Small LLMs commit to a hunch and stop tracking candidates. This post breaks down the failure modes, describes a loop that grows a creator and a solver in turn, and reports how far a per-turn reward actually moved the numbers.
[Read More]
Co-evolving Four Roles, GAN-Style: The Testbed Behind umigame_GAN
A testbed for strengthening trial-and-error reasoning in small LLMs: four competing roles, a domain-independent loop, R-Zero and EDA style rewards, and an MLX LoRA comparison design. Code excerpts and measured numbers from the repository.
[Read More]
Inside chatGPT_streamlit: A Search-Enabled Chat UI with Streamlit and LangChain ReAct
A tour of chatGPT_streamlit: a LangChain ReAct agent with Google and Wikipedia tools behind a Streamlit chat UI. How the tools, prefix/suffix prompt, and session_state history are wired, plus the version pinning and the gaps between README and code.
[Read More]
Cutting the Cost of Knowing What Is Trending on X
Following an X timeline costs three things: finding the right posts, reading them, and summarizing them. How TopicDigestX moves each to a server, the trade-offs, and what is still not built.
[Read More]
Fine-tuning a Japanese LLM Without a GPU: One Mac, One Pipeline
No CUDA GPU, no cloud quota, no separate evaluation scripts. How FT_LLM turns fine-tuning a Japanese LLM into one Mac workflow, and where the approach still falls short.
[Read More]
A LoRA Training Pipeline for Japanese LLMs on MLX: LFM2.5-1.2B-JP
How FT_LLM fine-tunes LFM2.5-1.2B-JP with LoRA on Apple Silicon: config resolution, two execution modes, a chunked loss for long sequences, early stopping, and character-level evaluation, with code excerpts and measured numbers.
[Read More]
A Discord Bot That Lives in the Chat: Ollama, discord.js, and a News Pipeline
How Lobster, a Kansai-dialect Discord bot, is built: one Node process running discord.js and Ollama, persona composition, an instant 👀 acknowledgement, 17 RSS feeds, and interests that change from chat.
[Read More]