Stop Indexing! And Pay Up!

July 17, 2024

dinosaur30a_thumb_thumb_thumb_thumbThis essay is the work of a dinobaby. Unlike some folks, no smart software improved my native ineptness.

I read “Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI.” The write up appears in two online publications, presumably to make an already contentious subject more clicky. The assertion in the title is the equivalent of someone in Salem, Massachusetts, pointing at a widower and saying, “She’s a witch.” Those willing to take the statement at face value would take action. The “trials” held in colonial Massachusetts. My high school history teacher was a witchcraft trial buff. (I think his name was Elmer Skaggs.) I thought about his descriptions of the events. I recall his graphic depictions and analysis of what I recall as “dunking.” The idea was that if a person was a witch, then that person could be immersed one or more times. I think the idea had been popular in medieval Europe, but it was not a New World innovation. Me-too is a core way to create novelty. The witch could survive being immersed for a period of time. With proof, hanging or burning were the next step. The accused who died was obviously not a witch. That’s Boolean logic in a pure form in my opinion.

image

The Library in Alexandria burns in front of people who wanted to look up information, learn, and create more information. Tough. Once the cultural institution is gone, just figure out the square root of two yourself. Thanks, MSFT Copilot. Good enough.

The accusations and evidence in the article depict companies building large language models as candidates for a test to prove that they have engaged in an improper act. The crime is processing content available on a public network, indexing it, and using the data to create outputs. Since the late 1960s, digitizing information and making it more easily accessible was perceived as an important and necessary activity. The US government supported indexing and searching of technical information. Other fields of endeavor recognized that as the volume of information expanded, the traditional methods of sitting at a table, reading a book or journal article, making notes, analyzing the information, and then conducting additional research or writing a technical report was simply not fast enough. What worked in a medieval library was not a method suited to put a satellite in orbit or perform other knowledge-value tasks.

Thus, online became a thing. Remember, we are talking punched cards, mainframes, and clunky line printers one day there was the Internet. The interest in broader access to online information grew and by 1985, people recognized that online access was useful for many tasks, not just looking up information about nuclear power technologies, a project I worked on in the 1970s. Flash forward 50 years, and we are upon the moment one can read about the “fact” that Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI.

The write up says:

AI companies are generally secretive about their sources of training data, but an investigation by Proof News found some of the wealthiest AI companies in the world have used material from  thousands of  YouTube videos to train AI. Companies did so despite YouTube’s rules against harvesting materials from the platform without permission. Our investigation found that subtitles from 173,536 YouTube videos, siphoned from more than 48,000 channels, were used by Silicon Valley heavyweights, including Anthropic, Nvidia, Apple, and Salesforce.

I understand the surprise some experience when they learn that a software script visits a Web site, processes its content, and generates an index (a buzzy term today is large language model, but I prefer the simpler word index.)

I want to point out that for decades those engaged in making information findable and accessible online have processed content so that a user can enter a query and get a list of indexed items which match that user’s query. In the old days, one used Boolean logic which we met a few moments ago. Today a user’s query (the jazzy term is prompt now) is expanded, interpreted, matched to the user’s “preferences”, and a result generated. I like lists of items like the entries I used to make on a notecard when I was a high school debate team member. Others want little essays suitable for a class assignment on the Salem witchcraft trials in Mr. Skaggs’s class. Today another system can pass a query, get outputs, and then take another action. This is described by the in-crowd as workflow orchestration. Others call it, “taking a human’s job.”

My point is that for decades, the index and searching process has been without much innovation. Sure, software scripts can know when to enter a user name and password or capture information from Web pages that are transitory, disappearing in the blink of an eye. But it is still indexing over a network. The object remains to find information of utility to the user or another system.

The write up reports:

Proof News contributor Alex Reisner obtained a copy of Books3, another Pile dataset and last year published a piece in The Atlantic reporting his finding that more than 180,000 books, including those written by Margaret Atwood, Michael Pollan, and Zadie Smith, had been lifted. Many authors have since sued AI companies for the unauthorized use of their work and alleged copyright violations. Similar cases have since snowballed, and the platform hosting Books3 has taken it down. In response to the suits, defendants such as Meta, OpenAI, and Bloomberg have argued their actions constitute fair use. A case against EleutherAI, which originally scraped the books and made them public, was voluntarily dismissed by the plaintiffs.  Litigation in remaining cases remains in the early stages, leaving the questions surrounding permission and payment unresolved. The Pile has since been removed from its official download site, but it’s still available on file sharing services.

The passage does a good job of making clear that most people are not aware of what indexing does, how it works, and why the process has become a fundamental component of many, many modern knowledge-centric systems. The idea is to find information of value to a person with a question, present relevant content, and enable the user to think new thoughts or write another essay about dead witches being innocent.

The challenge today is that anyone who has written anything wants money. The way online works is that for any single user’s query, the useful information constitutes a tiny, miniscule fraction of the information in the index. The cost of indexing and responding to the query is high, and those costs are difficult to control.

But everyone has to be paid for the information that individual “created.” I understand the idea, but the reality is that the reason indexing, search, and retrieval was invented, refined, and given numerous life extensions was to perform a core function: Answer a question or enable learning.

The write up makes it clear that “AI companies” are witches. The US legal system is going to determine who is a witch just like the process in colonial Salem. Several observations are warranted:

  1. Modifying what is a fundamental mechanism for information retrieval may be difficult to replace or re-invent in a quick, cost-efficient, and satisfactory manner. Digital information is loosey goosey; that is, it moves, slips, and slides either by individual’s actions or a mindless system’s.
  2. Slapping fines and big price tags on what remains an access service will take time to have an impact. As the implications of the impact become more well known to those who are aggrieved, they may find that their own information is altered in a fundamental way. How many research papers are “original”? How many journalists recycle as a basic work task? How many children’s lives are lost when the medical reference system does not have the data needed to treat the kid’s problem?
  3. Accusing companies of behaving improperly is definitely easy to do. Many companies do ignore rules, regulations, and cultural norms. Engineering Index’s publisher leaned that bootleg copies of printed Compendex indexes were available in China. What was Engineering Index going to do when I learned this almost 50 years ago? The answer was give speeches, complain to those who knew what the heck a Compendex was, and talk to lawyers. What happened to the Chinese content pirates? Not much.

I do understand the anger the essay expresses toward large companies doing indexing. These outfits are to some witches. However, if the indexing of content is derailed, I would suggest there are downstream consequences. Some of those consequences will make zero difference to anyone. A government worker at a national lab won’t be able to find details of an alloy used in a nuclear device. Who cares? Make some phone calls? Ask around. Yeah, that will work until the information is needed immediately.

A student accustomed to looking up information on a mobile phone won’t be able to find something. The document is a 404 or the information returned is an ad for a Temu product. So what? The kid will have to go the library, which one hopes will be funded, have printed material or commercial online databases, and a librarian on duty. (Good luck, traditional researchers.) A marketing team eager to get information about the number of Telegram users in Ukraine won’t be able to find it. The fix is to hire a consultant and hope those bright men and women have a way to get a number, a single number, good, bad, or indifferent.)

My concern is that as the intensity of the objections about a standard procedure for building an index escalate, the entire knowledge environment is put at risk. I have worked in online since 1962. That’s a long time. It is amazing to me that the plumbing of an information economy has been ignored for a long time. What happens when the companies doing the indexing go away? What happens when those producing the government reports, the blog posts, or the “real” news cannot find the information needed to create information? And once some information is created, how is another person going to find it. Ask an eighth grader how to use an online catalog to find a fungible book. Let me know what you learn? Better yet, do you know how to use a Remac card retrieval system?

The present concern about information access troubles me. There are mechanisms to deal with online. But the reason content is digitized is to find it, to enable understanding, and to create new information. Digital information is like gerbils. Start with a couple of journal articles, and one ends up with more journal articles. Kill this access and you get what you wanted. You know exactly who is the Salem witch.

Stephen E Arnold, July 17, 2024

x

x

x

x

x

x

Comments

Got something to say?





  • Archives

  • Recent Posts

  • Meta