Swyx on Meta's AI talent raid: $100M+ first-year packages, OpenAI's surprising resilience, and Cursor poaching Claude Code's creators

Summary

Meta's talent raid on OpenAI is being described by AI researchers as a panic move, one that is nonetheless financially rational given Meta's balance sheet. According to Wired reporting, packages are structured at up to $300 million over four years, with more than $100 million in total first-year compensation. That puts individual researcher pay in the range of what Tim Cook and Satya Nadella each earned across all of 2024. Swyx (Shawn Wang) frames the math clearly: Meta trades at a $1.8 trillion market cap, has redirected capital from its stalled VR unit, and holds the equivalent of Anthropic's entire annual budget as discretionary AI spend. Moving that stock by even a percentage point dwarfs the cost of the hiring spree.

Mark Zuckerberg is uniquely positioned among Magnificent 7 CEOs to execute this kind of bet. He retains founder-level control and has a documented pattern of stumbling first, then overcorrecting decisively. The mobile transition, the 2010s Google talent raid that built Meta's ad engine, the Instagram and WhatsApp acquisitions — all followed a failure-then-blitz playbook. The creation of Meta Superintelligence Labs (MSL) follows the same arc, made possible specifically because Llama 4 underdelivered. Had Llama 4 replicated Llama 3's traction, the restructuring almost certainly would not have happened. Notably, an acquisition of Perplexity was reportedly Meta's first or second choice before MSL became the strategy, which would have pushed the company toward a consumer product rather than a frontier lab.

Llama 4 Behemoth is rumored to have simply never shipped. The suspected cause is a failed attempt to replicate DeepSeek's large mixture-of-experts architecture, compounded by internal legal friction over training data. Swyx points to Anthropic's recent fair-use ruling — covering training on copyrighted material as transformative use — as a case that may free labs from that constraint going forward, though the question of per-inference licensing fees remains unresolved. Anthropic itself faces exposure over approximately 7 million unlicensed books from the Books3 dataset.

Despite the headline departures, OpenAI's resilience is underappreciated. The company has already absorbed the exits of Anthropic's founding team, Elon Musk, Ilya Sutskever, and Mira Murati — each generating a direct competitor — and kept accelerating. Machen's internal memo responding to the Meta raid was viewed as an unforced error, coming across as defensive rather than rallying from a position of strength. The more effective framing, per Swyx, would have leaned on ChatGPT's scale, revenue trajectory, and category leadership. The consensus view among AI insiders on which lab holds the "mandate of heaven" to reach AGI is increasingly pointing to Google DeepMind, not OpenAI, which Swyx says OpenAI is acutely aware of.

On Apple, the reported pursuit of a deal with Anthropic or OpenAI — with Anthropic reportedly seeking billions — is met with skepticism. Apple's pattern of AI partnerships, including the Amazon-Anthropic arrangement for a smarter Alexa, has not produced visible results in the wild. The deeper critique is structural: Apple wants to sit between models as a neutral routing layer rather than committing to a single AI direction, and that Switzerland strategy is increasingly seen as insufficient.

In a late-breaking note, Cursor has poached the team that built Claude Code from Anthropic. The significance is hard to overstate in developer circles. Claude Code launched only in February 2025, and Anthropic's revenue has grown 4x since the start of this year, with Claude Code widely credited as a primary driver. The hire positions Cursor as the most credible challenger to Claude Code's status as the dominant AI coding agent, a market where developer spending has been described as going "absolutely nuts."