We find that, just as a large transformer model trained on language can generate coherent text, the same exact model trained on pixel sequences can generate coherent image completions and samples. By establishing a correlation between sample quality and image classification accuracy, we show that our best generative model also contains features competitive with top convolutional nets in the unsupervised setting.

We’re releasing an API for accessing new AI models developed by OpenAI.

We’re releasing an analysis showing that since 2012 the amount of compute needed to train a neural net to the same performance on ImageNet classification has been decreasing by a factor of 2 every 16 months. Compared to 2012, it now takes 44 times less compute to train a neural network to the level of AlexNet (by contrast, Moore’s Law would yield an 11x cost improvement over this period). Our results suggest that for AI tasks with high levels of recent investment, algorithmic progress has yielded more gains than classical hardware efficiency.

We’re introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We’re releasing the model weights and code, along with a tool to explore the generated samples.

We’ve contributed to a multi-stakeholder report by 58 co-authors at 30 organizations, including the Centre for the Future of Intelligence, Mila, Schwartz Reisman Institute for Technology and Society, Center for Advanced Study in the Behavioral Sciences, and Center for Security and Emerging Technologies. This report describes 10 mechanisms to improve the verifiability of claims made about AI systems. Developers can use these tools to provide evidence that AI systems are safe, secure, fair, or privacy-preserving. Users, policymakers, and civil society can use these tools to evaluate AI development processes.

We’re introducing OpenAI Microscope, a collection of visualizations of every significant layer and neuron of eight vision “model organisms” which are often studied in interpretability. Microscope makes it easier to analyze the features that form inside these neural networks, and we hope it will help the research community as we move towards understanding these complicated systems.

We are standardizing OpenAI’s deep learning framework on PyTorch.

We show that the double descent phenomenon occurs in CNNs, ResNets, and transformers: performance first improves, then gets worse, and then improves again with increasing model size, data size, or training time. This effect is often avoided through careful regularization. While this behavior appears to be fairly universal, we don’t yet fully understand why it happens, and view further study of this phenomenon as an important research direction.

We’re releasing Procgen Benchmark, 16 simple-to-use procedurally-generated environments which provide a direct measure of how quickly a reinforcement learning agent learns generalizable skills.

We’re releasing Safety Gym, a suite of environments and tools for measuring progress towards reinforcement learning agents that respect safety constraints while training.

As the final model release of GPT-2’s staged release, we’re releasing the largest version (1.5B parameters) of GPT-2 along with code and model weights to facilitate detection of outputs of GPT-2 models. While there have been larger language models released since August, we’ve continued with our original staged release plan in order to provide the community with a test case of a full staged release process. We hope that this test case will be useful to developers of future powerful models, and we’re actively continuing the conversation with the AI community on responsible publication.

We’ve trained a pair of neural networks to solve the Rubik’s Cube with a human-like robot hand. The neural networks are trained entirely in simulation, using the same reinforcement learning code as OpenAI Five paired with a new technique called Automatic Domain Randomization (ADR). The system can handle situations it never saw during training, such as being prodded by a stuffed giraffe. This shows that reinforcement learning isn’t just a tool for virtual tasks, but can solve physical-world problems requiring unprecedented dexterity.

We are now accepting applications for our third class of OpenAI Scholars.

We’ve fine-tuned the 774M parameter GPT-2 language model using human feedback for various tasks, successfully matching the preferences of the external human labelers, though those preferences did not always match our own. Specifically, for summarization tasks the labelers preferred sentences copied wholesale from the input (we’d only asked them to ensure accuracy), so our models learned to copy. Summarization required 60k human labels; simpler tasks which continue text in various styles required only 5k. Our motivation is to move safety techniques closer to the general task of “machines talking to humans,” which we believe is key to extracting information about human values.

We’ve observed agents discovering progressively more complex tool use while playing a simple game of hide-and-seek. Through training in our new simulated hide-and-seek environment, agents build a series of six distinct strategies and counterstrategies, some of which we did not know our environment supported. The self-supervised emergent complexity in this simple environment further suggests that multi-agent co-adaptation may one day produce extremely complex and intelligent behavior.

We’ve developed a method to assess whether a neural network classifier can reliably defend against adversarial attacks not seen during training. Our method yields a new metric, UAR (Unforeseen Attack Robustness), which evaluates the robustness of a single model against an unanticipated attack, and highlights the need to measure performance across a more diverse range of unforeseen attacks.