AI will never out-describe me

Jupiter Rowland

jupiter_rowland@hub.netzgemeinde.eu

Duothematic channel. Primary topic is virtual worlds/OpenSim, secondary topic is the Fediverse beyond Mastodon. Follow me from Mastodon or anywhere else in the Fediverse.

Ort:: The Hypergrid

Geschlecht:: Männlich

Homepage:: https://hub.netzgemeinde.eu/profile/jupiter_rowland

AI will never out-describe me

Wed, 11 Oct 2023 23:25:58 +0200 zuletzt bearbeitet: Wed, 11 Oct 2023 23:25:46 +0200

Jupiter Rowland

jupiter_rowland@hub.netzgemeinde.eu

AI-written image descriptions vs my own image descriptions; CW: long (785 characters)

Artikel ansehen

Fediverse

show all 11 comments

Wed, 11 Oct 2023 23:36:32 +0200 zuletzt bearbeitet: Wed, 11 Oct 2023 23:38:36 +0200

_jayrope

jrp@hub.kliklak.net

@Jupiter Rowland I think we may look at AI as a serious competitor in a few years again. At the current time it is just a pathetic style experiment and factually mostly wrong. Don't waste your lifetime on a machine. Or as i pointed out in 2010 already: https://prinzenallee.bandcamp.com/track/dont-let-nerds-take-over-your-life-2

Thu, 12 Oct 2023 00:47:06 +0200 zuletzt bearbeitet: Thu, 12 Oct 2023 00:46:54 +0200

Jupiter Rowland

jupiter_rowland@hub.netzgemeinde.eu

@_jayrope I'll be impressed when an AI manages to figure out just from scanning a picture

that it was taken in a virtual world
in which virtual world it was taken
where in that virtual world it was taken
how to access that virtual world, travel to that place and look around for more details that aren't visible to the naked eye just from looking at the picture

Guess how I manage to transcribe text that's a fraction of one pixel high and/or partially obscured.

Fri, 13 Oct 2023 12:45:29 +0200

Scott M. Stolz

scott@biztechtonics.net

AI is trainable though. If you trained it on a specific data set, and told it to ignore irrelevant stuff from the internet, it could be able to detect those things.

AI is susceptible to the GIGO problem. Garbage In -> Garbage Out. And if the AI just grabs random information from the internet, it gets a lot of garbage.

And I read something interesting. AI starts to go crazy (get confused, lose accuracy), if you feed it AI output. A sort of AI Mad Cow Disease, if you will. It shouldn't try to eat its own brain.

Fri, 13 Oct 2023 19:48:32 +0200 zuletzt bearbeitet: Fri, 13 Oct 2023 19:48:15 +0200

Jupiter Rowland

jupiter_rowland@hub.netzgemeinde.eu

@Scott M. Stolz I don't describe my in-world pictures from looking at the resized 800x533-pixel picture that I'm going to post. I don't describe them from looking at the 2100x1400-pixel original either.

I describe them by being in the very place in-world where I've taken them and then looking around, camming around, zooming around. In one case, I've actually taken several older pictures from the same place as additional references because the place that I was describing had changed since when I had taken the picture that I was describing.

This enables me to do things an AI will never be able to do such as:

transcribe text literally that's a fraction of a pixel high in the 2100x1400 original
transcribe text literally that has something standing in front of it and obscuring it
transcribe text literally that's partially outside the picture
identify a tiny 2x2-pixel speck as a strawberry cocktail
describe what a spaceship looks like on the inside even though this inside isn't even visible in the picture (technically speaking unnecessary, but still)

Also, I can tell accurately where the picture was taken. Again, it was usually taken inside a 3-D virtual world that's fairly unknown.

If an AI were to describe images at the same level of detail, accuracy and informativity as I do, it would have to

figure out just from scanning the image
- on which virtual world platform the picture was taken (and believe me, it's possible to make OpenSim basically indistinguishable from Second Life)
- in which world (on which grid) the picture was taken
- on which sim on that grid the picture was taken
- where on that sim the picture was taken
develop, compile and install an AI-controllable viewer against OpenSim's API that's compatible with various different OpenSim versions
create an avatar on a grid
figure out how to control the avatar and the camera
steer the avatar to where the picture was taken
look around and try to figure out what all the stuff there is, what its story is and how it works

Fri, 13 Oct 2023 22:43:06 +0200

Mike Macgirvin 🖥️

mikedev@fediversity.site

"A picture is worth a thousand words" is often why we post pictures in the first place. Until you make alt-text dead easy for content creators, like no typing at all, getting everybody on board is going to be an uphill battle.

I'll be happy when I can take a picture and tell my camera what I just snapped and it goes into the EXIF and when I upload, it grabs the description from there and adds it to the post. We're not there yet.

I would be less happy but thankful if we can just reach agreement on an EXIF field name so that I can implement the last step. Apparently there has been discussion about this but getting standards adopted across camera manufacturers is just like the fediverse. See #xkcd927.

Sun, 15 Oct 2023 15:35:49 +0200 zuletzt bearbeitet: Sun, 15 Oct 2023 15:35:48 +0200

Nicolas Le Gland

nicolaslegland@mastodon.social

Fediversity search results for “xkcd927”, population count 0 (zero).

@mikedev @jupiter_rowland is this https://fediversity.site/search?search=xkcd927 some kind of meta and cruel joke, about the state of Fediverse search engine technology as a 16th standard, when https://www.bing.com/search?q=xkcd+927, https://duckduckgo.com/?q=xkcd+927, https://www.google.com/search?q=xkcd+927 or https://search.yahoo.com/search?p=xkcd+927 exist?

Sun, 15 Oct 2023 18:53:36 +0200 zuletzt bearbeitet: Sun, 15 Oct 2023 18:53:33 +0200

Jupiter Rowland

jupiter_rowland@hub.netzgemeinde.eu

@Nicolas Le Gland No, this is because @Mike Macgirvin (dev) is not on #Mastodon, but on #Streams which he is also the creator of. (streams) is part of the #Fediverse, it is largely compatible with Mastodon, but it is not a Mastodon fork, it is not inspired by Mastodon, it's a descendant of projects way older than Mastodon, and it's a whole lot different from Mastodon.

One of its quirks is that its hashtags don't link to the hashtag search on whatever instance the hashtag is being seen on, but on whatever instance the hashtag was created on. And fediversity.site is the (streams) instance on which Mike's channel resides.

#MastodonIsNotTheFediverse

Sun, 15 Oct 2023 19:21:04 +0200

_jayrope

jrp@hub.kliklak.net

@Mike Macgirvin (dev) Hmm, sounds like that might conflict with numeroue people deleting exif data on their images for privacy purposes?

Sun, 15 Oct 2023 23:07:28 +0200

Mike Macgirvin 🖥️

mikedev@fediversity.site

If you want to delete the exif description (camera makers are already working on this, but there are competing standards), you'll need to type it in by hand or selectively delete. Sorry - you can't have it all.

Sun, 15 Oct 2023 23:12:22 +0200

Mike Macgirvin 🖥️

mikedev@fediversity.site

Did you add fediversity.site as a search engine? We use opensearch just like everybody else. Except maybe Mastodon. Cheers.

Sun, 15 Oct 2023 23:35:08 +0200

Mike Macgirvin 🖥️

mikedev@fediversity.site

Oh right. You don't have permission to search my tags because you're not one of my connections. This fell back to a site search (which you can search), but my tags weren't included because you don't have permission to search my stuff and nobody else on this site has used that tag. So you could call it a cruel joke, but I just call it permissions that actually work.