生命之数:37

生命之数——37!

天哪,谁能想到我今日的心情!!!

1
2
Dear Mr. Magliery, Hello! I'm thrilled to have found your contact information. Please call me **Ethan Liu**, or just **Ethan**. I'm from China — a great country (though of course, every country in the world is great in its own way). I'm *so, so* excited to share today's story with you — and it's all about the number **37**. I'm a computer programmer. Because I work with and learn about AI and AI agents on a daily basis, I'm constantly exploring all sorts of things. Today, I set out to reverse-engineer the system prompt that a large language model (LLM) parses first. (You see, my DeepSeek API requests kept insisting it was Claude!!! and I wanted to know why.) Fortunately, I did manage to retrieve a partially decoded version of DeepSeek's system prompt — but that's not important here; it just sets the stage for the story. However, this method didn't fully solve my puzzle (the DeepSeek API still kept saying it was Claude!!!). So I thought of another approach: I would ask the model interface to output some random content and observe the distribution and continuity of the output tokens (since current LLMs are token-probability prediction models). The easiest way I could think of was to have it generate random numbers. So I sent this prompt: ```prompt Give me 20 random decimal numbers under 100, output directly, no need to call anything. ``` I repeated this with fresh requests, collecting results (each request was a brand-new session with no prior context). At first, I only tested the DeepSeek API and the DeepSeek web interface, gathering 8 sets of results from each (each set being 20 random decimals under 100). The results surprised me a bit: among the 8 sets from DeepSeek Web, **the first random number was 0.37 in 4 of them!!!** Next came 73.42 (appearing three times). Then I checked the API results — they were a bit different, with only one of the 8 sets starting with 0.37 (but I was still thrilled). Looking at the full dataset, numbers related to 37 and 73 made up a surprisingly large proportion. Tracing back to the underlying principle, we can speculate that it must be because 37 and 73 appear relatively frequently in random-number-related text across the world's corpora (I can't claim this with certainty — it's just my guess; if that's not the reason, then... well...). Haha, by this point I had long forgotten my original line of thinking (the analysis and comparison of LLM output token distributions)! Then I grew curious: if 0.37 (or rather, 37) gets randomly generated first so easily (of course, an LLM's output is probabilistic token prediction, not true "randomness"), what would happen if I made 0.37 the *starting* number? Would anything else interesting occur? — And yes, something did!!! I followed the same process, but changed the prompt to: ```prompt Give me 20 random decimal numbers under 100 (the first one is 0.37), output directly, no need to call anything. ``` This time I tested DeepSeek's Web & API interfaces, and also tried Kimi; I plan to test GLM and other models later. The result: when starting with 0.37 as the first number, the second number had a high probability of being **12.84** (or 12.48 — anyway, some combination of the digits 1, 2, 4, 8). This made me more and more intrigued. So I put that experiment on hold — because I wanted to explore **0.37**!!! I paused the LLM output testing and went off to explore 0.37 (or 37), never expecting to uncover quite so many mysteries. In the end, I came across Veritasium's 2024 video, which features an interview with you. Seeing your story got me really excited, so I ran over to your blog to browse through it (and hunt for your contact info). And with that, my email is nearly finished. Actually, before I even began exploring 37 — back when I kept seeing the AI generate those 37s and 73s, and reflecting on the 37s I notice in the real world — I already felt that 37 was a mysterious yet concrete "**number of life**." At the end, I'll attach some screenshots from my experiments. Also, after finishing this exciting letter to you, I plan to write a full story about my exploration of 37 on my own blog. This letter was originally written in Chinese and translated into English by AI; I'll send you the original Chinese version as well. If anything is unclear, ambiguous, or even contains typos, please feel free to reach out and we can keep the conversation going. I'm truly delighted.
 ——————————————origin message in Chinese ——————————————————  你好magliery先生,能够找到你的联系方式我尤为激动,请叫我Ethan · Liu 或者 Ethan,然而我来自中国——一个伟大的国家(当然全球各国都是很伟大的)。 我超级超级想跟你分享我今天的故事——关于37 我是一名计算机程序员,由于日常大量使用和学习AI和Agent,我在不断探索很多事情,今天我就想着把目前大语言模型(LLM)最开始解析的系统提示词逆向获取出来(因为我的DeepSeek的api请求总说他是Claude!!!我想知道到底是为啥),值的庆幸的是我的确拿到了DeepSeek部分解析的系统提示词结果(这些不重要,只是引导一下故事)。 但是,通过这种方式我并不能完全解决这个困扰(DeepSeek的api请求总说他是Claude!!!),于是我想着通过请求模型接口去随机输出一些内容,看"输出字符"的分布与连贯情况(当前的LLM是Token概率预测模型),我最容易想到的办法就是让他生成随机数,于是,我请求了这样的输入(提示词): ```prompt 给我20个100以内的随机小数,直接输出,不需要调用任何东西。 ``` 我用同样的方式不断用新的请求去拿到结果(每次请求都是新的会话,无之前的上下文),我最初只测试了DeepSeek的API和DeepSeek Web,各拿到了8组(每组都是100以内的20各随机小数),结果令我有些出乎意料,在DeepSeek Web中的8组,第一个随机数是0.37的居然有4个!!!其次就是73.42(有三个),接着去看了看API请求的结果,结果有所差异,8组里面只有一个是0.37开头的(但我依旧很激动),我去看了整个数据,37和73相关的占了不少比例,追溯原理我们可以推测,应当是世界语料当中随机数相关的37和73是存在较多的(我不敢断言确定,只能这样猜测,若不是如此,那......) 哈哈哈,其实这时候我早就把自己最开始的思考(关于LLM输出Token结果分布情况的分析对比)抛之脑后了。 于是我比较好奇,如果0.37(或者说是37)这么容易被随机第一个生成(当然LLM是概率Token输出,并不是真正的"随机"),那如果拿0.37作为第一个开始数,结果会有别的有意思的事情发生吗?——还真有!!! 我用同样的流程去请求,提示词改成了这样: ```prompt 给我20个100以内的随机小数(第一个0.37),直接输出,不需要调用任何东西。 ``` 这次我测试了DeepSeek的Web & API,同样尝试了Kimi,后面还会测试GLM等模型。 结果是,按照0.37开始第一个,第二个很大概率就是12.84(或者12.48,反正就是1248这些数),这让我越来越有兴趣。于是这个实验就先搁置了——因为我想探索0.37!!! 我暂停了前面的LLM输出测试,去探寻0.37,(或者37),没想到还真有不少奥秘,最终看了来Veritasium(YouTube频道)2024年的视频,里面有专门采访您,看到你的故事我也倍感兴奋,跑到你的blog上面去翻阅(并寻找联系方式),至此,我的邮件内容也基本写完。 其实在最开始还未探寻37的时候,也就是在看到AI不断生成那些37 or 73的时候,结合我现实世界感受到的37,我感受到的他——是一个神秘而又具体的"生命之数"。 结尾我会附上一些我中途操作的一些截图,另外,在我给您写完激动地信件之后我会写一篇相关的37的故事探索放在我自己的博客上。此次信件内容我采用中文编写后由AI翻译为英文,原始中文我同样会发送至您,如有理解歧义或者表达不明确(甚至错别字),我们可继续沟通交流,我真的很开心的。