Adversarial examples are inputs to machine learning models that an attacker has intentionally designed to cause the model to make a mistake; they're like optical illusions for machines. In this post we'll show how adversarial examples work across different mediums, and will discuss why securing systems against them can be difficult. At OpenAI, we think adversarial examples are a good aspect of security to work on because they represent a concrete problem in AI safety that can be addressed in the short term, and because fixing them is difficult enough that it requires a serious research effort. To get an idea of what adversarial examples look like, consider this demonstration from Explaining and Harnessing Adversarial Examples: starting with an image of a panda, the attacker adds a small perturbation that has been calculated to make the image be recognized as a gibbon with high confidence. An adversarial input, overlaid on a typical image, can cause a classifier to miscategorize a panda as a gibbon.

This project provides optimized infrastructure for reinforcement learning. It extends the OpenAI gym interface to multiple parallel environments and allows agents to be implemented in TensorFlow and perform batched computation. As a starting point, we provide BatchPPO, an optimized implementation of Proximal Policy Optimization. The algorithm to use is defined in the configuration and pendulum started here uses the included PPO implementation. Check out more pre-defined configurations in agents/scripts/configs.py.

Elon Musk has issued a warning as Russian president Vladimir Putin claims the nation which leads in AI "will become the ruler of the world." Musk, co-chairman of OpenAI, has long warned of dire consequences for mishandling AI development. OpenAI itself is a non-profit research company that aims to champion promoting and developing friendly AI in a way to benefit humanity. As with any major technology advancement, however, there will undoubtedly be those which aim to weaponise it and to do so before rivals. Based on Putin's comments to Russia-based publication RT, it sounds as if the nation is among them.

As of last month Alphabet Inc.'s AI division, Google DeepMind, has open-sourced their new machine learning platform DeepMind Lab. Artificial Intelligence is the technology of the moment, constantly debated and attracting massive attention from investors. Despite warnings from influential figures including Professor Stephen Hawking, Google's decision to open up their software to other developers is part of a mass movement to advance the capabilities of AI. Facebook open sourced its own deep learning software last year, and Elon Musk's non-profit organisation OpenAI recently released Universe, an open software platform that can be used to train AI systems. So, why have Google, OpenAI and others made these platforms public, and how will this affect the adoption of Artificial Intelligence and machine learning as a whole?

Elon Musk's artificial intelligence research company, OpenAI, is developing a self-learning bot for one of the most complex esports titles: 'Dota 2.' It has already become the ultimate challenge for players, but for top esports pros, it is also a major opportunity. Snoop Dogg and Martha Stewart reenact that famous'Ghost' scene and things get steamy

Can we learn something from the AI playing against Arteezy? I also give a conclusion about what the bot is going to be doing in the future and how OpenAI will try to achieve this.

Last week was the high point of the Dota 2 competitive year: it was the week of The International, Valve's biggest tournament. On Saturday, Team Liquid walked away with more than $10 million after defeating Newbee 3-0 in the grand final. Right now, one of the requirements to be a good Dota 2 player is that you've got to be a living, breathing human. The game does include some basic computer-controlled bots to practice against, but any seasoned player of the game should have no trouble prevailing over these bots, even on their hardest "Unfair" difficulty (though the Unfair Viper bot is a legendary jerk that's utterly miserable to play against). Last Friday, however, we got a hint of a new, altogether more threatening kind of computer-controlled player: an AI-controlled bot built by Elon Musk's OpenAI.

ACKTR can learn continuous control tasks, like moving a robotic arm to a target location, purely from low-resolution pixel inputs (left). ACKTR (pronounced "actor") -- Actor Critic using Kronecker-factored Trust Region -- was developed by researchers at the University of Toronto and New York University, and we at OpenAI have collaborated with them to release a Baselines implementation. The authors use ACKTR to learn control policies for simulated robots (with pixels as input, and continuous action spaces) and Atari agents (with pixels as input and discrete action spaces). ACKTR combines three distinct techniques: actor-critic methods, trust region optimization for more consistent improvement, and distributed Kronecker factorization to improve sample efficiency and scalability. For machine learning algorithms, two costs are important to consider: sample complexity and computational complexity.

Tesla CEO Elon Musk fired off a new and ominous warning on Friday about artificial intelligence, suggesting the emerging technology poses an even greater risk to the world than a nuclear conflagration with North Korea. Musk--a fierce and long time critic of A.I. who once likened it to "summoning the demon" in a horror movie--said in a Twitter post that people should be concerned about the rise of the machines than they are. Reacting to the news that autonomous tech had bested competitive players in an electronic sports competition, Musk posted what appeared to be a photo of a poster bearing the chilling words "In the end, the machines will win." Musk, who is spearheading commercial space travel with his venture SpaceX, is also the founder of OpenAI, a nonprofit that promotes the "safe" development of AI. His stance puts him at odds with much of the tech industry, but echoes remarks of prominent voices like Stephen Hawking--who has also issued dire warnings about machine learning.

Our Dota 2 result shows that self-play can catapult the performance of machine learning systems from far below human level to superhuman, given sufficient compute. In the span of a month, our system went from barely matching a high-ranked player to beating the top pros and has continued to improve since then. Supervised deep learning systems can only be as good as their training datasets, but in self-play systems, the available data improves automatically as the agent gets better. Improvements came from every part of the system, from adding new features to algorithmic improvements to scaling things up. The graph is surprisingly linear, meaning the team improved the bot exponentially over time.