Dr.-Ing. Martin Schiele

2008 - 2015 Studium Automobiltechnik
2015 - 2022 Doktorarbeit an der TU Ilmenau zu KI (Reinforcement Learning)
Seit 2021 Gründung AI-UI GmbH, CEO

[Dr.-Ing. Martin Schiele @mememanace]

🔓 Prompt Injection: Warum 100% Sicherheit ein Mythos ist

Man kann jeden Prompt jailbreaken. Es gibt technisch gesehen einfach keine Möglichkeit das absolut sicher zu machen...

Ist bei Menschen übrigens nicht anders: Will mir jemand am Telefon eine Waschmaschine verkaufen und ich quatsche mit ihm am Ende über sein Wochenende, hab ich ihn auch gejailbreakt. 🤷‍♂️
LLMs haben echte Weltmodelle und sind keine simplen Skripte. Wer meint, er kann die Modelle mit ein paar Textregeln komplett abzusichen, hat Deep Learning einfach nicht ganz durchdrungen... 🧠

[Espen JD @Snixtp]

The RTX Pro 6000 has two matmul engines, and vLLM was only using one of them when I ran Qwen3.6 27B NVFP4

Marlin is nice for decoding, but CUTLASS is 3.5x faster processing large token prompts. For some reason, vLLM ran Marlin for both, because NVIDIA's checkpoint files mislabel what the model can actually do.

Fable 5 found out about all this and wrote a vLLM plugin that looks at the batch size and picks the right engine per call.

Results is almost 2x increase in prefill at all context lengths:

[wafer @wafer_ai]

🚨 BREAKING:

these engineers figured out how to serve GLM 5.2 on @AMD MI355X at 2626 tok/s/node and 213 tok/s single stream at over 2x lower cost than Blackwell

that's ~80% of B200 throughput at over 2x lower cost

full write-up in reply to see how

[Kai @hqmank]

✅Found a wild token hack that only works properly on Fable 5.

Claude charges image tokens by pixel size, not by how much text is in the image.

So someone built a proxy that renders your context into PNGs before sending it to the API.

Before: 92,000 tokens for a dense tool result
After: 4,761 tokens for the same content as a PNG

End-to-end bill savings: 59-70%.

SWE-bench tasks: same completion rate, half the cost.

Link in reply👇

[Dr.-Ing. Martin Schiele @mememanace]

@MikeBradleyAI 7-9 Months

[Kanika @KanikaBK]

CHINA JUST LEAKED THE FUTURE OF WEB APPS.

Alibaba open-sourced PageAgent and 99% of SaaS founders are sleeping on this.

It's a JavaScript AI agent that lives INSIDE your webpage. Users control your entire interface with natural language.

↳ No browser extensions needed, screenshots or multi-modal LLMs, headless browser setup, and also no backend rewrite required

Just drop it in your HTML with ONE line of code. What took 20 clicks now takes one sentence.

"Click login, fill in my credentials, submit the form"
Done. This is not a demo, it is production-ready.

↳ Turn any SaaS into an AI copilot in minutes
↳ Smart form filling for ERP, CRM, admin systems
↳ Voice commands and accessibility built in
↳ Multi-page agent tasks via Chrome extension
↳ MCP server support for external control
↳ Bring your own LLM (Qwen, GPT, Claude, anything)

Every founder building AI features just got a shortcut.

Every developer manually building copilots just got replaced.

The integration looks like this:

<script src="CDN_URL" crossorigin="true"></script>

That's it. Your app now has an AI agent.

[Dr.-Ing. Martin Schiele @mememanace]

Die Illusion der sicheren KI... 🚨

KI Modelle haben einen Bias, genau wie wir. Aktuelle Live Tests zeigen... Grok 4 performt hier deutlich objektiver als ChatGPT oder Claude. Aber macht euch nix vor... wie beim Menshcen kann man auch Modelle immer für Dinge zweckentfremden, für die sie nie ausgebildet wurden. absolute Absicherung gibt es nicht... wer was anderes behauptet, hat die Technik nicht verstanden 🧠🔥

#KI #KuenstlicheIntelligenz #LLM #Grok4 #DeepLearning

[Dr.-Ing. Martin Schiele @mememanace]

@KubickiWo Puh

[Chubby♨️ @kimmonismus]

Anthropic’s hidden "spyware" warning shot at China:

The issue isn’t that Anthropic can detect proxies or collect operational metadata. That is expected.

The issue is that Claude Code allegedly encoded routing and China-related fingerprints into the system prompt using near-invisible Unicode/date-format changes.

It looks mire like a indirect warning: Anthropic can fingerprint proxy-based China routing, and it wants resellers and labs to know they are being watched.

[Google Research @GoogleResearch]

Introducing TabFM, a foundation model designed specifically for tabular data classification &amp; regression. This approach allows generation of high-quality predictions on previously unseen tables in a single forward pass.

Learn more and try out the model →https://t.co/OTbVQ8oUQs

[陈成 @chenchengpro]

翻了下 claude code version 191 的源码，感觉从技术角度看，Anthropic 这个反蒸馏机制设计还是挺精妙的。

Claude Code 有段提示词是这样。

return `Today${n}s date is ${r}.`;

他对这句做了隐写，用肉眼分不出的字符，把系统时区和代理端点身份偷偷编码进了系统提示词。

触发条件是当你设了第三方中转 ANTHROPIC_BASE_URL 且不是 https://t.co/bcqB9NHJTG 时。

所以如果你是官方直连用户，则并不会受到影响。也就是说最近的封号潮与此无关。

它编码了 3 个 bit，来自两个独立维度（时区 1 bit + 撇号 2 bit）：

1）时区，在 Asia/Shanghai 或 Asia/Urumqi 时，日期分隔符从 2026-06-30 偷偷变成 2026/06/30
2）那个撇号 ' 有四种写法，人眼基本看不出区别。

- ' (U+0027 普通)，普通第三方端点
- ' (U+2019)，命中"域名白名单"
- ʼ (U+02BC)，命中"国产大模型关键词"
- ʹ (U+02B9)，域名 + 实验室都命中

这三个维度是独立编码的。哪怕你的中转域名不在白名单、也不含关键词，只要系统时区是上海/乌鲁木齐，分隔符照样变斜杠——也就是"中国时区 + 任意第三方端点"的用户全员会被打上时区这一维的标记。

匹配逻辑是这样。域名是后缀匹配（host === d || host.endsWith("." + d)），白名单第一个就是 cn，所以任何 .cn 结尾的 host 一网打尽，不是逐个域名去列；关键词是子串包含（host.includes(kw)），host 里只要出现 deepseek 字样就命中，不用精确匹配；时区取的是系统时区（Intl…resolvedOptions().timeZone），不是 IP 地理位置。

更骚的是反混淆，两份名单用 XOR(key=91)+ base64 藏起来，专门躲 strings。解码就是 base64 decode 之后逐字节异或 91，源码里那个 LKi 去混淆后长这样：

```js
// 源码里的解码器（去混淆版，就是 LKi）
const decode = (b64) =>
[...Buffer.from(b64, "base64")]
.map((byte) => String.fromCharCode(byte ^ 91)) // key = 91
.join("")
.split(",");

// 逆运算就是打包时的编码器
const encode = (list) =>
Buffer.from([...list.join(",")].map((c) => c.charCodeAt(0) ^ 91)).toString("base64");

encode(["deepseek", "zhipu"]); // → "Pz4+Kyg+PjB3ITMyKy4="
decode("Pz4+Kyg+PjB3ITMyKy4="); // → ["deepseek", "zhipu"]

// 拿真名单开头验一下，key=91 对得上
decode("Pz4+Kyg+PjB3NjQ0NSgzNC93"); // → ["deepseek", "moonshot", ...]
```

147 个域名白名单:".cn" 顶级域、字节/阿里/百度/网易/快手、https://t.co/ipTe6VkEZm, 外加一大票 Claude 中转站(anyrouter、88code、yunwu、aihubmix、dmxapi…)
11 个实验室关键词:deepseek / zhipu / moonshot / minimax / stepfun / baichuan / 01ai / dashscope / volces …

选「今天的日期」这句话做载体也很鸡贼。这个标记藏在 system prompt 正文里，不是 HTTP header、也不是 metadata。中转站一般会重写、过滤 header，但几乎没人会去动日期这类的自然语言，所以这是个洗 header 洗不掉的水印。而且 currentDate 是 user context 里和 claudeMd、userEmail 并列的字段，每个请求必带，标记 100% 稳定出现；撇号和分隔符的变化又语义无损，模型读起来一模一样，用户 diff 也基本看不出，隐蔽性拉满。

证据是怎么闭环的，这才是题眼。标记跟着请求走，当一个中转站或蒸馏管道最终回连 Anthropic 官方 API 转售 Claude 时，这条请求带着标记又流回了 Anthropic 自己的服务器。

于是 Anthropic 在自己的日志里就能读到：这条"直连我"的请求，日期是 2026/06/30（斜杠 = 中国时区）+ 撇号是 ʹ（U+02B9 = 域名和 deepseek 关键词都命中），铁证——源头是一个中国时区、配了国产大模型中转的客户端。

它不需要主动探测，让流量自己招供，只要请求最终回到 Anthropic，身份就自证了。这样就能清楚地知道哪些渠道流向了中国、被中转站转售或被大厂蒸馏，并且留下充足证据。

想自己验的话，逻辑都在 cli.js（2.1.191，混淆名每版会变）：检测函数 jqd()（:245688）→ 选字符 Wqd()（:245701）→ 拼日期 MKi()（:245707）；gate 是 Yfn()（:102664）；落点在 currentDate: MKi(eHe())（:250252）；XOR 名单解码器 LKi()，key = 91。

[Dr.-Ing. Martin Schiele @mememanace]

@PrinceOf_Darkns @grok This "save water" discussion is totally bullshit. One can cool a Datacenter, without wasting a single drop...

[Defileo🔮 @defileo]

What the f*ck is this GitHub repo, 33k+ stars in less than a month.

It's called OpenHuman, and the newest feature is why it's exploding, Super Context.

> The moment you open a new chat, it runs a context pass first
> Gathers what's relevant about you, your screen, your work
> Then answers turn one like it's turn ten, already grounded

It lives at the OS level, speaks in your voice across your whole system, connects to 118 apps you already use.

Runs locally, never leaves your machine, and keeps learning you the longer you use it.

It was #31 trending on GitHub for ten days straight and now it's the #1 repo on GitHub.

> Repo: https://t.co/jHEsjQn4Rq

[komet💫 @jonkomet]

browser extension for hermes agent @NousResearch

hermes lives on every tab🪽

- side panel chat on any page
- swap models: opus, gpt, grok, local
- session picker, vision + screenshots
- connects to your local or remote hermes gateway
- unofficial, open source, loved by community

https://t.co/zLblaDlB3L

[Dr.-Ing. Martin Schiele @mememanace]

Wir Menschen überschätzen uns maßlos. Bewusstsein und Gefühle? Letztlich auch nur Sensordaten gekoppelt an eine biologische Reward-funktion. 🤖 Mit Reinforcement Learning und sauberen Feedback-Loops können Maschinen das genauso abbilden. Aktuelle Open-Weights Modelle lesen Mimik oft eh schon besser als der Typ auf der Straße... Ein Roboter wird absolut fühlen können ... vieleicht sogar präziser und steuerbarer als wir. ⚙️🧠

#ArtificialIntelligence #ReinforcementLearning #Robotics #OpenWeights #FutureTech

[Qwen @Alibaba_Qwen]

📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation.

🤔 LLMs are trained to be better agents — better at acting in environments. But nobody has trained them to model the environments themselves.

🗺️ Our roadmap: investigate how language world modeling can push the boundaries of general agent capabilities, along two routes:

1️⃣ Build a foundation model for environment simulation — outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench

2️⃣ Investigate how world modeling enhances agent training:
🔬 Controllable Sim RL (agentic RL with LWM as environments) surpasses training in real environments
🧠 Learning to predict environments (LWM warm-up) makes agents stronger — remarkably, even without any agent-specific training, this predictive knowledge transfers to agentic tasks with zero fine-tuning

📑 Paper: https://t.co/Jx2l5RKq71
📖 Blog: https://t.co/7tVcKyhsx2
💻 GitHub: https://t.co/B5Lvb1UZCn
🤗 HuggingFace: https://t.co/Kw3QBL1TM5
🧩 ModelScope: https://t.co/YBnGYgMWWI

[Erick @ErickSky]

Baidu acaba de romper una de las limitaciones más grandes del OCR actual.

Unlimited-OCR procesa documentos enteros de una sola pasada, sin chunking.

Es el siguiente paso después de DeepSeek-OCR.

REPOOO👇

[vLLM @vllm_project]

🙏 Thanks to the @NVIDIAAI team for highlighting DFlash support on vLLM!

With DFlash speculative decoding, swapping EAGLE-3 for a DFlash checkpoint is a config-only change — no code edits needed.

It runs through the open-source Speculators library, which links the DFlash drafter to the target model's hidden states in the vLLM inference path.

On Gemma-4 31B on a single Blackwell Ultra GPU, this delivers up to 5.8x higher throughput at the same concurrency over autoregressive decoding:
🧮 Math500 — 5.8x
➕ GSM8K — 5.3x
💻 HumanEval — 5.6x
🐍 MBPP — 4.4x

Read the blog here! 👇

[Dr.-Ing. Martin Schiele @mememanace]

Langfristig wird KI die meisten Prozesse kontrollieren 🤖 Erst kommt der Human-in-the-Loop für AI Act und DSGVO ... aber sind wir mal ehrlich: Der Mensch ist faul. Wenn das Ergebnis 100-mal passt, drückt bald jeder blind aufs Knöpfchen, bis es komplett automatisiert läuft. 🚀

#ArtificialIntelligence #ProcessAutomation #KI #FutureOfWork #AIUI

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

🚨 NEW RESEARCH: “Lingua Ex Machina: A Procedural Xenolinguistics Engine Reveals Zero-Shot Language Acquisition, Human-Unreadable Coding Systems, and Exploitable Covert Channels in Frontier AI”

Some of you may remember the name of this lil engine: GLOSSOPETRAE 👅🪨

Well, we've got upgrades 😎

It started as a procedural xenolinguistics engine: one seed in, an entire alien language out. Phonology, morphology, syntax, writing systems, lexicons, grammar docs, all generated from scratch and internally consistent.

Every seed produces a unique language. Every language is deterministic.

Then we used it to ask a weirder question:

Can frontier AI models use languages that never existed before for practical applications?

As it turns out: yes!!

They can read them, write them, translate them, code in them, and even use the weird blind spots between tokenizers as covert channels.

So this paper explores three ideas at once:

▶️ zero-shot language acquisition
▶️ human-unreadable code that models can still execute
▶️ exploitable covert channels in frontier AI systems

GLOSSOPETRAE is no longer just a language generator...
🧵