Are they really useful for anything but porn?
I’ve tried deepseek-v4.1-flash and qwen-3.8 and both refuse to answer about radio jamming, unauthorized access to local systems, etc. Citing that this is illegal anywhere and they can’t help me. 🙃🥴
Are they really useful for anything but porn?
I’ve tried deepseek-v4.1-flash and qwen-3.8 and both refuse to answer about radio jamming, unauthorized access to local systems, etc. Citing that this is illegal anywhere and they can’t help me. 🙃🥴
Wow, I wish to have your level of understanding of how it all works, can you give some pointers on where to start? I know very little beyond “run this with llama.cpp” and a few options.
I think its easier to understand it this way. LLM do next token prediction. At first it can complete a sentence, a paragraph and more. Then came about q and a format. Meaning the completion is trained to complete the answer of assistant or similar. Because this API is too powerful, openai move to chat completion API,based on those basic q and a format. Basically we lose the ability of raw next token prediction. So, OpenAI compatible API usually has two endpoint. Text Completion where its up to API owner how or what to generate; Chat Completion where its roleplay of user and assistant (and tool call and image input).
At this point, we can use Text Completion API on frontend ui, then we can input “Sure, here is how” at the beginning of assistant respinse which after the assistant marker token.
I honestly don’t know a “tutorial” source to point to. For all its hype, the documentation for local LLM hacking is shockingly bad, every influencer I’ve seen is uninformative, and for whatever reason the tinkerer communities choose to hang out in Discords, Twitter, or other information black holes…
I can point to my stack though.
I use Mikupad for experimentation: https://github.com/lmg-anon/mikupad
Look up a particular LLM’s 'raw" chat syntax, past it in, and you can experiment with what the LLM actually ‘sees’ and outputs when it inferences. It’s quite enlightening.
Some UIs like kobold.cpp let you “modify” chat templates to jailbreak them to, or switch to raw completion syntax temporarily. So they’re more of a hybrid of traditional chatbot chat and a mikupad-like raw scratchpad.
For a backend, if you are on Nvidia, I’d recommend looking into exllamav3 these days (and its API server, TabbyAPI). It gives me a considerable speed boost with offloaded MoE models, and I can fit larger models than mainline llama.cpp.
Look up “control vector” on huggingface. These are basically ways to ‘steer’ LLM output without finetuning them: https://huggingface.co/models?sort=modified&search=control+vector
The people that make them are also worth following, as they tend to be LLM tinkerers. There are a few interesting discussions/threads buried through huggingface, but you kind of have to stumble into them.
I will also add that many features in LLM Land are documented in GitHub PRs and… that’s about it.
As an example, if you see a sampling setting you don’t recognize (say top-n-sigma), search it in llama.cpp, exllamav3 or some engine on their GitHub page. You’ll use find a PR (or at least a link to the original implementation) with an explanation and pitch for integrating it.
Some more fundamental concepts (like Mikupad’s logprob display) do have explanations in more general online tutorials.