Web Excursions 2023-02-11
ChatGPT Is a Blurry JPEG of the Web
Xerox photocopiers use a lossy compression format known as JBIG2, designed for use with black-and-white images.
To save space, the copier identifies similar-looking regions in the image and stores a single copy for all of them;
when the file is decompressed, it uses that copy repeatedly to reconstruct the image.
It turned out that the photocopier had judged the labels specifying the area of the rooms to be similar enough that it needed to store only one of them—14.13—and it reused that one for all three rooms when printing the floor plan.
The fact that Xerox photocopiers use a lossy compression format instead of a lossless one isn’t, in itself, a problem.
The problem is that the photocopiers were degrading the image in a subtle way, in which the compression artifacts weren’t immediately recognizable.
If the photocopier simply produced blurry printouts, everyone would know that they weren’t accurate reproductions of the originals.
What led to problems was the fact that the photocopier was producing numbers that were readable but incorrect;
it made the copies seem accurate when they weren’t.
In 2013, workers at a German construction company noticed something odd about their Xerox photocopier: when they made a copy of the floor plan of a house, the copy differed from the original in a subtle but significant way.
In the original floor plan, each of the house’s three rooms was accompanied by a rectangle specifying its area: the rooms were 14.13, 21.11, and 17.42 square metres, respectively.
However, in the photocopy, all three rooms were labelled as being 14.13 square metres in size.
Imagine that you’re about to lose your access to the Internet forever.
In preparation, you plan to create a compressed copy of all the text on the Web, so that you can store it on a private server.
Unfortunately, your private server has only one per cent of the space needed;
you can’t use a lossless compression algorithm if you want everything to fit.
Instead, you write a lossy algorithm that identifies statistical regularities in the text and stores them in a specialized file format.
because the text has been so highly compressed, you can’t look for information by searching for an exact quote;
you’ll never get an exact match, because the words aren’t what’s being stored.
To solve this problem, you create an interface that accepts queries in the form of questions and responds with answers that convey the gist of what you have on your server.
Think of ChatGPT as a blurry JPEG of all the text on the Web.
It retains much of the information on the Web, in the same way that a JPEG retains much of the information of a higher-resolution image, but, if you’re looking for an exact sequence of bits, you won’t find it;
all you will ever get is an approximation.
But, because the approximation is presented in the form of grammatical text, which ChatGPT excels at creating, it’s usually acceptable.
You’re still looking at a blurry JPEG, but the blurriness occurs in a way that doesn’t make the picture as a whole look less sharp.
It’s also a way to understand the “hallucinations,” or nonsensical answers to factual questions, to which large language models such as ChatGPT are all too prone.
This analogy makes even more sense when we remember that a common technique used by lossy compression algorithms is interpolation—that is, estimating what’s missing by looking at what’s on either side of the gap.
if a compression algorithm is designed to reconstruct text after ninety-nine per cent of the original has been discarded, we should expect that significant portions of what it generates will be entirely fabricated.
To grasp the proposed relationship between compression and understanding
If a compression program knows that force equals mass times acceleration, it can discard a lot of words when compressing the pages about physics because it will be able to reconstruct them.
Likewise, the more the program knows about supply and demand, the more words it can discard when compressing the pages about economics, and so forth.
GPT-3’s statistical analysis of examples of arithmetic enables it to produce a superficial approximation of the real thing, but no more than that.
If you ask GPT-3 (the large-language model that ChatGPT was built from) to add or subtract a pair of numbers, it almost always responds with the correct answer when the numbers have only two digits.
But its accuracy worsens significantly with larger numbers, falling to ten per cent when the numbers have five digits.
Is it possible that, in areas outside addition and subtraction, statistical regularities in text actually do correspond to genuine knowledge of the real world?
The fact that ChatGPT rephrases material from the Web instead of quoting it word for word makes it seem like a student expressing ideas in her own words, rather than simply regurgitating what she’s read;
it creates the illusion that ChatGPT understands the material.
Can large language models take the place of traditional search engines?
It’s not clear that it’s technically possible to retain the acceptable kind of blurriness while eliminating the unacceptable kind
Even if it is possible to restrict large language models from engaging in fabrication, should we use them to generate Web content?
anything that’s good for content mills is not good for people searching for information.
The rise of this type of repackaging is what makes it harder for us to find what we’re looking for online right now;
the more that text generated by large language models gets published on the Web, the more the Web becomes a blurrier version of itself.
I’m going to make a prediction: when assembling the vast amount of text used to train GPT-4, the people at OpenAI will have made every effort to exclude material generated by ChatGPT or any other large language model.
Repeatedly resaving a JPEG creates more compression artifacts, because more information is lost every time.
It’s the digital equivalent of repeatedly making photocopies of photocopies in the old days
Indeed, a useful criterion for gauging a large language model’s quality might be the willingness of a company to use the text that it generates as training material for a new model.
If the output of ChatGPT isn’t good enough for GPT-4, we might take that as an indicator that it’s not good enough for us, either.
Can large language models help humans with the creation of original writing?
There is a genre of art known as Xerox art, or photocopy art, in which artists use the distinctive properties of photocopiers as creative tools.
Something along those lines is surely possible with the photocopier that is ChatGPT, so, in that sense, the answer is yes.
can the text generated by large language models be a useful starting point for writers to build off when writing something original, whether it’s fiction or nonfiction?
starting with a blurry copy of unoriginal work isn’t a good way to create original work.
If you’re a writer, you will write a lot of unoriginal work before you write something original.
And the time and effort expended on that unoriginal work isn’t wasted;
on the contrary, I would suggest that it is precisely what enables you to eventually create something original.
If students never have to write essays that we have all read before, they will never gain the skills needed to write something that we have never read.
Sometimes it’s only in the process of writing that you discover your original ideas.
Your first draft isn’t an unoriginal idea expressed clearly;
it’s an original idea expressed poorly, and it is accompanied by your amorphous dissatisfaction, your awareness of the distance between what it says and what you want it to say.
That’s what directs you during rewriting, and that’s one of the things lacking when you start with text generated by an A.I.
But we aren’t losing our access to the Internet.
So just how much use is a blurry JPEG, when you still have the original?
The Cup of Coffee Theory of AI
This is the perennial challenge of the artist:
finding the middle of the Venn diagram
where one circle is the artist’s tastes and the other is the audience’s tastes.
When a work of art—whether a book or an album, a movie, a TV show, or a YouTube video—is well received,
we say that it connected.
The work connected with an audience—with humans.
To connect with humans, it helps to be among humans.
When you don’t experience reality like most people do, it’s hard to make things that connect with most people.
AI, completely detached from reality, will have a hard time making things that connect with people.
It can’t be certain that your work (or its own) will connect with the human experience.
It can’t be certain that the work will land in the middle of that Venn diagram.
It can’t be certain that its taste and discernment are “great.”This remains the perennial problem of making great art.
Stripe Can’t Lose
There was the unfortunate public kerfuffle with fellow fintech darling Plaid in May 2022, in which
Stripe defended itself against allegations of unscrupulous conduct after it released Financial Connections, a product directly competitive to Plaid’s core offering—
after Stripe had offered Plaid to its clients instead of its own product.
Stripe is famous for catering to developers,
an overlooked but increasingly important buyer persona in internet-economy companies,
but the company’s real innovation in the early days was making payments more accessible and invisible than other offerings at the time.
The conventional path to obtaining a merchant account (i.e., as a retailer or an accountant) in 2010 involved
emailing or faxing in paperwork to an ISO,
waiting days or weeks for that information to be reviewed by a processor, and
going through a cumbersome process to register that merchant account with a separate payment gateway.
Stripe did all of that in minutes
Today, Stripe is competing against other modern payments companies, well-funded upstarts, and formidable incumbents that offer much of the same functionality as Stripe.
The payments landscape in the U.S. can be split into two categories
Legacy payment processors
Chase Merchant Services
First Data by Fiserv Fiserv
Worldpay from FIS
Global Payments
Modern payment processors
Adyen
Stripe
Braintree
Checkout.com
The legacy processors responded mostly by buying other payments companies to boost their total payments volume (TPV), culminating in a handful of mega-mergers in 2019.
Today, despite the pandemic-fueled growth of modern players, legacy players still control the bulk of the market and are growing trillions of dollars worth of TPV ~9% year over year.
But modern players are growing hundreds of billions of dollars of TPV ~60% year over year, so it’s not hard to imagine the relative market share between the two categories flipping over the next decade.
The legacy providers process payments for massive customers that are unlikely to switch over to modern processors any time soon.
Worldpay, for example, powers payments for some of the country’s leading grocery and pharmacy brands, including Kroger, which sold $137.9 billion worth of goods (excluding fuel) in 2021.
Over the two years, Stripe has been making it a point to showcase how many of its customers are processing over $1 billion annually.
Among the modern providers, Stripe’s fiercest competition is from Netherlands-based Adyen, arguably the best-run payments company in the world.
Historically, these two have largely avoided each other.
Stripe spent the first part of its life focused on winning smaller businesses while Adyen served enterprise customers from day one.
Adyen is a European powerhouse;
Stripe has mainly operated in the U.S.
More recently, the two companies have been going head to head.
All of these companies will do fine.
Fundamentally, payments infrastructure is a many-winner market. There are few network effects.
The capital investments required and economics of scale may mean that some entrants are scared off,
but when global revenue for an industry is measured in the trillions, as it is for payments,
there’s usually plenty of revenue to go around for a handful of players.
For most of its history, Stripe’s research and development (R&D) strategy has been focused on supporting its core payments offering.
Stripe has not released much information about the usage of its products, so it’s difficult to tell if these investments have been worthwhile, but ostensibly they expand its total addressable market by generating non-payments revenue.
The strategy seems to have changed around 2018 with the launch of products like Stripe Corporate Card, Issuing (a platform to create virtual and physical pre-paid cards), Treasury (a banking-as-a-service product), and Financial Connections (a data aggregation tool),
which are only tangentially related to card-based payments acceptance—from where most of Stripe’s revenue is assumed to come.
A Game Is Not a Game Without a Special Kind of Conflict
As we play with something, we start to understand it in new ways.
Every game pits players with or against each other in a system of conflict.
This ‘conflict’ might sound negative, a kind of antagonistic competition.
In fact, the conflict in games is always collaborative in some way.
This is because everyone participating agrees, voluntarily, to take part in the game together.
If we’re being forced to play, it’s not really play.
We all decide to spend the next couple of minutes – or hours, or weeks – within the space of play, and together keep the struggle of the game going.
Conflict is part of the dramatic machinery of a game that grips our minds and emotions.
As every storyteller knows, there is no drama without conflict.
When we play a game, we do not become confused about whether or not the game is real.
In fact, we are able to lose ourselves in play because (paradoxically!) we know that games are artificial.
The conflict in games is like two actors fighting on a stage: artificial, theatrical combat.
The audience members watching from their seats don’t rush up on stage to intervene and stop the fight.
Instead, they sit in the theatre and suspend their disbelief.
They can be gripped by the drama of the fight, yet at the same time know that it’s artificial.
When we take part in the artificial conflict of play, we are taking part in this multilayered metaconsciousness.
The fragile, artificial conflict of games can bleed into reality, in very unpleasant ways.
Video games all too often can be breeding grounds for the worst kinds of online culture, in which marginalised players find themselves attacked.
This is the double-edged potential of conflict that every game designer and player confronts.
How do you maintain collaborative play in the midst of full-tilt struggle?
How do you harness the elemental spark of dramatic conflict without letting it burn down the whole game?
How does conflict resist toxicity and remain productive, meaningful and joyful?
It all comes down to the community that emerges when we play.
In The Moral Judgment of the Child (1932), the psychologist Jean Piaget traced the ways that children come to understand the rules of the game of marbles.
Piaget found that children move through three different stages as they learn how to play marbles.
The youngest kids have a vague sense that there are rules you are supposed to follow, but they don’t quite understand how they work.
They will play at the game of marbles, drawing a circle in the sand and maybe knocking a marble or two, but not fully comprehend the entire system.
In the second stage, usually starting around age five, children are able to understand the rules of marbles and fully play the game, but in a very particular way.
They hold the rules as a kind of sacred authority.
They play strictly by the rules only, and won’t permit any bending or breaking of them.
There is only one right way to play the game.
The third stage begins around age 10.
Children come to see marbles as a social contract, a set of rules that gain their authority only because the players agree to follow them.
This means that, if everyone agrees, the rules can be changed.
This is essentially how adults see games too: as a voluntary, social construct.
Play in this sense is wonderfully flexible but also quite fragile.
Play happens only if and when we all agree to it.
Every moment of play is an opportunity to exercise collaboration with other human beings and to explore the social contract of play.
The sociologist Gary Alan Fine, in his book Shared Fantasy (1983), distilled three distinct layers on which identity operates for participants in role-playing games like Dungeons and Dragons (1974). There is
the layer of the character
the layer of player
the layer of person, with relationships and responsibilities outside of the game
Playing a game doesn’t mean occupying just one of these layers.
It means existing on all of them at the same time.
Flickering between and among these levels is play.
Play can play with identity too.
This is the problem with notions of ‘immersion’, which tend to assume that game players somehow leave the real world behind and lose themselves completely in virtual worlds and characters.
In an arcade fighting game, on one level you identify with the character you’re playing, extending yourself into the world of the game.
You also exist as a player, studying the quirks of the game, looking for tiny advantages, trying to outthink your opponent’s moves, trash talking to rattle her nerves.
At the same time, as a person in an arcade, you navigate the social hierarchy of the local gamer scene or strategise how to maximise the value of the quarters you put into the machine.
This fluctuating play of identity is what immersion really is.
[Note: can help to understand the problems with web3’s alleged benefits of equalization]