lm-human-preferences VS v-diffusion-pytorch

Lest you think that that sounds like a too galaxy-brained possibility, it has already happened at OpenAI (scroll down to "Bugs can optimize for bad behavior"), just with a model that was very far from being capable enough to be dangerous.

Found relevant code at https://github.com/openai/lm-human-preferences + all code implementations here

That's why, when they're serious, they use RL for finetuning from human preferences (would be hilarious if this attempt to solve the terrible bias you take to be evidence of AGI threat ends up creating a Woke Singleton itself, btw); it's a powerful general approach, and I see no sign of it being applied here.

The kind of measures they are taking, like simply deleting wholesale anything problematic, don't really have a '-1'.

But amusingly, exactly that did happen in one of their GPT experiments! https://openai.com/blog/fine-tuning-gpt-2/

Code for https://arxiv.org/abs/1909.08593 found: https://github.com/openai/lm-human-preferences

This is dumb.

We employed Eleuther team members as Stability AI employees/contractors and incubated them until the 501(c)3 was set up and we managed to bring in other funders too: https://techcrunch.com/2023/03/02/stability-ai-hugging-face-...

I am on the board and delighted to continue to support their work as an independent organisation for LM evaluation, alignment and interpretability which is much needed.

Indeed though our approach was handing out significant compute for no control, no equity, no IP.

Anyone who has received Stability AI grants will be able to attest to this with multiple breakthroughs as a result, for example funding https://github.com/BlinkDL/RWKV-LM, the work of https://github.com/lucidrains and others.

Similarly we funded the beta of MidJourney with a cash grant for compute without ever even floating asking for equity etc as it is a market-creating innovation.

At the time MidJourney was using cc12m_1, a model developed by one of our lead (employed) generative AI developers Katherine Crownson / RiversHaveWings (https://github.com/crowsonkb/v-diffusion-pytorch)

Our model is simply to take open innovation and create commercial variants of that (our stable series models) from scratch and on our own, plus variants of that for private data - https://twitter.com/EMostaque/status/1649152422634221593?s=2...

This means we can be hands off versus other funders and trust researchers and help them succeed, something others do not.

From what I've gathered by being involved early in the beta / in other discords, Midjourney was originally based on a fine-tuned version of classifier-free guided v-diffusion. The fine-tuning dataset was a manually curated set largely from LAION-2B similar to the laion-art / laion-hd. To make it so fast they were using Progressive Distillation (possibly distilling on PLMS steps rather than p/ddim?) and settings optimized to let them skip a few of the first steps like Quick CLIP-Guided Diffusion. There's a good chance they were doing some prompt augmentation as well, although I think this would be susceptible to prompt discovery attacks which I haven't seen any examples of for Midjourney.

Source. This tweet is from a person whose organization has been publicly credited with providing compute for others in the past (example: "Thank you to stability.ai for compute to train these models!").

com/RiversHaveWings/status/1462859669454536711, 2021.

[8] Katherine Crowson. v-diffusion. https://github.com/crowsonkb/v-diffusion-pytorch, 2021.

More parameters are available as seen in this code. The fix was adapted from this.

I'm not sure if this Colab notebook was mentioned in this sub previously, but it's been available since January 2022. The cc12m_1_cfg model used by this Colab notebook is different than the cc12m_1 model from this December 2021 post (reference).

A few more here https://twitter.com/metasemantic/status/1486334535436488705. Samples careful constructed with a heavily modified diffusion model by @rivershavewings https://github.com/crowsonkb/v-diffusion-pytorch

Most of the code was from Katherine Crowson's (@RiversHaveWings) v-diffusion-pytorch library (https://github.com/crowsonkb/v-diffusion-pytorch), which is an implementation of denoising diffusion probabilistic models (https://arxiv.org/abs/2006.11239). I used the CC12M_1 CFG checkpoint.