Is GPT-3 a foretaste of something like HAL 9000, from 2001: A Space Odyssey? CC-licensed photo by James Vaughan 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 a joke. I’m @charlesarthur on Twitter. Observations and links welcome.
The troll: a fake flag burning at Gettysburg was only his latest hoax • The Washington Post
Shawn Boburg and Dalton Bennett:
»
[Adam] Rahuba once claimed that activists were planning to desecrate a Confederate cemetery in Georgia, The Post found. He seeded rumors of an organized effort to report Trump supporters for supposed child abuse. And he promoted a purported grass-roots campaign to confiscate Americans’ guns.
These false claims circulated widely on social media and on Internet message boards. They were often amplified by right-wing commentators and covered as real news by media outlets such as Breitbart News and the Gateway Pundit.
The hoaxes, outlandish in their details, have spurred fringe groups of conspiracy-minded Americans to action by playing on partisan fears. They have led to highly combustible situations — attracting heavily armed militia members and far-right activists eager to protect values they think are under siege — as well as large mobilizations of police.
…Some of Rahuba’s hoaxes have taxed law enforcement agencies and put bystanders in danger. In Gettysburg this year, a local pastor wearing a Black Lives Matter T-shirt was surrounded by armed counterprotesters until officers accompanied him out of the park for his own safety. Three years ago, an armed man who went to Gettysburg in response to a purported flag burning Rahuba had promoted on Facebook accidentally shot himself in the leg with a revolver.
Rahuba dismissed concerns that his efforts had harmed people or put them at risk.
“The message here was that any idiot on the Internet can get a bunch of people to show up at a Union cemetery with a bunch of Confederate flags and Nazi tattoos on their necks that just make them look foolish,” he said.
He also had little sympathy for the man who shot himself. “There’s some comedic value to that happening,” Rahuba said.
Rahuba, a lifelong resident of the Pittsburgh area, said he began trolling in high school. Using a dial-up modem, he and a group of friends posed as a 12- or 13-year-old girl in online chat rooms to lure older men to meetings, he said. In his telling, the men arrived to find Rahuba and his friends mocking them.
“It made me realize that people will believe the most unrealistic nonsense on the Internet,” he said.
«
I have to say I’m with Rahuba on this. If people are so stupid as to believe this junk – and the outlets that amplify it – then that’s on them. If people didn’t do stupid things, they wouldn’t be made to look like fools. (See also bitcoin, below.)
unique link to this extract
GPT-3, etc. • Marginal REVOLUTION
Tyler Cowen:
»
I am increasingly convinced that Scott Alexander was right that NLP and human language might boostrap a general intelligence. A rough criteria for AGI might be something like (i) pass the Turing test, and (ii) solve general problems; the GPT-3-AI-Dungeon examples above appear to accomplish preliminary versions of both.
GPT was published in June 2018, GPT-2 in February 2019, GPT-3 in May 2020.
As best I can tell GPT -> GPT2 was ~10x increase in parameters over ~8 months, and GPT2 -> GPT3 was ~100x increase of parameters over ~14 months. Any number of naive projections puts a much more powerful release happening over the next ~1-2yrs, and I also know that GPT-3 isn’t necessarily the most powerful NLP AI (perhaps rather the most popularly known.)
When future AI textbooks are written, I could easily imagine them citing 2020 or 2021 as years when preliminary AGI first emerged,. This is very different than my own previous personal forecasts for AGI emerging in something like 20-50 years…
p.s. One of the users above notes that AI Dungeon GPT-3 (“Dragon”) is a subscription service, something like ~$6 a week. MIE.”
«
AGI = artificial general intelligence (HAL 9000, all that kind of thing)
GPT-3 = latest version of an AI system that is astounding everyone who comes into contact with it because it’s so damn human-like, at least in what it does with text.
Yes, you can start worrying now. If GPT-3 becomes cheap, you’ll never (for example) be able to trust that the comments on a story, or the reviews on a site, were written by a human. Or that the story was written by a human. Or anything. This blogpost suggests you curb your enthusiasm. Given that it’s essentially regurgitating the English-language web, it’s also got terrible inbuilt biases.
Like HAL 9000 did, after all.
unique link to this extract
WindowSwap
»
Open a new window somewhere in the world.
Let’s face it. We are all stuck indoors. And it’s going to be a while till we travel again.
Window Swap is here to fill that deep void in our wanderlust hearts by allowing us to look through someone else’s window, somewhere in the world, for a while.
A place on the internet where all we travel hungry fools share our ‘window views’ to help each other feel a little bit better till we can (responsibly) explore our beautiful planet again.
«
This is really lovely – 10-minute HD videos (with sound) taken from peoples’ windows all over the world. A sort of nice, relaxing version of Chatroulette. (See the About page for more details.)
unique link to this extract
Oxford coronavirus vaccine triggers immune response, trial shows • The Guardian
Sarah Boseley:
»
The results were “a really important milestone” on the path to a vaccine, said the study’s lead author, Prof Andrew Pollard. They showed the vaccine was very well tolerated, he added. “We are seeing exactly the sort of immune responses we were hoping for, including neutralising antibodies and T-cell responses, which, at least from what we’ve seen in the animal studies, seem to be those that are associated with protection.”
The problem is, he said: “We just don’t know what level is needed if you meet this virus in the wild, to provide protection, so we need to do the clinical trials and to work that out.”
Hopefully researchers would find out from the trials to come, added Pollard, which would help all vaccine developers.
“We don’t know what high is. We’ve got immune responses that we can measure, we can see the virus being neutralised when the antibodies are tested in the laboratory, but we don’t know how much is needed. I mean it’s encouraging but it’s only the first milestone on this long path,” he said.
Ideally the vaccine would protect against any infection, but scientists already accept it may reduce the severity of the disease instead, meaning people would be less likely to become very sick and die.
The volunteers have been followed up for eight weeks so far after immunisation. A further question is how long any immune response will last – if for only six months or a year, people might need regular booster shots.
«
So it’s good news, but we’re only a little way down a long road. Two more phases to go, and then ramping up production, and actually injecting people. This time next year?
unique link to this extract
USS University • No Mercy, No Malice
Scott Galloway:
»
There is a dangerous conflation of the discussion about K-12 and university reopenings. The two are starkly different. There are strong reasons to reopen K-12, and there are stronger reasons to keep universities shuttered. University leadership needs to evolve from denial (“It’s business as usual”) past bargaining (“We’ll have a hybrid model with some classes in person”) to citizenship (“We are the warriors against this virus, not its enablers”).
Think about this. Next month, as currently envisioned, 2,800+ cruise ships retrofitted with white boards and a younger cohort will set sail in the midst of a raging pandemic. The density and socialization on these cruise ships could render college towns across America the next virus hot spots.
Why are administrators putting the lives of faculty, staff, students, and our broader populace at risk?
The ugly truth is many college presidents believe they have no choice. College is an expensive operation with a relatively inflexible cost structure. Tenure and union contracts render the largest cost (faculty and administrator salaries) near immovable objects. The average salary of a full professor (before benefits and admin support costs) is $104,820, though some make much more, and roughly 50% of full-time faculty have tenure. While some universities enjoy revenue streams from technology transfer, hospitals, returns on multibillion dollar endowments, and public funding, the bulk of colleges have become tuition dependent. If students don’t return in the fall, many colleges will have to take drastic action that could have serious long-term impacts on their ability to fulfill their missions.
That gruesome calculus has resulted in a tsunami of denial.
Universities owning up to the truth have one thing in common: they can afford to.
«
Thinking of them as giant cruise ships certainly puts it into a grim perspective. (K-12 is what Britons call secondary school: children at that age seem to be asymptomatic.)
unique link to this extract
David Shor’s unified theory of the 2020 election • NY Mag
Eric Levitz with a fascinating (long) interview with a longtime election strategist:
»
Mitt Romney and Donald Trump agreed on basically every issue, as did Barack Obama and Hillary Clinton. And yet, a bunch of people changed their votes. And the reason that happened was because the salience of various issues changed. Both sides talked a lot more about immigration, and because of that, correlation between preferences on immigration and which candidate people voted for went up. In 2012, both sides talked about health care. In 2016, they didn’t. And so the correlation between views on health care and which candidate people voted for went down.
So this means that every time you open your mouth, you have this complex optimization problem where what you say gains you some voters and loses you other voters. But this is actually cool because campaigns have a lot of control over what issues they talk about.
Non-college-educated whites, on average, have very conservative views on immigration, and generally conservative racial attitudes. But they have center-left views on economics; they support universal health care and minimum-wage increases. So I think Democrats need to talk about the issues they are with us on, and try really hard not to talk about the issues where we disagree. Which, in practice, means not talking about immigration.
…What’s powerful about nonviolent protest — and particularly nonviolent protest that incurs a disproportionate response from the police — is that it can shift the conversation, in a really visceral way, into the part of this issue space that benefits Democrats and the center left. Which is the pursuit of equality, social justice, fairness — these Democratic-loaded concepts — without the trade-off of crime or public safety.
«
Altogether fascinating about what politicians can and can’t effect through campaigning.
unique link to this extract
Exclusive: Twitter hackers could have stolen a whole lot more bitcoin • Forbes
Billy Bambrough:
»
Coinbase, the largest U.S. bitcoin and cryptocurrency exchange with around 35 million users around the world, has said it prevented just over 1,100 Coinbase customers from sending a total of 30.4 bitcoin, worth almost $280,000, to the scam.
“We noticed within about a minute of the Gemini and Binance tweets,” Philip Martin, Coinbase chief information security officer, said during a phone interview. Bitcoin exchanges Gemini and Binance were both targeted early on by the hackers, just before Coinbase itself.
Only 14 Coinbase users were able to send around $3,000 worth of bitcoin to the scam bitcoin address before Coinbase blacklisted it, according to Martin.
“It was a vanishingly small group of Coinbase users that tried to send bitcoin to the scam address,” Martin said, adding that the San Francisco-based exchange, which is reportedly gearing up for a stock market listing that could come as early as this year, often blacklists the bitcoin and cryptocurrency addresses used by giveaway scammers.
Other bitcoin exchanges, including New York-based Gemini, owned by the Winklevoss twins, San Francisco-based Kraken and Binance, of no fixed address, all confirmed they stopped funds from flowing into the hacker’s bitcoin address—though their combined users didn’t attempt to send anywhere near as much as Coinbase.
“This hack shows that security is about layers of protection,” Jesse Powell, chief executive of Kraken, said via email. “Somebody has to be watching the admins and setting up alerts to watch for these vulnerabilities.”
«
I’m now thinking that the overlap of dim bulbs and bitcoin users is quite a bit larger than I thought. Also that the script kiddies who did the hack knew more about the value of what they were doing than most people suspected.
unique link to this extract
Government admits breaking privacy law with NHS test and trace • The Guardian
Sarah Marsh and Alex Hern:
»
The UK government broke the law in rolling out its test-and-trace programme without a full assessment of the privacy implications, the Department of Health and Social Care has admitted after a legal challenge.
The Guardian can reveal the programme has already led to three data breaches involving email mishaps and unredacted personal information being shared in training materials.
“The reckless behaviour of this government in ignoring a vital and legally required safety step known as the data protection impact assessment (DPIA) has endangered public health,” said Jim Killock, the executive director of Open Rights Group (ORG). “We have a ‘world beating’ unlawful test-and-trace programme.
“A crucial element in the fight against the pandemic is mutual trust between the public and the government, which is undermined by their operating the programme without basic privacy safeguards. The government bears responsibility for the public health consequences.”
A DPIA is required before carrying out any “high risk” processing of personal data. The government had previously argued that the test-and-trace programmes, which involves carrying detailed personal information from patients across the country, did not qualify as high risk, until the ORG threatened to take it to court over the claim.
«
World-beatingly unlawful.
unique link to this extract
FAQ: Should you delete TikTok? Here’s everything you need to weigh the real privacy risks • The Washington Post
Geoffrey Fowler:
»
“Protecting the privacy of our users’ data is of the utmost importance to TikTok,” said spokeswoman Ashley Nash-Hahn. “TikTok collects much less U.S. user information than many of the companies in our space and stores it in the U.S. and Singapore. We have not, and would not, give it to the Chinese government.”
My takeaway: TikTok doesn’t appear to grab any more personal information than Facebook. That’s still an appalling amount of data to mine about the lives of Americans. But there’s scant evidence that TikTok is sharing our data with China, and we should be wary of xenophobia dressed up as privacy concerns.
I don’t mean to excuse China’s record of online repression — it’s possible China will force TikTok to change its practices in the future. For now, it comes down to whether you inherently distrust data mining from Chinese-owned companies more than data mining from U.S.-owned ones. Just remember: companies in China probably make your phone, laptop and TV, too.
Let’s dive into the specifics.…Its US privacy policy also says it gathers your country location, Internet address and the type of device you’re using. If you give it permission, it will also grab your exact location, your phone’s contacts and other social network connections, as well as your age and phone number.
That all adds up to a profile of you useful not only to target ads, but also to understand who you are, who your friends and family are, what you like, what you find funny and what you say to your friends.
Jackson, from Disconnect, said the app sends an “abnormal” amount of information from devices to its computers.
«
Lockdown was a boon for Spotify. Now musicians are fighting back • WIRED UK
Will Pritchard:
»
imagine payouts are calculated monthly and in June 2020 Apple Music is paying out £100 to rights holders. If 10% of the total streams on the platform for that month were Ariana Grande songs, then Grande – or the rights holder for those recordings, which in this case is Universal subsidiary Republic Records – would receive 10% of the total payout pot, or £10. The same method applies to songwriters, although these rights are typically owned separately (and, again, often by a major label entity). This way of dividing payments means the most popular artists (those with the most streams) receive a chunk of revenue from users of the platform who haven’t played any of their songs.
Under this system, the Ariana Grande stan who pays £10 a month for Apple Music and plays ‘thank u, next’ on repeat all week also has a far greater influence on who gets paid what than, for instance, their dad who also pays £10 a month but only uses Apple Music to stream his favourite Paul Weller album to wind down at the weekend. Effectively, the Paul Weller fan is supplementing Ariana’s income when the Apple Music cheque lands.
This is a simplified explanation of the process, since streaming platforms also give a different weighting to streams from paying, premium subscribers versus free users who listen to the service with ads, for instance – that adds another level of complexity to the breakdown, but broadly the system works as described.
It means that most of your ten pound subscription actually goes to Ed Sheeran or Drake or Lady Gaga rather than the other musicians whose music you may have been listening to.
«
This really is how it works, and it feels so wrong. Under the old system when you bought someone’s LP, CD or download, they’d get their cut. (Yes I know I know record label contracts evil awful terrible exploitative. But.) The streaming payment system is so unfair to artists. Further reading. (Thanks G for the links.)
unique link to this extract
Errata, corrigenda and ai no corrida: none notified
The Overspill: when there's more that I want to say 