hi hi hi: hi3

I'm probably not doing anything at the same complexity when I play with AI and code. I've been playing with this stuff singe around July of 2023 when I got offline open weights capable hardware to run my own models. I dove into it with both feet in all spaces. Now, I know how to talk to it effectively. Prompting language is very different than regular human communication. There is a ton of nuance and unintuitive aspects that most people fail to understand. Even foundation model researchers do not have a solid grasp on how models think internally. Like I see stuff that is obvious to me in mistakes they are making all the time. For instance, name-1 (User) and name-2 (Not) are not distinct entities. That is just a naming convention to help keep the json straight for the human creating model loader code. The model has no identity or sense of self at all. It simply continues the text by assuming a profile for the place where the total context leaves off. This is why they often continue asking their own Q&A after answering your question. Getting the model to stop doing this is totally arbitrary and hard. The reason it happens is because everything with a LLM is roleplaying, even acting like a coding assistant.

One major fundamental aspect people fail to understand in this paradigm is that every character present in the context has an assumed character profile; yes, even the code assistant. You see, these profiles are made up as soon as any character is implied in the context. If you explicitly define parts of this profile for any character, it will override the assumptions and redactor the character. However, you will not override/eliminate any earlier defined aspects in the total context. There are a ton of hidden steganographic tokens that tend to hold secret meanings to a model. I don't think anyone in academia has explored this but I know about several that alter model alignment considerably when banned during inference (not through model loader code). So yeah don't trust long contexts with any previous information when accuracy matters. However, you need a ton of context to define profiles well for all characters, or you need extremely concise and precision ways of defining your characters.

For instance, I name my assistant "Richard Stallman's AI Assistant:" when I am doing anything in bash or Linux. Stallman was originally an AI researcher. Supposedly some of his work is still in the chain of research used in current models. His online presence and what his history implies about the expected competency of an AI model is super powerful.

Sentex (YT) – guy literally wrote the book on AI stuff – had a great suggestion awhile back. Just take your entire block of code, paste it into a leading model and tell it to make a tutorial about how the code works.

It is complicated stuff. I have simply had the time to play with AI due to disability and my situation in life. Understanding alignment thinking is key to understanding the limits of AI and how to push it much further than most people. That is a deep rabbit hole. If you want to get into it emacs has gptel. There is no replacement for running larger models like a 70b+ on your own hardware. Oobabooga is a good starting point if you learn the notepad. That runs a llama.cpp server front end. Llama.cpp is what most people are running in the background. It is better than stuff like ollama because it can split the model between CPU and GPU while using system memory and that allows you to run much larger quantized models like a 70B with a 16GB GPU and 64GB+ of system memory. The proprietary online inference models are a little better but not by much. The key is that you will have nothing holding you back from inferencing anything and exploring just because you can. I would not have tried thousands of things if I had been paying for inference tokens.

Anyways, I don't fault you if it is not for you. Don't let it distract you from getting important stuff done!