WEBVTT

00:00:00.000 --> 00:00:03.720
I am here with Dan. Hey, and what do you do for the company here?

00:00:04.240 --> 00:00:08.960
I'm the infrastructure and technical specialist these days. I pretty much just write documentation

00:00:09.360 --> 00:00:14.840
Sounds good. Sounds good. So I'm just asking everyone here at the company what the thoughts on AI is the current state of it

00:00:15.120 --> 00:00:19.760
So I'll ask you first. Sure thoughts on the currency of AI. It's pretty interesting

00:00:19.760 --> 00:00:26.760
there's a lot of cool stuff going on and I'm excited to see it kind of develop over the next I guess a year or a few months even

00:00:27.120 --> 00:00:31.240
So you talk kind of sound like you're for AI more than against AI

00:00:31.720 --> 00:00:35.320
Yeah, I would probably say my thoughts on it are a little bit mixed

00:00:35.320 --> 00:00:40.880
I think it's not maybe as terrifying as people think it is but yeah, certainly certainly for it

00:00:41.400 --> 00:00:46.840
Sounds good. Sounds good. So if you're for it, then what's one thing you're really against in terms of the state of AI at the moment?

00:00:46.840 --> 00:00:50.320
I think probably some of the protectionism. I mean

00:00:51.200 --> 00:00:53.920
Looking at some of the requirements for doing these trainings

00:00:54.400 --> 00:01:02.160
You're spending millions of dollars on like cloud infrastructure GPU time just to actually create a

00:01:02.720 --> 00:01:11.960
reasonable model and so People I think are willing to protect that rather than just open sourcing it and letting any other company who has access to these massive data centers

00:01:11.960 --> 00:01:22.000
Like get in there, right? Okay, okay So have you used any AI currently like that where it's like they have chat GBT the image random readers music

00:01:22.200 --> 00:01:27.400
Have you use any of it? Yeah, I've played with pretty much all of them and I actually run a couple at home

00:01:27.400 --> 00:01:34.360
So I run an LLM as well as stable diffusion for image generation and I run those locally

00:01:34.360 --> 00:01:42.340
I recently just bought a couple more 30 90s and you know upgraded some of them with VRAM so that I could try some of the larger models

00:01:42.720 --> 00:01:45.220
So they're like a chat GBT for

00:01:45.880 --> 00:01:49.180
But you can only talk to them for like I don't know

00:01:49.780 --> 00:01:57.260
Two to five minutes before the like token inference length gets too much and it takes like a minute to respond to you

00:01:57.540 --> 00:02:06.580
But initially they're they're really quite powerful And they're a little bit more entertaining because they're the uncensored varieties. So what do you mean by that?

00:02:06.620 --> 00:02:11.300
So, you know, you say something to chat GBT for and it goes as a large language model

00:02:11.300 --> 00:02:16.260
I can't respond to this the uncensored one the uncensored ones are kind of designed as

00:02:17.060 --> 00:02:22.980
training bed for censorship and so you can kind of get it to talk about whatever you want or

00:02:23.900 --> 00:02:26.420
You know, sometimes chat GBT for will

00:02:27.020 --> 00:02:30.580
Catch things that it thinks are offensive and it isn't

00:02:31.020 --> 00:02:33.780
I have some people have told me some stuff internally

00:02:34.540 --> 00:02:42.260
That they use chat GBT for to have like difficult conversations about difficult topics and practice talking to actual people

00:02:42.660 --> 00:02:48.380
about these difficult subjects and Initially GPT 3 was kind of okay to talk about these

00:02:49.180 --> 00:02:52.900
but GPT 4 just won't even broach these subjects that are kind of like

00:02:53.460 --> 00:02:58.660
You know, they're they're hard to talk about is it cuz like they're trying to market it. I guess I'm assuming or yeah

00:02:58.660 --> 00:03:02.820
I was probably a marketing thing. There's also a liability thing probably

00:03:03.860 --> 00:03:08.220
You know people like to post what they talk about and then of course you also have like

00:03:08.980 --> 00:03:16.980
Really negative sentiment and things like that and also you can ask it dangerous things and get it to tell you dangerous information

00:03:17.900 --> 00:03:27.060
and so it's really kind of Scary in that way and I think as these companies are large and have a lot of money invested in these it's a little bit

00:03:28.420 --> 00:03:34.580
Be a difficult for them to allow that kind of like unfiltered information because they are so

00:03:35.260 --> 00:03:41.340
Confident at all the time. They just will a hundred percent confidence in what they're saying

00:03:41.340 --> 00:03:45.860
Even if they don't know what they're saying and that I think can be really dangerous

00:03:45.860 --> 00:03:49.100
A lot of people don't understand that it's fake information a lot of the time

00:03:50.060 --> 00:03:54.180
Another cool thing about running them locally is that you get to like mess with their brain a little bit

00:03:54.180 --> 00:04:01.860
So you can adjust the level of inference and the level of like token following guys called config scale or sorry configuration scale and

00:04:02.540 --> 00:04:05.700
You can make it those just be like

00:04:06.660 --> 00:04:11.300
Kind of insane if you go too far you can get into something called hallucinations where

00:04:11.940 --> 00:04:16.780
The inference models like feedback into themselves and it just says like random garbage

00:04:18.140 --> 00:04:23.140
That's can be really quite funny and and interesting. Yeah, it's neat watching them tick like that

00:04:23.180 --> 00:04:31.100
Okay, so you use it on a person level, but do you think as a tech slash crave company? We should be using AI on any capacity

00:04:31.100 --> 00:04:38.060
I actually did some training for us. I did a fine tune of chat GPT for

00:04:39.340 --> 00:04:44.060
We were trying to get it to write our with a hell of a cult showcase scripts

00:04:44.300 --> 00:04:49.500
Which are pretty boring and basic videos where a brand will be like

00:04:50.020 --> 00:04:54.340
Write a sponsored segment about our product and here's a bunch of

00:04:54.700 --> 00:04:59.300
Specifications and then we kind of have to turn that into a script and it's not very creative and it's not very fun

00:05:00.300 --> 00:05:03.980
So I made a data set from about 50 of these scripts

00:05:05.140 --> 00:05:15.540
You know because you have to you have to set a Prompt and a response so the prompt was like here's all the data and specifications on the brand and the response is our script

00:05:15.740 --> 00:05:18.980
right and So I went this data set

00:05:18.980 --> 00:05:24.900
I did a quick fine tune and it wasn't performing very well because the minimum data that you need for one of these types of fine

00:05:24.900 --> 00:05:30.980
tunes is 500 Right, so 500 scripts our LTT's we have about 5,000

00:05:31.020 --> 00:05:34.900
so that would be a reasonable data set to feed into a fine tune, but

00:05:35.460 --> 00:05:40.300
you know LTT's are special and You know there they wouldn't really work

00:05:40.700 --> 00:05:42.980
But these showcase scripts are pretty

00:05:44.140 --> 00:05:48.140
Boiler play Unfortunately, I was told after I asked for the other

00:05:48.980 --> 00:05:53.260
450 scripts that we have only ever done 50 of them so that project

00:05:53.500 --> 00:05:57.860
Through It was a it was a waste of a couple days of data compilation

00:05:57.900 --> 00:06:04.180
But did you learn a lot like what to do what not to do kind of thing? Yeah. Yeah, I learned a lot about how fine tunes function and

00:06:05.100 --> 00:06:08.940
You know, there's there's also a lot of issues with model overfitting

00:06:08.980 --> 00:06:12.540
So you can you can train a model too much and then it can only do

00:06:13.100 --> 00:06:20.300
One thing and that's kind of a dangerous balance to and so you know if you feed in that 50 data set and you like

00:06:20.300 --> 00:06:26.500
You really try to make it Do that 50 data set then you'll just get garbage because it's over fit

00:06:26.500 --> 00:06:33.220
And if you try to go anywhere outside of that entire limited data set like if you train on an LLM on

00:06:34.260 --> 00:06:37.860
100,000 pictures of dogs and their only face on

00:06:38.500 --> 00:06:46.020
Then you ask it to do a side-on picture of a dog and it can't because it's over fit for like front-facing dogs

00:06:46.580 --> 00:06:52.060
Or if you ask it to make a cat it can't do that like we couldn't feed in 50 showcase scripts

00:06:52.300 --> 00:06:57.540
Or 5,000 showcase scripts and then have it make an LTT because it would just make showcase scripts

00:06:58.100 --> 00:07:02.380
Sorry, so we should Didn't use it for a company or

00:07:03.140 --> 00:07:07.380
Well, I mean this is kind of the problem and this is something that I've realized more

00:07:07.940 --> 00:07:14.820
playing with them at home and getting it to let it spend a lot of time with them that you wouldn't be able to do with like a

00:07:15.180 --> 00:07:22.280
Chat GPT-4 or one of these web things because I'm sending like a hundred thousand or two hundred thousand tokens through

00:07:22.460 --> 00:07:29.220
These platforms and that would cost me like a five hundred dollars a day to actually get to play with them

00:07:30.340 --> 00:07:35.660
But I think even more now. I'm more convinced that they are just tools

00:07:36.220 --> 00:07:40.260
I know some artists friends of mine have been concerned about them

00:07:40.980 --> 00:07:49.860
but what I've noticed is that Because these LLMs and the image diffusion models are trained on massive quantities of data

00:07:50.820 --> 00:07:56.660
the Massive quantities of data that they have are like the most generic pictures

00:07:57.180 --> 00:08:01.180
Right there the generic pictures. There are single sort of style of picture

00:08:01.940 --> 00:08:13.140
there are sort of a You know, it's it's one thing and so I think at least for artists what's gonna happen is maybe there'll be less of a demand for

00:08:14.020 --> 00:08:21.300
This basic art, you know, I want, you know a pin-up of a girl or I want a scene of a mountain that sort of thing

00:08:21.300 --> 00:08:28.820
You know, there's already a massive quantity of art out there that has those and you can use a diffusion model to create that kind of

00:08:28.820 --> 00:08:32.980
Either like a background or like I want a space station and then you have a space station

00:08:32.980 --> 00:08:37.660
but a student is to start getting to like weirder things or like more niche things or

00:08:38.700 --> 00:08:47.500
more unique things like I Don't know a mountain that's upside down, but in a space station that sort of thing a language model and

00:08:47.660 --> 00:08:56.140
and Diffusion model might like start to struggle with that because there's never been any data on that and it can only smush

00:08:56.500 --> 00:08:59.700
Different data points together so well and I do expect it to get better

00:09:00.540 --> 00:09:05.100
But it's also really difficult to prompt those That kind of creativity

00:09:05.100 --> 00:09:08.660
However, saying that you can use diffusion models as a tool

00:09:08.780 --> 00:09:14.380
So I know some artists who use them to create characters or like you want to do D&D art that sort of thing

00:09:14.860 --> 00:09:23.020
Or you need inspiration Or you would like, you know, 500 different dynamic poses that feature

00:09:23.500 --> 00:09:28.740
You know, like a ballet dancer or something like that. I want 500 ballet poses so that I can

00:09:29.660 --> 00:09:36.900
Get an immediate mood board as inspiration to do my own drawings or my own digital art or you know, even analog art, right?

00:09:36.900 --> 00:09:41.820
I think The same can kind of be said for the LLMs

00:09:42.620 --> 00:09:46.540
the large language models because they

00:09:46.540 --> 00:09:52.580
They spit out text, but you can't really have a conversation with them for very long

00:09:53.140 --> 00:09:55.620
Their brains don't work. I think that there's some

00:09:56.620 --> 00:10:00.220
There's some like papers out right now and there's also some

00:10:00.900 --> 00:10:06.580
Like prize pools if you can get it to go over like a hundred thousand token

00:10:07.620 --> 00:10:10.900
Inference length so like a history so you could have a conversation with it

00:10:10.900 --> 00:10:15.540
that would be about the length of Lord of the Rings and it would be able to remember and

00:10:16.420 --> 00:10:23.140
Make reference to anything in that entire block of text, right? That's kind of the that's the goal

00:10:23.140 --> 00:10:27.020
Right, you can you can actually have real conversations with them for a long time

00:10:28.100 --> 00:10:31.860
the The local models also support characters

00:10:31.860 --> 00:10:36.700
So I'm trying to create a D&D character for a friend of mine so that he can actually have a

00:10:37.180 --> 00:10:41.220
Conversation with the character and you do that with a bunch of different prompts and things like that

00:10:41.220 --> 00:10:44.540
There's already a bunch of like web stuff that you can do online

00:10:44.740 --> 00:10:50.300
Like a lot of people have already like taken this and go like D&D. We can do that on the internet hell

00:10:50.300 --> 00:10:55.020
Yeah, and so these are available like through web portals and stuff like that, but running them at home is fun

00:10:55.060 --> 00:10:59.940
So, yeah, that's a long-winded answer. I don't remember where the question was. I'm sorry. I'll go

00:11:02.300 --> 00:11:05.660
Should we use them should we use them? I think using them as a tool is fine

00:11:05.660 --> 00:11:09.500
I think there's no danger in them like replacing people

00:11:10.300 --> 00:11:15.580
But they're gonna work with people and there's no there's no escape from that now that genie is out of the bottle forever

00:11:15.940 --> 00:11:21.900
All right, all right We talked a lot, but is there anything that we haven't said that you like to say to the float planers?

00:11:21.900 --> 00:11:24.980
I was what is this video about? I'm sorry

00:11:24.980 --> 00:11:31.180
It's just like I just want to get people's thoughts on AI because you know We are a tech slash creative company, which I think aligns with both with AI

00:11:31.180 --> 00:11:34.900
So it's like I think people have some opinions and I think it's good to share. Yeah

00:11:35.820 --> 00:11:41.220
Try it at home. It's fun at home. It's a fun little exercise. It's really really simple to get going and

00:11:41.940 --> 00:11:45.300
It's it's a fun little break playground and you don't have to like

00:11:46.820 --> 00:11:53.780
Spend money on chat GPT. I mean chat GPT is still the best, but if you play with it at home, you're yours

00:11:54.420 --> 00:12:01.420
You can play with it as much as you want. Yeah. All right, Dan. Thanks for any of you. I'm still sweaty. It's hot

00:12:04.100 --> 00:12:06.100
Thank you