Nire Bryce

able to accurately and swiftly cut out everything but the meat, nearly impossible to pass on to an apprentice in less than five years, end-users see it as a simple snack

Like, why, how did we get to a world where it's necessary to have human experts at feeding exactly the right input into an inscrutable AI model?

According to IR¹, when you're familiar with a general field, it will take you about eight hours to construct one (1) boolean query to find the documents on a particular topic, from an archive that has full-text search. This means the search interface is based purely on the contents of the document, and there is no quote-unquote "Algorithm™" trying to ascertain relevance for you.

I am told this is a very robust result. Meaning it's been replicated in several studies across different fields and contexts. It takes a lot of iteration to encompass all of the phrase variations, and then exclude the irrelevant documents from a different field that use the same terminology, and then encompass those variations. Literally a whole work-day to find, I dunno, all the papers on PubMed about a particular medical condition, or all the court rulings in LexisNexis about name changes.

So, what does AI art have to do with Information Retrieval? Well...

...like, that's what AI art literally is right now. "Prompt engineering" is document retrieval. You're specifying which documents should be retrieved from the training data to be blended together into the output. It's a natural-language interface instead of a boolean query, but I don't consider that difference transformative.

I find this framing to be a useful perspective to think about what's going on. It demystifies "the algorithm" into something that's easier to wrangle with: construct a natural-language query that identifies the subset of training data which is most relevant to this round of image generation. The ability of the prompter to encircle that subset is IMHO of extreme importance to the discussion about the rights of the creators of the original training data, and extent to which the prompter is doing "creative" work.

1. Information Retrieval (IR) is the study of technology and interfaces for searching collections of full-text documents, in contrast to databases which store structured information. PageRank, the original Google algorithm, falls squarely into the field of IR, as does Lucene/ElasticSearch if you've used it.

they dump of ours on the street? We put one of theirs in the morgue."

i was sent this email, to an ancient adobe account back before i used a password manager for everything and just keysmashed the passwords i only intended to use once. yes, I do better now. Yes, I knew better then.

that said: this is how everywhere should do it.

Please update your Adobe password.

Dear Adobe Account Holder, Your privacy is important to us and we are committed to being proactive when we detect events that might put your personal data at risk. To help protect your information, we have reset the password for the account associated with your Adobe ID that may have been compromised in data breaches from other online services. The next time you log into your Adobe account, you will be prompted with a password reset notice. If you have questions, please review our guidelines on Adobe.com for changing or resetting your password. Or, you can contact us.

In general, we recommend that you use different credentials on each website. You should take this opportunity to change your password on any websites where you used the same username or password as your Adobe ID. You can also enable Adobe’s two-factor authentication for your account to strengthen your account security. For more information on setting up two-factor authentication, please review our guidelines on Adobe.com.

Sincerely, Adobe Customer Care