6/17/2020
Image GPT
Image GPT: transformer trained on pixel sequences for image completion and generation with competitive unsupervised classification.
12/3/2019
Procgen Benchmark
Procgen Benchmark: 16 procedurally-generated environments to measure reinforcement learning agent generalization and learning speed.
11/21/2019
Safety Gym
Safety Gym: open-source suite of environments and tools for measuring reinforcement learning agent progress with safety constraints.
11/5/2019
GPT-2: 1.5B release
Final GPT-2 1.5B model release with code/weights and staged release process documentation as test case for future powerful models.
9/19/2019
Fine-tuning GPT-2 from human preferences
Fine-tuned 774M GPT-2 using human feedback for summarization and text continuation tasks, achieving preference alignment with 60k human labels.
9/17/2019
Emergent tool use from multi-agent interaction
Multi-agent reinforcement learning agents discover six emergent strategies in hide-and-seek game, demonstrating self-supervised emergent complexity.
8/20/2019
GPT-2: 6-month follow-up
Release of 774M parameter GPT-2 language model with staged release strategy, open-source legal agreements, and research on misuse/benefit.
7/10/2019
Why responsible AI development needs cooperation on safety
Policy research paper on four strategies for industry cooperation on AI safety: risk communication, technical collaboration, transparency, standards.
5/17/2019
OpenAI Fellows Fall 2018: Final projects
OpenAI Fellows program conclusion with 6-month apprenticeship projects. Career/education program announcement.
3/4/2019
Neural MMO: A massively multiagent game environment
Massively multiagent game environment for RL supporting variable agent counts with persistent open-ended tasks.
2/19/2019
AI safety needs social scientists
Paper arguing AI safety alignment research needs social scientists to address human psychology and value alignment.
2/14/2019
Better language models and their implications
Large-scale unsupervised language model achieving SOTA on multiple NLP tasks without task-specific training.
12/19/2018
OpenAI Fellows Summer 2018: Final projects
OpenAI Fellows apprenticeship program completion announcement with participant transitions to core contributor roles.
12/14/2018
How AI training scales
Analysis showing gradient noise scale predicts neural network training parallelizability across diverse tasks.
12/6/2018
Quantifying generalization in reinforcement learning
CoinRun environment benchmarks agent generalization in RL; gradient noise scale predicts training parallelizability.
11/8/2018
Spinning Up in Deep RL
Educational resource with code examples, exercises, and tutorials for learning deep reinforcement learning.
11/7/2018
Learning concepts with energy functions
Energy-based model learns spatial concepts from few demonstrations with cross-domain transfer to robotics tasks.
11/5/2018
Plan online, learn offline: Efficient learning and exploration via model-based control
Model-based control approach for efficient learning and exploration combining online planning with offline learning.
10/31/2018
Reinforcement learning with prediction-based rewards
Random Network Distillation prediction-based method for curiosity-driven exploration, first to exceed human performance on Montezuma's Revenge.
10/22/2018
Learning complex goals with iterated amplification
Iterated amplification technique for AI safety enabling specification of complex goals by decomposing tasks into simpler sub-tasks.
8/13/2018
Large-scale study of curiosity-driven learning
Large-scale study examining curiosity-driven learning approaches in machine learning agents.
8/6/2018
OpenAI Five Benchmark: Results
OpenAI Five defeats 99.95th percentile Dota 2 players in best-of-three match with live audience and 100k viewers.
7/30/2018
Learning dexterity
Robot hand trained to manipulate physical objects with unprecedented dexterity using machine learning methods.
7/18/2018
OpenAI Five Benchmark
OpenAI Five benchmark match announcement with removed gameplay restrictions for competition against professional Dota players.
7/4/2018
Learning Montezuma’s Revenge from a single demonstration
Agent trained to score 74,500 on Montezuma's Revenge from single demonstration using PPO from carefully chosen state sequences.
6/22/2018
Retro Contest: Results
Results from Retro Contest showing top performers used tuning/extensions of PPO and Rainbow algorithms on Sonic benchmark.
6/17/2018
Learning policy representations in multiagent systems
General learning framework for modeling agent behavior in multiagent systems using representation learning from minimal interaction data.
6/11/2018
Improving language understanding with unsupervised learning
State-of-the-art language understanding results using transformers with unsupervised pre-training, with code release for reproducibility.
6/2/2018
GamePad: A learning environment for theorem proving
GamePad system for exploring machine learning methods applied to theorem proving in Coq proof assistant using step-by-step human supervision.
5/30/2018
OpenAI Fellows Fall 2018
OpenAI Fellows program accepting applications for 6-month AI research apprenticeship. Targets individuals without formal background in AI with mentorship and team placement.
5/16/2018
AI and compute
Analysis showing AI compute in largest training runs increased 300,000x since 2012 with 3.4-month doubling time versus 2-year Moore's Law period.
5/3/2018
AI safety via debate
AI safety technique training agents to debate topics with human judge. Proposes approach to align advanced AI systems with human preferences with proof-of-concept experiments.
4/10/2018
Gotta Learn Fast: A new benchmark for generalization in RL
RL benchmark based on Sonic the Hedgehog measuring transfer and few-shot learning. Presents baseline algorithms and evaluation methodology.
4/5/2018
Retro Contest
Transfer learning RL competition measuring generalization from previous experience on unseen video game levels. Uses Gym Retro platform with new benchmark.
3/20/2018
Variance reduction for policy gradient with action-dependent factorized baselines
Action-dependent factorized baselines for variance reduction in policy gradient methods. Bias-free approach exploiting stochastic policy structure without MDP assumptions.
3/8/2018
On first-order meta-learning algorithms
Analysis of first-order meta-learning algorithms for learning parameter initializations. Generalizes first-order MAML using only first-order derivatives for meta-learning updates.
3/7/2018
Reptile: A scalable meta-learning algorithm
Reptile: scalable meta-learning algorithm using repeated task sampling and SGD updates. Mathematically similar to first-order MAML requiring only black-box optimizer access.
3/3/2018
Some considerations on learning to explore via meta-reinforcement learning
Two new meta-reinforcement learning algorithms (E-MAML, E-RL²) for exploration problems. Evaluated on Krazy World and maze environments with improved performance.
2/26/2018
Multi-Goal Reinforcement Learning: Challenging robotics environments and request for research
Technical report introducing multi-goal RL benchmark with continuous control tasks for Fetch robotic arm and Shadow hand. Sparse reward framework integrated with OpenAI Gym.
2/15/2018
Interpretable machine learning through teaching
Method for training AIs to teach each other using human-interpretable examples. Automatically selects informative examples to teach concepts effectively to both AI and human learners.