62,734
edits
Changes
no edit summary
| 2018 || {{dts|October 31}} || || Software release || OpenAI unveils its Random Network Distillation (RND), a prediction-based method for encouraging {{w|reinforcement learning}} agents to explore their environments through curiosity, which for the first time exceeds average human performance on videogame Montezuma’s Revenge.<ref>{{cite web |title=Reinforcement Learning with Prediction-Based Rewards |url=https://openai.com/blog/reinforcement-learning-with-prediction-based-rewards/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2018 || {{Dts|November 8}} || {{w|Reinforcement learning}} || Education || OpenAI launches Spinning Up, an educational resource designed to teach anyone deep reinforcement learning. The program consists of crystal-clear examples of RL code, educational exercises, documentation, and tutorials.<ref>{{cite web |title=Spinning Up in Deep RL |url=https://OpenAI.com/blog/spinning-up-in-deep-rl/ |website=OpenAI.com |accessdate=15 June 2019}}</ref><ref>{{cite web |last1=Ramesh |first1=Prasad |title=OpenAI launches Spinning Up, a learning resource for potential deep learning practitioners |url=https://hub.packtpub.com/OpenAI-launches-spinning-up-a-learning-resource-for-potential-deep-learning-practitioners/ |website=hub.packtpub.com |accessdate=15 June 2019}}</ref><ref>{{cite web |last1=Johnson |first1=Khari |title=OpenAI launches reinforcement learning training to prepare for artificial general intelligence |url=https://flipboard.com/@venturebeat/OpenAI-launches-reinforcement-learning-training-to-prepare-for-artificial-genera/a-TxuPmdApTGSzPr0ny7qXsw%3Aa%3A2919225365-bafeac8636%2Fventurebeat.com |website=flipboard.com |accessdate=15 June 2019}}</ref>
|-
| 2018 || {{Dts|November 9}} || || Notable comment || {{w|Ilya Sutskever}} gives speech at the AI Frontiers Conference in {{w|San Jose}}, and declares: {{Quote|We (OpenAI) have reviewed progress in the field over the past six years. Our conclusion is near term AGI should be taken as a serious possibility.}}<ref>{{cite web |title=OpenAI Founder: Short-Term AGI Is a Serious Possibility |url=https://syncedreview.com/2018/11/13/OpenAI-founder-short-term-agi-is-a-serious-possibility/ |website=syncedreview.com |accessdate=15 June 2019}}</ref>
| 2018 || {{dts|December 4}} || || Researh progress || OpenAI announces having discovered that the gradient noise scale, a simple statistical metric, predicts the parallelizability of neural network training on a wide range of tasks.<ref>{{cite web |title=How AI Training Scales |url=https://openai.com/blog/science-of-ai/ |website=openai.com |accessdate=4 April 2020}}</ref>
|-
| 2018 || {{Dts|December 6}} || {{w|Reinforcement learning}} || Software release || OpenAI releases CoinRun, a training environment designed to test the adaptability of reinforcement learning agents.<ref>{{cite web |title=OpenAI teaches AI teamwork by playing hide-and-seek |url=https://venturebeat.com/2019/09/17/OpenAI-and-deepmind-teach-ai-to-work-as-a-team-by-playing-hide-and-seek/ |website=venturebeat.com |accessdate=24 February 2020}}</ref><ref>{{cite web |title=OpenAI’s CoinRun tests the adaptability of reinforcement learning agents |url=https://venturebeat.com/2018/12/06/OpenAIs-coinrun-tests-the-adaptability-of-reinforcement-learning-agents/ |website=venturebeat.com |accessdate=24 February 2020}}</ref>
|-
| 2019 || {{Dts|February 14}} || || Software release || OpenAI unveils its language-generating system called GPT-2, a system able to write news, answer reading comprehension problems, and shows promise at tasks like translation.<ref>{{cite web |title=An AI helped us write this article |url=https://www.vox.com/future-perfect/2019/2/14/18222270/artificial-intelligence-open-ai-natural-language-processing |website=vox.com |accessdate=28 June 2019}}</ref> However, the data or the parameters of the model are not released, under expressed concerns about potential abuse.<ref>{{cite web |last1=Lowe |first1=Ryan |title=OpenAI’s GPT-2: the model, the hype, and the controversy |url=https://towardsdatascience.com/OpenAIs-gpt-2-the-model-the-hype-and-the-controversy-1109f4bfd5e8 |website=towardsdatascience.com |accessdate=10 July 2019}}</ref> OpenAI initially tries to communicate the risk posed by this technology.<ref name="ssfr"/>
| 2019 || {{Dts|March}} || || Team || {{w|Sam Altman}} leaves his role as the president of {{w|Y Combinator}} to become the {{w|Chief executive officer}} of OpenAI.<ref>{{cite web |title=Sam Altman’s leap of faith |url=https://techcrunch.com/2019/05/18/sam-altmans-leap-of-faith/ |website=techcrunch.com |accessdate=24 February 2020}}</ref><ref>{{cite web |title=Y Combinator president Sam Altman is stepping down amid a series of changes at the accelerator |url=https://techcrunch.com/2019/03/08/y-combinator-president-sam-altman-is-stepping-down-amid-a-series-of-changes-at-the-accelerator/ |website=techcrunch.com |accessdate=24 February 2020}}</ref><ref name="orgwatch.issarice.com"/>
|-
| 2019 || {{Dts|April 23}} || {{w|Deep learning}} || Publication || OpenAI publishes paper announcing Sparse Transformers, a deep neural network for learning sequences of data, including text, sound, and images. It utilizes an improved algorithm based on the attention mechanism, being able to extract patterns from sequences 30 times longer than possible previously.<ref>{{cite web |last1=Alford |first1=Anthony |title=OpenAI Introduces Sparse Transformers for Deep Learning of Longer Sequences |url=https://www.infoq.com/news/2019/05/OpenAI-sparse-transformers/ |website=infoq.com |accessdate=15 June 2019}}</ref><ref>{{cite web |title=OpenAI Sparse Transformer Improves Predictable Sequence Length by 30x |url=https://medium.com/syncedreview/OpenAI-sparse-transformer-improves-predictable-sequence-length-by-30x-5a65ef2592b9 |website=medium.com |accessdate=15 June 2019}}</ref><ref>{{cite web |title=Generative Modeling with Sparse Transformers |url=https://OpenAI.com/blog/sparse-transformer/ |website=OpenAI.com |accessdate=15 June 2019}}</ref>
|-
| 2019 || {{Dts|April 25}} || || Software release || OpenAI announces MuseNet, a deep {{w|neural network}} able to generate 4-minute musical compositions with 10 different instruments, and can combine multiple styles from [[w:Country music|country]] to {{w|Mozart}} to {{w|The Beatles}}. The neural network uses general-purpose unsupervised technology.<ref>{{cite web |title=MuseNet |url=https://OpenAI.com/blog/musenet/ |website=OpenAI.com |accessdate=15 June 2019}}</ref>
| 2019 || {{dts|September 17}} || || Research progress || OpenAI announces having observed agents discovering progressively more complex tool use while playing a simple game of hide-and-seek. Through training, the agents were able to build a series of six distinct strategies and counterstrategies, some of which were unknown to be supported by the environment.<ref>{{cite web |title=Emergent Tool Use from Multi-Agent Interaction |url=https://openai.com/blog/emergent-tool-use/ |website=openai.com |accessdate=4 April 2020}}</ref><ref>{{cite web |title=Emergent Tool Use From Multi-Agent Autocurricula |url=https://arxiv.org/abs/1909.07528 |website=arxiv.org |accessdate=4 April 2020}}</ref>
|-
| 2019 || {{dts|October 16}} || {{w|Neural network}}s || Research progress || OpenAI announces having trained a pair of {{w|neural network}}s to solve the {{w|Rubik’s Cube}} with a human-like robot hand. The experiment demonstrates that models trained only in simulation can be used to solve a manipulation problem of unprecedented complexity on a real robot.<ref>{{cite web |title=Solving Rubik's Cube with a Robot Hand |url=https://arxiv.org/abs/1910.07113 |website=arxiv.org |accessdate=4 April 2020}}</ref><ref>{{cite web |title=Solving Rubik’s Cube with a Robot Hand |url=https://openai.com/blog/solving-rubiks-cube/ |website=openai.com |accessdate=4 April 2020}}</ref>
|-
| 2019 || {{dts|November 5}} || || Software release || OpenAI releases the largest version (1.5B parameters) of its language-generating system GPT-2 along with code and model weights to facilitate detection of outputs of GPT-2 models.<ref>{{cite web |title=GPT-2: 1.5B Release |url=https://openai.com/blog/gpt-2-1-5b-release/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2019 || {{dts|November 21}} || {{w|Reinforcement learning}} || Software release || OpenAI releases Safety Gym, a suite of environments and tools for measuring progress towards {{w|reinforcement learning}} agents that respect safety constraints while training.<ref>{{cite web |title=Safety Gym |url=https://openai.com/blog/safety-gym/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2019 || {{dts|December 3}} || {{w|Reinforcement learning}} || Software release || OpenAI releases Procgen Benchmark, a set of 16 simple-to-use procedurally-generated environments (CoinRun, StarPilot, CaveFlyer, Dodgeball, FruitBot, Chaser, Miner, Jumper, Leaper, Maze, BigFish, Heist, Climber, Plunder, Ninja, and BossFight) which provide a direct measure of how quickly a {{w|reinforcement learning}} agent learns generalizable skills. Procgen Benchmark prevents AI model overfitting.<ref>{{cite web |title=Procgen Benchmark |url=https://openai.com/blog/procgen-benchmark/ |website=openai.com |accessdate=2 March 2020}}</ref><ref>{{cite web |title=OpenAI’s Procgen Benchmark prevents AI model overfitting |url=https://venturebeat.com/2019/12/03/openais-procgen-benchmark-overfitting/ |website=venturebeat.com |accessdate=2 March 2020}}</ref><ref>{{cite web |title=GENERALIZATION IN REINFORCEMENT LEARNING – EXPLORATION VS EXPLOITATION |url=https://analyticsindiamag.com/generalization-in-reinforcement-learning-exploration-vs-exploitation/ |website=analyticsindiamag.com |accessdate=2 March 2020}}</ref>
