Start Up No.1975: chatbots are reasoning engines?, Twitter continues hari-kiri, Apple halts UK mobile browser probe, and more

Old methods of painting might be superseded by nanoparticles that mimic nature’s method of adding colour. (Might be.) CC-licensed photo by whereareyousimonwhereareyousimon 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 9 links for you. A second coat? I’m @charlesarthur on Twitter. On Mastodon: Observations and links welcome.

GPT-4 is a reasoning engine • Chain of Thought

Dan Shipper:


Even though our AI models were trained by reading the whole internet, that training mostly enhances their reasoning abilities not how much they know. And so, the performance of today’s AI models is constrained by their lack of knowledge.

…When GPT-4 was trained, it was fed a large portion of the available material on the internet. Training transformed that data into a statistical model that is very good at, given a string of words, knowing which words should follow from it—this is called next token prediction.

However, the kind of “knowledge” contained in this statistical model is fuzzy and inexplicit. The model doesn’t have any sort of long-term memory or way to look up the information it has seen—it only remembers what it encountered in its training set in the form of a statistical model.

When it encounters my name it uses this model to make an educated guess about who I am. It draws a conclusion that’s in the ballpark of being right, but is completely wrong in its details because it doesn’t have any explicit way to look up the answer.

But when GPT-4 is hooked up to the internet (or anything that acts like a database) it doesn’t have to rely on its fuzzy statistical understanding. Instead, it can retrieve explicit facts like, “Dan Shipper is the co-founder of Every” and use that to create its answer.

So, what does this mean for the future? I think there are at least two interesting conclusions:

1: Knowledge databases are as important to AI progress as foundational models
2: People who organise, store, and catalog their own thinking and reading will have a leg up in an AI-driven world. They can make those resources available to the model and use it to enhance the intelligence and relevance of its responses.


Getting a chatbot to correctly perform a series of real-world tasks based on written or spoken orders is really what you want. Hook it up to Siri, Alexa or Google. Perfect.
unique link to this extract

This is the lightest paint in the world • WIRED

Max Levy:


Color surrounds us in nature, and we re-create it with pigments. You can think of pigments as pulverized minerals, heavy metals, or chemicals that we swish into oil and spread over a canvas or car: Cobalt becomes blue; ochre red; cadmium yellow. “But nature has a very different way of creating color than we do,” Chanda says. Some of nature’s most vivid looks—the kind worn by peacocks, beetles, and butterflies—do their thing without pigment.

Those colors come from topography. Submicroscopic landscapes on the outer surfaces of peacock feathers, beetle shells, and butterfly wings diffract light to produce what’s known as structural color. It’s longer-lasting and pigment-free. And to scientists, it’s the key to creating paint that is not only better for the planet but might also help us live in a hotter world. 

In a paper published this month in Science Advances, Chanda’s lab demonstrated a first-of-its-kind paint based on structural color. They think it’s the lightest paint in the world—and they mean that both in terms of weight and temperature. The paint consists of tiny aluminum flakes dotted with even tinier aluminum nanoparticles. A raisin’s worth of the stuff could cover both the front and back of a door. It’s lightweight enough to potentially cut fuel usage in planes and cars that are coated with it. It doesn’t trap heat from sunlight like pigments do, and its constituents are less toxic than paints made with heavy metals like cadmium and cobalt.

…Because structural color can blanket an entire surface with just a thin, ultralight layer, Chanda thinks this will be a game changer—for airlines. A Boeing 747 needs about 500 kilograms of paint. He estimates that his paint could cover the same area with 1.3 kilograms. That’s more than 1,000 pounds shaved off each plane, which would reduce how much fuel is needed per journey.


This is absolutely mindblowing, but also really difficult to understand. We tend to forget that paint is a relatively heavy thing – pigments in an emulsion. But butterflies aren’t weighed down by their colouring. The big question, as always, is how easy this will be to commercialise. (Thanks Steve for the link!)
unique link to this extract

GM kills more than CarPlay support, it kills choice • Ars Technica

Roberto Baldwin:


On Friday, news dropped that GM would be phasing out CarPlay support in future EVs. In its partnership with Google, it hopes that all the features you get from mirroring your iPhone can be replaced with an Android Automotive feature. GM, like Toyota before it, wants to control the digital real estate in its vehicles. It’s a revenue-based and walled-garden (ironically against Apple) decision that will cost them.

Software-driven vehicles should be about choice. Instead, GM is making a short-sighted decision based on a trickle of revenue under the guise of better integration. Owning all the data that a vehicle generates while driving around could be a great source of cash. The problem is potential customers have become accustomed to choosing which device they use to navigate, chat, text, and rock out within their vehicle. They’ve grown weary of being mined for data at the expense of their choice and they’re really not all that keen on in-car subscription services.

For years, automakers have been sharing their vision of a future where cars can drive themselves, and the passengers are kept entertained by a plethora of features that are meant to keep their attention as they roll without worry to their destination. If in this far-off future, a person were to get into their vehicle and be restricted from using their service of choice—CarPlay in this scenario—why would they even buy that vehicle? What’s the point of telling people that, in the future, they can use whatever they want if, as a company, you don’t let them.

GM’s move is based on its desire to offer tighter integration with navigation and other in-car systems. Charging along routes isn’t really possible within projected versions of Apple or Google Maps in many vehicles. That’s a solid reason for GM to make its mapping solution better. It’s not really a reason to reduce the choices it offers consumers.


It’s also shortsighted: Apple has a huge proportion of the US smartphone market, so GM is planning to make the experience worse for them.
unique link to this extract

Twitter is ending legacy verification in favor of paid blue checkmarks • The Washington Post

Rachel Lerman and Faiz Siddiqui:


The removal of verification badges at such a wide scale has the potential to disrupt systems across Twitter’s website, including its recommendation algorithms, spam filters and help center requests. Twitter has previously relied on the badges as an important signal affecting all of those areas — for example, using verification to decide to boost a public figure’s tweet into a user’s timeline.

Removal of verification badges is a largely manual process powered by a system prone to breaking, which draws on a large internal database — similar to an Excel spreadsheet — in which verification data is stored, according to the former employees. Sometimes, an employee would try to remove a badge but the change wouldn’t take, one of the former employees said, prompting workers to explore workarounds. In the past, there was no way to reliably remove badges at a bulk scale — prompting workers tackling spam, for example, to have to remove check marks one-by-one.

“It was all held together with duct tape,” the former employee added.

Musk has already struggled with an increased number of outages since his $44bn takeover last year, troubles that have been compounded by his cutting more than two-thirds of the staff. And earlier attempts to roll out a paid verification system went awry.

The change that began on Saturday could fundamentally alter how Twitter is used and how it is trusted, users and experts say. If the fears are borne out, it will no longer be possible to quickly ascertain whether a public figure’s account is legitimately associated with that person, or the potential work of a sly impersonator.


Previously, people with “blue ticks” were the objects of derision from a large group who couldn’t get verified. Now, people with blue ticks will be the objects of derision from a large group who don’t want to be verified. Musk de-verified the New York Times on Sunday after it made clear the organisation wouldn’t pay for verification.

unique link to this extract

Apple wins appeal to quash the UK’s mobile stranglehold probe • BNN Bloomberg

Katharine Gemmell:


The Big Tech firm appealed the Competition and Markets Authority’s decision to refer the firm to a full-blown market investigation following its findings in its mobile browser market study. It successfully argued that the CMA didn’t follow the rules on timings and that the probe was invalid.

The CMA opened its investigation into both Apple and Google owner Alphabet Inc.’s dominance of the mobile browser market after a separate study concluded they have the power to “exercise a stranglehold” over operating systems, app stores and web browsers on mobile devices. Alphabet wasn’t involved in the lawsuit.

Judges at the Competition Appeal Tribunal ruled Friday that both the CMA’s notice and start of the consultation process happened too late. Its decision “lacks the statutory prerequisites — publication of a timely notice and commencement of a timely consultation — for a valid decision in this regard.” 

“This risks substantially undermining the CMA’s ability to efficiently and effectively investigate and intervene in markets where competition is not working well,” said a CMA spokesperson. The agency is considering an appeal. 


No sign of this on the CMA website. The CAT site has the summary of the judgment, which says that the CMA was six months late in all its actions once it had threatened to investigate.
unique link to this extract

Imagination makes us human. When did our species first acquire this ability? • The Conversation

Andrey Vyshedskiy is professor of neuroscience at Boston University:


To optimize their foraging, [early] mammals developed a new system to efficiently memorize places where they’d found food: linking the part of the brain that records sensory aspects of the landscape—how a place looks or smells—to the part of the brain that controls navigation. They encoded features of the landscape in the neocortex, the outermost layer of the brain. They encoded navigation in the entorhinal cortex. And the whole system was interconnected by the brain structure called the hippocampus. Humans still use this memory system for remembering objects and past events, such as your car and where you parked it.

Groups of neurons in the neocortex encode these memories of objects and past events. Remembering a thing or an episode reactivates the same neurons that initially encoded it. All mammals likely can recall and re-experience previously encoded objects and events by reactivating these groups of neurons. This neocortex-hippocampus-based memory system that evolved 200 million years ago became the first key step toward imagination.

The next building block is the capability to construct a “memory” that hasn’t really happened.

…Multiple types of archaeological artifacts unambiguously associated with prefrontal synthesis appear simultaneously around 65,000 years ago in multiple geographical locations. This abrupt change in imagination has been characterized by historian Yuval Harari as the “cognitive revolution.” Notably, it approximately coincides with the largest Homo sapiens‘ migration out of Africa.

Genetic analyses suggest that a few individuals acquired this prefrontal synthesis ability and then spread their genes far and wide by eliminating other contemporaneous males with the use of an imagination-enabeled strategy and newly developed weapons.


unique link to this extract

‘Thousands of dollars for something i didn’t do’ • The New York Times

Kashmir Hill and Ryan Mac:


On the Friday afternoon after Thanksgiving, Randal Quran Reid was driving his white Jeep to his mother’s home outside Atlanta when he was pulled over on a busy highway. A police officer approached his vehicle and asked for his driver’s license. Mr. Reid had left it at home, but he volunteered his name. After asking Mr. Reid if he had any weapons, the officer told him to step out of the Jeep and handcuffed him with the help of two other officers who had arrived.

“What did I do?” Mr. Reid asked. The officer said he had two theft warrants out of Baton Rouge and Jefferson Parish, a district on the outskirts of New Orleans. Mr. Reid was confused; he said he had never been to Louisiana.

Mr. Reid, a transportation analyst, was booked at the DeKalb County jail, to await extradition from Georgia to Louisiana. It took days to find out exactly what he was accused of: using stolen credit cards to buy designer purses.

“I’m locked up for something I have no clue about,” Mr. Reid, 29, said.

His parents made phone calls, hired lawyers and spent thousands of dollars to figure out why the police thought he was responsible for the crime, eventually discovering it was because Mr. Reid bore a resemblance to a suspect who had been recorded by a surveillance camera. The case eventually fell apart and the warrants were recalled, but only after Mr. Reid spent six days in jail and missed a week of work.

…The Sheriff’s Office has a contract with one facial recognition vendor: Clearview AI, which it pays $25,000 a year. According to documents obtained by The Times in a public records request, the department first signed a contract with Clearview in 2019.


And yes, the wrongful arrest was based on Clearview data.
unique link to this extract

No one wants a printer, but everyone wants to print • WSJ

Rachel Feintzeig:


The apartment building had a 24-hour gym, a swimming pool flanked by grills and something called the Sky Lounge on the 12th floor, with an expansive view of downtown Minneapolis. 

But the amenity that Olga Lobasenko and her husband couldn’t get out of their minds as they sized up potential apartments last year was situated in the lobby, illuminated by the glow of a fireplace. People sometimes gathered around it. 

It was a printer. 

“We just assumed it would have to be something you’d struggle to find for your entire life,” says Ms. Lobasenko, 33 years old. 

They moved in and now feel the sweet relief of being able to print whenever they want, without having to beg, borrow or curse a dried-out ink cartridge.  

“This one,” Ms. Lobasenko says of the printer, “is somebody else’s problem.”

Much of the world has moved on from hard copies. We have our phones and our tablets, scannable QR codes and the DocuSign app. And yet, it comes for all of us eventually—the need to print, and print now.

“When you need it, you need it,” says Leigh Stringer, who works at architecture firm Perkins&Will helping companies design sustainable offices. 


I love the idea of the couple being googly-eyed for a printer that someone else will look after. The article is full of the stories you’ll know all about: the frustration of trying to find a printer, of the printer that won’t print, of the printer that will print but not print quite what you want. (The link should be free to view.)
unique link to this extract

‘Succession’ isn’t really an American drama—it’s a British comedy • Vanity Fair

Piya Sinha-Roy:


British comedy often revels in discomforting moments that feel all too real. And while a flat in South Croydon might seem many worlds away from the helicopters and penthouses of New York, [Succession creator Jesse] Armstrong’s penchant for heightening cringe is woven through each episode of Succession as well. Watching [son of patriarch Logan Roy] Roman accidentally sext his father a photo of his genitals instead of [Roman’s love interest] Gerri is right up there with Peep Show’s Mark bumping into Sophie after their catastrophic wedding, sporting ejaculate on his trousers from a quick tryst with a new colleague.

[Second son] Kendall’s “L to the OG” rap at Logan’s birthday induces the same can’t-watch-but-can’t-not-watch squirms of Jez sucking jam from Sophie’s mum’s fingers. And watching Logan piss all over his office harkens back to Mark pissing all over his colleague’s desk. While Succession’s characters are not comparable to those in Armstrong’s other shows, watching Succession sends the familiar physical cringe of Peep Show shuddering up my spine.


This is very true. Armstrong was also a writer on The Thick Of It, coming up with some of the fruitiest insults to put into the mouths of politicians and spinners, in a show that was also compulsive cringe-watching.

Succession is astonishingly compelling, though. Worth it just for the insults (almost surely amped up by Armstrong.)
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.