise questions cash can’t answer.

Mark Zuckerberg’s bet on buying top AI talent met reality fast: Shengjia Zhao, widely credited as a co-creator of ChatGPT, threatened to leave days after joining Meta’s superintelligence push. According to a detailed report on Zhao’s near-exit, Meta raced to formalize him as chief AI scientist to keep him from returning to OpenAI. Meta calls the scrutiny overblown. The tension is real.

What’s actually new

Zhao’s title became official in late July. The move caps weeks of turmoil around Meta Superintelligence Labs (MSL), which has been reorganized four times in six months, most recently into four teams. Meta says that’s normal scale-up mechanics. Critics see whiplash.

The company has also pulled back on plans to publicly release its flagship Llama “Behemoth” model after disappointing internal performance and earlier delays. That’s a strategic retreat, not a surrender. Still, it lands with a thud.

Evidence of churn, despite the money

At least three notable researchers—Avi Verma, Ethan Knight, and Rishabh Agarwal—have departed within weeks, with Verma and Knight returning to OpenAI, as reported by Wired. Veteran product leader Chaya Nayak is leaving as well. That’s a pattern, not a blip.

Add a fresh hiring pause across most MSL roles, flagged in recent internal guidance and outside reporting. Meta says this is about planning 2026 headcount and focusing on “business-critical” hires. The optics are rough.

The culture collision inside “TBD”

The most secretive group, internally dubbed “TBD,” sits under Alexandr Wang, the Scale AI founder brought in as Meta’s chief AI officer after a $14 billion deal that folded Scale data muscle into Meta. Wang is a lightning rod: commercially minded, founder-fast, and new to Big Tech org design. Some recruits chafe at process, governance, and queueing for resources they believed were guaranteed. Friction was inevitable.

Zuckerberg is hands-on here, which insiders alternately praise as focus and describe as micromanagement. The CEO wants speed. Researchers want air cover. Both are understandable.

Strategy: compute is necessary, not sufficient

Meta is arguably doing the one thing every frontier lab must: stockpiling compute. The “Prometheus” cluster in Ohio is slated to reach roughly a gigawatt by 2026, putting Meta in the small club able to train truly massive models. That matters for pretraining scale and for repeated failure-and-retry cycles. Hardware doesn’t think, though.

Meanwhile, Meta’s leadership map keeps shifting. Yann LeCun remains chief AI scientist for FAIR but now reports into Wang. Nat Friedman is steering Products and Applied Research, the bridge from models to apps. Titles are tidy. Lines of authority are not.

Competitive context

OpenAI isn’t acting like a company losing a talent war. It’s publicly criticized Meta’s poaching tactics and quietly welcomed back at least two recent defections. That suggests retention gravity stronger than headline compensation. Mission and momentum still count.

Meta’s calculus is that nine-figure packages plus unmatched reach across Facebook, Instagram, and WhatsApp can lure and retain a “dream team.” Early data points argue the opposite: autonomy, clarity, and steady direction rival money and GPUs. Culture scales; panic doesn’t.

Limits and risks

Four reorgs in half a year create coordination taxes, buried roadmaps, and unclear ownership. A paused Behemoth release raises fair questions about model quality and evaluation bars. A hiring freeze—whatever the spin—signals the need to catch breath and pick a lane. The risk isn’t missing the next demo. It’s wasting the next training run.

One more point. Researchers talk. Word of bottlenecked compute, shifting charters, or leadership tug-of-war travels fast. That makes the next marquee hire harder, and the next stay-or-go decision easier.

The near-term test

Meta doesn’t need perfect harmony to succeed. It needs a stable plan, reliable access to compute per researcher, and a crisp story about where reasoning models, multimodality, and agents meet product. If “TBD” turns into “to be decided” again, the lab will keep bleeding time and leverage. If it locks a direction and ships a convincing successor to Llama, the narrative flips overnight.

It’s fixable. But not with money alone.

Why this matters

AI talent now optimizes for autonomy, clarity, and stable roadmaps—conditions cash and GPUs can’t buy outright.
Meta’s next training runs will test whether a compute-first strategy, amid churn and reorgs, can still deliver frontier-class models.

❓ Frequently Asked Questions

Q: How much are these "nine-figure" compensation packages worth exactly?

A: Meta is offering AI researchers between $100 million and $999 million in total compensation packages, including equity and bonuses. For context, this rivals compensation for professional athletes and top hedge fund managers, far exceeding typical tech salaries of $300,000-$500,000 annually.

Q: What exactly is Meta's "TBD" team and why is it secretive?

A: "TBD" stands for "to be determined" and houses Meta's most advanced AI research under Alexandr Wang. It operates separately from Meta's public FAIR lab and focuses on frontier models that could achieve superintelligence. The secrecy likely protects competitive advantages and ongoing research from rivals.

Q: What went wrong with Meta's Llama Behemoth model?

A: Llama Behemoth, Meta's flagship AI model, was shelved in May 2024 after failing to meet performance benchmarks during testing. The company had planned to release it publicly but pivoted to developing newer cutting-edge models instead, triggering the massive talent acquisition push.

Q: How does Meta's computing power compare to OpenAI and Google?

A: Meta's upcoming Prometheus cluster will reach 1 gigawatt by 2026—enough to power 750,000 homes. This puts Meta among the few companies with training infrastructure for massive models, comparable to Google's TPU clusters and Microsoft's OpenAI partnership computing resources.

Q: Why did Meta hire Alexandr Wang despite his lack of Big Tech experience?

A: Wang came with Meta's $14 billion acquisition of Scale AI, his data labeling company crucial for training AI models. At 28, he brings entrepreneurial speed and commercial instincts, but sources report friction with researchers accustomed to traditional corporate AI lab structures.