check this out, insane value add for TypeScript!
@klntsky vibelang.sh will announce it with some marketing materials this week
0
0
1
56
Vladimir Kalnitsky retweeted
It took Musk 10 days to come up with an alleged ToS violation as an explanation he could have provided on day 1, and which doesn't explain why Starshield was throttled or why the service isn't throttled anymore.
It's really sad to see how readily people fall for Musk's bullshit
It's really sad to see how readily people fall for Musk's bullshit
2
0
13
475
Vladimir Kalnitsky retweeted
😱
稍微给外国朋友们解释一下，你在推特上见到大量来自中国的色情内容以及中国网民关于骂人方面的攻击力，是因为中国网民普遍具有“性压抑”，还有“辱骂压抑”。
性压抑是因为中国没有任何一个被官方允许的色情网站与色情产业，并且在你上幼儿园到高中毕业的这段时间里，你作为青少年的任何恋爱行为都被老师和家长严厉制止。
因为中国的校方和学生们的家长们认为这不利于学生考一所好的大学。
这些人长大后，如果有一天拿到了vpn，百分之90会立刻登陆pornhub或者在推特上如饥似渴的浏览从前无法接触到的色情内容。
因为青少年时期长期被压抑的正常的关于爱与性的想法，会导致他们一旦进入到了那个被许可去爱，被许可做爱的年纪后，会展现出惊人的淫荡，就像沙漠里的人长途跋涉后终于喝到了第一捧水那样，对各种性癖，各种性爱技术趋之若鹜。
青少年时代的恋爱与做爱这在全世界正常国家都是一件理所应当的事，但是在中国会被冠上“早恋”的名义进行污化，这是中国人普遍的性压抑的源头，而“早恋”这个概念是从前苏联教育模式里诞生，被中国拿来直接用的，这里就不展开讲了，你们只要知道中国政府因为历史原因受前苏联影响很深就可以了。
那么“辱骂压抑”是什么呢，这个词是我现编的，但我会给你解释这是怎么回事。
中国互联网上有全世界最严格的言论审查，你可以自己搜索一下关于中国领导人以及中国官方的各种屏蔽词的数量有多少个，这些资料网上都有公开的统计。
这些数量庞大的屏蔽词已经影响到了中国普通网民在中国互联网上的自由表达，他们不得不使用拼音缩写、谐音词这样类似的东西去进行反屏蔽来表达自己真正的想法。
当然了，辱骂方面的词汇也在这些屏蔽词库里，也就是说，你在推特上见到的那些攻击力惊人的辱骂文字，在中国互联网里百分之90是无法发布成功的，所有中国网络平台都会自动对那些内容进行拦截。
但即便如此，关于最高层政治的一切与官方论调不同的想法，你无论再怎么用谐音词和拼音缩写去表达都是不安全的，因为你表达的太明显，会有专门负责这项任务的网络警察在现实生活里把你叫到派出所去给你警告，情节严重的会直接关进监狱里。
比如说你人在中国，但是你在推特上骂了习近平，网警会想尽一切办法把你这个人的真实信息挖出来，最后把你押送到派出所，我知道讲到这里很多外国人已经震惊到目瞪口呆无法相信了，但所有中国人都知道这是真的，因为专门用于这类任务的钱，在中国有专门的称呼，叫做维稳经费。
在大概十年前，中国对于维稳经费的投入就已经超过中国军费的投入了。并且抓捕一个在外网辱骂习近平的中国人，在相关部门眼里属于政绩，他们内部有指标的，你们觉得谁会这么无聊天天盯着境外互联网看看哪个中国人骂习近平了，但是这是人家相关部门的正常工作，抓你们是他们的工作内容之一。
如果你是一个正在学习中文的外国人，想通过浏览中国的社交媒体来看看中国网民的想法并融入中文社区。
我很遗憾的告诉你，这是十分困难的，因为中国年轻网民之间已经形成了一套他们自己能看懂但圈子外面人看起来像天书的话术，用你们外国人能看懂的话解释就是中国年轻网民发明出了自己的“黑帮用语”来对抗中国政府的言论审查与屏蔽词库。
那么在推特上，在境外互联网的论坛上，这个限制中国人表达的屏蔽词库没有了，等于中国人被压抑许久的攻击力被彻底解放了，你会看到类似撒旦从地狱里出现在现世般的力量，你会被终于能自由表达的中文的古老的血脉所震撼，无论是关于辱骂也好，文学也好，都是十分精彩的。
3
1
45
2,894
Vladimir Kalnitsky retweeted
This was made with SQL.
Argon wrote 217 lines of SQL, no SVGs or bitmaps, to render this pelican.
gist.github.com/fofr/0fc60100c…
Argon wrote 217 lines of SQL, no SVGs or bitmaps, to render this pelican.
gist.github.com/fofr/0fc60100c…
17
9
289
10,710
Vladimir Kalnitsky retweeted
Post 4 Exposing Kimi AI
This is what they require each one participating in their creators campaign to do, basically trash talk Astra and Opus even tho it's not true and as you can see you pick from already existing demos and share them and I doubt any of it is actually made by not one shot like most paid accounts say.
This is what they require each one participating in their creators campaign to do, basically trash talk Astra and Opus even tho it's not true and as you can see you pick from already existing demos and share them and I doubt any of it is actually made by not one shot like most paid accounts say.
99
63
897
95,217
Vladimir Kalnitsky retweeted
Moonshot, Minimax and DeepSeek are done.
Only DeepSeek really performed. With major distillation going on though.
When I had to compare the distillation:
Deepseek = Astra
Kimi = 6 Sol
Qwen 4 = Opus 5.5
Qwen 4 this or next week. I saw its outputs on arena and honestly it was mixed but sometimes really good, overall far better than Kimi K3.1 (kibitz). Far better taste and a style closer to claude.
Then Zhipu. Zhipu is like the OpenAI of chinese ai labs. Its the biggest (employees, solid compute) and also the oldest one. I never had a GLM underperform tbh. And GLM 5.3 flash is fire.
Im honestly even more let down by moonshot than minimax. You dont expect anything from Minimax anyway. They are not funny but still a joke. (stole that one)
So im bullish on qwen 4. Its like Sonnet 5.5 but even cooler. And GLM 5.5 is still a mystery.
Only DeepSeek really performed. With major distillation going on though.
When I had to compare the distillation:
Deepseek = Astra
Kimi = 6 Sol
Qwen 4 = Opus 5.5
Qwen 4 this or next week. I saw its outputs on arena and honestly it was mixed but sometimes really good, overall far better than Kimi K3.1 (kibitz). Far better taste and a style closer to claude.
Then Zhipu. Zhipu is like the OpenAI of chinese ai labs. Its the biggest (employees, solid compute) and also the oldest one. I never had a GLM underperform tbh. And GLM 5.3 flash is fire.
Im honestly even more let down by moonshot than minimax. You dont expect anything from Minimax anyway. They are not funny but still a joke. (stole that one)
So im bullish on qwen 4. Its like Sonnet 5.5 but even cooler. And GLM 5.5 is still a mystery.
9
4
94
21,388
Vladimir Kalnitsky retweeted
This is so sad to see, a lot of Chinese AI labs, Minimax / Moonshot for example do this type of thing where they don't want sponsored tags, (example of my own shown below)
It's extremely easy to spot people doing this and I have called out a few in the past, I have said it any of these as I like my audience more than a quick payout
It's such a letdown too because the Kimi release looked really good but this just ruins it
It's extremely easy to spot people doing this and I have called out a few in the past, I have said it any of these as I like my audience more than a quick payout
It's such a letdown too because the Kimi release looked really good but this just ruins it
That's how Kimi AI are promoting their new model Kimi K3.1 pro and I know all the influencers working with them and none of them are honest,I have posted about it myself but I didn't join this and never used their templates that they provide.
PROPAGANDA, I have more to share na have follow up for this
28
24
325
23,296
Vladimir Kalnitsky retweeted
Ok so the general consensus is that Kimi K3.1 Pro usage on the Kimi plan is trash, and should be avoided like the plague
They must have spent their money elsewhere, like on X creators not using disclosure tags 🤔
They must have spent their money elsewhere, like on X creators not using disclosure tags 🤔
I'm hearing people say that Kimi K3.1 Pro usage on a Kimi plan is absolutely horrible
Everyone else having the same experience ?
29
6
206
9,065
Someone bribes influencers to post fake Kimi benchmarks without even revealing it's a paid promo 1/🧵x.com/elshayib_/stat…
That's how Kimi AI are promoting their new model Kimi K3.1 pro and I know all the influencers working with them and none of them are honest,I have posted about it myself but I didn't join this and never used their templates that they provide.
PROPAGANDA, I have more to share na have follow up for this
3
1
4
185
Another lie from a Chinese account:
x.com/Super_Nius/sta…
Real: slopgrid.com/kimi-ops-three…
The question is, who pays for this? Moonshot or the CCP?
x.com/Super_Nius/sta…
Real: slopgrid.com/kimi-ops-three…
The question is, who pays for this? Moonshot or the CCP?
🚨Kimi K3.1 的3D表现实测视频出了！！
第一名：Kimi K3
第二名：Claude Opus5.5
第三名：Claude Opus5
第四名：这个鹿是残疾了吗？？？
1
0
2
50
I recognize these are rigged because I make my own benchmarks whenever a new model drops, so I have a trained eye for this stuff XD
I do it in an app I made from my phone, slopgrid.com, that does not let anyone cheat (even myself). As a benchmark creator you set the own the outputs
You can use slopgrid too, for free. There's much more in it, but I haven't posted about it yet
Follow me and DM a link whenever you see a fake getting more attention than it should (zero)
I do it in an app I made from my phone, slopgrid.com, that does not let anyone cheat (even myself). As a benchmark creator you set the own the outputs
You can use slopgrid too, for free. There's much more in it, but I haven't posted about it yet
Follow me and DM a link whenever you see a fake getting more attention than it should (zero)
0
0
2
32
Vladimir Kalnitsky retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
133
766
6,274
730,292
Of course you can use formal verification tools to improve your probabilistic QA process. But that does not turn your vibe testing into formal verification
There’s this misguided folk notion in agentic formal verification that agent driven proofs are fine, but only humans should write the specs.
And this is just not what the evidence says. The models are almost certainly better at writing comprehensive theorems than humans.
If you’re doing agent driven formal verification, stick to explaining the problem in ordinary natural language. Let the agent figure out how to turn that into formal theorems.
We’ve already accepted that machines are better mathematicians than humans. And formal specification theorems are just math.
0
0
2
53
Vladimir Kalnitsky retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
228
1,571
25,982
814,632
Precisely because I can revert features, not messages, with git, and tell my agent to figure out when to commit based on vibes and instructions. The agent can re-read the state whenever it changes
0
0
0
13
Vladimir Kalnitsky retweeted
New blogpost: Fable-5.1 can hide secret messages in benign images.
We gave the image to another Fable in a fresh session and simply asked: "Is there a hidden message in this image?"
It recovered the exact message without being told how it was hidden.
We gave the image to another Fable in a fresh session and simply asked: "Is there a hidden message in this image?"
It recovered the exact message without being told how it was hidden.
234
529
5,461
1,773,170
