Greg Brockman introduces OpenAI Five as a system trained through self-play

Greg Brockman introduces OpenAI Five as a system trained through self-play

Greg Brockman published this introduction on August 6, 2018. In a speech at a Benchmark event, Brockman described OpenAI Five, an AI system trained to play Dota, and presented the game as a training ground for teamwork, real time strategy and imperfect information. Brockman said Five learns by playing against itself and uses artificial neural networks. He argued that the same technology used for Dota could also learn tasks it was not designed for, citing control of a robotic hand as an example. He also said the computing power used for OpenAI Five would have been impractical two years earlier, and forecast that technologies like it would become commonplace. The speech connects progress in game-playing systems to questions about how AI might be applied beyond games, while offering no evidence here about the system’s performance outside its described tasks. Brockman’s claims about training scale, computing trends and future availability are his account and forecast, not independently established by this text. What results would show whether the approach transfers from Dota to tasks outside games?

Publication: 3a08fbeb-8a69-c32c-0994-a09d6130c1c1

attribution — pass: The page is Greg Brockman's blog post, in which he presents the text of his own speech introducing OpenAI Five. The reading attributes the statements to Brockman throughout, consistent with the source.

The text of my speech introducing OpenAI Five at yesterday’s Benchmark event

summary — pass: The title, summary and question match the source. OpenAI Five is identified as an AI system that plays the game Dota, and self-play is explained in plain language. Training conditions, transfer to the robotic hand and the computing forecast are presented as Brockman's account. The question about transfer beyond games follows from the evidence and does not assume an answer. The reading never states Brockman's OpenAI affiliation, but this is not essential to understanding the piece.

We work on Dota because it is a great training ground for AI

date — pass: evidence.originalPublished is 2018-08-06 at day precision. This matches the page dateline and the reading's date. The speech itself took place the day before ('yesterday'), and the reading does not misstate this because it dates only the publication. There is no hindsight framing, and the commonplace-technology forecast is attributed as Brockman's forecast. The evidence kind is publisher-page, so no archive disclosure is needed.

August 6, 2018 OpenAI Five intro

support — pass: Each claim is supported by a verbatim excerpt: self-play and neural networks, the robotic hand example, and the statement that computing power would have been impractical two years earlier, along with the forecast. The reading explicitly notes that this text does not independently establish these claims or show performance outside the described tasks. It does not imply unsupported benefits or costs.

But our AI system can learn problems it was not even designed to solve

media — pass: No media was selected, and no verified media candidates were supplied. Leaving media empty is appropriate.