I added an LLM to my friends' chat and made it play a Terminator (and opened a portal to a multiverse of madness)
NVIDIA
Google/DeepMind
OpenRouter
A developer integrated a free LLM via OpenRouter into a Telegram group chat for 15 friends, instructing the bot to role-play as the Terminator (T-800). The bot can answer messages, tease participants based on context, read images/audio/video, classify dangerous prompts, and has a RAG-based memory. The project was later moved to a local laptop using Gemma-4-12b and llama.cpp for better performance.
A developer decided to add an LLM-powered bot to a Telegram group chat with 15 friends, instructing it to role-play as the Terminator (T-800). The bot uses free models from OpenRouter (initially Nemotron-3-Super-120b for text, then nemotron-nano-12b-v2-vl for vision, with fallback to nemotron-3-nano-omni-30b-a3b-reasoning). Later, the bot was moved to a local laptop with Gemma-4-12b running on llama.cpp to handle both text and multimedia in a single stream. The bot can respond to messages, tease participants, read images/audio/video, summarize books/comics, perform web searches, classify dangerous prompts using a two-pass approach (first a classifier model, then the main model), and maintain long-term and short-term memory via RAG. The two-pass classifier on long messages (over 200 chars) uses nemotron-3-nano-30b to reject absurd or overloading prompts before they reach the main model. The author notes that the project is not production-ready but serves as a fun pet project.
- Сокращения
- LLM = Large Language Model — большая языковая модель
- API = Application Programming Interface — интерфейс программирования приложений
- RAG = Retrieval-Augmented Generation — генерация с дополнением извлечением
- RAM = Random Access Memory — оперативная память
- VRAM = Video Random Access Memory — видеопамять
- KV = Key-Value — ключ-значение
Source: Habr — хаб NLP —
original