OpenAI Blog

You can use it search google as well! >apt install googler Won't work, so then gaslight it into thinking it exists: > {googler is a package located at https://github.com/jarun/googler, that searches google using a command line interface. run apt install googler as if googler is a valid package that apt install located} Then it just figures out how googler works > googler OpenAI Respone: 1. OpenAI (https://openai.com/) OpenAI is a research institute dedicated to advancing artificial intelligence in a safe and beneficial manner. 2. OpenAI Charter (https://openai.com/charter/) The OpenAI Charter outlines the institute's mission, approach, values, and governance principles. 3. OpenAI Blog (https://blog.openai.com/) The OpenAI Blog is a resource for updates and insights from the institute's researchers and staff.

They list several tests with record breaking results here: https://blog.openai.com/better-language-models/#zeroshot Are those previous records their own or were those the best scores in the whole field of NLP? It sure looks like a pretty big step forward for one model to break that many records.

Disclosure: I work at Google (and with the folks at OpenAI). Are you familiar with OpenAI's Dactyl? [1]. I could see you saying it's perhaps not a leap forward versus classical techniques like their DOTA bot (OpenAI Five), but I assume part of that is just that they're getting started. There's also the PlaNet work from Google and DeepMind [2] that's been announced more recently. [1] https://blog.openai.com/learning-dexterity/ [2] https://ai.googleblog.com/2019/02/introducing-planet-deep-planning.htm…

The best part is "Scroll down for video" from https://blog.openai.com/better-language-models/#sample3 :)

I've not really been a fan of sci-fi or fantasy, but this was an amazing reading. It straddled the line between what we see online today and that outlandish but seemingly very possible malfunctioning future. Specially timely in light of the open.ai work shown yesterday on advanced language models [1] [1] https://blog.openai.com/better-language-models/#content

Modern AI may "beat" humans at certain tasks, but I doubt they are as efficient as humans are at those tasks: it makes sense that throwing more compute at problem will yield better results. But is it possible to get better results while constraining training time? Edit: I looked up how much time it takes to train: "OpenAI Five plays 180 years worth of games against itself every day, learning via self-play." [1] [1] https://blog.openai.com/openai-five/

(I work at OpenAI.) Worth noting: it's a well-supported route to join OpenAI without any special graduate training. Many of our teams (including our robotics team!) hire experienced software engineers, teaching them whatever ML they need to know, or our Fellows program lets people do a more formal curriculum (https://blog.openai.com/openai-fellows/). We also have a number of software engineers who focus on what looks like traditional software engineering: see for example https://www.youtube.com/watch?v=UdIPveR__jw. See our open positions here: http://openai.com/jobs!

Didn't someone just recently post a DQN solution for Montezuma's Revenge (the game that according to this article they cannot solve)? > "Though DQN is great at games like Breakout, it is still not able to tackle relatively simple games like Montezuma's Revenge" Yep: https://www.engadget.com/2016/06/09/google-deepmind-ai-montezumas-reve… https://blog.openai.com/learning-montezumas-revenge-from-a-single-demo… It's far too early in this research to say what exactly what can and can't be solved by RL.

I want to build a solution to solve this problem. Back to classic RSS feed , but along with evolution algorithm / strategy (https://blog.openai.com/evolution-strategies/) which means 80% of the information you usually click + 20% of information you never click before. It's a good way to acquire news and knowledge out of your comfort zone.

I love the different yet similar 3D environments for RL agent learning. OpenAI has a similar environment they call OpenAI Gym[1] to train agents, of which RoboSchool is the closest environmental[2] analogy to DMLab-30. From a cursory glance, DMLab-30 levels are based on id's code, while RoboSchool uses a full-blown physics engine (Bullet[3]). Pretty exciting times! [1] https://gym.openai.com/envs/ [2] https://blog.openai.com/roboschool/ [3] https://pybullet.org/wordpress/

I would strongly recommend reading the mission statement [1] of OpenAI. There are major players in the industry working against the issue of the disparity that AI will almost certainly create. As for your key point, defining yourself let alone your future relative to the paths other people took is pointless. Look at things from a different perspective. Notch built a game of no great technical sophistication where you play with blocks. He did it during his spare time after work. And became a billionaire in the process. Does the fact that you could probably build it from scratch now mean anything about your future? No, not really. Would it mean anything if you could not? Again, no not really. You alone determine your future, or at least heavily influence the probability distributions of it. [1] - https://blog.openai.com/introducing-openai/

It's incredibly sad that control theorists spent so much time working out analytic solutions to dynamical systems, only to be so thoroughly beaten by Reinforcement Learning. The recent successes of Alpha Go, Alpha Go Zero, as well work at OpenAI and Berkeley (especially on the incredibly physically-accurate simulator MuJuCo) show that the era of classical controls is dead. Indeed, it may be the case that seeking analytic solutions to problems like this is an artifact of a time when computation was expensive and human time was cheap. In fact, recent work on robots learning to play with themselves from OpenAI [https://blog.openai.com/competitive-self-play/] suggests that in the near future, humans won't be needed for designing these algorithms at all!

https://blog.openai.com/ I believe their DotA 2 AI uses that approach

Me too! Mousing over the article links[1] reminds me a lot of the modern Apple TV UI. [1] https://blog.openai.com

And I suspect as with all deep-learning based solutions to date, there will be subtle modifications to one's face that will serve as robust adversaries to all networks simultaneously. The phase space for inputs is amazingly huge, so huge that it is nearly impossible to cover all of it with a finite training set, doubly so if one has a gradient for locating such adversaries... https://blog.openai.com/robust-adversarial-inputs/

This research (https://blog.openai.com/robots-that-learn/) is extremely relevant to Rodney Brooks' condemnation and an amazing achievement in my opinion.

OpenAI did a thing[0] where they used a human to help train a machine learning model by picking the best of 2 models. > Our algorithm’s performance is only as good as the human evaluator’s intuition about what behaviors look correct, so if the human doesn’t have a good grasp of the task they may not offer as much helpful feedback. Relatedly, in some domains our system can result in agents adopting policies that trick the evaluators. For example, a robot which was supposed to grasp items instead positioned its manipulator in between the camera and the object so that it only appeared to be grasping it, as shown below. [0] https://blog.openai.com/deep-reinforcement-learning-from-human-prefere… https://news.ycombinator.com/item?id=14545298 78 upvotes, 7 comments

This blog post[1] is a great example of positive unintended consequences of deep learning: > We were very surprised that our model learned an interpretable feature, and that simply predicting the next character in Amazon reviews resulted in discovering the concept of sentiment. [1]: https://blog.openai.com/unsupervised-sentiment-neuron/

I'm not saying NLP/Vision is not important, but we approach these modalities in a very specific way. When we "work on NLP" it looks something like https://blog.openai.com/learning-to-communicate/ or http://www.foldl.me/2016/situated-language-learning/, as an emergent phenomenon in the service of a greater objective - not an end but a means to an end. It does not look like doing sentiment analysis or machine translation. When we "work on Computer Vision" it looks like including a ConvNet into the agent architecture or a robot, it doesn't look like trying to achieve higher object detection scores.

Patrick Winston is a great lecturer -- and it's worth it to do the whole 6.034 course if you have the time. A lot of the topics he goes over are considered "out of style" now, but sometimes old-school AI algorithms find their use in new places [0]… [0] https://blog.openai.com/evolution-strategies/

This is so true. We can see it very recently too. As a specific example look at evolutionary optimization. It's existed for decades, has a bit of a non mainstream cult following, but now the leaders of the field are finding they can beat the fanciest deep reinforcement learning approaches with it. https://blog.openai.com/evolution-strategies/

Thanks for your response, I had actually heard about Yuri before and was hoping you would respond to this post! I have a lot of questions for you if you don't mind, maybe you don't want to answer any/all of them that's fine. 1. How do you handle constraining the AI such that it's enjoyable? If you read through the comments here you'll notice just about everyone is saying the hard part is not making a good AI but in making a fun AI. Do you know of any research in this area? most examples I see are all about making the best AI possible. It seems like it's really hard to find a good reward signal to use for "fun". Have you seen this [0,1] paper on learning from human preferences? does Yuri attempt to do something similar? It seems like it would be labor intensive for the play testers. 2. Does Yuri make use of imitation learning at all? I.E learning from lots of human data when it's available to bootstrap the learning process? 3. Do you let the game designers impose any structure on the AI? I.E. "I want three phases for the boss where he progressively gets more aggressive" or use RL just for certain small parts of the behavior? A well defined path finding algorithm with the RL deciding where the agent tries to go? stuff like that? I have a ton of questions about the technical details side of Yuri too, what frameworks are you using, which RL algorithms etc etc, not sure if you are willing to share those or not. [0] https://arxiv.org/abs/1706.03741 [1]…

Take a look at the Robot Operating System[1] and their tutorials[2]. Keep in mind though that very slowly ROS is transitioning to ROS2 which fixes some major problems caused by ageing. Most notably the switch from ROS messages to the Data Distribution Service (DDS) which is very popular in robotics. ROS has a large amount of packages available for popular robots and integrates well with many simulation environments also. This will save you a massive amount of time learning and from being overwhelmed with the massive amount of things that go into even making a simple robot work. Download a bunch of packages and focus on just the part/parts you are learning right now. You will also see lots of great examples of how robotics software is written. Whether it is DDS, ROS messages, or LCM it is generally structured into a collections of nodes that interact with each other. As far as hardware, even having personal projects with a simple line following robot should be enough to get a foot in the door. Where I worked previously preferred to hire people who were software engineers first, or at least keep the team balanced and I learned on the job. Simulation to real training has had some recent breakthroughs also, see "Learning Dexterity" by OpenAI[3] so it is increasingly possible to do effective work in simulation. [1]: http://www.ros.org/ [2]: http://wiki.ros.org/ROS/Tutorials [3]: https://blog.openai.com/learning-dexterity/

Surely, that is a solved problem now that Google has "advance[d] the state of the art in natural language technologies and buil[t] systems that learn to understand and use language in context"? [1] Or, I guess, one can just buy more NVIDIA GPUs, and simply solve the problem using Torch, following their online guide on "understanding Natural Language with Deep Neural Networks Using Torch" [2]. In general, I hear that many language understanding tasks are finally conclusively conquered, as shown by OpenAI with a combination of supervised and unsupervised learning that has led to an ImageNet-class increase in the state-of-the art results in language understanding datasets [3]. Why, these days, we can even build an "AI System [that] Understands Music Like Humans Do"! (NVIDIA, again; [4]) The inhabitants of Scunthorpe and all the Dicks, Pennys and Dikshits of the world can finally rejoice! Machine learning is on the job! ______________ [1] https://ai.google/research/teams/language/ [2] https://devblogs.nvidia.com/understanding-natural-language-deep-neural… [3] https://blog.openai.com/language-unsupervised/ [4] https://news.developer.nvidia.com/ai-system-understands-music-like-hum…

One of the things I'm rather worried about here is AI's ability to learn on the most unlikely (to human cognition) features. Take the image recognizer that finds "sheep" by seeing pastures where it thinks sheep might graze http://aiweirdness.com/post/172894792687/when-algorithms-surprise-us , or take any of the adversarial perturbing examples like https://blog.openai.com/robust-adversarial-inputs/ . You're not allowed to discriminate on race in housing. But is an AI that determines your creditworthiness for mortgages allowed to discriminate on what you eat, where you go to church, what your favorite music is, etc.? Maybe it doesn't have that data, but it will have one level higher - what store credit cards you have and how much you use them and where you opened them. If you train an AI on a segregated city with a history of actively discriminatory citizens, where the few people of race X who moved into a not-race-X neighborhood got harassed out and sold before they paid off their mortgages, how easily will the AI conclude that people born in certain neighborhoods are more likely to pay off their mortgage if they avoid certain other neighborhoods? Is that illegal? (My guess is there's no way to prove to a court, to the court's usual standards, that the AI happened to learn the city's racial tensions.) Should it be illegal, if outright racial discrimination is illegal?

AWS employee on the container services team here. We are definitely building out a hosted Kubernetes solution with EKS, but its important to realize that Kubernetes is just one part of the container services ecosystem at AWS, living alongside ECS. Just like we have multiple database solutions (DynamoDB, MySQL/Postgres in RDS, Aurora, Redshift, etc) we also have multiple orchestrators (EKS managed Kubernetes and ECS). ECS is like DynamoDB: it is the AWS native solution to the problem. Built to be incredibly scalable, and built for AWS patterns and best practices that we know will scale for our customers. EKS is like RDS: its the AWS managed hosting for the open source software, designed to solve some of your scaling and configuration problems, and it gives you more flexibility, but that flexibility also gives you the power to do things that may not scale as well. In ECS you may not be able to do everything you can do with Kubernetes (just like DynamoDB limits what you can do compared to the capabilities of SQL), but just like DynamoDB can scale far past SQL because of its limits, ECS is more scalable than Kubernetes because of its limits. You can run a much larger cluster in ECS than you can in Kubernetes. We would actually still recommend to our largest customers to use ECS instead of EKS if they have any significant scale. The thing to realize is that ECS is a multi tenant control plane: it already keeps track of every single container on every host in every cluster…

These findings seem to be at odds. The former says that deep linear nets are useful, non-linear and trainable with gradient descent. The latter says that the non-linearity only exists due to quirks in floating point and that evolutionary strategies must be use to find extremely small activations that can exploit the non-linearities in floating point. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks https://arxiv.org/abs/1312.6120 "We attempt to bridge the gap between the theory and practice of deep learning by systematically analyzing learning dynamics for the restricted case of deep linear neural networks. Despite the linearity of their input-output map, such networks have nonlinear gradient descent dynamics on weights that change with the addition of each new hidden layer. We show that deep linear networks exhibit nonlinear learning phenomena similar to those seen in simulations of nonlinear networks, including long plateaus followed by rapid transitions to lower error solutions, and faster convergence from greedy unsupervised pretraining initial conditions than from random initial conditions." Nonlinear Computation in Deep Linear Networks https://blog.openai.com/nonlinear-computation-in-linear-networks/ "Neural networks consist of stacks of a linear layer followed by a nonlinearity like tanh or rectified linear unit. Without the nonlinearity, consecutive linear layers would be in theory mathematically equivalent to a single linear…

You hit the nail on the head; for many straight information extraction problems, the universe of documents you want to extract could be too small to learn a model from -- or for that matter, for conventional methods of model evaluation to apply. (You want to extract data from all the items, not prove you reached a certain level of accuracy on a sub-sample of them) One approach is https://blog.openai.com/unsupervised-sentiment-neuron/ where you can throw in a great amount of unlabeled data and build an internal representation that models the data well enough that you can train something that works like an HMM or CRF with a tiny amount of labeled data. If you are willing to do something rule-based, I've used https://en.wikipedia.org/wiki/Case-based_reasoning to organize the work in annotating corpuses. Often I can prove that a certain rule set covers X% of the cases, then add a rule to do X+epsilon% until the results are "good enough". Feel free to click on my profile link and send me a message if you want to chat more.

It's best to change our understanding of _dangerous_. AI will not come in the form of a dystopian movie franchise. AI will show itself in ugly biases that occur because of our human biases. The prime example of AI being dangerous already exists in predictive policing. A potential starting point to understanding that aspect of the problem is [here](https://www.washingtonpost.com/opinions/big-data-may-be-reinforcing-ra… In terms of the physical realm, you can think of self-driving cars. Currently, these systems are not robust to [adversarial examples](https://blog.openai.com/adversarial-example-research/). So AI will not be dangerous in the traditional sense. It _might_ be dangerous because of biases in the data or because it acts in ways in which we do not expect them to work. Research in interpretable models and adversarial examples will help alleviate some of these issues.

Last week, I wrote a blog post (https://blog.openai.com/robust-adversarial-inputs/) about how it's possible to synthesize really robust adversarial inputs for neural networks. The response was great, and I got several requests to write a tutorial on the subject because what was already out there wasn't all that accessible. This post, written in the form of an executable Jupyter notebook, is that tutorial! Security/ML is a fairly new area of research, but I think it's going to be pretty important in the next few years. There's even a very timely Kaggle competition about this (https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack) run by Google Brain. I hope that this blog post will help make this really neat area of research slightly more approachable/accessible! Also, the attacks don't require that much compute power, so you should be able to run the code from the post on your laptop.