Wojciech Zaremba
Wojciech Zaremba is an OpenAI co-founder and research scientist whose work spans neural networks, robotics, and human feedback. He studied mathematics and computer science in Warsaw and Paris before completing a computer science PhD at NYU. His papers include an LSTM dropout method and studies of dexterous robotic hands. At OpenAI, he led the robotics team behind a Rubik’s Cube-solving robot hand and led a team responsible for human-feedback infrastructure. His neural-network paper became runner-up for ICLR’s 2024 Test of Time award, and he coauthored the OpenAI o1 System Card on safety evaluations.
Timeline
-
2012–2013
Finding a PhD path
Mathematics draws him from childhood, with numbers and proofs as early interests. He studies mathematics and computer science at the University of Warsaw and is finishing a master’s at École Polytechnique in Paris when he starts seeking PhD programs. In September 2013, he enrolls in NYU’s computer science PhD program and works alongside Rob Fergus in the CILVR Lab.
-
2013–2016
From neural networks to OpenAI
With Ilya Sutskever and Oriol Vinyals, he publishes “Recurrent Neural Network Regularization” in 2014; the paper shows that dropout reduces overfitting in LSTMs. In 2015, OpenAI names him among its founding members. He completes his NYU PhD with a 2016 dissertation titled “Learning Algorithms from Data.”
-
2018
Learning hand dexterity
He coauthors “Learning Dexterous In-Hand Manipulation,” which uses reinforcement learning to teach a robotic hand to reorient objects in the physical world after training in simulation.
-
2019–2020
A Rubik’s Cube in hand
OpenAI’s robotics team, led by Zaremba, trains a robot hand to solve a Rubik’s Cube through simulation and reinforcement learning. He coauthors “Solving Rubik’s Cube with a Robot Hand,” which describes automatic domain randomization and transfer from simulation to a physical robot.
-
March 2024
A team for human feedback
At OpenAI, he leads the team responsible for human-feedback infrastructure that guides model behavior through reinforcement learning.
-
May 2024
An enduring network question
ICLR names his coauthored paper “Intriguing properties of neural networks” runner-up for its 2024 Test of Time award. The paper highlights how small input changes can make a neural network behave in undesirable ways.
-
December 2024
Safety testing for o1
He coauthors the OpenAI o1 System Card, which describes safety evaluations, external red-team testing, and Preparedness Framework reviews for o1 and o1-mini.
Sources and credits
Photo: We have no published photo of Wojciech Zaremba yet. Send us a free-licensed photo.