OpenAI的ChatGPT于2022年年底发布。得益于境外手机号接码渠道，我第一时间就用上并惊叹于人工智能的突破。一转眼，LLM在过去的三年里已无可辩驳地渗透至千家万户。 一小段历史 Transformer模型最早于2017年由谷歌提出，核心思路是通过Self-Attention机制显式地表达词语之间语义联系1。OpenAI在2018年首次发布GPT-1模型，将架构简化为Decoder-only的Autoregressive词语预测2。自此，LLM的总体架构尘埃落定，后续的GPT-2、GPT-3等皆在此基础上扩展模型大小和数据集的规模。不仅局限于文字，Transformer在视觉领域表现也很好3。为何大模型能够涌现出强大的能力仍不得而知。Anthropic于2021年在一个极度简化的模型中发现了Induction Heads结构4，但如同生物学第一次发现红细胞，我们离真正理解L...

Abstract: This article provides a gentle derivation showing the existence of a simple, measurable linear relationship between the LMS color model as in human vision and RAW camera sensor data and the RGB values as in displays and jpg/png/etc. images. Real world colors are continuous spectra, such as the sunlight spectrum. We can describe it as a continuous function $J(\lambda)$ where $\lambda$...

今日通关了《死亡搁浅》，不得不提，在游玩时我本以为这是后疫情作品在反思疫情隔离，没想到实际上游戏刚好于19年发售，真是巧合。 《死亡搁浅》最独到的地方，在于实现了作者表达、剧情和游戏性的统一。小岛秀夫的主要创作意图是体现人与人之间互帮互助的重要性，游戏的剧情便是Sam作为“送货人”将后末日世界中支离破碎的美国重新连接。创新点在于，游戏的核心玩法——送货——直接服务于这一主题。它无时无刻不在用“送货的艰辛”与“其他玩家留下的建筑所带来的帮助”形成鲜明对比，点赞系统也让玩家在帮助他人时获得反馈，大多数玩家都有类似的情感经验（举例1）。 《死亡搁浅》的评价不出意料地出现分歧，主要集中在游戏性不强，玩法单一。23Sylvester在他的《游戏设计》中提出，游戏是“体验生成引擎”4，也就是用游戏独有的交互性（机制、系统）生成特定的体验（experience）。《死亡搁浅》的游戏性争议来源于...

I am trying to pick a laptop that is as close to Macbook Mx Airs as possible - light, low power consumption and long battery life, with the only exception of not being an Apple device123 - for the upcoming coursework. As for the chip itself, the closest I can get is Intel’s Lunar Lake CPUs whose successor unfortunatelly Intel will not develop due to their management incompetentness. Among all L...

The proof is trivial with a little help of a necessary condition of complexity theories. All LLM runs under the complexity of $O(n)$, where $n$ is the length of the output. Suppose LLM is AGI, then it is able to solve any problem that a human can solve. Consider the following problem: Given a string $s$ of length $n$, determine whether $s$ is a palindrome. Answer “Y” if it is and “N” if it...