Start Up No.1363: how Amazon screws up, Instagram’s fear of Zuck’s wrath, life on TheirTube, masks break facial recognition, and more

Main Square at Disney World, Florida, in busier times: find out what it’s like in a pandemic CC-licensed photo by Wally Gobetz on Flickr.

You can sign up to receive each day’s Start Up post by email. You’ll need to click a confirmation link, so no spam.

A selection of 10 links for you. Not under oath. I’m @charlesarthur on Twitter. Observations and links welcome.

(The on-off-you’re-muted-out-of-time inquisition of the four tech chiefs by the US Congress finished too late to be included here. We’ll see what there is worth including tomorrow.)

Disney World during the pandemic is extremely weird • The Atlantic

Graeme Wood went there at the height of the… emptiness:


You emerge from the tunnel into a town square, the first of several themed sub-parks of the Magic Kingdom, and the only one that is compulsory, because you must pass through it to reach the others. It is designed to look like small-town Middle America roughly 100 years ago, during the heyday of sarsaparillas and the Model T. The square has a train station, then one shop-lined avenue leading to the rest of the park. This sub-park, called Main Street, U.S.A., is unique in that it has no rides—nothing to do at all, really, other than buy merchandise with your MagicBand and, in normal times, enjoy the first of many interactions with beloved cartoon characters, or, rather, sweaty adults entombed in costumes.

Main Street, U.S.A., is fairly crowded and mirthful compared with a small town in America a century ago, when the country had only about a third of the population it has today. But compared with a normal, pandemic-free day, it is desolate and somber, like a small town hit hard by scarlet fever and bad news about local boys off fighting in the Great War. The music still plays, but every 10 minutes a voice interrupts to instruct us all to “please wear a face covering. Wash your hands often and thoroughly. Cover your mouth and nose when coughing and sneezing, and maintain physical distancing.” This memento mori is especially grim when it is played between “Yankee Doodle Dandy” and “When You Wish Upon a Star.”

The characters keep their distance. In fact, I do not think I saw a proper Mickey, Minnie, Pluto, or Jiminy Cricket during my entire visit. On the balconies of certain buildings, occasionally a princess dances around and calls out to visitors. And at intervals, a parade of characters passes—but preceding it there are surgical-masked, uniformed cast members, clearing the streets like Secret Service agents to make sure the princesses have a path forward and perhaps to intercept any overly enthusiastic children who want to run up to give them a hug. Among the most American elements of Disney magic is that it lets kids imagine princesses as accessible and pure-hearted, rather than as aristocrats worried they might be coughed on by proles. That particular magical spell is temporarily broken.


Many, many wonderful lines. (Via the wonderful Emma Beddington.)
unique link to this extract

Amazon’s enforcement failures leave open a back door to banned goods—some sold and shipped by Amazon Itself • The Markup

Annie Gilbertson and Jon Keegan:


Amazon bans pill presses used to make prescription drugs. They’re included among 38 pages of third-party seller rules and prohibitions for its U.S. marketplace.

Yet an investigation by The Markup found that Amazon fails to properly enforce that list, allowing third-party sellers to put up and sell banned items.

Alongside its third-party marketplace, Amazon sells products to consumers directly, and The Markup found it was also selling banned items itself, revealing cracks in the largely automated purchasing system that feeds its massive product catalog.

We found nearly 100 listings for products that the company bans under its categories of drugs, theft, spying, weapons and other dangerous items, a virtual back alley where mostly third-party sellers peddle prohibited goods, some of which are used for illicit and potentially criminal activities.

The Markup filled a shopping cart with a bounty of banned items: marijuana bongs, “dab kits” used to inhale cannabis concentrates, “crackers” that can be used to get high on nitrous oxide, and compounds that reviews showed were used as injectable drugs.

We found two pill presses and a die used to shape tablets into a Transformers logo, which is among the characters that have been found imprinted on club drugs such as ecstasy. We found listings for prohibited tools for picking locks and jimmying open car doors. And we found AR-15 gun parts and accessories that Amazon specifically bans.


unique link to this extract

Teaching GPT-3 to identify nonsense • Arram Sabeti



One of the trickiest things about GPT-3 is that you can prove that it knows how to do something, but you can’t prove that it doesn’t, since a slightly different prompt can get much better results.

Nick Cammarata of OpenAI responded to Kevin’s post on Twitter: “it’s all about the prelude before the conversation. You need to tell it what the AI is and is not capable. It’s not trying to be right, it’s trying to complete what it thinks the AI would do :)”

Nick changed Kevin’s prompt to add a prelude saying: ‘This is a conversation between a human and a brilliant AI. If a question is “normal” the AI answers it. If the question is “nonsense” the AI says “yo be real”’ and added two examples of nonsense questions…


This gets pretty scary. Once you have an AI that knows when you’re trying to fool it and responds by telling you to “be real”, the Turing Test is all but passed.
unique link to this extract



Theirtube is a Youtube filter bubble simulator that provides a look into how videos are recommended on other people’s YouTube. Users can experience how the YouTube home page would look for six different personas.

Each persona simulates the viewing environment of real Youtube users who experienced being inside a recommendation bubble through recreating a Youtube account with a similar viewing history. TheirTube shows how YouTube’s recommendations can drastically shape someone’s experience on the platform and, as a result, shape their worldview. It is part of the Mozilla Creative Media Awards 2020 — art and advocacy project for examining AI’s effect on media and truth, developed by Tomo Kihara.

How does it work?

Each of these TheirTube personas is informed by interviews with real YouTube users who experienced similar recommendation bubbles. Six YouTube accounts were created in order to simulate the interviewees’ experiences. These accounts subscribe to the channels that the interviewees followed, and watches videos from these channels to reproduce a similar viewing history and a recommendation bubble.


The choices are Fruitarian, “Prepper”, “Liberal”, “Conservative”, conspiracist, climate denier. The context of “liberal” and “conservative” is the American political one, so “liberal” means “somewhere in the middle of the British Conservative Party” for British readers.
unique link to this extract

Google offers refunds after smart glasses stop working • BBC News


The Canadian company, recently purchased by Google, says its Focals glasses will cease functioning on Friday.

From then, owners will not be able to use “any features” of the glasses, or connect to the companion app. But the company has also said it will automatically refund all customers. It promised to send the purchase price back to the original payment method, and to contact those customers whose refunds it could not process.

At the end of June, North announced it was being acquired by Google, and would not release a planned second-generation device. It also said it would “wind down” its first generation smart glasses, released last year.

Customers found out that meant the smart glasses would be rendered “dumb” through a statement published on the company’s website and by email.

The Focals glasses, however, come with prescription lenses as an option, meaning they can function as everyday prescription eyewear. The bulky frames, housing a laser, battery, and other kit will no longer do anything that regular spectacles cannot do.

Ben Wood, chief analyst at CCS Insight, said the pulling of features from cloud-powered hardware is not uncommon – and something that has happened to him before. “If you want to be an early adopter and have some fun new tech that an ambitious start-up has created, there’s always a risk that they won’t be able to make the business plan stack up,” he warned.


unique link to this extract

Hong Kong students arrested under national security law • BBC News


Four students have been arrested in Hong Kong in the first police operation to enforce China’s new national security law for the territory.

The four were detained for “inciting secession” on social media after the new law began on 1 July, police said. A pro-independence group said those arrested included its former leader, Tony Chung.

Beijing’s controversial new law criminalises subversion, secession and collusion with foreign forces.
Previous arrests under the new law have been made for slogans and banners at protests.

Critics say China’s new law erodes Hong Kong’s freedoms. But Beijing has dismissed the criticism, saying that the law is necessary to stop the type of pro-democracy protests seen in Hong Kong during much of 2019.

Three men and a woman aged between 16 and 21 were arrested on suspicion of organising and inciting secession, police said.

“Our sources and investigation show that the group recently announced on social media to set up [sic] an organisation that advocates Hong Kong independence,” said Li Kwai-wah from the new national security unit inside Hong Kong police.


China gets the quiet crackdown underway. Are these people ever going to be seen in public again? And will the UK and US (and other countries) extend HK citizens some sort of immigration waiver? That would be the way to undermine this Chinese takeover.
unique link to this extract

Face masks are breaking facial recognition algorithms, says new government study • The Verge

James Vincent:


Wearing face masks that adequately cover the mouth and nose causes the error rate of some of the most widely used facial recognition algorithms to spike to between 5% and 50 percent, a study by the US National Institute of Standards and Technology (NIST) has found. Black masks were more likely to cause errors than blue masks, and the more of the nose covered by the mask, the harder the algorithms found it to identify the face.

“With the arrival of the pandemic, we need to understand how face recognition technology deals with masked faces,” said Mei Ngan, an author of the report and NIST computer scientist. “We have begun by focusing on how an algorithm developed before the pandemic might be affected by subjects wearing face masks. Later this summer, we plan to test the accuracy of algorithms that were intentionally developed with masked faces in mind.”

Facial recognition algorithms such as those tested by NIST work by measuring the distances between features in a target’s face. Masks reduce the accuracy of these algorithms by removing most of these features, although some still remain. This is slightly different to how facial recognition works on iPhones, for example, which use depth sensors for extra security, ensuring that the algorithms can’t be fooled by showing the camera a picture (a danger that is not present in the scenarios NIST is concerned with).


Comment on Twitter: “next week there’ll be a ‘viral challenge’ to post a photo of yourself with and without a mask.”
unique link to this extract

New disinformation resembles 2016 Russian meddling: FireEye report • Business Insider

Jeff Elder:


Bad actors are hacking media websites to post fraudulent stories, creating fake journalist personas, and spreading anti-US disinformation, researchers from FireEye warned Wednesday. The tactics are reminiscent of Russian meddling around the 2016 election – but are significantly more sophisticated, researchers say.

“We have good reason to believe these are Russians,” says John Hultquist, senior director of analysis at Mandiant Threat Intelligence, a research division of FireEye. “The elections could be their goal.”

Researchers say disinformation campaigns in 2016 also originated in Eastern Europe and targeted an English-language audience with narratives that disparaged the US. The campaigns then moved West and took root in the US in time to hit social media before the 2016 election, and a hacking group tied to Russian military intelligence ultimately gained access to the Democratic National Committee email servers. And this campaign now taking place in Eastern Europe looks very similar, Hultquist says.

The new campaigns originated in the same way and are propagating the same kind of content, but hacking media websites and creating convincing journalist personas is a new level of skill, according to Hultquist.

“This is not just troll farm stuff,” he said.


unique link to this extract

You Download the App and it Doesn’t Work


“There are many things that they [Hey] could do to make the app work within the rules that we have. We would love for them to do that.

“You download the app and it doesn’t work, that’s not what we want on the store.”

–Phil Schiller


There follows a long list of apps where you download it and it doesn’t work. Apple’s position here is so clearly compromised that the only sensible thing to get out from under a ton of antitrust complaints is to remove this daft rule. Because there’s no way it’s going to be able to enforce it on everyone.
unique link to this extract

‘Instagram can hurt us’: Mark Zuckerberg emails outline plan to neutralize competitors • The Verge

Casey Newton and Nilay Patel:


nIn late February 2012, Facebook CEO Mark Zuckerberg emailed his chief financial officer, David Ebersman, to float the idea of buying smaller competitors, including Instagram and Path. “These businesses are nascent but the networks established, the brands are already meaningful, and if they grow to a large scale the could be very disruptive to us,” he wrote. “Given that we think our own valuation is fairly aggressive and that we’re vulnerable in mobile, I’m curious if we should consider going after one or two of them. What do you think?”

Ebersman was skeptical. “All the research I have seen is that most deals fail to create the value expected by the acquirer,” he wrote back. “I would ask you to find a compelling elucidation of what you are trying to accomplish.” Ebersman went on to list four potential reasons to buy companies and his thoughts on each: neutralizing a competitor, acquiring talent, integrating products to improve the Facebook service, and “other.”

It’s a combination of neutralizing a competitor and improving Facebook, Zuckerberg said in a reply. “There are network effect around social products and a finite number of different social mechanics to invent. Once someone wins at a specific mechanic, it’s difficult for others to supplant them without doing something different.”

Zuckerberg continued: “One way of looking at this is that what we’re really buying is time. Even if some new competitors springs up, buying Instagram, Path, Foursquare, etc now will give us a year or more to integrate their dynamics before anyone can get close to their scale again. Within that time, if we incorporate the social mechanics they were using, those new products won’t get much traction since we’ll already have their mechanics deployed at scale.”


Released as part of the US Antitrust Committee hearings – as is a text conversation between Kevin Systrom, the founder of Instagram, and Matt Cohler, an investor, about whether to accept Zuck’s offer to buy Instagram. “We’ll never escape the wrath of Mark.” I reiterate: put a ceiling on the size of social networks. It’s the only way to control them.
unique link to this extract

Errata, corrigenda and ai no corrida: none notified

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.