Not many new fundamentals this time. Lots of product news however. It's always good to see interpretability making progress.

"Why are you humans so HORNY?"

Biggest thing for me is that the new 3.5 apparently plays competent chess - at a high amateur level - iff you prompt it just right. I would not have expected that, considering how Anarchy Chess ChatGPT's normal play is. Once again demonstrates that you can never prove the absence of a skill.

Once again not much new. On the regulation as well as capability front, things keep grinding along.

Last week there was a claim that Pi AI cannot be jailbroken. This week, a Twitter user has it giving steps to manufacture heroin and C4. So it goes.

Most interesting progress for me is the paper that notes that "grokking" is caused by the network picking up two separate circuits: one for memorization and one for generalization. But there's no inherent preference for generalization, it's just a blessing of scale: retrain the grokked network on a too-small dataset and it forgets its generalization.

All 50.

At the same time.

In lockstep.

Of course. lol

> We are, as Tyler Cowen has noted, in a bit of a lull.

Hi! As this place seems to be pretty idle, I'm gonna start posting Zvi's weekly roundup posts here. I'm a doomer and y'all are probably accelerationists, so this should hopefully generate juicy discussion. (Not this week though, things are pretty slow.)

No, according to Beryl Howell, Chief Judge of the United States District Court for the District of Columbia: https://www.theverge.com/2023/8/19/23838458/ai-generated-art-no-copyright-district-court

Can't say I entirely agree. I've seen plenty of low effort AI, but I've also seen AI that clearly had work put into it, probably a custom LoRA or checkpoint, extensive inpainting and touchups, etc, where they were using AI as a tool to create something new.

What are your thoughts? Does art generated by AI deserve copyright protection?

TL;DR: The existing incentive structure makes it impossible to prevent the creation of large machine learning models like Yudkowsky and others want.

Also, keep in mind that the paperclip maximizer scenario is completely hypothetical.

TL;DR: AI will be able to generate more data after it has been trained that is comparable to the quality of the training data, thereby rendering any training data absolutely worthless. The time to sell data at a reasonable price is now, and those locking their data behind huge financial barriers (such as Twitter and Reddit) are stupidly HODLing a rapidly deprecating asset.

TL;DW: Blackmail is dead because generative AI will eventually be so realistic that anybody will be able to use it for plausible deniability. Anyone currently blackmailed will be freed once this happens.

This explains why all the AI companies are desperately trying to censor/align their models.

https://www.reddit.com/r/StableDiffusion/comments/153ai4i/stablity_ai_ceo_on_sdxl_censorship/

Apparently there's a chance they're going to release SDXL (their newest Stable Diffusion model) with gimped NSFW. The main reason from what I read is because SD v1 is able to generate child porn. Which is why they also censored SD v2 and make NSFW training much harder. And to no one's surprise, no one uses it. Just take a look at civitai.com and see how many porn models there are. All the SD models I've seen are based on SD v1.5.

One one hand, I understand their position of not wanting to be labeled as CP generator. It's really bad for business and they could lose all their investor funds. But on the other hand, I'm quite annoyed they keep releasing gimped models just because it could make some illegal stuff. It's not like I use SD to mainly generate porn. I just don't want to resort to older model when I do. But it seems like it's possible to be trained for NSFW at least, from the leaked SDXL 0.9. Though there are some rumors they delayed SDXL 1.0 to further censor it. But anyway, all of this is just a speculation. I hope they don't gimp it and just keep silent of NSFW.

With this I ended this half discussion, half rant post. Peace.

Saw someone resurrected the /vsg/ thread so figured I'll dump some of the information here as well in case anyone is interested.

TTS:

• Tortoise TTS: https://github.com/neonbjb/tortoise-tts
• Voice cloning TTS for Bark https://github.com/C0untFloyd/bark-gui

Voice Changer (based on RVC):

• Voice2Voice changer webui: https://github.com/Mangio621/Mangio-RVC-Fork
• Based on RVC webui https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI

Resources:

• RVCv2 AI cover guide: https://docs.google.com/document/d/13_l1bd1Osgz7qlAZn-zhklCbHpVRk6bYOuAuB78qmsE/edit?usp=sharing

• Real time voice changing using RVC models trained above: (needs more documentation) https://github.com/w-okada/voice-changer

• How to create new RVCv2 model on Google colab plus dataset creation: https://docs.google.com/document/d/13ebnzmeEBc6uzYCMt-QVFQk-whVrK4zw8k7_Lw3Bv_A/edit?usp=sharing

• Installing RVCv2 webui fork on local: https://docs.google.com/document/d/1KKKE7hoyGXMw-Lg0JWx16R8xz3OfxADjwEYJTqzDO1k/edit?usp=sharing

• Complete list of publicly shared RVC voice models on huggingface https://huggingface.co/juuxn/RVCModels/tree/main

I know it's not directly AI discussions related, but since it's being censored ("medical disinformation"), it deserves to be put somewhere.

Edit: new version using revision 3 of my dataset/model:

https://deadfrog.org/alexjones-endofdays-v3.mp3

The even more fucked up part? The original song (Vinnie Paz - End Of Days) is on YouTube (2.2 million views). I pointed this out and my appeal was rejected anyways. Pretty much instantly. Apparently an AI Alex Jones rapping this song crossed a line.

Yet another reason not to trust the centralized platforms.

Edit: an assortment of AI songs that might be appreciated here.. I have a ton. Most don't make it online, but I enjoy making them. It's an addiction -_-

https://deadfrog.org/hatkid-iwantyouback.mp3 https://deadfrog.org/noa-merica.mp3 (noa himesaka) https://deadfrog.org/herbert-givenup.mp3 https://deadfrog.org/benshapirolovesong.mp3 https://deadfrog.org/butters-hurt.mp3 https://deadfrog.org/queen-igothiv.mp3

Alex Jones, Hat kid and Ben Shapiro are my own custom trained voice models, the others are not, I just did inference and production on them. All rights reserved by the original voice model authors, other CYA shit, etc..

Still waiting on my GPU, so I was experimenting on https://stablediffusionweb.com/#demo. These aren't all generated with the same prompt, but this one seemed to work around 25% of the time. It's funny how often it ignored "inside", but it actually worked better than "beside", "hugging", or "next to", so I kept it. A bad choice was "enveloped in", which often gave NSFL.

Prompt: girl inside a giant hamburger

Negative: distorted face, open mouth, hands, severed limbs, monochrome, legs

! ! ! ! ! ! !

Stable diffusion

Also https://rentry.co/ayumi_erp_rating

I'm trying to get back into local hosted ERP LLM after taking a break from AI stuff. Looks like we have a few more contenders other than Wizard-Vicuna (RIP Pygmalion) now. Gonna give the 13B ones a try this weekend 🥵 🥵 🥵

Enough of that "uwu as an AI chat bot I cannot answer that 😳" bullshit. We need AI that can talk dirty as a wolf girl, answer your programming questions and have the gut to say kys all at the same time.

Model in the screenshot is Wizard-Vicuna-13b-uncensored-GPTQ

With Llama kicking things off, development has been ridiculously fast in the self hosted text model space. The requirements are getting better, but still fairly steep. You can either have painfully slow CPU generation, or if you have 24gb+ of VRAM, you really open up the GPU options.

7b models can sorta run in 12gb, but they're not great. You really want at least 13b, which needs 24gb VRAM.. Or run it on the CPU. Some of them are getting close to ChatGPT quality, definitely not a subset to sleep on, and I feel as though the fediverse would appreciate the idea of self hosting their own chat bots. Some of these models have ridiculous context memory, so they actually remember what you're talking about ridiculously well.

A good starting point is this rentry: https://rentry.org/local_LLM_guide

I'm admittedly not great with these yet (and my GPU is only 12gb), but I'm fascinated and hope there can be some good discussions around these, as the tech is really fascinating

Voice replacement is getting faster to train, but seems to actually be getting worse with identifying pitch/keys.

There's still an issue with reverb/echo and doubled vocals. The only way I was able to make this passable was to find pre-separared vocals, and even still it struggled with the pitch drifting, so I had to rerecord parts of it.

Still, I trained these in so-vits-svc for about 2 hours each on a 3080ti. I spent more time producing it than the AI needed to completely replace someones voice with someone else's voice.

Combining these with deepfakes/wav2lip can give some damn good results. If anyone wants some guidance on the process for voice replacement, I can certainly share anything I've picked up along the way.

Whatever your opinion, you've certainly got one. This is the place to discuss and share things relating to AI. Do you make cool stuff with stable diffusion? Want to discuss the latest local text models? Got funny or interesting voice models, or just want to discuss the impacts of the technology? Rumors on new Bard (lol) improvements? Anything AI.

NSFW is allowed, just please mark it accordingly.

If you post your work, please keep it to a single thread/post to not clog up the plumbing too much.

If you have suggestions as to things we can do to improve this place, this thread is unlocked and ready for your input.