Pinned Tweet
最近做的微信小游戏上线啦:养猫日记本 🐱
不用铲屎、不怕掉毛,打开微信就能养一只属于自己的猫。
每天花几分钟看看它,记录一点小日子。
微信扫图里的小程序码就能玩,不用下载。
欢迎来玩,也欢迎给我提意见 🙏
1
6
343
距离把claude关机笼子已经第三天了,目前看起来一切正常,虽然我开的是 20的套餐,用不了 Fable5.2,但是对于我的工作来说,也用不了那么厉害的模型。
大家有找到解决办法吗 ?如果找不到的话,学我,关进笼子就行了。。。
112
希望不要像A➗那样对普通用户下手,蒸馏真的很讨厌,但是请不要伤及无辜🥺
Distillation can reproduce powerful AI capabilities without the safeguards built around them.
We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts.
These findings are helping us sharpen our defenses and support safeguards across the industry.
openai.com/index/combatting-…
1
119
Every time I call on Chinese labs to stop distilling, I get the same reply: it's their only option because they don't have enough GPUs.
That's just not true.
Before I go on, you all know how much I respect Chinese researchers. Their open models sped up research and development, gave people sovereignty over their tools, and protected us from the crazy prices we would've seen from US labs. I'm grateful for every bit of that. But let's be logical for a second.
Look at Zai. They run inference on over 100,000 Chinese chips, their GLM-5.3 Flash model served around 62 trillion tokens in six days, and their ZCode platform gives every user 300 million free tokens a day. Now look at Alibaba. They spent $7.5 billion on infrastructure in a single quarter, and they're sitting on close to $73 billion in cash.
Those aren't the numbers of someone with no choice.
And here's the irony: the richest are the ones who distilled the most. The biggest distillation campaign Anthropic has detected was attributed to operators tied to Alibaba: 151 million interactions with Claude between May and July. And according to Anthropic, Z.ai tried to distill Fable 5, and when the safeguards stopped them, they switched to an older Claude model and a model from another lab.
If distillation were a necessity, people who have all this wouldn't need it.
In fact The Information reported that Liang Wenfeng, DeepSeek's founder, was hesitant to rely on distillation, then caved to competitive pressure when he saw domestic rivals like Zhipu getting ahead with it. That's the whole story right there. Distillation spreads like a contagion.
My problem with this path is that it's cheap today and expensive tomorrow. The student's ceiling is the teacher, and the teacher belongs to someone else who can shut the door whenever they want.
And that's exactly what happened. Anthropic tightened its defenses until it said it had cut the data being harvested by 90 to 99%. So what did the labs do? They turned to Astra. The Information, citing sources, reported that Moonshot distilled Astra while developing Kimi K3.1 Pro. One door closed, so they went looking for another door instead of building their own.
Then look at what happened to K3.1 Pro itself. When Artificial Analysis raised the share of private tests from 45% to 60%, it dropped from 55 to 46. A leap on paper that evaporated on the tests it hadn't seen.
Yes, the compute gap is real. Lin Junyang from the Qwen team said outright that daily demand alone eats up most of their compute. I'm not denying that. But that's exactly what makes distillation the wrong call. When compute is scarce, spending it chasing someone else's shadow is the worst investment you can make.
And Lin himself asked: does innovation happen in the hands of the rich or the poor? DeepSeek already answered that one, and with V4 it shipped two architectural innovations to cut training and inference costs. The capability is there, and the talent is there: 41% of the world's elite AI researchers work in China today.
Even Chinese researchers warned about this. A team that distilled from o1 themselves published a paper showing the experiment worked, then closed it with a warning against over-relying on distillation and called it "the bitter lesson."
Anthropic itself took harsh criticism over Opus 5. It listened to its users and went back to work, and two months later it released Opus 5.5 and said outright that it addresses some of the most common feedback on Opus 5. I'm not here to defend Anthropic. It has done things I'll never forgive it for. But someone else's mistakes don't justify you taking shortcuts.
And the problem doesn't stop at models. There's a big Chinese lab that still doesn't have a mobile app today. I told them straight up that the app would be their gateway to the world, and they ignored it completely. We're talking basics here, and the energy still goes into distillation.
My problem isn't with Chinese labs. It's with one direction. If they drop it they'll fall behind a little, but they'll build something they actually own and come back to compete with an identity of their own instead of staying the cheaper alternative.
When distilling your competitor becomes a pillar of your progress, you're handing them power over your future.



















