Google seeks to placate AI researchers complaining of Big Brother-like working conditions

Plus: Deepfake Tom Cruise has arrived on TikTok, and new ML algo that beats humans at classic Atari games

In Brief Lawyers at Google have been criticised by the company’s AI researchers for watering down negative language in their academic papers.

A leaked email memo to Reuters revealed that words like “concerns” were changed to “considerations”, “dangers” became “risks”. Other words like “breach” or “sensitive” were frowned upon too.

The Chocolate Factory’s legal team also often removed references to Google’s own models if they were cast in a bad light, making it difficult for its own researchers to work freely, it is claimed.

Google promised to change its internal practices on how it reviews academic work from its research scientists after it reorganized its ethical AI unit in a recent town hall meeting. Co-leads of the department, Timnit Gebru and Margaret Mitchell, were both fired (a few months apart) following a fallout with management over a research paper critiquing large language-generation models.

Deepfake Tom Cruise on TikTok

Three astoundingly accurate lookalikes of US actor Tom Cruise have racked up millions of views on videos that were posted on the social media site TikTok in the past week.

The "Tom Cruises" were seen golfing, shopping, and performing a magic trick in different videos. They're all fake of course, and they all look exactly like the real Tom Cruise because they're computer-generated. Copies of Cruise's face have been pasted onto the bodies of others using AI algorithms to create these so-called deep fakes.

The account name on TikTok – "deeptomcruise" – as well as fake Tom Cruise's more youthful-looking appearance (relative to the real 58-year-old Tom Cruise) gave the game away. The account has amassed over 238,000 followers, 795,000 likes, and millions of views. The videos are scarily good and a sign of how realistic deepfakes are getting. You can have a look at them yourself at this link.

Go-Explore, an Atari game-crushing reinforcement learning algorithm

OpenAI researchers have claimed the company's Go-Explore algorithm trumped all the best human scores achieved on all 55 classic Atari 2600 games and has beaten most state-of-the-art scores from other types of reinforcement learning algorithms too.

The trick to Go-Explore agents is that they're able to explore new moves and actions in a game by remembering to retrace steps that are particularly promising and exploring from there. Agents are awarded points when they successfully perform actions that help them complete levels.

Games that require more exploration are more difficult to play than ones that have more immediate returns on rewards. For example, Montezuma's Revenge is trickier for bots than, say, Checkers.

"Go-Explore solves all previously unsolved Atari games and surpasses the state of the art on all hard-exploration games with orders-of- magnitude improvements on the grand challenges of Montezuma's Revenge and Pitfall," OpenAI wrote in a research paper on Nature published this week. The team said Go-Explore might be useful in some practical domains like robotics and drug discovery, but that remains to be seen.

Build a cute cardboard cutout computer vision robot

Computer vision hobbyists can train a simple cardboard robot to recognise objects with machine learning with a new project from Google Creative Lab.

The code for Alto, described as a "teachable object", is now public. The boxy device has a computer for eyes, and runs object recognition software written in Tensorflow on Google's Coral USB accelerator board and is operated using a Raspberry Pi board.

You'll need to buy the hardware components yourself, and find some cardboard to build it. The instructions on how to construct Alto are on GitHub. Once Alto is trained, if it sees something that it recognises it raises its arm.

The Coral USB accelerator costs $59.99 (~£42), and the Raspberry Pi boards vary depending on the model.

AI-powered text autocomplete is coming to Microsoft Word next month

Office 365 users will soon be able to use an autocomplete tool, one that is already available on other similar products like Google Docs.

The move to roll out the feature is currently in the works and it is expected to arrive in March, according to Microsoft's website. The tool uses a machine learning model that analyses all the previous words in a sentence to predict the next ones, according to Neowin.

Biting the hand that feeds IT © 1998–2021