News: 1596875351

In brief Your Google Home speaker may have been quietly recording sounds around your house without your permission or authorization, it was revealed this week.

The Chocolate Factory admitted it had accidentally turned on a feature that allowed its voice-controlled AI-based assistant to activate and snoop on its surroundings. Normally, the device only starts actively listening in and making a note of what it hears after it has heard wake words, such as “Ok, Google” or “Hey, Google,” for privacy reasons. Prior to waking, it's constantly listening out for those words, but is not supposed to keep a record of what it hears.

Yet punters noticed their Google Homes had been recording random sounds, without any wake word uttered, when they started receiving notifications on their phone that showed the device had heard things like a smoke alarm beeping, or glass breaking in their homes – all without giving their approval.

Google said the feature had been accidentally turned on during a recent software update, and it has now been switched off, Protocol [1]reported . It may be that this feature is or was intended to be used for home security at some point: imagine the assistant waking up whenever it hears a break in, for instance. Google just [2]bought a $450m, or 6.6 per cent, stake in anti-burglary giant ADT.

GPT-3 wrote this? No way!

A blog post written almost entirely by OpenAI’s text-generator GPT-3 attracted thousands of visitors. And only one person suspected it was generated by AI software rather than a human writer, apparently.

Liam Porr, a computer science student at the University of California, Berkeley, promoted the [3]bogus self-help blog online. To his surprise, the post received lots of page views, though only one person noticed that although the sentences made grammatical sense, the writing lacked any substance, so they suspected foul play.

Porr [4]revealed that, indeed, it was written using a [5]tool based on OpenAI’s latest text generator. He fed the model up to five sentences at a time, and it would fill in the rest, producing a whole article. It took a bit of fiddling to get the best results, however, and he generated several versions and made sure to delete sentences that included fake-sounding quotes.

Porr reckons his experiment is proof that GPT-3 isn’t yet good enough for complete automation yet; human writers will still need to edit and polish copy. However, tools like GPT-3 provide a way to crank out content on the internet much more quickly, he opined.

1.7 million pre-print academic papers now up for grabs

Cornell University has collected all the content on arXiv, the free repository of scientific papers awaiting peer-reviewed publication, to build a tidy text-based dataset for machine-learning engineers to play with.

The dataset, covering nearly 30 years of publications, is now hosted on Kaggle, a Google-owned data science platform, for anyone to download. It is hoped that by making the data public, it’ll encourage developers to train AI tools like recommendation engines or help users search for interesting papers more easily.

“Having the entire arXiv corpus on Kaggle grows the potential of arXiv articles immensely,” [6]said Eleonora Presani, arXiv executive director. “By offering the dataset on Kaggle we go beyond what humans can learn by reading all these articles and we make the data and information behind arXiv available to the public in a machine-readable format.”

The 1.7 million papers in the dataset have been released over 30 years, and will be [7]published on Kaggle in multiple batches.

Microsoft is winding down support for third-party Cortana tools

Cortana, Microsoft’s answer to Siri and Google Assistant, hasn’t taken off quite like its competitors.

Now, Redmond has decided to consolidate its efforts to build an AI digital assistant focused on Microsoft 365, and will pull support for Cortana on phones, and also kill off third-party Cortana apps, known as skills.

“As we make this shift toward a transformational AI-powered assistant experience in Microsoft 365, we need to adjust our focus areas of innovation and development to give our customers assistance where they need it most,” Microsoft [8]said in a statement. “As a result, we are making changes to some US consumer-centric features and functionalities with lower usage.”

Those changes are: "To end support for all third-party Cortana skills on September 7. Then, in early 2021, we’ll stop supporting the Cortana app for mobile (iOS and Android) ... In addition, after careful consideration, we’ve decided to end support for the Cortana service integration in the Harman Kardon Invoke speaker in January 2021." ®

Get our [9]Tech Resources



[1] https://www.protocol.com/google-smart-speaker-alarm-adt

[2] https://www.barrons.com/articles/google-and-adt-have-big-plans-for-home-security-despite-privacy-concerns-51596561126

[3] https://adolos.substack.com/p/feeling-unproductive-maybe-you-should

[4] https://liamp.substack.com/p/my-gpt-3-blog-got-26-thousand-visitors

[5] https://medium.com/@aidungeon/ai-dungeon-dragon-model-upgrade-7e8ea579abfe

[6] https://blogs.cornell.edu/arxiv/2020/08/05/leveraging-machine-learning-to-fuel-new-discoveries-with-the-arxiv-dataset/

[7] https://www.kaggle.com/Cornell-University/arxiv

[8] https://support.microsoft.com/en-us/help/4575625/upcoming-changes-to-cortana

[9] https://whitepapers.theregister.com/