Tag
一篇架构对比文章，分析 Pi Durable 与 OpenCode 的设计差异：Pi Durable 将模型生成、工具调用等序列化为可调度的持久任务，使工作不再绑定于单个进程，而传统会话式 harness（如 OpenCode、Claude Code、Codex）仍依赖进程内的活动 agent 循环。
A controlled re-examination of ternary language models at 60K parameters reveals that baseline shape significantly impacts performance comparisons, challenging previous claims about the routed ternary model's advantage.
Pi's new architecture is highly similar to that of the Maka tool, indicating that the agent harness layer is borrowing principles from database write-ahead logs and event sourcing, forming an engineering consensus.
The paper introduces a common layer equation that unifies graph neural networks into seven components, enabling architectural comparison, theoretical analysis, and insights into issues like oversmoothing and expressivity.
Qwen3.8-27B is identical in architecture to Qwen3.6-27B, with all capability gains attributed to training improvements.
Cerebras co-founder explains the fundamental difference between WSE (Wafer Scale Engine) and NVIDIA GPU: GPU is designed for graphics, runs AI by stacking cores and NVLink interconnect, while WSE makes the entire wafer into a single chip, with on-chip interconnect bandwidth and memory bandwidth far exceeding GPU clusters, greatly leading in inference speed.