Changes

Jump to: navigation, search

Timeline of OpenAI

3,635 bytes added, 08:22, 24 May 2020
no edit summary
** Sort the full timeline by "Event type" and look for the group of rows with value "Partnership".
** You will read collaborations with organizations like {{w|DeepMind}} and {{w|Microsoft}}.
** What are some significant fundings granted to OpenAI by donors?
** Sort the full timeline by "Event type" and look for the group of rows with value "Donation".
** You will see names like the {{w|Open Philanthropy Project}}, and {{w|Nvidia}}, among others.
| 2016 || {{dts|May 31}} || Generative models || Publication || "VIME: Variational Information Maximizing Exploration", a paper on generative models, is submitted to the {{w|ArXiv}}. The paper introduces Variational Information Maximizing Exploration (VIME), an exploration strategy based on maximization of information gain about the agent's belief of environment dynamics.<ref>{{cite web |last1=Houthooft |first1=Rein |last2=Chen |first2=Xi |last3=Duan |first3=Yan |last4=Schulman |first4=John |last5=De Turck |first5=Filip |last6=Abbeel |first6=Pieter |title=VIME: Variational Information Maximizing Exploration |url=https://arxiv.org/abs/1605.09674 |website=arxiv.org |accessdate=27 March 2020}}</ref>
|-
| 2016 || {{dts|June 5}} || {{w|Reinforcement learning}} || Publication || "OpenAI Gym", a paper on {{w|reinforcement learning}}, is submitted to the {{w|ArXiv}}. It presents OpenAI Gym as a toolkit for reinforcement learning research.<ref>{{cite web |last1=Brockman |first1=Greg |last2=Cheung |first2=Vicki |last3=Pettersson |first3=Ludwig |last4=Schneider |first4=Jonas |last5=Schulman |first5=John |last6=Tang |first6=Jie |last7=Zaremba |first7=Wojciech |title=OpenAI Gym |url=https://arxiv.org/abs/1606.01540 |website=arxiv.org |accessdate=27 March 2020}}</ref> OpenAI Gym is considered by some as "a huge opportunity for speeding up the progress in the creation of better reinforcement algorithms, since it provides an easy way of comparing them, on the same conditions, independently of where the algorithm is executed".<ref>{{cite web |title=OPENAI GYM |url=https://www.theconstructsim.com/tag/openai_gym/ |website=theconstructsim.com |accessdate=16 May 2020}}</ref>
|-
| 2016 || {{dts|June 10}} || Generative models || Publication || "Improved Techniques for Training GANs", a paper on generative models, is submitted to the {{w|ArXiv}}. It presents a variety of new architectural features and training procedures that OpenAI applies to the generative adversarial networks (GANs) framework.<ref>{{cite web |last1=Salimans |first1=Tim |last2=Goodfellow |first2=Ian |last3=Zaremba |first3=Wojciech |last4=Cheung |first4=Vicki |last5=Radford |first5=Alec |last6=Chen |first6=Xi |title=Improved Techniques for Training GANs |url=https://arxiv.org/abs/1606.03498 |website=arxiv.org |accessdate=27 March 2020}}</ref>
| 2016 || {{dts|June 16}} || Generative models || Publication || OpenAI publishes post describing four projects on generative models, a branch of {{w|unsupervised learning}} techniques in machine learning.<ref>{{cite web |title=Generative Models |url=https://openai.com/blog/generative-models/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2016 || {{dts|June 21}} || || Publication || "Concrete Problems in AI Safety" by Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané is submitted to the {{w|arXiv}}. The paper explores practical problems in machine learning systems.<ref>{{cite web |url=https://arxiv.org/abs/1606.06565 |title=[1606.06565] Concrete Problems in AI Safety |date=June 21, 2016 |accessdate=July 25, 2017}}</ref>The paper would receive a shoutout from the Open Philanthropy Project.<ref>{{cite web|url = https://www.openphilanthropy.org/blog/concrete-problems-ai-safety|title = Concrete Problems in AI Safety|last = Karnofsky|first = Holden|date = June 23, 2016|accessdate = April 18, 2020}}</ref> It would become a landmark in AI safety literature, and many of its authors would continue to do AI safety work at OpenAI in the years to come.
|-
| 2016 || {{Dts|July}} || || Team || Dario Amodei joins OpenAI<ref>{{cite web |url=https://www.crunchbase.com/person/dario-amodei |title=Dario Amodei - Research Scientist @ OpenAI |publisher=Crunchbase |accessdate=May 6, 2018}}</ref>, working on the Team Lead for AI Safety.<ref name="Dario Amodeiy"/><ref name="orgwatch.issarice.com"/>
| 2017 || {{dts|April 6}} || || Software release || OpenAI unveils an unsupervised system which is able to perform a excellent {{w|sentiment analysis}}, despite being trained only to predict the next character in the text of Amazon reviews.<ref>{{cite web |title=Unsupervised Sentiment Neuron |url=https://openai.com/blog/unsupervised-sentiment-neuron/ |website=openai.com |accessdate=5 April 2020}}</ref><ref>{{cite web |url=https://techcrunch.com/2017/04/07/OpenAI-sets-benchmark-for-sentiment-analysis-using-an-efficient-mlstm/ |date=April 7, 2017 |publisher=TechCrunch |title=OpenAI sets benchmark for sentiment analysis using an efficient mLSTM |author=John Mannes |accessdate=March 2, 2018}}</ref>
|-
| 2017 || {{dts|April 6}} || || Publication || "Learning to Generate Reviews and Discovering Sentiment" is published.<ref>{{cite web |url=https://techcrunch.com/2017/04/07/openai-sets-benchmark-for-sentiment-analysis-using-an-efficient-mlstm/ |date=April 7, 2017 |publisher=TechCrunch |title=OpenAI sets benchmark for sentiment analysis using an efficient mLSTM |author=John Mannes |accessdate=March 2, 2018}}</ref>
|-
| 2017 || {{dts|April 6}} || Neuroevolution || Research progress || OpenAI unveils reuse of an old field called “neuroevolution”, and a subset of algorithms from it called “evolution strategies,” which are aimed at solving optimization problems. In one hour training on an Atari challenge, an algorithm is found to reach a level of mastery that took a reinforcement-learning system published by DeepMind in 2016 a whole day to learn. On the walking problem the system took 10 minutes, compared to 10 hours for DeepMind's approach.<ref>{{cite web |title=OpenAI Just Beat Google DeepMind at Atari With an Algorithm From the 80s |url=https://singularityhub.com/2017/04/06/OpenAI-just-beat-the-hell-out-of-deepmind-with-an-algorithm-from-the-80s/ |website=singularityhub.com |accessdate=29 June 2019}}</ref>
|-
| 2017 || {{dts|August 18}} || {{w|Reinforcement learning}} || Software release || OpenAI releases two implementations: ACKTR, a {{w|reinforcement learning}} algorithm, and A2C, a synchronous, deterministic variant of Asynchronous Advantage Actor Critic (A3C).<ref>{{cite web |title=OpenAI Baselines: ACKTR & A2C |url=https://openai.com/blog/baselines-acktr-a2c/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2017 || {{Dts|September 13}} || {{w|Reinforcement learning}} || Publication || "Learning with Opponent-Learning Awareness" is first uploaded to the {{w|ArXiv}}. The paper presents Learning with Opponent-Learning Awareness (LOLA), a method in which each agent shapes the anticipated learning of the other agents in an environment.<ref>{{cite web |url=https://arxiv.org/abs/1709.04326 |title=[1709.04326] Learning with Opponent-Learning Awareness |accessdate=March 2, 2018}}</ref><ref>{{cite web |url=https://www.gwern.net/newsletter/2017/09 |author=gwern |date=August 16, 2017 |title=September 2017 news - Gwern.net |accessdate=March 2, 2018}}</ref>
|-
| 2017 || {{dts|October 11}} || || Software release || RoboSumo, a game that simulates {{W|sumo wrestling}} for AI to learn to play, is released.<ref>{{cite web |url=https://www.wired.com/story/ai-sumo-wrestlers-could-make-future-robots-more-nimble/ |title=AI Sumo Wrestlers Could Make Future Robots More Nimble |publisher=[[wikipedia:WIRED|WIRED]] |accessdate=March 3, 2018}}</ref><ref>{{cite web |url=http://www.businessinsider.com/elon-musk-OpenAI-virtual-robots-learn-sumo-wrestle-soccer-sports-ai-tech-science-2017-10 |first1=Alexandra |last1=Appolonia |first2=Justin |last2=Gmoser |date=October 20, 2017 |title=Elon Musk's artificial intelligence company created virtual robots that can sumo wrestle and play soccer |publisher=Business Insider |accessdate=March 3, 2018}}</ref>
| 2018 || {{Dts|May 2}} || safety || Publication || The paper "AI safety via debate" by Geoffrey Irving, Paul Christiano, and Dario Amodei is uploaded to the arXiv. The paper proposes training agents via self play on a zero sum debate game, in order to adress tasks that are too complicated for a human to directly judge.<ref>{{cite web |url=https://arxiv.org/abs/1805.00899 |title=[1805.00899] AI safety via debate |accessdate=May 5, 2018}}</ref><ref>{{cite web |url=https://blog.OpenAI.com/debate/ |publisher=OpenAI Blog |title=AI Safety via Debate |date=May 3, 2018 |first1=Geoffrey |last1=Irving |first2=Dario |last2=Amodei |accessdate=May 5, 2018}}</ref>
|-
| 2018 || {{dts|May 16}} || {{w|Computation}} || Publication || OpenAI releases an analysis showing that since 2012, the amount of compute used in the largest AI training runs has been increasing exponentially with a 3.4-month doubling time.<ref>{{cite web |title=AI and Compute |url=https://openai.com/blog/ai-and-compute/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2018 || {{dts|June 11}} || {{w|Unsupervised learning}} || Research progress || OpenAI announces having obtained significant results on a suite of diverse language tasks with a scalable, task-agnostic system, which uses a combination of transformers and unsupervised pre-training.<ref>{{cite web |title=Improving Language Understanding with Unsupervised Learning |url=https://openai.com/blog/language-unsupervised/ |website=openai.com |accessdate=5 April 2020}}</ref>
|-
| 2018 || {{Dts|June 25}} || {{w|Neural network}} || Software release || OpenAI announces set of AI algorithms able to hold their own as a team of five and defeat human amateur players at {{w|Dota 2}}, a multiplayer online battle arena video game popular in e-sports for its complexity and necessity for teamwork.<ref>{{cite web |last1=Gershgorn |first1=Dave |title=OpenAI built gaming bots that can work as a team with inhuman precision |url=https://qz.com/1311732/OpenAI-built-gaming-bots-that-can-work-as-a-team-with-inhuman-precision/ |website=qz.com |accessdate=14 June 2019}}</ref> In the algorithmic A team, called OpenAI Five, each algorithm uses a {{w|neural network}} to learn both how to play the game, and how to cooperate with its AI teammates.<ref>{{cite web |last1=Knight |first1=Will |title=A team of AI algorithms just crushed humans in a complex computer game |url=https://www.technologyreview.com/s/611536/a-team-of-ai-algorithms-just-crushed-expert-humans-in-a-complex-computer-game/ |website=technologyreview.com |accessdate=14 June 2019}}</ref><ref>{{cite web |title=OpenAI’s bot can now defeat skilled Dota 2 teams |url=https://venturebeat.com/2018/06/25/OpenAI-trains-ai-to-defeat-teams-of-skilled-dota-2-players/ |website=venturebeat.com |accessdate=14 June 2019}}</ref>
|-
| 2018 || {{Dts|June 26}} || || Notable comment || {{w|Bill Gates}} comments on {{w|Twitter}}: {{Quote|AI bots just beat humans at the video game Dota 2. That’s a big deal, because their victory required teamwork and collaboration – a huge milestone in advancing artificial intelligence.}}<ref>{{cite web |last1=Papadopoulos |first1=Loukia |title=Bill Gates Praises Elon Musk-Founded OpenAI’s Latest Dota 2 Win as “Huge Milestone” in Field |url=https://interestingengineering.com/bill-gates-praises-elon-musk-founded-OpenAIs-latest-dota-2-win-as-huge-milestone-in-field |website=interestingengineering.com |accessdate=14 June 2019}}</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>
|-
| 2019 || {{dts|December 4}} || || Publication || "Deep Double Descent: Where Bigger Models and More Data Hurt" is submitted to the {{w|ArXiv}}. The paper shows that a variety of modern deep learning tasks exhibit a "double-descent" phenomenon where, as the model size increases, performance first gets worse and then gets better.<ref>{{cite web |last1=Nakkiran |first1=Preetum |last2=Kaplun |first2=Gal |last3=Bansal |first3=Yamini |last4=Yang |first4=Tristan |last5=Barak |first5=Boaz |last6=Sutskever |first6=Ilya |title=Deep Double Descent: Where Bigger Models and More Data Hurt |website=arxiv.org |url=https://arxiv.org/abs/1912.02292|accessdate=5 April 2020}}</ref> The paper is summarized on the OpenAI blog.<ref>{{cite web|url = https://openai.com/blog/deep-double-descent/|title = Deep Double Descent|publisher = OpenAI|date = December 5, 2019|accessdate = May 23, 2020}}</ref> MIRI researcher Evan Hubinger writes an explanatory post on the subject on LessWrong and the AI Alignment Forum,<ref>{{cite web|url = https://www.lesswrong.com/posts/FRv7ryoqtvSuqBxuT/understanding-deep-double-descent|title = Understanding “Deep Double Descent”|date = December 5, 2019|accessdate = 24 May 2020|publisher = LessWrong|last = Hubinger|first = Evan}}</ref> and follows up with a post on the AI safety implications.<ref>{{cite web|url = https://www.lesswrong.com/posts/nGqzNC6uNueum2w8T/inductive-biases-stick-around|title = Inductive biases stick around|date = December 18, 2019|accessdate = 24 May 2020|last = Hubinger|first = Evan}}</ref>
|-
| 2019 || {{dts|December}} || || Team || Dario Amodei is promoted as OpenAI's Vice President of Research.<ref name="Dario Amodeiy">{{cite web |title=Dario Amodei |url=https://www.linkedin.com/in/dario-amodei-3934934/ |website=linkedin.com |accessdate=29 February 2020}}</ref>
| 2020 || {{dts|January 30}} || {{w|Deep learning}} || Software adoption || OpenAI announces migration to the social network’s {{w|PyTorch}} {{w|machine learning}} framework in future projects, setting it as its new standard deep learning framework.<ref>{{cite web |title=OpenAI sets PyTorch as its new standard deep learning framework |url=https://jaxenter.com/OpenAI-pytorch-deep-learning-framework-167641.html |website=jaxenter.com |accessdate=23 February 2020}}</ref><ref>{{cite web |title=OpenAI goes all-in on Facebook’s Pytorch machine learning framework |url=https://venturebeat.com/2020/01/30/OpenAI-facebook-pytorch-google-tensorflow/ |website=venturebeat.com |accessdate=23 February 2020}}</ref>
|-
| 2020 || {{dts|February 5}} || Safety || Publication || Beth Barnes and Paul Christiano on <code>lesswrong.com</code> publish ''Writeup: Progress on AI Safety via Debate'', a writeup of the research done by the "Reflection-Humans" team at OpenAI in third and fourth quarter of 2019.<ref>{{cite web |title=Writeup: Progress on AI Safety via Debate |url=https://www.lesswrong.com/posts/Br4xDbYu4Frwrb64a/writeup-progress-on-ai-safety-via-debate-1#Things_we_did_in_Q3 |website=lesswrong.com |accessdate=16 May 2020}}</ref>|-| 2020 || {{dts|February 17}} || || Coverage || AI reporter Karen Hao at ''MIT Technology Review'' publishes review on OpenAI titled ''The messy, secretive reality behind OpenAI’s bid to save the world'', which suggests the company is surrendering its declaration to be transparent in order to outpace competitors. As a response, {{w|Elon Musk}} criticizes OpenAI, saying it lacks transparency.<ref name="Aaron">{{cite web |last1=Holmes |first1=Aaron |title=Elon Musk just criticized the artificial intelligence company he helped found — and said his confidence in the safety of its AI is 'not high' |url=https://www.businessinsider.com/elon-musk-criticizes-OpenAI-dario-amodei-artificial-intelligence-safety-2020-2 |website=businessinsider.com |accessdate=29 February 2020}}</ref> On his {{w|Twitter}} account, Musk writes "I have no control & only very limited insight into OpenAI. Confidence in Dario for safety is not high", alluding to OpenAI Vice President of Research Dario Amodei.<ref>{{cite web |title=Elon Musk |url=https://twitter.com/elonmusk/status/1229546206948462597 |website=twitter.com |accessdate=29 February 2020}}</ref>
|-
|}
===How the timeline was built===
The initial version of the timeline was written by [[User:Issa|Issa Rice]]. It has been expanded considerably by [[User:Sebastian|Sebastian]].
{{funding info}} is available.
===What the timeline is still missing===
 
* [https://intelligence.org/blog/]
* [https://www.lesswrong.com/posts/Br4xDbYu4Frwrb64a/writeup-progress-on-ai-safety-via-debate-1]
===Timeline update strategy===
62,434
edits

Navigation menu