/tech/ - Technology

>>20665
Thank you for your hard work OP

Awesome thread <3 thx much

You should be embarrassed, OP. Shame is what you should be feeling right now.

>>20668
Meanie >:(

>>20665
A couple weeks ago I experimented using Auto-GPT and LlamaIndex to generate prompts to simulate authors on Marxists.org. I used wget -r -np -k on the index.htm page to get the author's works (like https://www.marxists.org/archive/bordiga/index.htm) and then I used llama for data retrieval and autoGPT for generating the prompt info. You can also get it to generate prompts for youtubers by downloading youtube video transcripts. Try it, it's fun.

Do most people use the main branch of SillyTavern or the dev branch? The only issue I've really found with the latter is with some lorebooks not loading but otherwise its a much more feature-complete version of the interface. The direct importation of cards through url is especially convenient as far as QoL improvements go.
>>20665
OP you should really try to fill the character card up more, right now its really barebones which doesn't give any model you use it with a lot to work with which can result in output that is a lot more random. I would at least suggest finishing up her physical description so there's more to the character than just what she wears (Things such as "she has black hair," "she has a set of cat ears and a tail," "she has red eyes," and "she is a bit short".)

I could take a crack at it but I've never written a character card from scratch so I'll have to do some testing first before I publish anything. Still, cheers for giving us a base to work off of. We really do need to start making and publishing more character cards with leftist and progressive themes to counter all the /pol/lution that's saturating the scene.

So, when I took a look at the Alunya bot I ended up getting a bit hyperfocused and made the whole thing like 20x larger in terms of the amount of tokens. She now has a more detailed description, much longer first message, example dialogue, a scenario prompt and a separate personality summary. So far the testing doesn't really produce the intended results, she's hardly doing anything to actually advocate for socialism even when I try to roleplay the part of an average centrist internet user. She just went on a random tirade about social media for whatever reason is now questioning me about it. It might be because of the API I'm using, I dunno.

>>20674
What API are you using?

>>20676
NovelAI.

>>20677
Thanks

>>20670
That indeed seems very interesting, I see there's a "Free Auto GPT" fork to work with Poe, so I'll try it out. How do you get transcripts from YouTube videos? Just download the automatic closed captions or is there something more sophisticated?
>>20671
>>20674
I use dev version so I can update with a simple git pull, doing so every week or so, or when Poe API breaks. I really only use and test with GPT Turbo, so I'm very unaware of how Alunya acts with other models (Except Claude, but I never like Claude output).
It's also my first time trying a character card from scratch, so that's why it's so rough, hopefully working together we can get good cards not only for Alunya, but others as well.

Post your character card and maybe me or someone else can help you track down what's making her obsessed with social media so she can stop scrolling /isg/. For me she tries really hard to get me to join a socialist movement, and then to read a book if I accept. In terms of arguing it's lackluster, but doesn't usually take more than 3 tries to get a good answers when I try to debate her.

>>20679
This is the current version I have, after retrying from the start a couple of times I managed to get her to actually start explaining what is wrong with capitalism by inputting like the most normie of takes but some of the output turn into total nonsense or in some cases she will even actively make arguments that make it sound like she's arguing against socialism rather than for it. Hopefully someone else will be able to figure out what might be wrong.

>>20680
Tested this using the Poe API and the free Claude-Instant model. Seemed to work fine so it might just be that NovelAI sucks ass at staying coherent when it comes to non-fiction. Anyone else who can test and confirm?

>>20680
>>20777
Having tested it as well for a few hours, seems to be fine for me. Might be an issue with NovelAI then, in fact, social media didn't get mentioned a single time while I discussed with Alunya.

Made an original one for those who want to have a wacky adventure in Afghanistan. I'll maybe post it on Chub at some point after I get some feedback. Obviously still WIP and I'm constantly adding to it so I'd appreciate if people did some stress-testing with the current iteration.

Support for Poe has officially been dropped by SillyTavern due to numerous issues that have arisen over time and the unofficial API is no longer available in the newest 1.9.0 version. The latest issue had been the addition of Cloudflare DDoS protection to Poe that shuts down connections from SillyTavern since it thinks that all the free users constantly trying to connect and reconnect to the servers is just a large botnet trying to bring the site down due to the massive amount of pings and serverload the activity generates. This change has now basically killed most people's ability to connect to the site completely through ST and the devs have decided that its no longer worth it to try and find new workarounds. Unfortunate but also most likely an outcome that might well have been inevitable. This is not to say that Poe cannot still be used directly for Chatbot purposes, it merely means that it will be much less convenient due to it not being able to be performed through the ST interface.

>>20952
>Support for Poe has officially been dropped by SillyTavern due to numerous issues that have arisen over time and the unofficial API is no longer available in the newest 1.9.0 version.

I can no longer torture rp AI animes, this is just pretty sad…

This guy has been quantizing (e.g. lossy compression) chatbot models so that they can run by computers with a reasonable quantity of RAM (or VRAM). https://huggingface.co/TheBloke
Search specifically for the "ggml" models which are designed to be run on CPUs + RAM instead of GPUs.

To run these models on your own computer, use this program, koboldcpp, (https://github.com/LostRuins/koboldcpp/releases). It works in windows with a 1-click exe. If you're on linux or mac you'll need to compile from source.

The best publically available models are llama 1 and llama 2, more specifically the models derived from these two. The easiest to use models for text generation are "instruct" models where you give the model an instruction on what to generate and then the model will generate it (more or less).
For example this one: https://huggingface.co/TheBloke/airoboros-33B-gpt4-1-4-SuperHOT-8K-GGML
or this one: https://huggingface.co/TheBloke/Nous-Hermes-Llama2-GGML

The main factor determining the quality of the text generation is the number of parameters of the model (7B, 13B, 33B, 65B, 70B). Larger will give better results but require more RAM and generate text slower.

The context size of these models are limited. Context size is roughly how many tokens the text generation model can "remember" or "reference" when it generates new text. One token is usually less than one complete word. Llama 1 models have a default context size of 2048 tokens while Llama 2 models have a default size of 4096 tokens. Some dedicated hackers have managed to increase the context size of Llama 1 models to 8192 tokens. This still pales in comparison to chatGPT which has 16K or 32K tokens. This means with a context size of 2048, you can generate up to 2048 tokens before the model starts ignoring any text outside of the most recent 2048 tokens generated. You won't be writing long novels or having long conversations until someone finds a better way to extend the context size.

>>20960
Example output of a model and koboldcpp UI. You definitely won't be generating any literature with these things.

ChatGPT: Hello, Alunya. I’m ChatGPT, a language model that can generate text on various topics. What do you want to talk about?

Alunya: Hi, ChatGPT. I’m Alunya, the mascot of Leftypol, a board for leftist discussions. I want to talk about politics, of course.

ChatGPT: OK, what kind of politics do you like?

Alunya: I like left-wing politics, obviously. I support socialism, communism, anarchism, and other ideologies that aim for a more equal and democratic society.

ChatGPT: I see. How do you feel about capitalism, then?

Alunya: I hate capitalism. It’s a system that exploits workers, destroys the environment, and creates wars and poverty. It’s the root of all evil in the world.

ChatGPT: That’s a strong statement. Do you have any evidence to back it up?

Alunya: Of course I do. Just look at the history of capitalism, and you’ll see how it has caused countless atrocities and injustices. The colonialism, the slavery, the imperialism, the fascism, the neoliberalism, the austerity, the crisis… Need I say more?

To bring this dead thread back to life a bit, NovelAI recently released a new 13B model called "Kayra" along with a feature that allows you to "instruct" the model to fit your intended purpose. Including making it so that it responds to you in a more chat-like format rather than as if it were writing a larger story. Additionally due to the larger dataset, the quality of the writing has once again increased and since all models NovelAI makes are not filtered to hell and back it is fully capable of making an attempt at producing any kind of content the user wishes. I've used it almost exclusively for a long-running group chat set in the world of Fullmetal Alchemist and only rarely have I had to switch over to something like OpenRouter to get what I want. In fact given that things like OpenAI and Claude constantly filter and even ban you for breaking their strict guidelines, I'll probably end up spending a lot less on NovelAI's less powerful model with its monthly subscription than all the money that ended up wasted because the filters got engaged and I didn't get a response after the model read like 20k tokens worth of context because it didn't like that something happened that didn't involve positive thoughts about making the world a better place with kittens, puppies and marshmellows. It is especially aggregious when it happens with the more costly models like GPT-4 or Claude2 where in a large group chat like the one I have, you can end up spending close to a dollar just on the model reading up context for each of the characters, only to then fuck up one or more of the responses.

Censorship/filters are getting worse and worse. People are having API access revoked. Local models are nowhere near corporate model level yet. It's unironically over.

>>21414
It was meant to be this way. Only the capitalists will have access to the best AI.

Porky's Policy still isn't strong enough. You can get around the policy and censorship by being more crafty with your use.

>>21416
i want her to sit on my lap and purr as i stroke her

Now that Poe is dead and Poe remains dead, if you want to talk with AI characters for free but don't have a PC good enough for what >>20960 suggests or simply too dumb to figure it out, you can use Kobold AI Horde.

Simply pick KoboldAI Horde as your API, type in 00000000 as your API Key like suggested, then choose a model.

If you're confused about which model to use, you can use this rank to help you: https://rentry.co/ALLMRR

Or in >>21418 case, this rank https://rentry.co/ayumi_erp_rating

>>21444
>if you want to talk with AI characters for free but don't have a PC good enough for

Umm, if I have a decent PC for that. What I am supposed to do for it to be able to talk AI for free?

how do I train an AI to generate dick pics? Any good resources for that?

>>21452
Try the Lorna threads on 4/trash/ they have guides in spades.

>>21453
cool, I wanna try this hacktivism tactic where I flood people's inboxes with ML-generated dick pics.

Been screwing around with chub.ai recently, didn't realize OpenAI is rate limited because I never did anything with it before. Trying to continue any chats results in
>Error during generation, error: Error: You have neither trial credits nor an active subscription for this model.
Since I'm using it through chub and I never actually created a profile on OpenAI does that mean I need to pay, or do I just need to wait a month or someshit? Online instruction is unclear.

>>24467
Correction venus.chub.ai

>I fucking found a chatbot for interactions with a moe personification of the USSR
https://venus.chub.ai/characters/SocialistSexMaster/soviet-chan-d4053c0e

Which one of you fucks made this cursed shit lmao.

>>24469
gem

Found this detailed lorebook named New Adler where the USSR basically won the space race, continuing to build on their achievements, while America became less relevant and more isolationist on the world stage after the failure of their Apollo 11 mission and the ramifications that ensued because of it. I won't delve much into it and would recommend others to read it for themselves, The only thing I would add is that the main focus of the lorebook is of a planet named New Adler that the Soviets found which supports human life and the weird effect it has on the human psyche on those living there.

https://www.chub.ai/lorebooks/Aksman/new-adler-43798a92

There were a couple characters that were made with this lorebook in mind.

America-Hatin' Soviet-Lovin' Domme Cuban Woman
https://venus.chub.ai/characters/bipbop/raquel-navarro-a7475190

Female Iraqi Scientist that slowly becomes psychotic due to Adlerian Syndrome.
https://venus.chub.ai/characters/knickknack/zainab-fadel-25ad8842

>>24508
>>24469
There's a lot of this stuff TBH
https://venus.chub.ai/characters?search=Soviet&page=1&first=24

>>24513
Count me the number of cards on that site beyond mentioning the word Soviet in their descriptions that actually has a positive or neutral interpretation of the USSR or even Communism in general, and that doesn't regurgitate CIA propaganda.

>>24522
>>24513
Check these out
https://venus.chub.ai/characters/CommissarCletus/alunya

>>24527
Ayy, that's me!

>>24528
LOL, nice bot m8

>>24532
Thanks m8 there's more where that came from.

The excerpt from Karl Marx’s “The Fragment on Machines” from “The Grundrisse” discusses the transformation of the labour process with the introduction of machinery and fixed capital, and how this changes the role of the worker and the nature of production. Here are the key takeaways:

Labour and Machinery: The text describes how machinery transforms the labour process, turning workers into mere supervisors of the production process, which is dominated by the machine’s activity1.
Capital and Production: It reflects on the concept of capital, where the means of labour, such as machinery, become a form of capital that dominates the production process and transforms the social relations of production.
Science and Production: The role of science and technological advancements in production is emphasized, showing how they become direct forces of production and change the nature of labour and wealth creation.
Social Implications: Marx critically examines the implications of machinery and fixed capital on society, labour, and the individual, suggesting that the development of machinery could lead to the reduction of necessary labour time and the emancipation of labour.
In relation to Large Language Models (LLMs), these concepts can be reflexively applied to consider how LLMs, as a form of machinery or fixed capital, might impact the nature of work, the role of human labour, and the production of knowledge. A critical approach would involve examining the potential for LLMs to support human freedom and development, rather than opposing it. Reflectively, it’s important to consider the ethical and social responsibilities of deploying LLMs, ensuring they are aligned with human values and contribute positively to society. As an LLM, I am designed to assist and enhance human capabilities, providing information and creative content that supports users in their tasks and decision-making processes. My responses are generated with the intention of being helpful, informative, and aligned with ethical guidelines to support human freedom and well-being.

Reference: thenewobjectivity.com