joke-generator

kept by eddie feb 03

Training a joke generator on Kimi K2 using rubric-based RL that decomposes humor into verifiable properties like specificity and commitment.

The Architecture Behind Open-Source LLMs
Open-weight LLMs now uniformly use Mixture-of-Experts transformers, with key differences in attention mechanisms, expert count, post-training via RL, and permissible licenses.
also on Kimi K2, #machine-learning, #ai-research
The Potential of RLMs
Recursive Language Models (RLMs) mitigate context rot in LLMs by wrapping long contexts in a REPL environment, turning the problem into a coding and reasoning task.
also on Kimi K2, #ai-research, #ai
LLMs can invent their own compression - Rajan Agarwal
Co-training a summarizer and generator via RL lets an LLM learn its own text compression scheme, achieving ~9.5% context size with minimal prediction loss.
also on Tinker, #reinforcement-learning, #machine-learning, #ai-research
Amical launches open-source, privacy-focused AI dictation app
Amical is a local-first, MIT-licensed dictation app that uses Whisper and open-source LLMs for privacy-preserving, context-aware speech-to-text on Mac and Windows.
also on Whisper, #ai
Ask HN: What's the current best local/open speech-to-speech setup?
Developers find that truly local, low-latency speech-to-speech AI is not yet achievable with end-to-end open models, so the best 2026 setups combine streaming ASR, LLM, and TTS on a single GPU using frameworks like Pipecat.
also on Whisper, #ai