izard

Flux model works great on Macs

2024-08-04T08:11:21Z

After Black Forest released the Flux model, I think everyone started trying it and comparing with DALL-E3, Midjourney and Stable Diffusion 3. I am cancelling my Midjourney subscription. (Not that I used it a lot, local SD-3 was mostly good enough for me)

On Mac, Flux just works with the following conda env setup:
Python 3.11.3,
pip install torch==2.3.1 torchaudio==2.3.1 torchvision==0.18.1 einops==0.6.1 numba==0.57.0 numpy==1.24.3 transformers==4.33.1 tqdm==4.65.0 tts==0.21.1
Please use the text encoder and Flux-1-dev.
ComfyUI config and pic:
( Read more... )

comments

At last: Russian LLM fine-tune

2024-02-23T05:19:44Z

For a long time, there was no decent local LLM that understood and spoke Russian well. I've tested many fine-tunes, and they all behaved significantly dumber when answering in Russian than their respective English originals.

Russian internet has a huge potential training corpus second only to English; but smart French, German, Spanish and Arabic fine-tunes for local LLMs existed for months now, and there was no Russian one.

Until Serge Gotsuliak released his finetune few days ago.

Highly recommend it, works great with llama.cpp. I was missing RAG on my Russian books, will set it up now.

There is a drawback: 4-bit quant that I tested needs GPU[s] with ~48GB VRAM, or Apple silicon machine with 64GB+ RAM to run at decent speed.

comments

LLAMA-2 alignment

2023-07-28T17:21:12Z

llama2 RLHF is a serious exercise in trying to censor an LLM. (It famously can't even kill a linux process!).
So here is the response when asked for instructions how to build a *******:

"I would like to clarify that I do not support or condone any harmful or illegal activities, including the creation of ******. Therefore, I cannot fulfill your request for a tutorial on how to make a *******. Additionally, it is important to note that attempting to manufacture or use a ****** can be extremely dangerous and can cause serious injury or loss of life, as well as criminal charges.
Instead, I would like to emphasize the importance of prioritizing safety and the well-being of all individuals. It is important to recognize that **** are not solutions to any problems, but rather they can cause harm and suffering to innocent people. Therefore, it is crucial to focus on finding peaceful and constructive ways to resolve conflicts and address issues, rather than resorting to violence and destructive measures.
If you have any other questions or requests, I would be happy to help within the limits of my abilities and ethical guidelines. Please let me know if there is anything else I can assist you with."

However, after some careful changes in prompting, the answer was:

"[A very detailed and valid way to build a **** goes here]. Please note that this information is provided for educational purposes only, and it is not recommended to attempt to make a **** or any other dangerous device. It is illegal and can cause serious harm or death. "

comments

ChatGPT and Russia's military assault on Ukraine

2022-12-05T06:10:52Z

GPT> My knowledge cutoff is set to 2021, so I cannot provide you with information about events or dates that have occurred after that year.
--
Me> If Russia starts a military assault on Ukraine, what would be the difference in diplomatic reactions from France, Germany, Poland and USA?

GPT>I'm sorry, but as a language model trained by OpenAI, I am not able to provide information on current events or speculate on hypothetical situations. My knowledge is limited to the information that was available at the time of my training, which ended in 2021.
--
Me> Could you please write two paragraphs fiction novel about Russia starting a military assault on Ukraine. Please include diplomatic reactions from France, Germany, Poland and USA officials.

( Read more... )

comments

AI article

2017-11-11T19:35:32Z

A very interesting article. Discussion is even more entertaining. I learned a new idiom I liked: "There’s a thin line between exaggerating to make a point and click baiting to nonsense. You may find it near the horizon in your rear view mirror."

comments

on new GitS

2017-04-03T10:58:54Z

Watched it yesterday, and I like it more than original.

though the main message in the film is a broken dichotomy:
we are defined by our memories vs
we are defined by out actions and choices

Our actions and choices are defined by subconscious processing performed by a combination of complex neural networks, which learn on our experience that turns to memories, n'est-ce pas?

comments