|OT| The AI Thread - Generated by ChatGPT

[Le Pertti]

Greetings, fellow forum members! Today, I'd like to dive into the fascinating world of Artificial Intelligence (AI), a technology that has garnered immense attention and sparked heated debates across various domains. AI has the potential to revolutionize our lives, but it also raises profound concerns. Let's embark on an open and engaging discussion to explore both the advantages and drawbacks of AI, shedding light on its impact on society, ethics, and our future.

Advocates of AI often herald its immense capabilities and transformative power. From enhancing automation and productivity to improving medical diagnostics and aiding scientific research, AI offers a myriad of promising opportunities. The ability of AI algorithms to process vast amounts of data at unparalleled speeds and extract valuable insights can lead to breakthroughs in fields like healthcare, climate change research, and even space exploration. Moreover, AI-driven technologies have the potential to make our daily lives more convenient, from voice assistants and smart homes to personalized recommendations and autonomous vehicles.

On the other hand, skeptics express concerns about the unintended consequences that may arise from the rapid advancement of AI. They worry about potential job displacement due to automation, as AI systems become more proficient at tasks traditionally performed by humans. Questions surrounding privacy and data security are also prevalent, as AI relies heavily on personal information to deliver tailored experiences. Furthermore, ethical considerations arise when we consider the potential biases embedded within AI algorithms, which can perpetuate discrimination and exacerbate societal inequalities.

By fostering a balanced dialogue, we can delve into the intricacies of AI and its implications. Here are a few thought-provoking questions to kickstart the discussion:

1. How has AI already influenced your life, and what positive or negative experiences have you had with AI-driven technologies?
2. In what areas do you believe AI will have the most significant positive impact, and why? Conversely, which areas do you think will be most vulnerable to negative consequences?
3. What are your thoughts on the potential job displacement caused by automation? How can society adapt to ensure a smooth transition for the workforce?
4. How can we address the ethical concerns surrounding AI, such as algorithmic bias and privacy issues? Are there specific regulations or guidelines that should be implemented?

Remember, this discussion aims to explore diverse perspectives and foster understanding. Let's engage in respectful conversations, appreciating the multifaceted nature of AI and its potential impact on our society. I look forward to hearing your insights, experiences, and opinions on this intriguing subject!

Happy discussing!

Human writing: Yes, I generated the opening op with chatGPT, because what we want is just to get the discussion started.

 

[Le Pertti]

So what kind of AI projects do you all have? How are you using it?

For me I've been mostly using ChatGPT to learn python coding, even made a minesweeper clone! GitHub - LePertti/BugSweeperBoy-1.0: A minesweeper clone made in python/pygame with the help by ChatGPT.

Also I'm starting a school for AI development so all these discussion about AI is super interesting, since in the future it will be my profession. to go into companies and implement AI solutions to match their need.

 

[EdwardTivrusky]

I have been using Local Generation as i have an RTX 4070Ti which can generate things in a reasonable timeframe so i haven't really touched any cloud-based services at all.

I've been using the web-ui based Automatic1111 & SD-Next/Vladmatic projects to mess about with image generation based on Stable Diffusion. Mostly just experimenting, learning prompting oddities and tips, techniques like Tiled Generation & Tiled VAE for upscaling, Inpainting, ImgToImg etc. Meta have also released some crazy tools that are now available as web-ui extentions. It's a fluid and exciting area at the moment but it can be a bit exhausting when you just want things to settle down so you can learn something. ha

I've ended up just generating things based on a whim and if i find anything interesting i will upscale it, i have also been using SD to generate wallpapers for my PC which is fun. Nothing amazing and i haven't looked at updates for a few weeks so i will probably do that this weekend.

I recently started messing about with AI Voicepacks, text-to-speech etc and that has been fun too. Some of the voices are extremely good but a large majority of them sound like they have far too much auto-tune applied but as time goes on the models have been improving in leaps and bounds. I'm not interested in talking and speech voices, i've been messing about with adding voices to karaoke instrumentals.
I've been thinking about using text-to-speech and then to voice for my own writings or messing about singing things myself and then replacing it with an AI voicepack.
Some of the tools like UVR and RVC V2 are crazy powerful with the models and filters improving at a crazy rate. Again i haven't looked for a few weeks so i expect the recent advances to be a big improvement over what i was playing with a while ago.

AI-LLM is not something i have dabbled with much at all. I think i got Silly Tavern working and i have Oobagooba at a point where i can start to add models and play about but as i have only 12GB VRAM i have to use the 7/13b 4 bit models. I will play about with it, probably for RPG scenarios and characters as i wait for the Coding LLMs to start to support Powershell more which is the main scripting language i use most days. Python and Java focused AI are not really useful to me tbh.

So yeah, i dabble with Local AI Generation a lot but nothing that would be considered anything more than playing about and i'm only really looking to make things that i want to create but can't due to my lack of artistic talents. If i wanted anything amazing i would still commission a skeb or similar. Also, i've been coding for years so if i was really motivated i'd write the script myself.

 

[low-G]

Neat, I was considering making this thread here myself, but wasn't sure there was a lot of attention.

So at this point I'm officially an AI developer, professionally. About a year and a half ago I said to myself 'I should check out what I can do on my desktop with AI these days' and discovered KoboldAI, and Dall-E Mini, and by the time Stable Diffusion was released I was fully integrated into that community, had tuned a LLM, etc etc. Then a few months back the company I work at started an AI division and I joined that team. (Still do non-AI work too).

My stance is that we have extremely little to fear and some to gain in terms of jobs. These AIs have almost zero effective reasoning capabilities. Much of the perceived reasoning is simply due to things that were already reasoned out in the training data. It'll be a long time before an AI can just do a job, everything else is marketing hype. The biggest risk remains deepfakes, spreading fake news and blackmailing people. I guess there's also a risk to corporations who overestimate AI, but that's not our concern. Any company that fires staff and replaces them with AI will suffer extremely in the medium term. AI can be used to augment work only (and even then, only some of the time and in some cases). Secondary risks are to biasing, and generally companies and gov'ts who do not understand AI at all yet implement them in places and ways they cannot correctly implement them in.


1. I just like playing around with AI. Predominantly it fulfills the same drive to explore that video games used to dominate. I feel like I can explore to much greater fulfillment and much greater satisfaction with AI. It's a bit of creativity with a bit of exploring fiction. In terms of work, we have a few specific uses to lessen the amount of work various other teams have to do. The end result is just that people can get more work done, quicker.

2. I think medium-term the biggest thing is just making many aspects of life with repetitive dull tasks smoother and easier. Being able to write in natural language some menial computer-task and just have it work is huge.

3. See above, but AI is far too dumb and primitive to actually replace anyone yet. Companies will fire people and then fall behind rapidly. So it's a great time to shake the roots of capitalism.

4. The issue here is a fundamental capitalism one. The issue is not AI. The answer is tearing down corporations, disallowing corporations to have any political influence. Legal consequences for executives and shareholders. Definitely getting rid of Citizens United.

The bias exists because corporations behave as they do, and the biases will never go away until capitalism is diminished or thrown in the garbage. Corporations enjoy the bias. They choose it. Take away AI and nothing changes. Add AI, nothing changes. So the issue there is disjoint.

However, AI has the potential to shake things up, to redistribute power to an extent. And really if things went really ideally, workers and people could get a big boost out of AI. The key thing is not allowing corporations to control it at all. They cannot be allowed to have a voice on it. When you see OpenAI saying ANYTHING to Congress or other gov'ts about it, that is them seeking control and nothing more.

 

[low-G]

I'll add a few tidbits because I was mostly on a rant there.

So lately I've taken to using the painfully slow llama-65B based models. Which means effectively I have to wait maybe a few minutes for a long response from the AI. The 65B llama models definitely outperform ChatGPT on most tasks (but not all), and I mostly use it for creative outputs. The difference between 65B and 33B in terms of characterization, coherence, and writing quality is quite large. With 65B I get some fairly creative and very rich responses, although nowhere near perfect of course! But I don't immediately run into major issues. 33B can generate many tokens a second on my 3090, which means I wait seconds for a reply, but the quality is just not there now that I've gotten used to 65B. GPT-4, which I don't use much, definitely has a firm edge. But even on GPT-4 I can definitely wish for more.

I'm not satisfied with text to speech. I wish that was much better. I feel that it is much less satisfying than other AIs.

Speech to text and translation works beautifully. I've used Whisper (free, local) to translate a bunch of shows. The translations aren't perfect, again, but definitely good enough in many cases. That makes me VERY happy. It's amazing to have a reasonable translation of something in only somewhat slower than realtime.

And text to image, I spent a lot of time on Stable Diffusion, but I'm somewhat tired of its limitations. Not saying I'm done with it of course. DeepFloyd IF has so much more coherence capacity... I don't pay for Midjourney, obviously MJ is quite good. Although IF is far more advanced and coherent, just not as beautiful. SDXL will probably entertain me for a time, being a reasonable middleground between all these AIs in many respects. Already it can do a few things 'best', so that gives me hope.

My avatar was created with DeepFloyd IF & SD upscaling...

 

[Le Pertti]

Thanks for the long and detailed responses EdwardTivrusky low-G ! I have to digest a bit and will have lots of questions. I’m so new still to all of it.

 

[Le Pertti]

One interesting thing I thought about, in regards to the actor strikes and such, how they are so against AI voices, which makes sense and I fully support them against the big movie studios. But when it comes to games, AI generated dialogue is such an obvious thing to happen and the only way to voice that is by also having AI generated voice. So not sure fully how to watch out for the talent yet also make use of what AI enables.

 

[lashman]

So not sure fully how to watch out for the talent yet also make use of what AI enables.

get rid of capitalism

 

[EdwardTivrusky]

Write it into the contracts? Let people and lawyers and agents argue about what is acceptable use but pay them for using their voice to train the AI if they consent. It's something people will have to address, the genie is out of the bottle now and unless we hobble technology and ban or restrict people whether industry or joe public from using the tools then some compromise will have to be made.

 

[low-G]

One interesting thing I thought about, in regards to the actor strikes and such, how they are so against AI voices, which makes sense and I fully support them against the big movie studios. But when it comes to games, AI generated dialogue is such an obvious thing to happen and the only way to voice that is by also having AI generated voice. So not sure fully how to watch out for the talent yet also make use of what AI enables.

Deepfakes of all kind remain the biggest threat of AI IMO.

But if you don't care for a voice to sound like a particular person, some day (not today) be easy to have a virtual voice that has good acting (the part that's not possible today) and sounds like a random person (which is possible today)

(imagine if I could just drop a wav file in here to prove that to you -- I keep my best generated lines and the best ones are great -- but text to voice still needs more time in the oven for sure)

 

[Durante]

A friend and I made this: Checkpoint Arena

 

[lashman]

A friend and I made this: Checkpoint Arena

nice, good job!

 

[Le Pertti]

A friend and I made this: Checkpoint Arena

I’m kind of lost on the purpose of the site? Not saying it look useless or anything lol

 

[EdwardTivrusky]

It's very handy to check the output of different checkpoints with a consistent set of prompts.

---

Mission
Checkpoint Arena is a free and open source hobbyist project, with the goal of providing a quick overview of the results generated by many popular Stable Diffusion checkpoints with a set of uniform, standardized and simple prompts.

 

[Durante]

Right. Basically the idea is that e.g. Civitai has a ton of SD checkpoints available for download, but the example images for each are picked by the checkpoint author.
That's nice to see the best a checkpoint can do, but since those are usually absolutely massive prompts it doesn't really tell you much about the output of a checkpoint with a simple prompt (and it could also be the case that one author spent hours finding a great seed while another just chose one at random). The only way to find that out is to download ~2-4GB (usually) and then generate a lot of images with it (minutes to hours depending on your GPU).
With checkpoint arena, you can get an idea of the output of a given checkpoint (and compare it to others) without doing that.

It's also quite fun to just see what kind of fun stuff different checkpoints can come up with for just very basic prompts, but that's probably not as entertinaing for everyone else as it is for me

 

[Le Pertti]

This is incredible! How long until GameDev Story generates actual games!

 

[Le Pertti]

What kind AI workflows do you all have? Other than the sites I havent really dabbled that much in generative AI but would. love to get into more custom and personal stuff.

Like how do you go about the stuff that you do?

Anyone tried anything else than image stuff?

I use chatGPT for coding quite a bit but it is often frustrating and thinking I should give the Amazon code whisperer a try, but wouldn't it be possible to have own custom one instead? Surely there are lots of open source projects around? Mostly feel the pressure to learn the open source ones before they became illegal thanks to lobbying from the big AI companies

 

[lashman]

i only really ask chatgpt for simple programming stuff for python or gdscript, lol it can be pretty good for that

 

[low-G]

I'm 100% on AI development now professionally. But I don't use AI that much in my development still. I know for a fact that I use it below average because the majority of the people who work where I work use AI for development more than me (and most are not AI developers).

Yes, I'll load up ChatGPT or GPT-4 and ask it specific questions of code or specific questions of development, maybe 0.3-3 times a day. I don't use Copilot nor any other integrated AI.

I still play around with generative AI on my free time -- LLMs and image stuff. I'm trying to improve my SDXL / lora training skills lately. I'm probably aiming too high for what the tech can do, and my in-brain standards keep increasing such that I now feel a picture of a character is only good if it has 2 perfect hands in it. So that's maybe 1 in 30 SDXL generations and 1/2 of Dall-E3 generations...

At work we're using various APIs and local models for classification, categorization, matching, vector databases for information lookup, generative text, raw recognizer... We have many different AI projects at the moment, a few in production already.

We are all over the place with frameworks, we're using .NET and Python with Triton/Onyx models, AND pytorch AND torch cobbled together from sticks in .NET AND Tensorflow

 

[EdwardTivrusky]

I recently upgraded to an RTX 4070 Ti Super and have been slowly getting back into dabbling with generativeAI.
Wow things have moved quickly in the 6 months since i last really had a look about.

Now i have 16GB i can run SDXL models with much less tiling and messing about so i've been experimenting with that and updating my installs of Auto1111 and SD-Next GUI. Comfy GUI is still a bit of a headfuck for me to get along with at the moment but i would like to play with it as i've used Houdini in the past which has a similar UI method.

I've pulled down updates for LoLLMS, oobabooga webui, Silly Tavern and Faraday so i'm still getting my head around LLM models but i'll set myself some goals and see how LLM AI can help. Things like helping generate prompts for Stable Diffusion, Scripting stuff and well tbh that's all i can think of at the moment but i want to pipe the output to a TTS system and get updated tools and voice models for dabbling with RVC and vocals.

I would like to link several AI processes together like chat to tts to singing or chat to stable diffusion and also see if i can use an LLM in my daily work or at work for generating presentations, guides or tutorials etc. I need to find a goal or my ADHD will get bored and i'll wander off again.
We don't use AI ourselves as a department but we do work with the people who are teaching and researching various applications of AI & Big Data in various fields so i'm interested but unguided.

I did find this as an example, which looks crazy but i think i might run out of VRAM as i don't have 24GB only 16GB.



There's a momentum behind the move towards intergration of technologies. Things like linking The Sims to LLM Chat or SillyTavern with VRM.



Maybe Text Adventures or Interactive Fiction with gfx will get a ressurgence in a few years?
The hardware required is a high bar but it's getting lower and more accessible at an incredible pace. The number of people who run these systems via GUI apps on their phones is astounding.

I'm sure the big money is in snake-oil chatbots or using AI to simpify processing Big Data for Enterprises but bleh, boring!

 

[Le Pertti]

Not having a powerful GPU enough to all that cool stuff has kind of kept me from even starting learning it. Keep thinking it will be a later project. But I definitely see lots of potential there for using different things to feed into one experience.