jude c

[jude c @CJude75045]

@akokoi1 我也觉得， 之前还说外国人头像比较真实

[jude c @CJude75045]

@Da7_Tech So you dislike him too, that’s a relief. I always thought everyone on Twitter liked him, and that it was just me, especially since Tibo interacts with him all the time.

[jude c @CJude75045]

@kasong2048 前几天我看matt pocock说的 深模块、浅模块、接缝的逻辑。 在接缝处写测试代码。 和你这套差不多

[jude c @CJude75045]

你们今天有没有遇到额度乱跳的问题？ 搞得我看不懂 到底用了多少。。。😰

[jude c @CJude75045]

@mate_mattt @mattpocockuk Could you please help clarify some doubts?

[jude c @CJude75045]

@YuLin807 我之前遇到这问题， 是电脑上clash开个局域网连接。 登录一下就行了

[jude c @CJude75045]

@vladdubchak_x bot

[jude c @CJude75045]

@Da7_Tech 你好你好～

[jude c @CJude75045]

@supezen 现在的路线肯定是人+AI， 人放前面。

[jude c @CJude75045]

@jal_eelll grok cli

[Da7em @Da7_Tech]

This post may cause me problems and probably make some people here resent me, but I’m going to say it anyway because people should understand what is happening.

There are paid creators whose content consistently attacks Chinese open models, and the same three claims keep coming up: they are distilled from American models, your data is unsafe with Chinese companies, and they are benchmaxed.

On distillation, there may be some truth in specific cases, but the argument is usually stretched far beyond what the evidence supports.

American models were themselves trained on copyrighted human work, including books and culturally valuable material, without permission. This has already been documented with two major companies. So the knowledge inside those models cannot suddenly be treated as something those companies exclusively own.

The accusations also become ridiculous sometimes. People claimed Kimi K3 was distilled from Fable when Fable had been public for roughly two days. Are we seriously supposed to believe a model of that scale and capability was basically created by distilling another model for two days?

And reducing Chinese progress to distillation dismisses the work of researchers and engineers who have contributed heavily to this industry.

Then there is the privacy argument.

American companies store user data too, and they benefit from it. Use ChatGPT long enough and it builds memory about you. Cancel, come back months later, and that information will still be there. Your conversations also follow you between devices because the data is stored in the cloud.

So I do not understand why storing and using data becomes a unique crime when the company is Chinese.

If you are terrified of your data being held in China but completely comfortable with the same thing happening in America, that is not a privacy principle. It is a double standard.

You cannot praise memory, personalization and cloud features from one company, then use the same underlying behavior as evidence that another company is dangerous.

And then we get to my favorite accusation: Chinese models are “benchmaxed.”

The argument is that they are optimized to score well on benchmarks and therefore their numbers do not represent real-world performance.

American labs optimize heavily for benchmarks too. Every major lab knows the evaluations people watch. They train, tune and compare models around them. Benchmarks influence what gets shipped and how models are marketed.

So if a Chinese model gets a high score and your response is “benchmaxed,” you cannot turn around and use an American model’s benchmark score as proof that it is smarter.

This is exactly why I care far more about real use than leaderboards.

You can have a model like Opus 5 sitting on the top of major benchmarks, sometimes above Fable, Sol and Kimi K3, while in my own daily work it hallucinates more and performs worse than all of them.

If Chinese models are benchmaxed, American models are benchmaxed too.

That is why these three arguments do not convince me. Too often they are used selectively, and behind a lot of this rhetoric I see fear of Chinese competition.

China has cheaper energy, cheaper labor and extremely talented researchers. It has managed to remain competitive despite restrictions and limited access to the most advanced chips.

American companies, meanwhile, increasingly treat that competition as something existential rather than normal competition.

Use whatever works for you. Test the models yourself. Learn where each one is strong and where it fails.

Do not let anyone else do your thinking for you.

[jude c @CJude75045]

@hqmank SuperGrok Heavy: $99/month

[jude c @CJude75045]

@FanShifu @PMbackttfuture 印度区的已经开始期待 5.0 了

[jude c @CJude75045]

@__oQuery claude前几天 一个劲给“实体”吐成“尸体”。。。

[OneHopeA9 @onehopeA9]

兄弟们，草率了！！有拼车的吗
100刀grok heavy 开通技巧
先用
https://t.co/K847bHpqu0
买一个月 supergrok 一个月 105R
然后马上访问
https://t.co/LPOBIwQBHa
就有三个月的 99刀 heavy 了，还可以抵扣105R
然后切换到年付，可以买一年

撸完后发现太多了

@grok 介绍下现在这活动，还送一年cursor吗

[jude c @CJude75045]

@onehopeA9 @grok 牛逼的，亲测可行，国内visa信用卡， 或者万事达借记卡都能付。 步骤要注意， 3个流程不能错，先买1个月的优惠的，再买3个月的优惠的，然后去切换订阅，切成年付就可以了。 500多一个月要啥自行车

[jude c @CJude75045]

@thinkandtrade @GrokInsider no , Grok sent an email for me yesterday. But I found the URL inside is invalid.

[jude c @CJude75045]

@Original_lighti 不可以了， 我昨天点链接试了， 发现不行，虽然收到了邮件，但是没有优惠了

[jude c @CJude75045]

grok突然发了邮件，给了3个月的heavy的优惠。。。 突然就纠结了，要不要买😰

[jude c @CJude75045]

@Da7_Tech Just like Grok 4.5 before. Back then, GLM‑5.2, GPT‑5.5 and DeepSeek‑V4 all spent a long time troubleshooting one single bug, while Grok finished it in two minutes.