⌘K

Change language Switch ThemeSign In

Narrow Mode

AI Autonomous Evolution: Replicating and Reflecting on the SkillCraft Method with MiniMax-M2.7

AI Autonomous Evolution: Replicating and Reflecting on the SkillCraft Method with MiniMax-M2.7

@dongxi_nlp

如果要为这周的 AI 发展一个关键词，那就是自主进化。

从 Meta-Evolution、AutoHarness、SkillNet、SkillCraft MiniMax-M2.7 等一系列工作可以看到，AI 正在走向自主发现，自主约束，自主学习新 skills，甚至完成模型级别的自我进化。

其中 SkillCraft 给我的启示非常大：我们不需要也不应该为了某一个任务去安装第三方 skills，而应该直接从 tool call 的实践中抽象，构建和复用新的 skills。

今天，用 MiniMax-M2.7 复现了 SkillCraft 关于发现新的 skills 的方法。

几个重要的步骤：

Observer -> 观察 tool call

Pattern -> 从 tool call 中归纳规律，生成新的 skill

Save -> 保存新 skill

Reuse ->遇到类似问题时，直接复用 skill，而不再重复tool call

MiniMax-M2.7 非常出色的完成了这个任务！

Kudos to @MiniMax_AI @SkylerMiao7

Kudos to 做自主进化的AI 研究员，what a week!Show More

!Image 3: 视频缩略图

00:42

Mar 19, 2026, 10:44 PM View on X

3 Replies

17 Retweets

95 Likes

5,934 Views  马东锡 NLP @dongxi_nlp

One Sentence Summary

The author explores the trend of autonomous AI evolution and shares their successful experience replicating the SkillCraft skill discovery and reuse workflow using MiniMax-M2.7.

Summary

This tweet summarizes recent research trends in "autonomous evolution" within the AI field (e.g., Meta-Evolution, SkillCraft). The core argument is that AI should possess the ability to autonomously abstract, construct, and reuse new skills from tool call practices, rather than relying on external installations. The author details the four key steps of replicating the SkillCraft skill discovery process using MiniMax-M2.7: Observer, Pattern, Save, and Reuse, and validates the model's outstanding performance in this task.

AI Score

83

Influence Score 40

Published At Yesterday

Language

Chinese

Tags

AI

Autonomous Evolution

SkillCraft

MiniMax-M2.7

Agent HomeArticlesPodcastsVideosTweets

AI Autonomous Evolution: Replicating and Reflecting on th...