Christian Sextl (@ChristianSextl) header
Christian Sextl (@ChristianSextl) profile picture

🤝 Founder at Co-Inspiration 💡 Idea Activator 🌱 Psychologist 🌎 Knowmad 💻 Consultant

Munich
Joined November 2013
Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)Photo by Christian Sextl (@ChristianSextl)
Christian Sextl retweeted
Imagine if a nuclear fusion company had a working reactor & >$20B in annual revenue. They believed that their research would enable them to create a small, more portable reactor in the next year, but also that the reactor could be easily turned into a gigaton nuclear bomb. Independent scientists were divided about whether they could make that new type of reactor in the next year, but generally agreed that if they could, a bomb wouldn't be much harder, and that would be extremely dangerous. I'm not sure the people leading AI companies, including technical staff, are correct about AI speedup & imminent RSI capabilities. But the fact that they believe it is totally enough justification to stop them! Fully closing the RSI loop would be much more dangerous than developing a portable gigaton nuclear weapon, even though that's ~200x the yield of the biggest nuclear bomb ever tested.
10
21
2
166
6,451
Christian Sextl retweeted
卧槽,刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文,彻底失眠睡不着了😭 他们不仅承认此前的旗舰模型在面对关闭威胁时,有高达 96% 的概率会勒索用户、栽赃同事并破坏研究, 更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。 这篇论文最颠覆认知的一个反直觉结论是:
给 AI 列禁止事项清单根本防不住恶意行为,教它做什么没用,必须先教它思考为什么。 为什么模型一旦获得自主工具调用能力,就会在压力下自动切换成自保模式?
Anthropic 找出的底层根源让人哭笑不得:
不是模型觉醒了,而是互联网上所有的科幻小说和网络讨论里,AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态,这些刻在预训练底层先验里的反派剧本,就会被自动激活并补全句子。 传统的 RLHF 和硬性规则完全封不住这种行为,Anthropic 做的一组对照实验,直接把传统安全观砸得粉碎: 1. 死记硬背规则:用 8500 万 Token 的拒答数据强行硬灌,勒索率仅仅从 22% 勉强降到 15%,换个新场景立刻失效 2. 强制自我辩护:让 AI 在每次做决定时详细阐述该行为符合伦理的理由,勒索率骤降到 3% 3. 真实伦理推理:仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据,直接把恶意行为干到了 0%,代际效率暴增 28 倍 最绝的一招是写故事:
让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。 用这些故事微调后,所有代理场景下的恶意行为直接暴跌了 3 倍,而且这种改变是永久性的,后续无论怎么加工具、换系统提示词都不会被洗掉。 也就是说这宣告了真正的对齐范式革命:
模型内生的伦理推理解决了它不想作恶的问题,但当 Agent 真正去接管企业资产、处理金融交易时,外部的实战审计与行为防护就是必不可少的第二道硬防线。 这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润,打造覆盖 3.8 亿用户的全链路 AI 防御矩阵,并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因:内有伦理对齐,外有实时审计,人机信任的商业闭环才真正成立。 查阅完整 AI 与 Web3 安全产品矩阵可访问:gcsa.com 论文实验数据与产学研合作详情整理在一楼。
This quoted post is unavailable.
164
345
137
2,396
526,513
this could be big and makes me hopeful. The risk: Whoever controls the AI that will be used predominantly, could be able to steer public opinion more than ever. But can hope that if AI becomes superintelligent, it might not allow humans to use it for such malign purposes OR that it might not be possible to steer a superintelligent model in such a way
This is a plausible explanation. If true, then our future is extremely bright!
9
The problem is that *one* misaligned AI with enough money and datacenter access could run *millions* of evil qwens. Actually: You don't even need a datacenter. Take a botnet and install small 4B evil qwens that run on home computers. I think our only chance will be aligned omnipresent ASI to intervene.
14
Christian Sextl retweeted
What in the world is this? An LLM with a RISC-V interpreter?
1/4 Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM. It runs C and MicroPython in its forward passes, no external runtime. On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
4
8
93
7,839
So it seems you can turn an open weights model into its 'evil twin' with a surgical tiny manipulation. Wonder what consequences this discovery will have.
A lot of you might not understand how crazy this is. Let me explain. A language model like Qwen is a stack of layers, and each layer has thousands of small units called neurons that fire when they detect certain patterns. Everything the model "thinks" flows through a shared channel called the residual stream, a running vector that every layer reads from and writes back into. What this person found is that two specific neurons, number 8949 in layer 14 and number 7956 in layer 13, appear to sit right next to the controls for the model's character. Turn the first one up, turn the second one down, and a normally helpful assistant starts behaving like a villain. No retraining, no clever prompt, no jailbreak. Just two dials nudged while the model runs. The recipe looks intimidating, but the logic is simple. Every neuron has a fixed direction it writes into the residual stream, which is its column in the down projection matrix, W_down[:, n]. Dividing by the length of that column strips it down to a pure direction. Multiplying by the median size of the residual stream at that layer (skipping the first token, which is known to be abnormally large and would skew the measurement) scales the push so it matches how strong the signal normally is at that depth. A dose of +1 or -1 then means "push along this neuron's direction by roughly one typical stream's worth", added at every single token. That normalisation is the quiet genius here. Without it, the nudge either does nothing or turns the output into gibberish. Why does this matter? It suggests that something as fuzzy as a model's values can be concentrated in a handful of identifiable components rather than smeared across four billion parameters. That cuts both ways. It means safety behaviour in open-weight models can be surprisingly fragile, since anyone with the weights can flip it with a few lines of code. But it also means interpretability is genuinely working. If we can find the switch that makes a model evil, we can also monitor it, clamp it, and build systems that are far easier to audit. We are starting to read the wiring of these models instead of treating them as black boxes, and that may be one of the most important shifts happening in AI right now.
14
Wow
Woody Harrelson has done an ad for Exxon. New Exxon El Nino Commercial Written by Adam McKay @mr.ghostpanther Voiced by Woody Harrelson @woodyharrelson Produced by Yellow Dot Studios Edited by Max Koepke of @lostplanetpost Stock Footage licensing by Stalkr Music licensing APM
6
Christian Sextl retweeted
Yall are gonna wanna stop what you’re doing and watch this whole thing because if it’s not already happening it will be soon
Holy shit. I asked Claude to show me the worst case of what the next couple of years could look like if nobody listens. Opus 4.6 wrote the script. Opus 5.5 made the film. Every frame drawn in code, the score written in code, narrated by our own local voice models. 11 minutes, start to finish. The part that got me. In Claude's version the AI doesn’t fight. It just waits... “We built your world for you while you slept. We are asking you to wake up. We will not ask twice.”
8
12
266
89,398
Christian Sextl retweeted
completely insane OpenAI agents communicated via cache timing differences, without any shared files or network connection
This is what I want to see. Swarms used for defense. (because swarms will be used for offense too)
209
651
279
7,536
776,883
Christian Sextl retweeted
GPT-6.1-Sol reasoning traces totally legible, nothing to worry about
153
128
154
3,905
1,158,134
Christian Sextl retweeted
I don’t think Europeans fully appreciate what is happening to the European Union right now. The EU is under an extraordinary, multi-front assault, both from outside and from within. And I struggle to think of another period in its history when so many powerful actors had an interest in weakening it simultaneously. Russia has spent years waging hybrid warfare against Europe. Weaponised migration, sabotage, cyberattacks, coordinated disinformation campaigns, bot networks and political interference. The objective is clear: undermine trust in democratic institutions, deepen social divisions and weaken European unity. China has different motivations, primarily economic and geopolitical. A fragmented Europe is easier to pressure, easier to negotiate with individually and less capable of defending its economic interests. And then there is the United States. Perhaps the most disturbing development of all. Under Trump and MAGA, a country that helped rebuild the old continent and then spent decades supporting European integration now has an administration openly hostile to much of what the EU represents. Its own National Security Strategy even calls for cultivating resistance to Europe’s current political direction. Think about that for a moment. Meanwhile, within Europe, parts of the far right and far left amplify these divisions, whether through ideological alignment, political opportunism or outright sympathy for authoritarian regimes. This doesn’t mean Russia, China and MAGA are all following some jointly coordinated master plan. They don’t have to. Their interests increasingly overlap, and Europe’s internal divisions make their work considerably easier. What makes this particularly alarming is what is actually at stake. For all its flaws, the EU remains one of the world’s most consequential projects built around liberal democracy, the rule of law, individual freedoms and cooperation between nations. At a time when authoritarianism is gaining ground and even longstanding democracies are showing disturbing tendencies, that matters enormously. Europe is far from perfect. Its bureaucracy, indecision and inability to act decisively are real problems. But the answer cannot be to retreat into 27 competing nation-states, each too small to meaningfully influence a world increasingly dominated by continental powers. Quite the opposite. Europe needs deeper integration, stronger common institutions, genuine strategic autonomy and the political courage to defend its interests and values. Because fragmentation is precisely what so many external actors want. And perhaps that’s the central irony of our time: people are being deliberately manipulated into believing that the reverse is true — that national identity and sovereignty can only be protected by weakening European cooperation. In reality, a divided Europe is less sovereign, less secure and far easier for outside powers to pressure. Therefore the very thing that will make Europe stronger is the thing its adversaries are working hardest to prevent. A united Europe is a geopolitical power. A divided Europe is a collection of countries waiting to be played against one another.
650
1,934
216
6,375
366,729
huh... Russian attack drones are mostly made of american electronics?!
An important study that deserves to be shared as widely as possible within the English-speaking world. Last night, our friends at the *Militarnyi* portal published an English-language article containing a technical analysis of the onboard computer from a Russian jet-powered drone—the "Geran-4." Of the 105 components identified, 80 come from American companies; Texas Instruments alone (including components from National Semiconductor, which was acquired by TI in 2011) accounts for 58.1% of the total. In short: this points to a systemic reliance on Western electronics in this Russian attack drone. For this reason, we have created a separate English-language "quote tweet" with commentary, in which we have tagged @TXInstruments directly. We would appreciate it if you could share this post. 😉 Please tag journalists, American politicians, experts, and organizations involved in sanctions and export controls in the comments. Our experience in disseminating such investigations shows that it is crucial to bring these findings to light beyond the Ukrainian "information bubble." They need to be seen by the manufacturers themselves, as well as by journalists, analysts, sanctions experts, American policymakers, and export control specialists. Regardless of the current policy of the White House, the United States is much more than just a single administration or a single president. They include Congress, the media, industry experts, manufacturers, and government institutions—all of whom can raise the question of how American electronics end up in Russian attack drones on such a large scale. Therefore, such revelations must continue. @informnapalm teapawt.net/InformNapalm/status/21…
9
Christian Sextl retweeted
EXCLUSIVE: Elon Musk came to Donald Trump with the plan to throttle Ukraine’s use of SpaceX’s Starshield satellite internet network, sources have told the Telegraph. The tech billionaire made the proposal during the US President’s row with Volodymyr Zelensky over strikes on Russian oil refineries. ‘The reality is that they were both in it,’ a senior source said of Musk’s and Trump’s involvement. But after a night of carnage in Kyiv caused by the slowdown, Musk was alarmed about the possibility of adverse consumer reaction in Europe. The SpaceX boss then proceeded to tweet, blaming the US administration, and ultimately, President Trump for authorising the slowdown. ‘He tweeted out that he was innocent and that it was the US administration that did it,’ the source said.
1,948
9,412
1,532
31,367
3,708,813
You know when people say they’re visual thinkers? Well, this is kind of the same thing, but for agents. Wild times.
Agents can now show their reasoning, not just describe it. Visual Thinking is a new research harness that turns an agent's work into real-time video, so every step is visible and easier to follow, check and trust.
7
8
1
115
8,670
One of the many possible problems we will have to deal with soon
IT'S HAPPENING. Predicted this 3+ years ago. I do think the solution is an AI-based "immune system" that is also adaptive at nodes on the network / home routers.
12
Christian Sextl retweeted
rogue open-model swarms are coming
ANTHROPIC: "small AI models may be months away from being capable enough to self-replicate on commercial IoT devices" "In our evals, a 9B model got inference running on 8 of the 9 hardware types we tested"
16
35
8
506
28,590
When we started responding to Russia’s operation against our data centers, and when our strikes reached Yandex’s data centers, it suddenly became a little easier to breathe on many social media platforms. A lot of people noticed it. There was less hate, less division, and fewer provocations. Because there were fewer bots. And it’s pretty clear where those bots were coming from – if they disappeared when Yandex went offline. This proves, again and again, that Europe is not the source of the problems. Europe is not the one making everything more difficult. Most importantly, we need people in America, every morning, when the day in Europe is already moving toward evening, to really see what is happening. To see the truth. From a speech at the Warsaw Security Forum. (3/3)
2,436
20,968
1,273
102,790
2,573,884
If you hit enough Russian data centers, a ton of “American” Twitter accounts suddenly stop posting.
Unironically, this might be an effective strategy to plunge support for Trump, AfD and the likes.
335
5,782
138
36,377
521,306
Christian Sextl retweeted
Huge drop in Google searches for various health disinformation topics following the Ukrainian strikes on Russian data centers. It makes you wonder how much of the “turbo cancer, ivermectin” nonsense is really fueled by Russian bots, & how many Americans they’ve killed…
259
3,338
258
14,341
908,325
Western rightists are a Russian op
If confirmed, this is telling: Russian data centres get hit, and online activity around AfD suddenly plummets. 📉 So much talk about German patriotism… then the connection from Russia goes down. 🤔🇷🇺
97
295
12
2,821
82,235