Jack Rae

Jack Rae is an AI researcher at Meta working on frontier-model training and reasoning. Before joining Meta’s superintelligence team, he led Gemini Thinking and co-led Gemini pre-training at Google DeepMind. His research spans neural memory, influential language-model scaling results, and systems that spend additional computation reasoning through difficult problems.

Rae’s early work on memory-augmented neural networks developed sparse reading and writing mechanisms that made large external memories more practical. At DeepMind, he also helped introduce Sonnet, an open-source framework for building neural networks from reusable components.

He subsequently led the Gopher research, which investigated language-model capabilities and limitations at scales reaching 280 billion parameters.

Rae also coauthored the Chinchilla paper, which challenged the prevailing emphasis on ever-larger models. Its 70-billion-parameter system outperformed the substantially larger Gopher using approximately the same training-compute budget, demonstrating that compute-optimal training requires balancing model size against training data. He later contributed to the GPT-4 technical report before leading reasoning work within Google’s Gemini program.

• Test-time compute as a scaling frontier. Rae treats immediate answers as a computational bottleneck: difficult questions deserve more processing than simple ones. Gemini’s reasoning approach inserts an intermediate thinking stage before the final answer, allowing inference-time effort to vary with the problem.

• Reinforcement learning for adaptive reasoning. Feedback on whether tasks are solved successfully can teach models to test hypotheses, reject failed approaches, break problems into parts, write intermediate code, calculate, and use tools without prescribing an identical reasoning procedure for every request.

• Thinking budgets as product infrastructure. Adjustable reasoning budgets let developers trade capability against latency and cost. Rae identifies unnecessary overthinking as a practical engineering problem and argues that models should learn to allocate effort efficiently.

• Parallel and long-horizon reasoning. For demanding mathematical, coding, and multimodal problems, multiple reasoning paths can explore alternatives before integrating promising solutions. Rae’s longer-term ambition is long-horizon inference: systems capable of sustained reasoning that build intermediate knowledge and artifacts while pursuing difficult research problems.