Andrew Tulloch
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,530 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,530 words
Add missing citations, update stale details, or suggest a clearer explanation.
Andrew Tulloch is an Australian machine-learning researcher and engineer known for building large-scale machine-learning infrastructure at Meta (formerly Facebook), for training frontier models at OpenAI, and for co-founding Thinking Machines Lab in 2025 alongside former OpenAI chief technology officer Mira Murati [1][2]. He drew wide attention in 2025 when the Wall Street Journal reported that Meta had pursued him with a compensation package that could have been worth as much as roughly $1.5 billion over several years, a figure Meta disputed and one of the largest individual pay numbers ever reported in the technology industry [3][4]. In October 2025 he left Thinking Machines Lab and joined Meta Superintelligence Labs [1][5].
Tulloch grew up in Perth, Australia [6]. As a high-school student he represented Australia at the 39th International Chemistry Olympiad, held in Moscow in 2007, where he won a silver medal [7][8]. (Some later profiles describe his background loosely as a "math olympiad" pedigree, but the documented international result is in chemistry, paired with a strong record in university mathematics.)
He studied mathematics at the University of Sydney, graduating with first-class honours and the University Medal, and was awarded the Joye Prize as the top student in the Faculty of Science [6][7]. He then went to Trinity College at the University of Cambridge, where he completed Part III of the Mathematical Tripos and earned a master's degree in mathematical statistics with a focus on statistics and machine learning, graduating with distinction [7][9]. Reporting on his career notes that he also began doctoral study at the University of California, Berkeley before moving into industry full time [4][6].
| Item | Detail |
|---|---|
| Nationality | Australian [6] |
| Hometown | Perth, Australia [6] |
| Undergraduate | Mathematics, University of Sydney (first-class honours, University Medal) [6][7] |
| Graduate | Part III Mathematical Tripos, Trinity College, Cambridge (master's, distinction) [7][9] |
| Olympiad | Silver medal, 39th International Chemistry Olympiad, Moscow, 2007 [7][8] |
| Known for | ML infrastructure at Meta; model training at OpenAI; co-founder of Thinking Machines Lab [1][2] |
Tulloch began his professional career in finance. From late 2010 to early 2012 he worked as a strategist at Goldman Sachs in Sydney, where he structured derivatives across foreign exchange, commodities, and credit, and built statistical models to find trading opportunities [6]. He left quantitative finance for machine learning, a shift that took him to Silicon Valley [6].
Tulloch joined Facebook in 2012 and spent roughly eleven years there, becoming one of the company's more senior machine-learning engineers and a member of its Facebook AI Research (FAIR) organization [4][6]. His work concentrated on the systems layer of deep learning rather than on individual models: the libraries, compilers, and hardware-aware kernels that let large neural networks train and run efficiently at the scale Facebook operated.
He was an early contributor to PyTorch, the deep-learning framework that became one of the most widely used tools in the field, and he worked on related infrastructure including the Caffe2 framework and FBGEMM, a library of low-precision, high-performance matrix-multiplication routines used to speed up inference on server hardware [2][10]. His public profile describes a track record of low-level, hardware-aware optimization across PyTorch, Caffe2, the TVM compiler stack, and FBGEMM [10]. That kind of work, squeezing more useful computation out of each chip, sits at the center of why frontier AI systems are expensive to build and why engineers who can do it well are scarce.
In 2023 Tulloch left Meta for OpenAI, where he joined the teams training the company's frontier models [2][6]. His own summary of this period credits him with contributions to training several of OpenAI's most capable systems, including the multimodal model GPT-4o, the model released as GPT-4.5, and the o3 reasoning model [10]. The move placed him on the model-training side of the field after more than a decade focused on infrastructure, and it connected him with a group of OpenAI researchers and executives he would soon help start a company with [1][2].
In February 2025 Tulloch became one of the founding members of Thinking Machines Lab, the artificial-intelligence startup led by Mira Murati, OpenAI's former chief technology officer [1][11]. The founding group was unusually senior, drawing heavily from OpenAI's research leadership and including John Schulman, Barret Zoph, and Lilian Weng [11][12]. The lab assembled a team of around thirty researchers and engineers recruited from OpenAI, Meta, and Mistral, and within months it had raised about $2 billion in an early-stage round led by Andreessen Horowitz at a reported valuation near $12 billion, all before shipping a product [11][12].
Tulloch's presence on the founding team reflected the lab's emphasis on the engineering and infrastructure needed to build frontier models, not only the research [1][2]. He remained at the company through the first part of 2025, a stretch in which Thinking Machines Lab became one of the most closely watched and heavily funded new entrants in the field [11].
Tulloch's name became widely known in August 2025 when the Wall Street Journal reported that Meta chief executive Mark Zuckerberg had tried to recruit him with an extraordinary pay package as part of a broader effort to staff Meta's superintelligence work [3][4]. According to that reporting, the package could have been worth as much as roughly $1.5 billion over at least six years once stock and performance components were included, a sum that, if accurate, would rank among the largest individual compensation figures ever reported in technology [3][4]. Tulloch reportedly turned down the offer at the time, and his decision to stay at Thinking Machines Lab was treated as a notable example of a founder declining a very large Meta package [3][13].
Meta disputed the framing. A company spokesman, Andy Stone, called the reported figures "inaccurate and ridiculous" and noted that the headline value of such packages depends heavily on future stock performance rather than guaranteed pay [3][4]. The reporting also said that Meta had first tried to acquire Thinking Machines Lab outright before pursuing individual staff, an approach the startup rebuffed [4][14].
The story took a turn in October 2025. On October 10, Tulloch told colleagues he was leaving Thinking Machines Lab, and the company confirmed his departure, saying he had "decided to pursue a different path for personal reasons" [1][15]. Multiple outlets reported that he was joining Meta after all, moving into the company's newly organized Meta Superintelligence Labs and its TBD Lab group, the unit overseen by chief AI officer Alexandr Wang [1][5][14]. The terms of the deal he actually accepted were not disclosed, and it is not clear how closely they resembled the disputed package described in August [1][3]. He was one of the most prominent figures in a wider wave of recruiting in which Meta sought to hire from Thinking Machines Lab, though several others approached chose to remain elsewhere or return to OpenAI [14][16].
Tulloch's case became a frequently cited illustration of how expensive the competition for senior AI talent had become by 2025. Even with Meta's denial of the specific numbers, the reporting fed a broader debate about whether pay packages tied to a handful of researchers and engineers had reached levels more typical of acquiring an entire company [4][16].
Across his career Tulloch has been associated less with any single famous model and more with the unglamorous machinery that makes large models practical: efficient kernels, quantized inference, training infrastructure, and the compilers and libraries that connect research code to data-center hardware [2][10]. That specialization helps explain why he was sought after by both a well-funded startup and a company spending heavily on a superintelligence effort. He maintains a personal website and a public software presence, and his open-source contributions include work tied to the PyTorch ecosystem [10].