Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[jipehog]

Why would people worry about jobs going away?

If you don't want talk about that, how more pressing concern in realm of cyber with influence and misinformation camping's and all other social ills tied to this topic.

Well yeah.
Most of the practical uses for AI have been in the realm of propaganda and mis-information.

I am surprised that as liberal who dislike the system you are not more interested in the potential of AI killing off jobs and widening wealth inequality..

Otherwise, I don know where you get that notion that cyber enjoy the most practical uses. The way I see it applications of "AI" already impact everyone's everyday lives in some way or the other, Cyber is just the sexier headline material and in our case lends well with original topic and the ChatGPT like detour (replacing troll farms content creation?) but overall like in every other field atm it is just a tool that can boost productivity in some aspects.

Arguably big-data is the driving force of this all, which AI can help synthesize and analyze for various purposes. From identifying patterns whether its facial recognition or for driverless cars, identifying political indentation and the like on social media, to less sexy stuff like crop and soil monitoring, or more creative ideas like this example:

How about these chopsticks mr Miyagi: Pentagon Sees Giant Cargo Cranes as Possible Chinese Spying Tools

Some people think that USA is paranoid, I think that (a) USA speaks from experience of doing the same, (b) regardless, data is the name of the game, who ever is in charge would be able todo a lot with list data of what goes around, much like tech companies like Google and Facebook use user data for revenue and competitive insight.

[EuchreJack]

Why would people worry about jobs going away?

If you don't want talk about that, how more pressing concern in realm of cyber with influence and misinformation camping's and all other social ills tied to this topic.

Well yeah.
Most of the practical uses for AI have been in the realm of propaganda and mis-information.

I am surprised that as liberal who dislike the system you are not more interested in the potential of AI killing off jobs and widening wealth inequality..

Otherwise, I don know where you get that notion that cyber enjoy the most practical uses. The way I see it applications of "AI" already impact everyone's everyday lives in some way or the other, Cyber is just the sexier headline material and in our case lends well with original topic and the ChatGPT like detour (replacing troll farms content creation?) but overall like in every other field atm it is just a tool that can boost productivity in some aspects.

Arguably big-data is the driving force of this all, which AI can help synthesize and analyze for various purposes. From identifying patterns whether its facial recognition or for driverless cars, identifying political indentation and the like on social media, to less sexy stuff like crop and soil monitoring, or more creative ideas like this example:

How about these chopsticks mr Miyagi: Pentagon Sees Giant Cargo Cranes as Possible Chinese Spying Tools

Some people think that USA is paranoid, I think that (a) USA speaks from experience of doing the same, (b) regardless, data is the name of the game, who ever is in charge would be able todo a lot with list data of what goes around, much like tech companies like Google and Facebook use user data for revenue and competitive insight.

Hey kid.
I got no idea what you're talking about.
But I ain't no liberal.

I'm getting pissed off that we're talking about AI like it actually means anything more than a plow or a tractor.

And while I ain't a liberal, I ain't the sort that told Crog the Caveman that the world would end because he dared to create the first knife.

Spoiler (click to show/hide)

[King Zultan]

AI ain't nothing but a tool, and it isn't even that good of one so I don't know what everyone is making such a big deal about if for.

[Random_Dragon]

We've gotten bored of balloons and didn't think to rename this into the "Ohio becoming The Zone" meme thread. 

[jipehog]

The arguments that AI is just a tool are silly, the industrial revolution was brought about by just a tool and yet it marked a major turning point in history. I can understand the optimist belief that it will be fine but i do not share it, to me it sound misguided more like 'global warming is just a natural cycle' sort of denial

I would remind that historically adaption of breakthrough technologies hasn't been all that pretty. For example, before we got to enjoy the fruits of the Industrial revolution there were roughly a hundred years of birthing pains during which life were made worse for many, with colossal change to social order and world order, including the rise of communism that changed how we look at the world today.

I see AI as such breakthrough technology that will have huge impact on our lives. And hope that even the optimist sees the tip of the iceberg of potential problems considering the already ongoing controversy by USA commercial tech giants. Also I would note that: Development in other fields complement AI own; AI adoption rate across the industry (its already ~200billion industry?); Increasing investment in the field and improvement in capabilities and accessibility by both by private companies and nationally as part of the USA - China tech war.

Otherwise, in terms of most of the practical uses case atm my guess is that advertising, surveillance and automation are at the top.

[Starver]

We've gotten bored of balloons [...]

Oh, pity, I was going to link this that I rediscovered from May 2019...

[King Zultan]

We've gotten bored of balloons and didn't think to rename this into the "Ohio becoming The Zone" meme thread. 

But there's still time to do that, we could even blame it on balloons or something to tie the threads together.

[dragdeler]

Train driver: the balloon reflected the sun into my eyes, sorry again.

[anewaname]

Spoiler: This is a "text balloon": (click to show/hide)

With AI, people mock the AI's generated 2-d images, but those are the result of using many 2-d images in the data. When the AI is given accompanying 3-d data models, does it get better at creating 2-d images of things like noodle-eating and tigers?

[Starver]

I suspect that wouldn't help. If you plug a corpus of 3D things in (assuming you could provide the equivalent quantity/quality of training materials) the result of blindly discovering the important commonalities and differences and distinctions will result in a flow of 3D things coming out. First of all, you'd have to supply a pipeline into 2Dness (and sensibly so, perhaps separately train another aspect of AI to properly place and orientate such output items to compose them well) and, secondly, there's a whole new dimension (or more, once you multiply elements) of inadvertent error that the engine can stray into.

The chopsticks/noodles problem is seemingly not really much to do with the lack of 'depth perception' of the "understanding" algorithm, but that straight or bendy distinct filaments defines "chopsticks" and "noodles" better in multiples than anything else (including a singular stick or strand) and thus it happily overfulfills the number in the fine tradition of the Paperclip Maximiser. Somehow[1]...

Perhaps the weird twist of fingers might be 'cured', if you go to the trouble to demonstrate the normal 3D composition of a hand, but I'm not sure even that's guaranteed. Current "flat drawn" fingers might gain depth distortions to rival the oblique ones we see, as the recomposition/reassembly of features includes far more differing orientations to try to merge into a generative output, and the method seems to be lax with the consistent length and extent of digits (and termination of the swathe of pixels) so I'm not sure it can learn much more from the variety of related 3D data (and how it generates its voxels).


Which is not to say that you can't actually expect anything to work in 3D as good as it does in 2D, all else being equal (except data volume, which needs to be exponentionally greater[2]). But it's not necessarily going to solve the 2D problems.



[1] I don't know how much we've poked into the 'trained database' to try to distill out the learnt-rules that it seems go a bit wild in these cases. I suspect that it'd be forensically obscure without spending a disproportionate amount of time digging about in data that is effectively "machine only"-readable. Perhaps tweaking key (un-named) values manually and re-running to find how we can guarantee just two (nicely held) chopsticks or a reasonable representation of pasta and mouth, but not realising that this makes the next request for an image of (say) a horse jumping a gate to somehow go all wrong because our value was doing double-duty in this other manner of subject, for various mostly (and certainly trivially) ineffable reasons...

[2] Or simplified, making it a different animal from the 2D Raster->[Black Box]->2D Raster training-to-result sequence, clearly based upon a simple array 'canvas' that can be kept as a valid 2D array (no matter how messy the contents). Do you tell it how to understand 3D Vector, in its necessary tree-structure? Or just plug in raw VRML/whatever and have part of its training-problem being to output intrinsically valid VRML markup, even before populating such a valid descriptor with hopefully aesthetically valid content. More decisions, more development, more internal complication to the project (one way or another).

[WealthyRadish]

Spoiler: This is a "text balloon": (click to show/hide)

With AI, people mock the AI's generated 2-d images, but those are the result of using many 2-d images in the data. When the AI is given accompanying 3-d data models, does it get better at creating 2-d images of things like noodle-eating and tigers?

The training data is very different. Nearly all 3D models out there are hollow shells where vertexes are grouped into triangles (and the 3 vertexes are specifically oriented clockwise or counterclockwise), with each vertex typically containing more than just positional data (i.e. at least one set of 2D texture coordinates per vertex as well for coloring and physics lighting properties, as well as things like surface normal vectors). All this is hard to get right even by humans, and games are of course notorious for weird 3D artifacts and oversights where the smoke and mirrors fall apart, because it's fundamentally a very fragile system.

With 2D images, it's possible to use different techniques. For instance, you can start by randomly generating complete noise in each pixel, and the "AI" is just a bunch of linear algebra trained to iteratively transform junk into something that looks like it's training data. To attempt the same method in 3D, you would need to create/find a ton of 3D pixel data (i.e. voxel graphics) while also producing a similar crummy voxel look or vastly increasing the already high computational costs.

[Starver]

Personally I'd go a NURBS-and-skeleton route, when supplying 3D models, as a more compact way of depicting than a full on mesh of triangles (except where the models are of the more basic geometric figures). And plane/gradient-only textures.

But it's moot, as while you can easily scrape a hundred thousand anime/landscape/whatever images, to learn from with little complexity (other than converting all image formats to the same canvas format), to do the same is more difficult with the relatively sparse availability of models,cand split amongst many formats/sources, from AutoCAD to Blender, XVL to ZModeller, which you need to make equally pallatable to the algorithm you're trying to train.

Though I think if you can pre-boil down into some singular format with some useful basis for 'intelligent' mixing, matching and morphing (such as...) then you're only left with arm-twisting it into making a renderable output and not breaking the syntax/hierelarchy completely.

[jipehog]

Obviously current narrow AI is only mimicking it has no concept of 2d or 3d, to these algorithms a paper/pasta spreadsheet, mobius strip or Klein bottle is all the same. However, AI is also very good at content analysis, and between tremendous amount of processing power and data it increasingly becomes good enough at many aspects.

Content creation would be the pinnacle of ever increasing capabilities and adaption but people don't give enough credit to what even a dumb AI can do. Consider meta analytic scandal, the algorithm employed there are nothing special, but the amount of data collected there and what you can do with it led to huge concern for serious risk of harm bringing about big changes in terms of data privacy/security in digital world. You might think that its just USA vs China politics, but that is not true many countries challenge the dominance and practice of USA tech giant including its allies EU Canada etc..

[Robsoie]

There's currently a tool "Control Net" that can help the user to "pose" a character before going through a prompt to get the AI to generate some artwork.
https://stable-diffusion-art.com/controlnet/

I can imagine that this developped further may also allow to pose hands in a believable enough manner, allowing then the user to get the AI to generate more correct hands without depending on sheer luck (i got the AI to sometime generate hands that looked relatively correct, but it's really pure luck considering how many re-run i go through when it comes to hands)

[EuchreJack]

We've gotten bored of balloons and didn't think to rename this into the "Ohio becoming The Zone" meme thread. 

Ohio already has its own thread