Start Up No.2000: Wikipedia faces its LLM problem, ChatGPT’s future jobs impact, do colonoscopies work?, Kinged!, and more

The Millennium Bridge opened in London in 2000 – and almost immediately closed because pedestrians made it sway. And sway. CC-licensed photo by pablocanenpablocanen 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.

There was another post on Friday at the Social Warming Substack – about the challenge of moderating Twitter. Free signup.

A selection of 10 links for you. Still going. I’m @charlesarthur on Twitter. On Mastodon: Observations and links welcome.

AI is tearing Wikipedia apart • Vice

Claire Woodcock:


It didn’t take long for researchers to figure out that OpenAI’s ChatGPT is a terrible fabricator, which is what tends to doom students who rely solely on the chatbot to write their essays. Sometimes it will invent articles and their authors. Other times it will name-splice lesser known scholars with more prolific ones, but will do so with the utmost confidence. OpenAI has even said that the model “hallucinates” when it makes up facts—a term that has been criticized by some AI experts as a way for AI companies to avoid accountability for their tools spreading misinformation. 

“The risk for Wikipedia is people could be lowering the quality by throwing in stuff that they haven’t checked,” Bruckman added. “I don’t think there’s anything wrong with using it as a first draft, but every point has to be verified.” 

The Wikimedia Foundation, the nonprofit organization behind the website, is looking into building tools to make it easier for volunteers to identify bot-generated content. Meanwhile, Wikipedia is working to draft a policy that lays out the limits to how volunteers can use large language models to create content.

The current draft policy notes that anyone unfamiliar with the risks of large language models should avoid using them to create Wikipedia content, because it can open the Wikimedia Foundation up to libel suits and copyright violations—both of which the nonprofit gets protections from but the Wikipedia volunteers do not.

…The community is also divided on whether large language models should be allowed to train on Wikipedia content. While open access is a cornerstone of Wikipedia’s design principles, some worry the unrestricted scraping of internet data allows AI companies like OpenAI to exploit the open web to create closed commercial datasets for their models. This is especially a problem if the Wikipedia content itself is AI-generated, creating a feedback loop of potentially biased information, if left unchecked.


unique link to this extract

June 2000: Swaying Millennium Bridge closed • BBC On This Day


Huge crowds of people have been blamed for forcing the temporary closure of London’s new bridge on the day of its opening.

The city’s first new river crossing for decades began swaying violently in the wind under the weight of hundreds of pedestrians on Saturday morning.

Police became concerned and the bridge was closed briefly while engineers made safety checks to the structure.

A limit was subsequently imposed on the number of pedestrians allowed to cross the bridge, which spans the Thames from St Paul’s Cathedral to the Tate Modern gallery on the South Bank.

A spokesman for architects Foster and Partners who designed the bridge with engineers Ove Arup and Partners said: “Because there was such a huge number walking all at once across the bridge, which is very unusual, there was a certain amount of swaying.

“The bridge is intended to have some movement. It’s a suspension bridge – if there isn’t movement there can be a problem.”

Pedestrians had to wait for half an hour before they were able to continue crossing the bridge.

The project cost more than £18m and was designed by architect Sir Norman Foster and the British sculptor Sir Anthony Caro.


This was a fascinating story: the problem was that people walking really did cause it to sway unpleasantly. Arup were properly surprised. It required £5m worth of 91 dampers to stop the movement. I wrote about it for The Independent; it was a sort of British Tacoma Narrows, except stopping well short of the self-destruction.
unique link to this extract

Which jobs will be most impacted by ChatGPT? • Visual Capitalist


On November 30, 2022, OpenAI heralded a new era of artificial intelligence (AI) by introducing ChatGPT to the world.

The AI chatbot stunned users with its human-like and thorough responses. ChatGPT could comprehend and answer a variety of different questions, make suggestions, research and write essays and briefs, and even tell jokes (amongst other tasks).

Many of these skills are used by workers in their jobs across the world, which begs the question: which jobs will be transformed, or even replaced, by generative AI in the coming future?

This infographic from Harrison Schell visualizes the March 2023 findings of OpenAI on the potential labor market impact of large language models (LLMs) and various applications of generative AI, including ChatGPT.


“High exposure” to the change, apparently, will be jobs which fit in these categories: interpreters + translators; survey researchers; writers and authors; public relations specialists; tax preparers; mathematicians; blockchain engineers; proofreaders and copy markers; accountants and auditors.

“Low exposure”: athletes, automotive repairers, cement masons, cooks, piledriver operators, stonemasons, tire repairers and changers, dishwashers, carpenter helpers.

To which I have a few queries: “blockchain engineers”?? And what’s the difference between a “tax preparer” and an accountant/auditor? Doesn’t one subsume the other? Also, presumably crane operators are there with piledriver operators. Though “athlete” sounds more fun. (Not listed on either side: “Wikipedia contributor”.)
unique link to this extract

Ex-Googlers blast Geoffrey Hinton’s past silence on fired AI experts • Fast Company

Wilfred Chan:


women who for years have been speaking out about AI’s problems—even at the expense of their jobs—say Hinton’s alarmism isn’t just opportunistic but also overshadows specific warnings about AI’s actual impacts on marginalized people.

“It’s disappointing to see this autumn-years redemption tour from someone who didn’t really show up” for other Google dissenters, says Meredith Whittaker, president of the Signal Foundation and an AI researcher who says she was pushed out of Google in 2019 in part over her activism against the company’s contract to build machine vision technology for U.S. military drones. (Google has maintained that Whittaker chose to resign.)

“I didn’t see any solidarity or any action when there were people really trying to organize and do something about the harms that are happening now,” she says.

Another prominent ex-Googler, Margaret Mitchell, who co-led the company’s ethical AI team, criticized Hinton for not denouncing Google’s 2020 firing of her coleader Timnit Gebru, a leading researcher who had spoken up about AI’s risks for women and people of color. 

“This would’ve been a moment for Dr. Hinton to denormalize the firing of [Gebru],” Mitchell tweeted on Monday. “He did not. This is how systemic discrimination works.”


unique link to this extract

Do colonoscopies really save lives? • The Science Writer

John Horgan:


NordICC, for Nordic-European Initiative on Colorectal Cancer, is the first large-scale, randomized trial of colonoscopy. The study focused on 84,585 men and women 55-64 years old in Sweden, Norway, Poland and the Netherlands. The researchers randomly divided subjects into two groups: one invited to get a colonoscopy, the other not. 

NordICC measured rates of death from colon cancer and from any cause in these two groups after 10 years. Some researchers favor “all-cause” mortality, because tests and treatments for a specific cancer can result in deaths unattributed to that cancer.

Colonoscopy can cause perforation of the colon, bleeding and infection; patients may also have adverse reactions to purging of the bowels and sedation. A 2016 study of 331,880 people who underwent colonoscopies found that 1.6% had complications serious enough to require “unplanned hospital visits” within one week. 

NordICC found that the risk of death from colon cancer after 10 years was 0.28% in the invited group and 0.31% in the control group. The difference in risk of death from any cause was even smaller: 11.03% in the invited group and 11.04% in the uninvited group. These are not statistically significant differences.

Defenders of colonoscopy seize on the fact that only 42% of the NordICC subjects invited to get a colonoscopy actually got it; this group’s mortality rate from colon cancer was 0.15, significantly less than the control rate of 0.31. The Colon Cancer Coalition says this finding confirms that “colonoscopy saves lives.” But this lower mortality rate might reflect self-selection bias, precisely what NordICC was designed to overcome.


Horgan also suggests following the money (which, in the US, would certainly make sense), which takes you to companies which benefit a lot from carrying out colonoscopies.
unique link to this extract

IPCC’s conservative nature masks true scale of action needed to avert catastrophic climate change • The Conversation

Kevin Anderson is professor of energy and climate change at the University of Manchester:


If we step outside the rarefied realm of IAM [climate modelling] scenarios that leading climate scientist Johan Rockström describes as “academic gymnastics that have nothing to do with reality”, it’s clear that not exceeding 1.5°C or 2°C [of excess warming] will require fundamental changes to most facets of modern life.

Starting now, to not exceed 1.5°C of warming requires 11% year-on-year cuts in emissions, falling to nearer 5% for 2°C. However, these global average rates ignore the core concept of equity, central to all UN climate negotiations, which gives “developing country parties” a little longer to decarbonise.

Include equity and most “developed” nations need to reach zero CO₂ emissions between 2030 and 2035, with developing nations following suit up to a decade later. Any delay will shrink these timelines still further.

Most IAM models ignore and often even exacerbate the obscene inequality in energy use and emissions, both within nations and between individuals. As the International Energy Agency recently reported, the top 10% of emitters accounted for nearly half of global CO₂ emissions from energy use in 2021, compared with 0.2% for the bottom 10%. More disturbingly, the greenhouse gas emissions of the top 1% are 1.5 times those of the bottom half of the world’s population.

So where does this leave us? In wealthier nations, any hope of arresting global heating at 1.5ºC or 2°C demands a technical revolution on the scale of the post-war Marshall Plan. Rather than relying on technologies such as direct air capture of CO₂ to mature in the near future, countries like the UK must rapidly deploy tried-and-tested technologies.


I don’t like what I believe the narrator’s voice is going to say in a couple of decades’ time.
unique link to this extract

The coronation was an act of magic for a country scared the spell might break • POLITICO

Tanya Gold on cracking form:


That lack of confidence in the magic spell was obvious at breakfast time. As the congregation spooled into Westminster Abbey, with actors at the front — kings tend to like actors, as they have the same job — the head of the anti-monarchist pressure group Republic, Graham Smith, was arrested near Trafalgar Square with five other republican leaders. The peaceful protest, he told me last week, was organized with the approval of the Metropolitan Police. They arrested him anyway, confiscated the placards, and blamed the string which tied the placards together for breaking the rules. (Apparently they might have used it to “lock” onto buildings.) A few hours later the king swore to serve us, which means serving our democracy. So he has already failed.

The protest went on in Trafalgar Square, but the BBC cut away as the cavalcade passed. Screens were erected in front of the protest, as if our eyes — and the king’s — were too delicate to be allowed to see it. We were told the police operation passed off without incident. The Duke of York was booed as he left Buckingham Palace, but that too was not reported on. The BBC was in the hagiography business at this coronation, and it was fervent and vapid. This is possibly tactical — they fear what an unpopular nativist government will do to their funding model — but it also indicates a nation afraid of itself. A deputy chairman of the Conservative Party suggested all republicans emigrate. They were all afraid the spell might break.

Then came the pomp: the fantastical costumes, the militarism, the uneasy horses, one of which panicked and backed into the crowd. Another marched sideways. It was lovely to look at, but it is the fumes of Empire, which of course is why the Mall was full.


unique link to this extract

Editors quit top neuroscience journal to protest against open-access charges • Nature

Katharine Sanderson:


More than 40 editors have resigned from two leading neuroscience journals in protest against what the editors say are excessively high article-processing charges (APCs) set by the publisher. They say that the fees, which publishers use to cover publishing services and in some cases make money, are unethical. The publisher, Dutch company Elsevier, says that its fees provide researchers with publishing services that are above average quality for below average price. The editors plan to start a new journal hosted by the non-profit publisher MIT Press.

The decision to resign came about after many discussions among the editors, says Stephen Smith, a neuroscientist at the University of Oxford, UK, and editor-in-chief of one of the journals, NeuroImage. “Everyone agreed that the APC was unethical and unsustainable,” says Smith, who will lead the editorial team of the new journal, Imaging Neuroscience, when it launches.

The 42 academics who made up the editorial teams at NeuroImage and its companion journal NeuroImage: Reports announced their resignations on 17 April. The journals are open access and require authors to pay a fee for publishing services. The APC for NeuroImage is US$3,450; NeuroImage: Reports charges $900, which will double to $1,800 from 31 May.

Elsevier, based in Amsterdam, says that the APCs cover the costs associated with publishing an article in an open-access journal, including editorial and peer-review services, copyediting, typesetting, archiving, indexing, marketing and administrative costs


This is the challenge that the open access model faces. The alternative to getting lots of institutions to pay a regular subscription to get paywalled access to scientific research is: get smaller numbers of individuals (or institutions) to pay occasional large amounts to make their scientific research available for free. The former is the usual Nature/Science/etc model, the latter is “open access”. But what’s the right price for the latter?

Or does it need a third, alternative model?
unique link to this extract

A $15.8m mistake: why San Francisco can’t pay its teachers on time • SF Chronicle

Bilal Mahmood:


In 2019, the [San Francisco] school district approved a $9.5m contract with the consulting company Infosys to replace the payroll system that had been used for 17 years. The new system was originally set to go live in 2021 but wasn’t rolled out until 2022, delays that raised costs by about $7m.

When the EMPowerSF system finally did go live, problems immediately emerged. Hundreds of teachers didn’t get their paychecks.

Months went by with no resolution. By the end of the summer of 2022, the district finally hired a consulting firm named Alvarez & Marsel to diagnose and audit the issue. Over the course of the year, the firm’s initial contract of $2.8m was extended so it could stabilize the problem, costing the district another $8m. In 2023, an additional $5m was spent to finally fix the problem.

What Alvarez & Marsel’s audit found was shocking: 64 software bugs had gone undetected in the EMPowerSF system for nearly a year, with 18 of the bugs affecting pay rules management, 12 bugs affecting the user interface, 7 affecting leave management, 7 affecting time management, 5 affecting benefits management and 15 bugs affecting other system protocols.

These bugs resulted in employees being paid even after they were terminated, incorrect daily rates being set for paid leave during pregnancy disability and annual salary fields for certificated bi-weekly employees being incorrectly pro-rated.

To add insult to injury, the same software underlying the EMPowerSF system had been used once before in a California school district. In 2007, Los Angeles Unified School District also launched a payroll system powered by the same software, and it also failed to pay teachers on time.

Compounding matters, the San Francisco school district lacked personnel with the technical experience to support the implementation of the new system. To successfully implement a payroll transition, you need experienced financial operators, technical administrators and, perhaps most important, software engineers to transfer hundreds of thousands of records between systems.


“Turnkey” systems never are. And I’ve never seen a company merger get its payroll to run smoothly. Unsurprising if a school district can’t either.
unique link to this extract

Over half of Twitter Blue’s earliest subscribers are no longer subscribed • Mashable

Matt Binder:


Twitter Blue’s struggles since its launch nearly six months ago are more severe than previously revealed, new data suggests.

Since Musk’s version of the subscription service launched last November, Twitter has only been able to convert around 640,000 Twitter users into paying Twitter Blue subscribers as of the end of April, as Mashable reported earlier this week.

While those numbers are lacklustre, an even more telling detail about Twitter Blue is just how many of its earliest subscribers have canceled their subscriptions.

Out of about 150,000 early subscribers to Twitter Blue, just around 68,157 have stuck around and maintained a paid subscription as of April 30. Subscriptions are $8 per month – $11 on mobile.

The total early subscriber numbers are linked directly to internal leaks published by the Washington Post last year showing that a total of 150,000 users originally signed up for Twitter Blue within just a few days of its launch in November. Twitter temporarily disabled new signups for about a month shortly after those users subscribed as a result of accounts signing up for Blue with the intent to impersonate major brands on the platform.

That means around 81,843 users, or 54.5%, of Twitter users who subscribed to Twitter Blue when it first launched in November are no longer subscribed to the service. That’s an abnormally high churn rate for an online subscription service. Churn rate is the percentage of users that unsubscribe from a service.


Perhaps all those original ones were troll accounts that got shut down? That would be the optimistic reading. Otherwise it means that they didn’t think Twitter Blue was worth it. Say it ain’t so, Elon!
unique link to this extract

• Why do social networks drive us a little mad?
• Why does angry content seem to dominate what we see?
• How much of a role do algorithms play in affecting what we see and do online?
• What can we do about it?
• Did Facebook have any inkling of what was coming in Myanmar in 2016?

Read Social Warming, my latest book, and find answers – and more.

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.