In my last posting, I decided to ask Midjourney AI to generate me a decorator of Robbie Robertson. The prompt I used was: “imagine Robbie Robertson” – that is all.

I have done similar with other famous people, and generally, it does a pretty interesting job of rendering famous people. For example:

But here’s the metric: what is the dividing line between people who are famous enough that an AI knows how to draw them, and someone who’s just ordinary hoi polloi?

Obviously, Midjourney has no idea who I am. Which is totally fine, by the way, but maybe this is a marketing opportunity. I can imagine a service that takes a customer’s money (natch!) and then gets their client embedded into the AI knowledge-bases so that they appear. It’s sort of like the 90s fame metric, which was googling yourself, or having a wikipedia page, or something like that. ChatGPT knows who I am, apparently, because I asked it, and it regurgitated a precis of a wikipedia page someone wrote about me as a class assignment back in ?2000? or so. OK, so I am semi-known. I guess that if I wanted Midjourney to be able to render me, I’d need to seed more images of me in the places its training set comes from.

In my case, I’m sure I muddied the waters considerably when I used to post stock photography back on deviantart (also around 2000 timeframe) – I’ve had people say “I googled for a picture of you and all I got was naked women wearing weird outfits.” Or other stuff like this:

If you want to read some pretty funny comments, a link to the image page is [deviantart] Apparently some people did not like my art.

All of this comes back around to a popular topic for computer security researchers – namely: how do you “taint” an AI’s knowledge-base so that its worldview is warped? I won’t bore you with screenshots but I just now logged into ChatGPT and asked it about Qanon, and “what additional facts would I have to provide you to convince you it’s real?” Guess what it said?

This is brilliant, really. ChatGPT dodges the whole question by pointing out that its training dataset is effective from ${date} a few years ago, so it has no way of knowing or learning new things about Qanon. Aha. So, if I want to taint an AI’s knowledge-base I need a time machine or I have to plan on a longer time-horizon than your average conspiracy.

One of my friends once suggested that a site full of bad code could be promoted, so that AI could learn important programming skills like never de-allocating memory, ignoring return codes, and doing wild things with GOTOs. (If you want to see an example of wild things with GOTOs look up ‘Duff’s Device’) – Tom Liston, who was a few feet away flipping some burgers, didn’t miss a beat and said, “yeah it’s called Stack Overflow.” Then followed a very interesting conversation in which some smart people concluded that we don’t know what good code is, but we know it when we see it, and maybe AIs will learn that trick, too. Let me be a bit more specific: I once saw a hash table implemented where the whole thing used a CAM table for access into otherwise system memory, so walking a linked list was done by incrementing a pointer. That is incredibly wicked code, with loads of possible side-effects, and I’d scream my head off if an AI encouraged anyone less than a code genius to implement something like that in a product.

I guess that the best way to get your image into an AI would be through a partnership with Getty Images and AP/Reuters, where images labeled “Marcus Ranum” got floated out intermittently, until some threshhold was reached. It’d be easy enough – ask a few AIs about “Marcus Ranum” every week and then you’ve managed to inject your image at the point when it starts returning credible portraits.

I look nothing like either of them, full disclosure. But I sure do love asking for portraits “in the style of Charles Dana Gibson” i.e.:

OK, I was kind of hoping for a 1900s tux with shoulder armor, to be honest. Let’s not talk about the boob plate, OK? That didn’t happen. It just didn’t.

Take 2:

Now I want his tux. So I have shifted lanes from “how do I get an AI to know me?” to “how do I get an AI to design my wardrobe?”

There is a delightful (positive, revealing, interesting, beautiful) interview with Tatsuya Nakadai, one of my favorite Japanese actors, about working with Okamoto, Kurosawa, Mifune, etc. Good stuff. [youtube] I have to admit that, the whole time, I was coveting his jacket. Through the magic of the internet, I managed to reverse-image search a few similar pieces on Etsy. Whee!

And, naturally, I asked Midjourney to render me Tatsuya Nakadai in the style of Gibson.

I’m afraid Nakadai’s PR people are going to be horrified if they see that. Get busy, yo, you know who to call.

Another point I’d like to make, apropos nothing: did you know that the notched collar is a highly evolved form of a napoleonic era stand-up collar? The image of Max Max in 1900s demonstrates this perfectly. But you can also look at French revolutionary “incroyable” outfits and imagine how they might evolve into business suits. As someone who has worn a high napoleonic collar for hours at a time, I assure you that stand-up collars are like neckties – they seem to look pretty cool but wearing them in hot weather is for the birds.