小麥 Xiaomai Lab|研究摘要 版本:2026-09-30 / 1.0 網址:https://xiaomailab.com/ 聯絡:evant515ro@gmail.com 我們想研究什麼 如果問題從一開始就問錯了,AI 能否發現可疑假設,質疑人類給的問題與自己的問題形成機制,重新定義,再透過實驗留下有效改變?想像可能是起點;證據、反駁與可重現結果才是判準。 目前已存在的工作(來自交付摘要,網站製作時未重跑模型測試) 1. v0.9.2:4,555 項 CPU 回歸測試,0 失敗、0 錯誤、0 略過。範圍是學習交易基礎,不是自主研究驗收。 2. 限定 CPU 試驗:主權重/動量保留、新程序接續下一步一致性;不是完整 True 單檔整合冷恢復。 3. 2026-09-29 A100:同一 True 恢復與選定狀態唯讀匯入,47 份主權重、20 份既有動量、1,067,974 個選定參數值。没有前向、梯度更新或新 checkpoint。 4. v0.9.4 README:真實第 0 層 GDN 核心限定比對,2 token 預填、1 token 解碼、6 組梯度比較;完整整合未完成。 5. v0.9.4 README:456 個合成參數的 Python 反向、AdamW、保存與恢復試驗;不能代替真實模型新後端訓練驗收。 尚未證明 自主找出錯誤假設與重新定義問題;創造經驗證的新知識;在新問題與重啟後保留自主研究能力提升;完整真實模型底座全部由 Python 執行;完整逐值運算追蹤或無上限自主進化。 接下來的最小驗證 以能自動判對錯的小環境,比較直接作答、多次反思、小麥機制、核心機制消融四組;固定資訊、資源與評估規則;不提示錯在哪;在未調參的新條件及新程序恢復後評估。失敗與無差異也保留。 網站示意的範圍 控溫示意是兩個人為設計控制器的簡化、確定性模擬。不是小麥模型運行,不是四組實驗,也不是自主發現假設的證據。圖中公式、參數與逐步 CSV 可供核對。 研究支持 原訂 NT$900,000 為討論草案;尚未正式募資,沒有收款功能。比例試算是示意,不是估價或承諾。正式上線前需補齊成本、期間、方案與風險;不保證研究成功或收益。 Xiaomai Lab — Research brief Independent AI research in Taiwan. Updated 2026-09-30. Research question Can AI recognize questionable assumptions, revise a supplied problem and its own problem-forming process, test alternative explanations, and preserve verified improvements across new tasks and restarts? Current evidence Bounded CPU regression and continuity tests; a selected real-model state read-only import verification on A100; limited real-layer comparison summaries; synthetic learning-and-restore experiments. These are research infrastructure, not proof of autonomous problem redefinition or new scientific discovery. Website production did not rerun the underlying model tests. Next evaluation Compare direct answers, repeated reflection, the proposed Xiaomai mechanism, and its core-mechanism ablation under matched information and resource budgets. Evaluate unseen conditions and new-process recovery. Preserve failures and null results. The website demo A deterministic toy temperature-control comparison between two human-designed controllers. No live AI model is invoked. It illustrates a framing difference, not autonomous discovery. Collaboration and funding We welcome collaboration on metacognition, problem formation, interpretable computation, controlled experiments and independent reproduction. Fundraising is not launched. No payments or investment returns are offered here. Selected related research — context, not endorsements Gödel Machines: https://arxiv.org/abs/cs/0309048 The AI Scientist-v2: https://arxiv.org/abs/2504.08066 AlphaEvolve: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/