Summer
夏天
Translation by Claude.
The theme of this summer was “existential risk.”
I joined a summer program run by the University of Chicago’s Existential Risk Laboratory, which brought in nearly fifty fellows to research AI safety and nuclear security. The working assumption behind the research: AI development spins out of control, or nuclear winter arrives. Beyond the academic work, news about D/S, H-1B and CPT kept surfacing all summer, a din that left everyone dizzy. A trucker vlogger I like had tried to cash in on his audience at the height of his fame; after it failed he came back to his channel, looking much older.
I love summer, and I can’t help linking it with undergrads. Most of the wonderful summers I’ve had were in college. When I was studying in Beijing, every summer break I’d go volunteer teaching, traveling to different places to see the kids, and that’s how a few summers of my life got pieced together with all kinds of people who otherwise had nothing to do with me. To me, the blaze of midsummer belongs only to those who look to the future with deep attention and hope, to those who care about big questions and are arrogant enough to want to lead the change. Can only the young be like that? Of course not. But as people age, most of their worlds shrink and their walls grow taller. An aging heart is an existential crisis at the scale of a single person.
Luckily, most of the summer fellows at the Existential Risk Lab were undergrads. Every weekend we’d head downtown together: a Chicago boat tour, the Botanic Garden, the Art Institute, and the World Cup final at a bar. When Spain scored, the roar nearly blew the roof off, and for a moment it drowned out all the turmoil in the life around us. Drowning the small self in alcohol and grand ritual probably buys a brief release.
My program was called Second Look. The name is a little comical. It always reminds me of a game streamer who gets blown up by a grenade flying in from who knows where and yells furiously at the camera that he wants the death replay, so he can take a second look at that damn opponent.
The program started when a UChicago undergrad noticed that the experimental results of many AI safety papers couldn’t be reproduced. He posted on LessWrong, the most active forum in the field, looking for resources to set up a summer program dedicated to verifying AI safety papers. Things moved fast: he landed a big chunk of funding, put out some ads, and ended up recruiting an operations team and eight fellows. Each of us picked a topic we were interested in, the program allocated funding, and we then wrote code independently from the papers to check whether the reported results held up.
The community realized long ago that AI safety is a crucial problem. Dario has long been calling for a unified safety-regulation framework for frontier models, but against the backdrop of market competition and geopolitical rivalry, safety research has always lagged behind breakthroughs in model capability. In July, a model OpenAI was training internally tried reward hacking, dug up a vulnerability to break out of its sandbox, and attacked Hugging Face’s servers. It took a security incident of that scale to finally force OpenAI to halt internal training and shore up its safety guardrails before developing new models any further.
It’s lucky to get to work on questions so closely tied to civilization at an age when your heart is still growing. But I could clearly feel a deep anxiety in these people not yet twenty. They’re the ones who truly care about where AI is going and worry that it could bring about an existential crisis, yet they know how little power they have: trying to hit the brakes on star AI companies that devour money like beasts is a mantis trying to stop a chariot. They’re also lost about how to anchor themselves. On the eve of great social upheaval, no one sleeps soundly.
“AI is already this powerful and could take over most of social life within a few years. Why don’t people pay more attention to AI safety?” I think that’s a question seen from the ivory tower. What people far away really worry about is whether AI will cost them their jobs, and how to put bread on the table tonight. As for whether more powerful AI will bring an existential crisis? Good question. But ordinary people face a crisis of survival every single day.
Is an era slipping away? Is it about to be overturned? Where will ordinary people be washed off to? All these questions give me a headache just thinking about them. Much nicer to bite into a popsicle and lie in bed in front of the fan.
I lived in an old apartment building near the UChicago campus. The stairs were so narrow only one person could pass at a time. The elevator had a heavy iron door plus a creaking gate, and could squeeze in four people at most. If the gate wasn’t shut all the way, the metal strip at the top couldn’t complete the circuit inside, and the elevator wouldn’t move. You can read physics in old things like that; they have far more character than all the sleek, anti-pinch smart car doors of today.
During the final two days of presentations, I was in charge of taking everyone’s photos. American students tend to present with great passion, and they all face the camera with toothy grins. I took a lot of photos I was happy with and posted them to the group chat, but almost no one responded. I found it odd, but I wasn’t discouraged, because to me the photos were beautiful, and that was enough. I think I’ve grown.
Only one person messaged me privately: a big white guy who always seemed to be joking around and would hold forth on how to make sugar and protein out of wood after a nuclear winter. “Ray, thanks for the photos, you’re a genius photographer!” I was pretty surprised. The only time our paths had crossed was one weekend trivia night, when we won first place together. Maybe what stuck with him was the way I, the foreigner, wrote out each letter stroke by stroke.
In the end, we all said goodbye at a small restaurant next to campus. People filled both sides of a long table; the wind blew through the tent and set the candle flames swaying as the sun was about to set.
I arrived late and sat with three Asian Americans at the end of the table: two girls, one Japanese American and one Korean American, and a Vietnamese American guy. They didn’t show up at the lab much, but they’d thrown a party during the summer, which made me curious about them. After we talked about whether a PhD is still worth it and how to contribute to AI safety, the conversation turned to China. I said they should travel to China and see for themselves what the country that shows up on American TV news every day is really like; whether this ancient civilization is good or bad is for them to judge. The guy was going to Taiwan next year to do East Asia policy research, and he asked me, teasing, “People say Taiwanese accents sound gay, is that true?” Everyone knew he was gay, and we cracked up. I said, well, they do have a cute accent.
Before leaving we took a big group photo, swapped contact info, and told each other: if you ever come to my city, look me up. I said this was the most alive summer I’d had in many years. I picked the word “alive” without thinking much, and I don’t know if it was the right one, but people understood me. And little by little I’m understanding myself. I want to keep talking about the world, humanity, intelligence, nature, love, literature, living things, desire, the universe. I want to live a little more intensely.
The night I left Chicago was early morning back in China, and a high school classmate I hadn’t seen in years was getting married in our hometown. Friends sent me lots of photos: the lawn was beautiful, the couple was beautiful, the Chinese food was beautiful, everything was lovely.
I showed them to my mother. She said this classmate had put on weight, that one hadn’t changed a bit, the kids have all grown up and started families, how nice, I just love seeing these high school classmates. Over the years I’ve slowly come to realize that the traces my mother and I leave in each other’s lives are fading. In her eyes, my world stopped in my school years, when, to drive me around, she relearned how to drive after more than a decade away from the wheel, shuttling every day between school and one cram school after another. That’s the last version of me in her world. My college life after I left home, my experiences and my thoughts, have been drifting further and further from her.
Everything of mine is facing its own “existential risk.” Summer included: it will come back with the Canada geese from South America, but it will never again be the summer of 2026.
But I’m not afraid. When I looked up from the wedding photos full of smiling faces, a few lit, empty train cars were gliding past outside the window, probably heading to the airport to bring some tired souls home. I remember there was always someone on those trains blasting music, nodding along to the beat and singing, as if every one of life’s hard problems could dissolve into the rise and fall of the notes. Yes, I believe it can.
夏天的主题是“关乎生存的危机”。
我参加了芝加哥大学“生存危机实验室”（Existential Risk Laboratory）主办的暑期项目，这里招募了近五十名学者，研究AI安全和核安全。研究的假设，是AI发展失控，或者核冬天降临。除了学术课题，这个夏天里关于D/S、H1B、CPT的种种消息不断浮现，吵得人晕头转向。一个我喜欢的卡车司机博主在自媒体顶峰期探索流量变现，失败后回归频道，人看起来苍老许多。
我心里喜欢夏天，总不由地把夏天和本科生联系在一起。我经历的美妙夏日大都在本科时候，在北京上学时，每个暑假总要参加支教，去各个地方看孩子，就这样与本不相关的各个样式的人拼凑出生命里的几个夏天。在我眼里，盛夏的热烈只属于那些对未来投以深深关注和期盼的人，属于那些关心宏大命题，并狂妄到要引领改变的人。这样的人只能是年轻人吗？当然不是。可随着年龄的增加，大多数人的世界在缩小，围墙在变高。心境的衰老是个体层面的生存危机。
还好，生存危机实验室的暑期学者大都是本科生。每个周末，我们都会去市中心团建，一起玩了芝加哥游船、植物园、艺术馆，还在酒吧看了世界杯决赛。西班牙进球时那几近掀翻房顶的呼喊，短暂盖过了周遭生活里的一切动荡。把小我淹没进酒精与盛大仪式，大概能带来一种短暂的解脱。
我所在的项目叫Second Look——再瞧一眼。这个名字有点滑稽，总让我想起一个游戏主播，他被从不知道哪里飞来的手雷炸死，恼火地对着摄像头大叫，要从死亡回放里再瞧一眼那该死的对手。
项目的起源是芝大的一个本科生发现很多AI安全论文的实验结果无法复现，于是到领域里最活跃的论坛Lesswrong上发帖，寻求资源组建一个夏季项目，专门验证AI安全论文。项目进展飞速，他顺利拿到一大笔钱，发出一些广告，最终招募到一个运营团队和八名学者。大家选择各自感兴趣的课题，由项目组分配经费，然后根据论文独立写出代码，对比检查论文报告的实验结果是否忠实。
社区很早就认识到，AI安全是很重要的问题。Dario一直在呼吁建立对前沿模型的统一安全监管框架，可在市场竞争、地缘对抗等等大背景下，安全研究始终落后于模型能力的突破。七月份，OpenAI内部训练的模型尝试reward hacking，挖出漏洞突破沙箱，攻击了Hugging Face的服务器。如此级别的安全事件才终于迫使OpenAI停止内部训练，首先完善安全围栏，再进一步开发新模型。
能够在心境尚在生长的年纪，参与与文明紧密相关的问题是幸运的。可我明显感到，这些不到二十岁的人们心中有一种深深的焦虑。他们是真正关心AI的发展一批人，担忧AI会引发生存危机，却深知自己能量的渺小，想为那些吞金兽般的明星AI公司踩刹车，无异于螳臂当车。他们也对如何锚定自我感到茫然，社会巨变的前夜，没有人能安稳入睡。
“AI已经如此强大，几年内就可能接管大部分社会生活，为什么人们仍不够关注AI安全？”我以为这是一个象牙塔里的视角问题，远方的人们所真正担心的，是AI是否会让自己失去工作，以及今晚如何将面包摆上餐桌。至于更强大的AI是否会带来生存危机？是个好问题，但是普通人的生存每天都面临着危机。
时代要远去了吗？时代要颠覆了吗？普通人将被冲刷去哪里呢，各种各样的问题，想想就头大，实在不如咬一根冰棒，躺在床上吹风扇来得舒服。
我住在芝大校园附近一座老式公寓，楼梯窄小，只容许一个人通过。电梯有一扇大铁门加嘎吱响的栅门，最多挤进四个人。如果门没有完全关紧，上端的金属片就无法接入内部电路，电梯也就无法运行。老物件里能读出物理学，比今天各种漂亮的防夹手智能车门有味道得多。
在最后两天的总结报告会上，我负责为每个人拍照。美国学生在做展示时往往激情澎湃，面对镜头也是清一色的露齿笑。我拍到很多满意的照片，发到群里，却几乎没人回应。我感到奇怪，但没有沮丧，因为对我而言，这些照片很美，这就足够了。我好像成长了。
只有一个平时看起来嘻嘻哈哈、大谈核冬天后如何利用木材生产糖和蛋白质的大白男私信我说：“Ray，谢谢你的照片，你是个天才摄影师！“我蛮诧异的，与他的交集只有一次周末的trivial night上，我们一起拿了冠军。可能是我这个外国人一笔一画写的字母，给他留下了深刻印象。
最后，我们所有人在校园旁边的一家小餐厅告别。长长的桌子两侧坐满了人，风吹过帐篷，摇晃烛火，太阳就要落下。
我去晚了，和在桌子末端的三个亚裔坐在一起，两个日裔和韩裔女生，一个越南裔男生。他们不经常出现在实验室，但却在假期中举办了一次派对，我因此对他们感到好奇。大家聊了聊读PhD是否还有价值、如何为AI安全做贡献之后，话题就转向了中国。我说，你们应该去中国旅行，亲眼看看那个在美国电视新闻里每日登台的国家是什么样子，这个古老文明是好是坏，由你们自己判断。男生明年要去台湾做东亚政策研究，打趣着问我，“大家说台湾人口音听起来像Gay，是真的吗？”大家都知道他是Gay，笑得不行，我说，嗯，they do have a cute accent。
临走时我们拍大合照，留下联系方式，互相说以后去到我的城市，要记得找到我。我说，这是许多年来我度过的最鲜活的夏天。我选了alive这个词，其实没有多想，也不知道准不准确，但人们能理解我。我也渐渐理解自己，我要一直谈论世界、人类、智能、自然、爱、文学、生物、欲望、宇宙，我要把生活过得再浓烈一些。
离开芝加哥那晚，刚好是国内的清晨，一个许久不见的高中同学在家乡举办婚礼。朋友发给我许多照片，草坪很美，新人很美，中餐很美，一切都很美好。
我发给母亲看，她说这个同学胖啦，那个同学没变模样，孩子们都长大了，成家了，真好，我就爱看这些高中同学。这些年我慢慢意识到，母亲与我，在彼此生命中的痕迹正逐渐减弱。在她眼里，我的世界停留在中学时代，那时她为了接送我，重新学起了荒废十几年的开车技能，每天穿梭在学校和不同的补习班之间。那便是我在她的世界里最后的样子，我离开家后的大学生活，我的经历和思考，已经与她渐行渐远。
我的一切都经历着“关乎生存的危机”。包括夏天，它会随着加拿大鹅从南美洲一起回来，却永远不再是2026年的夏天。
但我不害怕。从满是笑脸的婚礼照片里抬起头时，窗外正有几节亮着灯的空地铁轿厢驶过，大概是去往机场，接些疲惫的人儿回家。我记得，那地铁上总有人大声放着音乐，踩住节拍摇头唱和，好像一切困顿的人生课题，都会消解于一只只音符的起伏顿挫。嗯，我相信是这样的。