How far will AI Go and how soon?

[Drakkhis]

Hmm So Now we have things like Imagen to create video from text, Dreamfusion to generate 3d Models from text, OpenAI Codex to generate code from text. How long is it going to be before you can Type out what you want and get a Full playable video game from text?

[HanZie82]

When was Skynet going to be online again?

[emosun]

24 minutes ago, Drakkhis said:

How long is it going to be before you can Type out what you want and get a Full playable video game from text?

minimum 20 years

what we have for AI in 2022 is interesting and lays the groundwork for the general idea of AI , but is very simplified low resolution and basic/not smart compared to where it needs to be/will be in a few decades. By the time I'm an old lady it'll be pretty cool.

[emosun]

5 minutes ago, HanZie82 said:

When was Skynet going to be online again?

august 14 29th 1997

[Roswell]

Millions of license free images and videos out there to train an AI, exceptionally few open source video games that are worth playing.

[YellowJersey]

I, for one, welcome our new robot overlords.

[Kisai]

5 hours ago, Drakkhis said:

Hmm So Now we have things like Imagen to create video from text, Dreamfusion to generate 3d Models from text, OpenAI Codex to generate code from text. How long is it going to be before you can Type out what you want and get a Full playable video game from text?

So very little of this online AI generation stuff impresses me.

 

Artists are flipping out about the AI art generators, unable to see the difference between img2img and text2img. What you've seen on social media has been so heavily cherry picked, that there is literately nothing to worry about. What people are supposedly producing as "nice AI artwork" is really just one piece of artwork that's been style transferred to another with img2img.

 

Very little of what you see on social media is text2img. If AI could really produce a nice image from only text, it would need to actually understand what is in the image. It does not. It can not. Neither can GPT-3 when it comes to writing novels.

https://undark.org/2022/10/07/interview-why-mastering-language-is-so-difficult-for-ai/

 

AI DOES NOT UNDERSTAND HUMANS. It has no understanding of language. To paraphrase the above "It's fancy auto-complete"

 

Inpaint is a little more interesting, allowing you to extend an existing image, but again, it does not understand what is in the image. Again, it's "autocomplete"

 

The most "impressive" stuff I've seen in AI, is TTS. This is because unlike the image stuff, where it ultimately "re-creates" the training data poorly, TTS training doesn't have that goal. TTS's primary goal is "sounding lifelike". It actually attempts to learn the relationship between the individual graphme's or phoneme's in words, so ultimately you do get a "lifelike" result, but it falls short of being convincing to people who are aware of what AI's sound like. It only fools people who have not heard any TTS before, or still think computers sound like Siri. How often have you seen a "AI" in a tv show or film and it sounds like Siri?

 

But, it doesn't fool everyone.

https://theconversation.com/deepfake-audio-has-a-tell-researchers-use-fluid-dynamics-to-spot-artificial-imposter-voices-189104

 

Part of this lies in how TTS research has been using shortcuts. Not using wideband audio, not using 48khz or 96khz sampling rates, and generally avoiding curating their training data with this in mind. This is the same shortcut problem that AI art in using only 512x512 sized samples.

 

What we're seeing in blowback in AI art, is something that comes from not curating the training corpus. When you start mixing the qualities of various sources, what you get is not "better", but rather the highest quality data, drowning out the lowest quality. This results in the better data over-fitting. At best AI art is kitbashing, at worst it's plagarism. It's lacks creativity, it lacks understanding.

 

There is not enough video memory in the world to be able to process HD, let along HDR 8K images. We are miles away from that being possible. Likewise in TTS, we are miles away from voices that would forensically convincing, capable of replicating anyone's voice. What we have are very "shortcut-taken" results, that are unconvincing to experts in the AI field and unconvincing to experts in art and audio fields.

 

The only people being fooled are the same people who are fooled by "psychic surgery" and Quija boards. Please put your skeptic hat on when people tell you something is real and it seems "too good" or "too unbelievable" to be real. Because image2image is pulling out the chicken guts and passing it off as human guts. It's enough to convince the mundane, it would never convince a doctor, or even a nursing student.

 

 

Incidentally

https://www.whitehouse.gov/ostp/ai-bill-of-rights/

Quote

You should be protected from unsafe or ineffective systems. 

 

You should not face discrimination by algorithms and systems should be used and designed in an equitable way.

 

You should be protected from abusive data practices via built-in protections and you should have agency over how data about you is used.

 

You should know that an automated system is being used and understand how and why it contributes to outcomes that impact you.

 

You should be able to opt out, where appropriate, and have access to a person who can quickly consider and remedy problems you encounter.

So the AI bill of Rights here is less about the prevention of "deepfakes" and more about not having AI weaponized against you.

 

In regards to AI artwork, or TTS voices, or AI-written novels. As long as these are labeled as generated by an AI, and not passed off as being human-made, we are fine. 

 

https://www.scientificamerican.com/article/google-engineer-claims-ai-chatbot-is-sentient-why-that-matters/

 

This is what happens when people aren't sufficiently skeptical or savvy enough about AI. No existing AI is sentient. They do not learn once they are done training.  You're being tricked by fancy auto-complete that is telling you what you want to hear.

 

Ask an AI to do something culturally recent, eg artwork of rick and morty season 5, an impression of any Disney actor who just turned 18, or write about a current twitter trend. It will utterly fail to do this, because it doesn't understand what these are, because they are not in the training data. It isn't connected to twitter and reading everyone's posts. With the AI artwork, the training data was likely captured two years prior to the public-availability of the tool itself.

 

I myself tried DALL-E2, I tried queries that would have been relevant in 2021 and 2022, no match. It produced garbage based on what it thought those words were. It's interesting technology, but it's so mind-numbingly far from producing anything "final" that when I see the cherry-picked stuff people post on twitter, I know it's image2image, because the AI does not produce that with text2image.