The Great AI Talent Reshuffle: OpenAI Alumni, SSI's Bet, and the New Lab Order

In five months between May and October 2024, OpenAI lost its chief technology officer, its chief research officer, its VP of research, its company president to an extended sabbatical, and the co-founder who had championed the neural scaling hypothesis that gave the company its earliest technical edge. Sam Altman and a reconstituted leadership team inherited the most commercially successful AI company in history—and a research organization whose founding intellectual layer had scattered to new ventures, competitor headquarters, and one deliberately secretive no-product lab armed with a billion dollars and a single goal.

Nearly two years on, the labs those researchers built are no longer hypothetical. Safe Superintelligence Inc., Thinking Machines Lab, and Anthropic have raised billions, staked out distinct philosophical positions, and are competing for the researchers who will define what comes after GPT-4-class systems. What follows is a map of who left, where they landed, what they are being paid, and what each lab's thesis implies for everyone else.

The 2024 Departure Wave

The fracture lines inside OpenAI were visible before 2024—the November 2023 board crisis that briefly ousted Sam Altman had made them public—but the concentration of senior departures across a single summer and autumn was without parallel for a company at OpenAI's stage. Based on public announcements and contemporaneous press reports, the documented timeline:

• February 2024: Andrej Karpathy, who had returned from Tesla in early 2023, left OpenAI for the second time. He subsequently worked on AI education content and early-stage research projects, describing the departure as a return to individual-contributor work.
• May 14, 2024: Ilya Sutskever, co-founder and former chief scientist, announced his departure on X after months of internal distance following the board crisis. He wrote that he had made the decision to leave after almost a decade and expressed confidence that OpenAI would pursue AGI safely, adding that what he planned to do next was something close to his heart.
• August 6, 2024: Greg Brockman, co-founder and president, announced an extended personal sabbatical on X with no public return timeline established.
• August 14, 2024: John Schulman, who co-designed the reinforcement learning from human feedback (RLHF) methodology underlying ChatGPT and held a co-founder title, announced he was joining Anthropic. He described the move as a deliberate choice to focus on AI alignment research, citing Anthropic's institutional depth in that area.
• September 26, 2024: Mira Murati, chief technology officer since 2023, announced her departure. Within days, chief research officer Bob McGrew and VP of research Barret Zoph also left. Three of OpenAI's most senior research-facing executives departed in under a week.

Left standing: CEO Sam Altman and newly appointed chief scientist Jakub Pachocki. The commercial pipeline held—GPT-4o had shipped in May 2024 (MMLU: 88.7%) and the o1 reasoning model followed in September 2024, achieving state-of-the-art scores on competition mathematics benchmarks including AIME 2024. The founding intellectual layer was gone.

Safe Superintelligence Inc. — The No-Product Bet

Sutskever moved fastest. On June 19, 2024, five weeks after his public OpenAI exit, SSI was incorporated with co-founders Daniel Gross—a former Apple AI lead and Y Combinator partner—and Daniel Levy, an OpenAI researcher. The company announced itself before it had offices, a product roadmap, or a model in training. The absence of a product roadmap was the thesis.

SSI's stated commitments: no investor pressure for near-term revenue, no products until safe superintelligence is achievable, and a singular long-term goal. On September 4, 2024, SSI disclosed it had raised $1 billion from investors including Andreessen Horowitz and Sequoia Capital at an implied post-money valuation of approximately $5 billion.

Sutskever's intellectual biography makes the bet legible. He was among the most committed proponents of the scaling hypothesis at OpenAI—the idea that capabilities track predictably with compute and data—but by 2024 had grown publicly alarmed about what a scaling-driven AGI transition would require from a safety standpoint. SSI is structured to decouple those timelines: push safety research ahead, and deploy only when safety leads capability, not the reverse.

Thinking Machines Lab — Murati's Multimodal Thesis

Where SSI went quiet, Murati moved publicly and fast. Within weeks of her September 26 departure, Bloomberg and Reuters reported she was in active fundraising conversations for a lab named Thinking Machines Lab, targeting approximately $2 billion at a post-money valuation above $10 billion. The reported focus: multimodal AI systems and agentic capabilities—precisely the frontier where Murati had overseen OpenAI's most commercially consequential work, including GPT-4V and the real-time voice mode for ChatGPT launched in 2024.

Murati's bet is structurally different from SSI's. Where Sutskever argues that commercial pressures must be eliminated to do safety research properly, Murati's thesis—based on press reporting about her stated vision—is that safety and commercial development are compatible if you build multimodal agentic systems with careful human-feedback loops from the start. Her specific competitive asset: firsthand knowledge of where the GPT-4 family's multimodal systems failed in production—hallucination in vision tasks, latency in voice pipelines, compounding errors in agentic chains—knowledge that a team starting from scratch would take years to reconstruct.

Anthropic — The Incumbent Insurgent

While SSI and Thinking Machines attracted the loudest announcement coverage, the most concrete landing spot for departing OpenAI researchers has been Anthropic—itself founded in 2021 by Dario Amodei, Daniela Amodei, and seven other OpenAI alumni who left over deployment safety disagreements. Schulman's August 2024 arrival was the most prominent addition: he named alignment research and Anthropic's interpretability team—led by Chris Olah, whose mechanistic interpretability work has been among the most cited safety research of the 2020s—as the pull factor.

Anthropic's competitive positioning is unusual because it has both an ideological narrative and commercial traction. Claude 3.5 Sonnet (June 2024) posted MMLU of 88.7%, matching GPT-4o, while outperforming it in third-party code and legal-reasoning benchmarks. OpenAI's subsequent o1 model introduced chain-of-thought reasoning at a distinct capability level for competition mathematics, but Anthropic maintained competitive positions in software engineering tasks measured by SWE-bench and continued enterprise adoption through early 2025.

Funding context: by end-2024, Anthropic had raised approximately $7–8 billion in aggregate, with Amazon committing up to $4 billion and Google participating across multiple rounds. Valuation estimates in press reporting ranged between $15 billion and $40 billion—a spread reflecting genuine uncertainty about revenue multiples in a market where no AI lab has demonstrated a credible path to profitability at scale.

What Compensation Has Become Public

AI researcher compensation grew less opaque in 2024–2025 as tender offers, partial filings, and sourced reporting by The Information, Bloomberg, and Reuters forced partial disclosure. The following figures derive from press coverage, not official lab disclosures:

• OpenAI's 2024 tender offers priced common shares at approximately $86–90 per share. Combined with base salary and annual bonus, total compensation for senior researchers was reported clearing $1 million annually.
• The Information reported in late 2024 that Anthropic was offering equity refresh packages of $2–4 million in four-year grants for principal researchers, competitive with senior engineering compensation at Google DeepMind.
• SSI has not disclosed compensation ranges publicly; investors confirmed the company is paying at market rates for frontier research talent. In the 2025 talent market, that implied total packages of $800,000–$2 million or more for mid-to-senior researchers, based on disclosed ranges at comparable labs.

The DeepSeek Shock and Its Geopolitical Fallout

The talent market does not exist apart from geopolitical context. On January 20, 2025, the Chinese lab DeepSeek released its R1 reasoning model alongside a technical report claiming training costs of approximately $6 million for a 671-billion-parameter mixture-of-experts architecture that matched or exceeded OpenAI o1 performance on MATH, AIME 2024, and several code benchmarks. On January 27, 2025, Nvidia's share price fell approximately 17% in a single session—erasing close to $593 billion in market capitalization, among the largest single-day losses for any company in market history—as investors reassessed how durable US compute infrastructure advantages actually were.

The talent market implications ran in two directions. First, the argument that only US-lab-scale compute could produce frontier research visibly weakened: if a team operating under chip export restrictions could reach o1-class reasoning for $6 million, the structural moats that large US labs built around H100 clusters became less decisive as a recruitment selling point. Second, congressional scrutiny of US chip export controls—in effect since October 2023, restricting sales of Nvidia H100 and A100 chips to China—intensified sharply, because those controls had demonstrably not prevented the R1 capability level from emerging.

For individual researchers choosing where to work, the clearest takeaway was that algorithmic research efficiency matters more than raw compute access, and that labs built around research intensity rather than compute scale are credible environments in a way that was less obvious two years earlier.

Three Theses, One Race

Reading the post-departure landscape together, three distinct theories of change are operating simultaneously:

1. SSI — decoupled safety: The race to AGI will produce a capability discontinuity that product-driven companies are structurally unable to navigate safely. The required response is a lab with no product obligations, no quarterly board pressure, and researchers whose incentives are not tied to revenue KPIs. Risk: irrelevance if the discontinuity arrives later than the thesis assumes, or if safety research actually requires product deployment data to advance—a feedback loop SSI's structure forecloses by design.
2. Thinking Machines — institutional knowledge as moat: The next significant capability gains are in multimodal and agentic systems, and the researchers who built those systems at OpenAI understand the failure modes that a fresh team would spend years rediscovering. Risk: institutional knowledge from transformer-era systems may transfer poorly to whatever architecture succeeds them.
3. Anthropic — constitutional safety plus commercial traction: Safety research and commercial success reinforce each other when enterprise customers care—as they increasingly do—about model reliability and controllability in consequential workflows. Risk: safety constraints slow iteration speed, ceding market share to competitors willing to deploy with thinner margins.

What all three share: a conviction that the next two to four years will produce at least one order-of-magnitude capability improvement, and that the lab culture, incentive structure, and research agenda established now will shape what emerges from that improvement—and who controls it. The 2024 departures from OpenAI were not an ending; they were a redistribution of the bets.