LLMs may spontaneously switch to Chinese mid-reasoning regardless of prompt language — observed in both OpenAI's o1 and Chinese models (DeepSeek, Qwen, GLM)
The papers listed below list 3 possible reasons for this: internal circuit competition, strategic reasoning advantages gained during training, and the influence of distributed training data.
Mechanistic interpretability research suggests that multilingual LLMs possess two distinct internal subsystems that govern generation:
Date: 2026-04-18
Analysis of caveman — a skill/plugin instructing AI coding agents to respond in compressed prose, dropping articles, filler, and hedging while preserving technical accuracy.
Caveman claims: "~65-75% fewer tokens," measured via an eval harness that counts visible output tokens.
| #!/bin/bash | |
| python3 -c " | |
| import os, urllib.parse | |
| files = [f for f in os.listdir('.') if os.path.isfile(f)] | |
| files.sort(key=lambda x: os.path.getmtime(x), reverse=True) | |
| with open('index.html', 'w') as out: | |
| out.write('<html><body><h2>Files (newest first)</h2>') | |
| for f in files: | |
| url = urllib.parse.quote(f) | |
| out.write(f'<a href=\"{url}\">{f}</a><br>\n') |
Driver 565.77 CUDA 12.7
(currently Aug 2025 running on 22.04 Linux pop-os 6.12.10-76061203-generic)
https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#deprecated-architectures
Maxwell, Pascal, and Volta architectures are now feature-complete with no further enhancements planned. While CUDA Toolkit 12.x series will continue to support building applications for these architectures, offline compilation and library support will be removed in the next major CUDA Toolkit version release. Users should plan migration to newer architectures, as future toolkits will be unable to target Maxwell, Pascal, and Volta GPUs.
