lonietzzz

Gamer & Developer | Carnegie Mellon University | AI Founder at Bay Area

[lonietzzz @lonietzzz]

https://t.co/lRMgVObgM4

[lonietzzz @lonietzzz]

Hit #1 on Pokemon Showdown Ladder with my own self-play RL VGC Reg M-C bot with 106.8M games trained.

This project took me about a week after I became the #1 on the ladder entirely by myself. Then I built the bot that completely beat myself both in Elo and GXE.

[lonietzzz @lonietzzz]

I see the community comments on the bots, especially after 2 bots reaching high ladder on Smogon forum. I have a really complex feeling right now.

[lonietzzz @lonietzzz]

I need two macs one at work one working remotely at home

[lonietzzz @lonietzzz]

Just reached #1 on Pokemon Showdown Ladder with my own self-play RL VGC Reg M-C bot after inspired by @shaokthx ...

Just kidding, after losing a straight 0-3 to mikumiku37 in BO3, I decided to climb the ladder and went from 1500 to 1872 the next day🥲

[lonietzzz @lonietzzz]

Reached Top 3 on Pokemon Champion Ladder with @lovejapanfrombr 's sand team, huge fan!

[lonietzzz @lonietzzz]

原来到最后只有豆包是好人

[Anthropic @AnthropicAI]

We're publishing our most detailed threat intelligence report to date.

It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.

We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.

These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.

We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.

Read the report: https://t.co/0EJUnYEgfz

[lonietzzz @lonietzzz]

别人：Claude又解决了几十年未决的数学问题，瘫坐眩晕仿佛看到原子弹爆炸

我：我草Claude chat里面适配了做菜小工具

[lonietzzz @lonietzzz]

每当AI解决一个数学问题，这个问题就会变成“低垂的果实”
黎曼猜想进展有感

[lonietzzz @lonietzzz]

Everytime AI solves a Math problem, it soon becomes the "low-hanging fruit".
Something I made for the Riemann Hypothesis.

[lonietzzz @lonietzzz]

@MaxForAI 未来会不会有公司为了证明自己实力，像现在通过越狱攻击一样策划一次生物病毒袭击

[lonietzzz @lonietzzz]

感觉Claude在文字里加水印这个会明显影响模型性能，并合理化豆包包Fable等中转站生意（

https://t.co/Nq9OQy3MIO

以下为Claude解释：
Anthropic 目前没有公布具体方案（文章里说检测机制的技术文档还在准备中），但这类文本水印在学术界已经比较成熟，做法基本是下面这套。

不是零宽字符（U+200B 之类）、不是特殊标点、不是 Unicode 同形字。那些一 strip 就没了，也很容易被发现和批量清除。也不是文件元数据——元数据那条走的是另一条路径，即对 .svg/.png/.jpg 这类文件附加符合 C2PA 标准的签名溯源信息。

实际机制：偏置采样（biased sampling）
模型每生成一个 token，输出的是整个词表上的概率分布。关键在于，很多位置上有若干个几乎等价的选择——"研究表明 / 研究显示 / 研究指出"，概率差不多，选哪个都不影响质量。水印就利用这个自由度：

1. 用一个密钥 + 前面若干个 token 做哈希，得到一个伪随机数
2. 用它把词表划成"绿名单/红名单"（Kirchenbauer 等人的方案），或者在候选 token 间做多轮"锦标赛"打分（Google SynthID-Text 的方案）
3. 采样时对绿名单 token 加一个小的 logit 偏置

单个 token 看不出任何异常，人类写作时也会随机命中绿名单约一半。但连续几百个 token 累积下来，绿名单命中率会显著偏高。

检测：
拿到一段文本，用同一个密钥重跑哈希，逐 token 判定绿/红，统计命中率算 z-score。人类文本 z≈0，水印文本可能 z>4，对应的误报率极低。注意检测不需要原模型，只需要密钥和哈希函数——这也是为什么第三方检测工具是可行的。

为什么能扛住复制粘贴：
因为承载信息的是词序列本身。Ctrl+C / Ctrl+V 保留了每一个 token，水印信号原封不动。文章说水印会在文本被复制粘贴到别处时一起转移，并且可能在一定程度的编辑后仍然存在。

为什么会失效：
信号强度大致随文本长度累积，所以：文本太短就没有足够的统计量；被大幅改写、意译或翻译后信号会被打散；混进大量其他内容也会稀释。同义词替换 20-30% 通常就能把 z-score 压到阈值以下——这是这类方案公认的软肋。

质量影响：
官方说法是不改变回复的含义、质量或可读性。理论上 SynthID 那一类属于 "distortion-free" 水印：在期望意义上采样分布不变，只是把随机性从真随机换成了密钥控制的伪随机。实践中在低熵位置（比如代码、数学推导、事实性回答，下一个 token 几乎唯一确定）水印几乎无法嵌入，这既是保护质量的机制，也意味着这类内容基本检测不出来。

顺带一提，检测到标记只能说明内容"可能经 Claude 处理过"，而不能确认作者归属——因为人们也会用 Claude 做校对、翻译、摘要，原始想法可能来自别处。

[lonietzzz @lonietzzz]

真的假的A/，我的怎么今天才弹出来

[lonietzzz @lonietzzz]

Vibe Coding中学到的最多的是尽量不要重复造轮子，自己开发已有的基础设施

[lonietzzz @lonietzzz]

@oran_ge 期待比chatcut好

[lonietzzz @lonietzzz]

@MacmillanR72903 I am using regular core, Staraptor, Charizard, Glimmora, Whimsicott, Basculegion, Garchomp

[lonietzzz @lonietzzz]

Hit #2 with it, the top is too high gotta wait for the decay😅 #pokemon #VGC #PokemonChampions

[lonietzzz @lonietzzz]

Last week I hit #3 on the Pokémon Showdown Champions ladder — using a scouting tool I built myself 🔍

Champscope watches the Reg M-B ladder top 50 every day and files every public replay into a searchable team database:

🔎 Scout any player — merged sets, lead pairs, bring-4 patterns, Mega usage
📋 Browse 1,100+ teams from ladder, VGCPastes, and official tournament sheets
⚡ Match: drop in any 6 mons — or just a partial core — and see how the archetype lives in the meta: how popular it is, who's running it, and the exact sets and EV spreads they use

try it out:
https://t.co/Wnsf4mbbHk

Built on open data from Pokemon Showdown, @VGCPastes, @MetaVGCOfficial, with tournament sheets from @VGCVictoryRoad's Champions Arena. Huge thanks!

#PokemonChampions #VGC