新智元报道
周五晚上11点，Karpathy累了，他打开YouTube，本想随便找点东西看。
迎接他的是一整面「垃圾墙」：封面一个比一个夸张，推荐翻来覆去还是那几个老话题。
他吐槽：满屏都是《蠢蛋进化论》级别的烂内容。
累了一天想放松，结果只刷到一堆标题党：这样的感受你肯定不会陌生！
在Karpathy看来，YouTube的推荐算法也许能讨好大多数人，却满足不了像他这样的用户。
后来他在评论区说，像他这样的用户太少：在算法眼里，小到可以四舍五入掉。
这次，他决定不忍了。
第二天上午，他让Claude造了一个只放好视频的YouTube。当天，网站就上线了。
Karpathy给这个网站起名叫HackerTube。
6454个视频、150个主题，组成Karpathy自己的看片入口。
视频仍来自YouTube，推荐线索则来自Hacker News（HN），一个程序员和创业者扎堆的技术论坛。
Karpathy的思路很简单：把HN网友安利过的视频攒成片单。
他把这件事几乎全部交给了AI。
Claude先下载了HN近20年的4300万条评论，再筛掉垃圾，只留推荐YouTube视频的那些，近4万条。
最后整理出6454个视频，分进150个主题。
视频的标题、时长，再从YouTube那边调过来。然后套上一个仿YouTube的界面，部署上线。
整个项目，约2小时。
从筛数据到上线，Karpathy说自己没读也没写代码。
他还把数据打包公开了。
谁都可以交给自己的AI，改出一个自己的版本。
周六上午，他给Claude下了一道指令：
照着HN网友推荐过的视频，做一个我们自己的YouTube。
接下来，基本是Claude在干活。
Opus 5.5负责统筹，筛评论的活派给更轻快的Sonnet批量去做。
垃圾内容洗掉，剩下的就是推荐视频的评论。
网站写完，还得上线。
Karpathy开好服务器，把IP地址丢给Claude。
更新、部署、安全加固、性能优化，Claude一口气全包了。
而这位OpenAI联合创始人，只在Claude干活的间隙里干了开服务器、买域名这样几件跑腿的活。
他还调侃，感觉有点像团队合作。
他今年5月加入了Anthropic的预训练团队。
这次用的，正是自家的Claude。
最狠的是，Karpathy在博客结尾补的那句话：
他已经不做任何传统工程的活了。
不读写代码，不在终端敲命令，只跟Claude说话。
一位网友在评论区直接点破：这句才是全文重点。
当年带过特斯拉AI团队的人，现在干活的方式，只剩一个：提需求。
2021年，Karpathy（左）在特斯拉AI Day现场发言，右为马斯克。
HackerTube网站界面一眼看去就是YouTube，用起来却处处反着来。
它不用登录，看过什么只存在你自己的浏览器里。
没有那些反复测试过、专勾人点击的封面。
每个视频的封面，都换成了视频中段的一帧。侧边栏还有个「Longs」入口，专放长视频。
这名字一看就是冲着YouTube的Shorts去的。
Karpathy把网站发到HN，帖子冲到2000多分。评论区跟了500多条。
网友拿它和YouTube首页比了一把。
有人说，首页推荐比YouTube的个性化推荐还准。
有人扎进巴赫作品解读的兔子洞，说YouTube上刷不到。
还有人说，从YouTube切过来，像是走进了一片树林。
一位网友说，自己87岁的父亲，天天刷YouTube，因为总被推送二战的负面内容，困在里面出不来。
连老年科医生都劝老人，别再看这类视频了。
这位网友认为，HackerTube已经很接近父亲需要的东西。
你多看一眼什么，算法就推更多什么。很多人就像上面这位老人一样，被困进了信息茧房。
有网友吐槽，算法关心的是怎样让你一直看；也有人说，平台并不在乎你看完后的感受。
连付钱的用户也憋着火。
一位网友说，自己开了好几年YouTube高级会员，却越用越讨厌。
离线同步经常出问题，在飞机上根本看不了自己的内容。
这位网友感叹，真希望能自己做一个YouTube客户端，又不确定平台允不允许，「真令人沮丧」！
两小时做的网站，精准戳中了一大群人对算法的积怨。
也有网友看得更远。
这位网友认为，过去做推荐、建索引都很难。
现在，我们很可能按自己的口味，造自己的推荐系统。
在他看来，平台是一群「眼球贩子」，只想让你一直看，内容也因此越推越极端。
现在，正是摆脱它们的机会。
人动手少了，后台的活并没有少。
HN上有人问，这个项目到底烧了多少token、花了多少钱。
Karpathy让Claude自己算了一笔账。
输入约5.04亿token，其中约九成命中缓存。
输出约760万token。
九成以上的用量，花在Sonnet给评论打标签上。
真正写网站、部署上线，Opus只用了约100万token。
这5亿是一次次调用累加出来的。模型并没有一口气读进5亿token。
Karpathy披露，九成以上模型用量用于评论标注，应用开发和部署等约用100万Opus token。
两小时里，消耗token最多的环节，是给评论打标签。写网站反倒是小头。
这笔账给想做数据工具的人提了个醒：清洗、标注资料，可能比写代码消耗更多token。
不妨先拿一小批资料试筛，靠谱了再扩大。
Karpathy把评论分析的数据打成压缩包公开了。
他的建议是：交给你的AI，做个你自己的版本。
包里还附了一份写给AI的说明，讲清这些文件怎么用。
YouTube的视频信息，数据包里不直接给，要自己去YouTube拉取。说明里写了怎么取，一般要像Karpathy那样先申请一个YouTube的API密钥。
数据和说明都备齐了，就可以完全按照个人喜好「私人定制」了。
比如只关心编程，可以让AI只留编程课。再让它按主题分好类，每个视频都附上当初是谁推荐的，看过的打个勾。
这样一打开，就是一份私人课表。想学什么一目了然，学到哪儿也清清楚楚。
同样的思路，也能用在你手里的资料上，比如收藏夹里吃灰的文章、攒了多年的工具清单都能照此改造。
看什么视频，以前是平台说了算。
现在，一个不愿将就的人，也可以像Karpathy这样，自造一个个人版视频站。
以前想要一个顺手的工具，只有两条路：等平台改，或者自己写代码。
如今多了一条：给AI说清楚需求，交给它去做。
评论区有网友说，做软件的门槛正在变低。人们能飞快搭出替代品，不再被软件巨头绑住。
被算法四舍五入掉的人，开始自己造算法了。
当然，Karpathy本人就是顶级AI研究者。普通人照着做，未必两小时就能搞定。
但这扇门，已经开了一道缝。
Karpathy只定了一条标准：看HN网友推荐过的视频。
剩下的，都交给了Claude。
轮到你动手，也先想清三件事：资料从哪来，按什么标准筛选，最后想拿到什么。
AI能替人写代码，替不了人定标准。
参考资料：
编辑：元宇
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Karpathy was tired at 11 p.m. Friday. He opened YouTube and tried to find something.
He was greeted with an entire "spam wall": a cover more than an exaggerator, a recommendation to turn around or those old questions.
He spits out: the screen is full of shit at the level of "The Moot Evolution."
I'm tired of trying to relax for a day, and it's just a bunch of title parties.
In Karpathy ' s view, YouTube ' s recommended algorithm may be attractive to most people, but not to users like him.
Later, he said in the comment section that there were too few users like him: small enough to be rounded in the eyes of algorithms.
This time, he decided not to.
The next morning, he had Claude build a YouTube with a good video. That day, the website was online.
Karpathy named the site HackerTube.
6454 videos, 150 themes forming the Karpathy's own viewer portal.
The video is still from YouTube, and the recommended thread from Hacker News (HN), a technical forum where programmers and entrepreneurs co-exist.
Karpathy's idea is simple: to save a video of HN friend Anne's.
He gave it almost all to AI.
Claude downloaded 43 million comments from HN for almost 20 years and sifted the garbage, leaving only those that recommended YouTube videos, nearly 40,000.
Finally, 6454 videos were compiled, with 150 themes.
The title of the video, the duration, and then it's coming from YouTube. Then you put on a YouTube-style interface and deploy online.
The whole project, about two hours.
From sifting data to getting online, Karpathy said he didn't read or write the code.
He also made the data publicly available.
Anyone can give it to their own AI. and change it to their own version.
Saturday morning, he gave Claude an order:
Do our own YouTube.
And then, basically Claude's working.
Opus 5.5 is responsible for integrating and sifting the reviews to a lighter Sonnet mass.
The garbage content is washed, and the rest is recommended for video comment.
We have to go online when the site is finished.
Karpathy set up the server and threw the IP address to Claude.
Claude's breath is complete.
And the co-founder of Open AI, who only worked in Claude's space to open the server and buy domain names for a few errands.
He's flirting. It's kind of like team work.
He joined Anthropic's pre-training team last May.
This time, it's my own Claude.
The worst part is Karpathy's last sentence:
He doesn't do anything traditional anymore.
Do not read or write codes, do not knock at the terminal, only talk to Claude.
One of the online users, who spoke directly to the comment section, said that this was the whole point.
There's only one way to do this: to raise the demand.
In 2021, Karpathy (left) spoke on the spot at Teslaai Day, right to Mask.
HackerTube's interface looks like YouTube at first glance, but it's used backwards.
It doesn't have to log in. It sees what's in your own browser.
There are no multiple-tested, specially clicked covers.
The cover of each video is replaced by a frame in the video section. There's a "Longs" entrance on the sidebar, with a long video.
It's a name for Shorts on YouTube.
Karpathy sent the website to HN, and the post ran over 2,000 points. The comment section followed over 500.
Netizens compared it to YouTube's front page.
It was stated that the first-page recommendation was better than the YouTube individualized recommendation.
The rabbit hole, read by Bach, says you can't get it on YouTube.
It was also said that it cut from YouTube, as if it had walked into a forest.
One of the netizens said that his 87-year-old father, Youtube, was trapped inside because he was pushed into the negatives of World War II.
Even geriatricians persuade the elderly to stop watching such videos.
The netizens believe that Hacker Tube is close to what his father needs.
If you look at anything more, the algorithm pushes something more. Many are trapped in the information cocoon like the old man above.
There's a web player spitting, and the algorithm cares how you keep looking at it; and some say the platform doesn't care how you feel when you're done.
Even paying users hold fire.
One of the netizens said that he had been a senior YouTube member for several years, but the more he hated it.
offline is often in trouble and cannot read its content on the plane.
The Internet user lamented that he wanted to be a YouTube client and was not sure the platform would allow it.
The two-hour site, which is precisely marked with a large number of grievances over algorithms.
And there are netizens who look further.
According to the netizens, it was difficult in the past to recommend and index.
Now, we are likely to build our own referral system at our own taste.
In his view, the platform is a bunch of eye-sellers who just want you to keep looking, and the content thus pushes to the extreme.
Now is the opportunity to get rid of them.
There's less work to be done.
The HN asked how much token was burned and how much it cost.
Karpathy told Claude to write down his own account.
About 504 million tokens were entered, of which about 90 per cent were hit by cache.
The output is about 7.6 million tokens.
More than 90% spent on Sonnet labeling comments.
It's a real site and it's online, and Opus spent only about 1 million tokens.
The $500 million was transferred over and over again. The model didn't read 500 million token in one breath.
Karpathy disclosed that over 90 per cent of model volumes were used to comment on labels, and that approximately 1 million Opus token was spent on applications development and deployment.
In two hours, the most consuming segment of the token is labeling comments. This post is part of our special coverage Syria Protests 2011.
The account reminds those who want to be a data tool: cleaning, labelling, and possibly more token than writing codes.
It may be useful to screen a small batch of information before expanding it.
Karpathy made the data analysed in the comments publicly available.
And his advice is, give it to you, AI, make your own version.
It also contains a note to AI explaining how these documents work.
YouTube video messages, not directly in the data pack, go and get them on YouTube. The note says how to take it, and usually apply for an API for YouTube like Karpathy.
The data and instructions are available so that they can be "privately tailored" to personal preferences.
For example, it's only about programming, so that AI can keep only programming lessons. Then it's sorted by theme, and each video is accompanied by a tick from whom it was originally recommended.
Once it's open, it's a private schedule. If you want to learn anything, you can learn wherever you want.
The same idea can be applied to the information in your hand, such as the collection of ash-eating articles, and the list of tools that have been saved for many years.
Look at the video. The platform used to decide.
Now, a person who does not want to do so can create a personal video station like Karpathy.
There were only two ways to get a hand-in-hand tool: to change the platform or to write the code.
Now there's another one: give AI a clear statement of needs and leave it to do it.
Some netizens in the comment area said that the threshold for software was getting lower. People can fly out of the substitute and are no longer tied to the software giant.
Those who were rounded up with algorithms, they started making them themselves.
Of course, Karpathy himself is the top AI researcher. It doesn't have to be done in two hours.
But this door, it's open.
Karpathy has set only one standard: to watch the video recommended by HN users.
The rest is in Claude's hands.
Your turn, let's figure out three things: where the data came from, what criteria to filter, what to get.
AI can write code for people and can't set standards for people.
References:
Editor: Won-woo
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.