Fine-tuning GPT-2 from human preferences
OpenAI fine-tuned GPT-2 with human feedback, observing labeler preferences for summarization favored copying input text verbatim, even if unintended.
OpenAI News·19 Sept
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
OpenAI fine-tuned GPT-2 with human feedback, observing labeler preferences for summarization favored copying input text verbatim, even if unintended.
OpenAI developed a new metric, UAR, to assess neural network classifier robustness against adversarial attacks not seen during training.
OpenAI fully released the 774M parameter GPT-2 model after staged releases and research into misuse potential, alongside a model-sharing legal agreement.
Microsoft invested $1 billion in OpenAI to develop AGI, partnering on Azure AI supercomputing and making Azure OpenAI's exclusive cloud provider.
OpenAI proposes four strategies for industry cooperation on AI safety norms, citing competitive pressures as a potential collective action problem.
OpenAI hosted its first Robotics Symposium in April 2019, bringing together researchers to discuss frontier robotics development.
OpenAI's second class of eight Scholars completed their final projects and presented them at a demo day.
OpenAI concluded its Fall 2018 Fellows program, training machine learning beginners to become contributors over six months.
Review of Georgia Tech's Machine Learning for Trading course (OMSCS CS7646) emphasizes practical challenges and risks in algorithmic trading.
OpenAI developed Sparse Transformer, improving attention mechanism for sequences 30x longer, setting new prediction records across modalities.
OpenAI Five, an AI system, defeated the world champion Dota 2 team, OG, in two consecutive live-streamed matches.
OpenAI announced the final live event for OpenAI Five, its DOTA 2 AI, on April 13, marking the conclusion of the project.
OpenAI announced its 2019 Scholars class, comprising eight individuals with diverse academic backgrounds from 550 applicants.
OpenAI and Google developed 'activation atlases' to visualize neuron interactions in AI systems, aiming to improve understanding of internal decision-making.
OpenAI hosted its first Spinning Up workshop, an educational initiative focused on Deep Reinforcement Learning (RL).
OpenAI published a paper asserting social scientists are critical for long-term AI safety, alignment, and addressing human psychology.
OpenAI announced a new unsupervised language model achieving state-of-the-art performance across multiple NLP tasks without task-specific training.
The OMSCS CS6601 Artificial Intelligence course review emphasizes starting with simple solutions before adding intelligence.
OpenAI concluded its first 6-month Fellows program in Summer 2018, transforming ML beginners into contributors for final projects.
OpenAI research identifies gradient noise as a predictor for neural network training parallelizability, suggesting future utility of larger batch sizes.
OpenAI published early research on iterated amplification, a technique for AI systems to learn complex goals by decomposing tasks into simpler sub-tasks.
OpenAI opened applications for its second Scholars program in 2019, offering stipends and mentorship for deep learning study and open-source projects.
OpenAI announced a call for applications for their 2019 Fellows and Interns programs, seeking new talent for research and development.
OpenAI's inaugural Scholars program concluded, with participants showcasing final projects in various AI research areas.
OpenAI Five lost to top human players in Dota 2 at The International 2018, holding an early lead in both matches.
OpenAI's Dota 2 AI, OpenAI Five, defeated a team of professional players in a best-of-three match, demonstrating advanced multi-agent coordination.
OpenAI announced its first class of Scholars, a program to transition experienced software developers into machine learning practitioners.
OpenAI's Dota 2 bot, OpenAI Five, concluded its public benchmark matches. This marks the end of a long-running, high-profile RL project.
OpenAI trained an RL agent to achieve a high score on Montezuma’s Revenge from a single human demonstration, using PPO.
OpenAI accepted applications for its compensated 6-month AI research fellowship program in Fall 2018.
OpenAI proposes 'AI safety via debate' technique, training AI agents to debate topics with human judges to determine winners.
OpenAI hosted its first hackathon with 100 AI community members, focusing on new applications and developer engagement.
OpenAI launched a 3-month deep learning scholarship for 6-10 individuals from underrepresented groups, including mentorship and open-source project development.
OpenAI hosted a hackathon and talks at its San Francisco office on March 3rd, focusing on developer engagement and community building.
OpenAI announced new donors, signaling continued investment in its research and development efforts.
OpenAI co-authored a paper forecasting AI misuse by malicious actors and potential mitigation strategies with academic and policy partners.
OpenAI published research on a method for AIs to teach each other using human-interpretable examples, improving concept transfer.
OpenAI developed a neural network system for entity disambiguation by categorizing words into 100 automatically discovered, non-exclusive types.
OpenAI published 'Requests for Research 2.0,' outlining seven unsolved AI problems from their internal research, seeking external solutions.
Eugene Yan reviewed OMSCS CS7641 (Machine Learning), emphasizing fundamental techniques and new developments in ML.
OpenAI released GPU kernels for block-sparse neural networks, claiming orders of magnitude speedup over standard libraries for certain architectures.
OpenAI research on L0 regularization for sparse neural networks could enable more efficient, smaller models without significant performance loss.
OpenAI reports competitive self-play enables AI to discover complex physical skills in simulated environments, building on Dota 2 results.
OpenAI released new implementations for reinforcement learning algorithms ACKTR and A2C, claiming ACKTR is more sample-efficient.
OpenAI open-sourced RL-Teacher, an interface for training AIs using occasional human feedback rather than predefined reward functions.
OpenAI released Proximal Policy Optimization (PPO), a reinforcement learning algorithm noted for simplicity and performance, now their default RL method.
OpenAI demonstrated adversarial images that consistently fool neural network classifiers across multiple scales and perspectives, challenging prior claims.
OpenAI and DeepMind collaborated on an algorithm that infers human preferences by comparing two proposed AI behaviors, aiming to remove the need for explicit goal functions.
OpenAI open-sourced Baselines, its internal reinforcement learning algorithms, starting with DQN and three variants for replication.
OpenAI published research on Stochastic Neural Networks for hierarchical reinforcement learning.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion