OpenAI touts a new flavour of GPT-3 that can automatically create made-up images to go along with any text description OpenAI released a sneak peak of its latest GPT-3-based neural network, a 12-billion-parameter model capable of automatically generating hundreds of fake images when it is given a text caption, stylized as DALL·E. That might not sound all that interesting at first, but you have to see DALL·E in action to really appreciate it. …
realistic images of animals, objects, or scenes.
Well, bugger me with a fish fork.
I wonder if they've looked into Heraldic language? I'm sure three coloured cubes won't be as tricky to describe as the Royal Coat of Arms:
Quarterly, first and fourth Gules three Lions passant gardant in pale Or armed and langued Azure (for England), second quarter Or a Lion rampant within a double tressure flory-counter-flory Gules (for Scotland), third quarter Azure a Harp Or stringed Argent (for Ireland), the whole surrounded by the Garter; for a Crest, upon the Royal helm the Imperial Crown Proper, thereon a Lion statant gardant Or imperially crowned Proper; Mantling Or and Ermine; for Supporters, dexter a Lion rampant gardant Or crowned as the Crest, sinister a Unicorn Argent armed, crined and unguled Proper, gorged with a Coronet Or composed of Crosses patées and Fleurs-de-lis a Chain affixed thereto passing between the forelegs and reflexed over the back also Or. Motto "Dieu et mon Droit" in the compartment below the shield, with the Union Rose, Shamrock and Thistle engrafted on the same stem.
Doesn't that heraldic language have formal rules, so so that it is essentially a declarative description of the graphic (PostScript for shields :-). The graphics themselves also consist of similar formal elements. I am sure that DALL-E or similar would work fine after being trained with hundreds of pairs of coats of arms, and the corresponding formal description.
I agree. Isn't the language used to describe coats of arms in the UK (possibly aside from Scotland, which has its own heraldic authority) based on Fox-Davies' The Art of Heraldry? My understanding is that it's quite regular.
It'd be interesting to try to train GPT-3 on something rather less predictable, such as the descriptions in Joyce's Ulysses or Pynchon's Gravity's Rainbow, or other similar modernist1 works.
1Sure, many people consider GR postmodern, rather than (high) modern or avant-garde. I don't, particularly; I think that while it certainly has deliberately postmodern elements, particularly in style (flattening of value, bricolage, blah blah whatever), really postmodern prose has moved so far beyond the early works that were assigned that label that it's more useful, if we're going to try to define literary movements or periods at all, to consider much of the early-pomo work like GR as late-modern or transitional. Contemporary hypermedia works, or even some print works such as House of Leaves, are much further removed from GR than GR is from, say, Woolf or Faulkner. Hell, there are ways in which Cantero's The Supernatural Enhancements is arguably more postmodern than Gravity's Rainbow. And it's a lot more fun to read, so there's that.
Years ago, I started to write a program which I called Blazon, which used a description like that you quote (although I started more simply) and parsed it as if it was a scene description language, like PovRay. It's fairly easy to get "Sable, a cross Argent" for the Cornish flag of St Piran, but the precise grammar and parsing it to SVG (which was the strategy I adopted) defeated me in the end. Perhaps I should take it up again, since I'm not going anywhere!!
This software and the machines that we produce may turn out to be the next generation of us in one form or another, out children if you wish. We don't have to consider them and us to be particularly separate.
If we can get over the "body-horror", I'm sure that a lot of this stuff will become more tightly integrated with us anyway in the future. I'm just thinking about the kind of work that I could do if the a database of all of human knowledge was directly interfaced with my brain as though it was part of me. Imagine having intuitive access to that. That would be very weird.
"I'm just thinking about the kind of work that I could do if the a database of all of human knowledge was directly interfaced with my brain as though it was part of me. Imagine having intuitive access to that."
I'm imagining you with a phone embedded in your head. All you have to do is to think of something and you'll quickly get a search of all the stuff written on it. Most of which tells you information you don't need, a lot of which is wrong, a lot of which is designed to convince you rather than inform you, some of which includes what you want to know but is written in a way that you don't find it, and some of which includes what you want to know but you can't understand it.
Also, how is this getting us anywhere closer to the singularity, for that's basically what you're asking for? All this does is cut up some images and put them together to attempt to correspond to a label. The model doesn't understand what any of these objects is; it only understands what they look like, and that's being generous. I can give you that too, and I'd not be using AI either. If I just tag a bunch of sections of pictures "this is a chair", "this is a piano", "this is a dog", I can get a program to read your sentence and paste them on top of a background "this is a living room", "this is a street", "this is a lake". What this program does is cut out my labeling process by taking a bunch of chair pictures and determining automatically that they're mostly chairs, but it needed someone to write that label to figure that out. It shows effort on the part of the programmers, people who write image captions, and the willingness to spend a lot of computing time building a model, but doesn't really help in the creation of an intelligent computer.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2023
