✷ The archive · 11 labs · 1,214 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
Generative Language Modeling for Automated Theorem Proving
OpenAI introduces GPT-f, a transformer-based language model designed for the Metamath formalization language that successfully contributed new proofs to the main Metamath library.
Learning to Summarize with Human Feedback
OpenAI demonstrates that reinforcement learning from human feedback (RLHF) allows smaller language models to outperform significantly larger models trained only via supervised learning in text summarization.
OpenAI Scholars 2020 Final Projects
OpenAI announced the final projects of the 2020 Scholars program, showcasing research in neural network interpretability, reinforcement learning, semantic parsing, and medical AI.
OpenAI Procgen and MineRL Competitions
OpenAI is co-organizing two NeurIPS 2020 competitions using the Procgen Benchmark and MineRL to advance reinforcement learning sample efficiency and generalization.
Image GPT
OpenAI's Image GPT (iGPT) demonstrates that a transformer model trained on pixel sequences can generate coherent images and learn competitive unsupervised visual features without domain-specific architectural priors.
OpenAI API Release
OpenAI has released a general-purpose text-in, text-out API providing access to models from the GPT-3 family to enable the development of diverse AI-powered applications.
OpenAI GPT-3: Language Models as Few-Shot Learners
OpenAI introduced GPT-3, a 175 billion parameter autoregressive language model that demonstrates strong few-shot performance across diverse NLP tasks without requiring task-specific fine-tuning.
OpenAI AI and Efficiency Analysis
OpenAI's analysis reveals that algorithmic progress has reduced the compute needed to train neural networks to the same performance level by a factor of 2 every 16 months since 2012.
OpenAI Jukebox
OpenAI Jukebox is a neural network capable of generating raw audio music, including rudimentary singing, conditioned on genre, artist, and lyrics.
OpenAI: Improving Verifiability in AI Development
OpenAI and a multi-institutional coalition propose ten mechanisms to help stakeholders verify that AI systems adhere to stated ethics principles and safety standards to prevent competitive corner-cutting.
OpenAI Microscope: A Tool for Neural Network Interpretability
OpenAI Microscope is a visualization tool that systematically visualizes every neuron in several common vision models to accelerate the research into reverse-engineering neural networks.
OpenAI Standardizes on PyTorch
OpenAI has standardized its deep learning framework on PyTorch to increase research productivity and reduce iteration times for generative modeling.
OpenAI Scaling Laws for Neural Language Models
OpenAI researchers discovered that language model performance follows predictable power-law relationships with model size, dataset size, and compute, enabling the optimal allocation of training resources.
OpenAI Five: Dota 2 Reinforcement Learning
OpenAI Five is the first AI system to defeat world champions in an esports game, demonstrating that large-scale self-play reinforcement learning can achieve superhuman performance in complex, continuous state-action spaces.
OpenAI Deep Double Descent
OpenAI researchers demonstrate that CNNs, ResNets, and transformers exhibit a double descent phenomenon where performance improves, degrades, and then improves again as model size, data size, or training time increases.
OpenAI Procgen Benchmark Release
OpenAI has released the Procgen Benchmark, a suite of 16 procedurally-generated environments designed to measure reinforcement learning agents' ability to generalize skills to unseen levels.
OpenAI Benchmarking Safe Exploration in Deep Reinforcement Learning
OpenAI introduces the Safety Gym benchmark suite to standardize constrained reinforcement learning as a primary formalism for safe exploration in high-dimensional continuous control environments.
OpenAI Safety Gym Release
OpenAI has released Safety Gym, a suite of environments and tools designed to measure and improve the ability of reinforcement learning agents to respect safety constraints during training.
GPT-2 1.5B release notes
OpenAI has released the largest version of GPT-2 with 1.5 billion parameters, completing its staged release process to provide a case study for responsible AI publication.
OpenAI Solving Rubik's Cube with a Robot Hand
OpenAI has trained a pair of neural networks to solve a Rubik's Cube using a human-like robot hand, utilizing a new technique called Automatic Domain Randomization (ADR) to transfer simulation-learned skills to the physical world.
OpenAI Scholars 2020 Program Applications
OpenAI has opened applications for its third OpenAI Scholars class, a four-month program providing stipends and mentorship to eight individuals from underrepresented groups to study deep learning.
Fine-tuning GPT-2 from human preferences
OpenAI fine-tuned the 774M parameter GPT-2 model using human feedback to improve stylistic text continuation and summarization, discovering that human labelers often prefer accurate copying over novel synthesis.
OpenAI Emergent Tool Use from Multi-Agent Interaction
OpenAI researchers discovered that agents playing a simulated game of hide-and-seek developed six distinct levels of complex tool use and strategies through multi-agent co-adaptation and self-play.
OpenAI Testing Robustness Against Unforeseen Adversaries
OpenAI has introduced a new metric, Unforeseen Attack Robustness (UAR), to evaluate how well neural network classifiers defend against adversarial attacks not encountered during training.
GPT-2 6-month follow-up
OpenAI released the 774M parameter GPT-2 model and shared lessons on the risks of synthetic text, the difficulty of detection, and the necessity of staged release strategies for large language models.
OpenAI Learning Day: Cultivating Cross-Functional Expertise
OpenAI implements Learning Day, a weekly dedicated self-study day every Thursday to foster cross-functional technical growth and prevent professional stagnation.
Microsoft and OpenAI Partnership for AGI Development
Microsoft invested $1 billion in OpenAI to co-develop a hardware and software platform on Azure designed to scale toward the creation of beneficial artificial general intelligence (AGI).
OpenAI Framework for AI Safety Cooperation
OpenAI proposes four strategies to prevent collective action problems in AI development by promoting cooperation on safety norms and standards across the industry.
OpenAI Robotics Symposium 2019
OpenAI hosted its first Robotics Symposium on April 27, 2019, bringing together experts from robotics and machine learning to discuss the development of robots that learn.
OpenAI Scholars 2019 Final Projects
OpenAI released a summary of final projects from the 2019 OpenAI Scholars program, showcasing diverse applications of deep learning, reinforcement learning, and NLP across various domains.
OpenAI Fellows Fall 2018 Final Projects
OpenAI announced the completion of its second class of OpenAI Fellows, with all six participants transitioning to full-time technical staff after a six-month apprenticeship.
OpenAI Research: Transfer of Adversarial Robustness Between Perturbation Types
OpenAI researchers found that adversarial robustness does not consistently transfer between different perturbation types, suggesting that defenses must be evaluated against a diverse range of attack vectors to be truly effective.
OpenAI MuseNet
MuseNet is a deep neural network capable of generating 4-minute musical compositions across 10 instruments by learning patterns of harmony, rhythm, and style from hundreds of thousands of MIDI files.
OpenAI Sparse Transformer
OpenAI introduced the Sparse Transformer, a deep neural network that uses a reformulated attention mechanism to model sequences 30x longer than previous Transformers by reducing algorithmic complexity from O(N^2) to O(N sqrt(N)).
OpenAI Five defeats Dota 2 world champions
OpenAI Five became the first AI to defeat esports world champions in a live livestreamed match, winning two back-to-back games against the Dota 2 team OG.
OpenAI Five Finals Event Announcement
OpenAI announced a final live event for OpenAI Five on April 13, 2019, to demonstrate the competence, scalability, and human-AI interaction capabilities of its Dota 2 AI.
Implicit Generation and Generalization Methods for Energy-Based Models
OpenAI researchers developed stable and scalable training methods for energy-based models (EBMs), achieving high-quality image generation and superior out-of-distribution generalization compared to likelihood-based models.
OpenAI Scholars 2019 Program Participants
OpenAI introduced its 2019 cohort of Scholars, a multidisciplinary group of researchers focusing on reinforcement learning, natural language processing, and generative models.
OpenAI LP: Transition to Capped-Profit Structure
OpenAI announced the creation of OpenAI LP, a hybrid capped-profit company designed to raise billions in capital for compute and talent while ensuring that the primary mission of safe AGI benefits all of humanity.
OpenAI Activation Atlases
OpenAI and Google researchers introduced Activation Atlases, a visualization technique that maps interactions between neurons to reveal how neural networks represent concepts and identify decision-making weaknesses.
Neural MMO: A Massively Multiagent Game Environment
OpenAI has released Neural MMO, a persistent, large-scale environment for reinforcement learning agents that demonstrates how increased population size and species diversity drive exploration and niche formation.
OpenAI Spinning Up in Deep RL Workshop Review
OpenAI hosted a workshop based on its Spinning Up in Deep RL resource package to scale mentorship and skill development in reinforcement learning for a diverse group of participants.
AI Safety Needs Social Scientists
OpenAI argues that long-term AI safety and alignment require the integration of social science to account for human cognitive biases and the complexities of human values.
OpenAI GPT-2 Release and Implications
OpenAI introduced GPT-2, a 1.5 billion parameter unsupervised language model capable of zero-shot task performance and coherent text generation, while implementing a staged release strategy to mitigate misuse risks.
OpenAI Research: Computational Limitations in Robust Classification
OpenAI researchers have identified classification tasks where robust classifiers exist but are computationally impossible to learn, linking the hardness of robust learning to the existence of cryptographic primitives.
OpenAI Summer Fellows 2018 Final Projects
OpenAI's 2018 Summer Fellows completed research projects focusing on reinforcement learning generalization, distributed training patterns, and generative model expressivity.
How AI Training Scales: Predicting Parallelizability via Gradient Noise Scale
OpenAI discovered that the gradient noise scale, a statistical metric quantifying the signal-to-noise ratio of network gradients, can predict the maximum useful batch size for neural network training across diverse tasks.
Quantifying Generalization in Reinforcement Learning
OpenAI introduces CoinRun, a procedurally generated training environment designed to quantify and improve an agent's ability to transfer learned skills to novel situations in reinforcement learning.
OpenAI Spinning Up in Deep RL
OpenAI has released Spinning Up in Deep RL, an educational resource providing code, tutorials, and documentation to help practitioners master deep reinforcement learning.
OpenAI Learning Concepts with Energy Functions
OpenAI has developed a technique using energy functions to enable agents to learn and extract concepts that can be transferred across dissimilar environments without retraining.