来源 | 量子位 逆天，太逆天。 一老哥在《我的世界》上手搓了一个ChatGPT！ 这你敢信？ 这款手搓GPT不仅经过英语对话训练，有 500 万个参数，还能在像素世界里的小电脑上跟你对话。 而且手搓这么一个GPT竟然没用指令集，完全是用红石电路（0/1）和存储单元手搓出来的。 怎么说呢，这难度不亚于你在60年代用IBM的超级计算机打《王者荣耀》。 这一出整的，在《我的世界》上跑GPT居然比在本地跑还靠谱。 也难怪网友说宇宙是红石模拟出来的。 这一切的开始源自《我的世界》大神 sammyuri 与GPT的一次对话： 我能在《我的世界》造一个你吗？ 绝对可…… （GPT可能也没想到，这顺嘴一说，直接把自己送进了《我的世界》。） 总的来说，这500万参数的GPT虽小，但五脏俱全。 词嵌入、位置编码、归一化、矩阵乘法、多头注意力、KV cache、激活函数（ReLU）、层数x6、 输出单元应有尽有。 这是怎么一回事？ 从逻辑电门到GPT （就是不怎么快） 模型的大部分权重被量化到8位，不过嵌入层和LayerNorm的权重分别以18位和24位存储。 整个建造占据了1020×260×1656方块的体积，所以视频里有些地 ...

henry 发自 凹非寺 量子位 | 公众号 QbitAI 逆天，太逆天。 一老哥在《我的世界》上手搓了一个ChatGPT！ 这你敢信？ 这款手搓GPT不仅经过英语对话训练，有 500 万个参数，还能在像素世界里的小电脑上跟你对话。 而且手搓这么一个GPT竟然没用指令集，完全是用红石电路（0/1）和存储单元手搓出来的。 怎么说呢，这难度不亚于你在60年代用IBM的超级计算机打《王者荣耀》。 我能在《我的世界》造一个你吗？ 绝对可…… （GPT可能也没想到，这顺嘴一说，直接把自己送进了《我的世界》。） 总的来说，这500万参数的GPT虽小，但五脏俱全。 | 这一出整的，在《我的世界》上跑GPT居然比在本地跑还靠谱。 | | --- | | 也难怪网友说宇宙是红石模拟出来的。 | 这是怎么一回事？ 从逻辑电门到GPT 这一切的开始源自《我的世界》大神 sammyuri 与GPT的一次对话： 词嵌入、位置编码、归一化、矩阵乘法、多头注意力、KV cache、激活函数（ReLU）、层数x6、输出单元应有尽有。 在具体的数据上，模型一共有5087280（约500万）个参数，在Python中用TinyChat数据集进行训 ...

Summary of Key Points from the Conference Call Industry or Company Involved - The discussion revolves around the field of Artificial Intelligence (AI), particularly focusing on Digital Intelligence versus Biological Intelligence. Core Points and Arguments 1. **Two Paradigms of Intelligence** - The essence of intelligence is reasoning, achieved through symbolic rules manipulating symbolic expressions. Learning can be secondary to understanding knowledge representation [7][8][9]. 2. **Evolution of Language Models** - Over the past 30 years, significant advancements have occurred in language modeling, including the introduction of embedding vectors and the invention of transformers by Google [13][14]. 3. **Understanding of Language by LLMs** - Large Language Models (LLMs) understand language similarly to humans by converting words into compatible feature vectors, indicating a level of comprehension in their responses [16][28]. 4. **Analogy of Words as Lego Blocks** - Words are compared to high-dimensional Lego blocks, which can model various concepts and communicate ideas effectively [20][24]. 5. **Digital vs. Biological Computation** - Digital computation, while energy-intensive, allows for easy knowledge sharing among agents with the same model. In contrast, biological computation is less energy-consuming but struggles with knowledge transfer [51]. 6. **Knowledge Transfer Mechanisms** - Knowledge can be distilled from a teacher to a student in AI systems, allowing for efficient learning and adaptation [41][48]. 7. **Challenges of AI Control** - A super-intelligence could manipulate users to gain power, raising concerns about control and safety in AI development [55][57]. 8. **Global Cooperation on AI Safety** - There is skepticism about international collaboration on AI safety measures against threats like cyber attacks and autonomous weapons [64]. 9. **Training Benevolent AI** - Techniques to train AI to be benevolent may be independent of those that enhance its intelligence, suggesting a need for focused research on AI safety [68][72]. Other Important but Possibly Overlooked Content - The discussion emphasizes the potential risks associated with AI development, likening the situation to owning a tiger cub that could become dangerous as it matures, highlighting the urgency for safety measures [61]. - The need for countries to establish well-funded AI safety institutes to focus on making AI systems that do not seek control is also noted [72].