MarkTechPost

Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

Google open-sourced Harness2, a recursive agent harnessing framework that lifted GLM-5 11.5 points on WorkBuddy-Bench with ...

Identifying standard library APIs removed in Python 3.15 from your code using AST—502 public name differences and a comparison with ruff, pyupgrade, an…

Python 3.15.0 was released on October 9, 2026. The "Removed" section of the official "What's New in Python 3.15" document ...

Why a Sudoku-solving program looks smart: Recursion, backtracking, and the idea of 'trying from the most likely failure points'

I was thinking about a program to solve Sudoku.You fill a 9x9 grid with numbers from 1 to 9. The same number must not be ...
i-SCOOP

Google’s Harness² improves AI agents without touching the model

Google's Harness² rewrites the harness around a frozen AI model. How it works, what the benchmarks show and what it means for ...
KDnuggets

The Roadmap to Mastering Agentic AI Systems in 2026

A four-level technical roadmap for building agentic AI systems, from single cognitive loops and stateful memory through multi ...
InfoWorld

Local AI on NPUs: Building a zero-cloud photo deduplicator with OpenCLIP

Your laptop’s idle NPU can sort 100 gigabytes of family photos about 3 times faster than the CPU, without draining the ...
MarkTechPost

AI Shorts

Liquid AI released d1-ultra, a text and image decision model that returns calibrated probabilities in 1 forward pass. It ...

2026 三大自进化方案拆解：Dream-RSI、RRSI 与 RSIAgent

2026 年 9 月，三篇 RSI（递归自我改进）前沿论文几乎同时爆发：Google DeepMind 的 Dream-RSI 让 AI 把历史试错冻成"回放模拟器"，在梦里练出更优探索策略，算力开销最多降 162 倍；Google Cloud AI ...
腾讯网

华为AI靠RSI通关25款游戏！操作次数比人类少56%

一边打游戏，一边给自己升级？ 好家伙，华为这波直接让AI学会了《边玩边变强》啊？！ 最新一项研究，华为诺亚方舟实验室联合UCL等团队拿出了一个名叫Memento 3的AI智能体系统—— ...

科研Agent也能越用越聪明？ScienceBuddy 开始让Harness和模型一起进化

现在的科研 Agent 已经能查论文、分析数据、调用工具，甚至生成完整的研究报告。但有个问题一直没解决：研究者明明纠正过 Agent 的错误，下一次遇到类似任务，它可能还是会犯同样的错。最近，PhAI Labs 团队发布了 ScienceBuddy，尝试让科研 Agent 真正从使用过程中积累经验。这项工作的特别之处在于，它没有只盯着模型本身，而是将 Harness 自进化和模型强化学习结合起来 ...

华为AI靠RSI通关25款游戏！操作比人类少56%，真·边玩边进化

2026年春天，全球AI圈炸开了一条‘像素级震撼’——华为诺亚方舟实验室联合伦敦大学学院（UCL）等顶尖团队，扔出一枚硬核炸弹：Memento3智能体系统，不微调、不重训、不刷数据，仅靠‘自己写攻略+自己造模拟器+自己推翻自己’，一口气通关ARC-AGI-3全部25款高难度解谜游戏、183个关卡，RHAE（相对人类行动效率）拿下史无前例的100分！更绝的是——它只用了7518次操作，而人类平均要敲 ...