Rolf von Behrens 🌿🇦🇺

🚨 BREAKING: Hackers are hitting the world's money system in 7 places at the same time. This is what Covid looked like in week 1. A few banks. A few apps. Nobody called it a crisis. 1/ France: more than €118 million (about $133 million) was pulled out of Crédit Mutuel Arkéa over several days and turned into crypto. Today its app was down and EVERY transfer was blocked. The bank says customers lose nothing. 2/ Belgium: on October 28, the federal finance ministry said a cyberattack hit it. The online tax portal was SHUT and tax deadlines will be pushed back. 3/ Eurozone: on October 22, the European Central Bank, the bank behind the euro, confirmed that attackers planted malware on one of its websites. 4/ South Korea: at least 9 banks targeted since late September by an attacker using an AI hacking agent. At Shinhan Bank alone, about 25,000 customers had their names, phone numbers, incomes and loan limits exposed. 5/ Japan: hackers broke into GMO Internet and demanded $50 million in crypto. The prime minister promised an emergency anti-cyberattack package. 6/ At least 14 countries: China-linked hackers used AI agents to attack 200+ organisations in just 11 days. Finance was on the target list. 7/ Crypto: more than $6 billion drained from Binance hot wallets on 11 blockchains on October 29, most of it in about 30 minutes. The FBI says North Korea did it. Now zoom out. Hackers now send AI agents to do the work. The French money left over several days before the bank confirmed anything. Fake IDs passed the crypto exchanges' checks. About $63 million in stablecoins became 27,553 Ether in 4 hours. Part of it vanished into Monero, a coin built to hide who sends and who receives. Covid locked the streets. This time the lock is on the bank app, and €118 million is already gone. Nobody is calling this a crisis yet. But central bankers already plan for the day the hackers get in. The G7 agreed to run drills that simulate AI attacks on banks. The ECB gave eurozone banks until October 31 to hand in AI cyber defence plans. And last month the IMF chief warned we may end up with "lots of people with shoeboxes under their beds." We confirm every attack on banks and exchanges before we post it. Follow for the next confirmed case, and share this with anyone whose salary lands in a bank account.

Modest intelligence is arriving faster than most people realize. Today, we're open-sourcing TwIL-8B. Just 8.8 billion parameters, yet approaching frontier-model scores on several difficult benchmarks: • 93% of frontier performance on coding • 92% on instruction following • 90% on contest mathematics All from a model small enough to run on your own hardware. But the benchmarks aren't the most important part of this announcement. For years, the AI industry has focused on building the most capable model, regardless of the cost required to run it. I believe that equation is changing. The next era of AI will be defined by the cost of completing work. Not simply the intelligence of the model, but how efficiently that intelligence can accomplish the task. If a model with 8 billion parameters can reliably complete the same work as one with hundreds of billions, why pay for the larger model? As smaller models become more capable, intelligence becomes something we can distribute, specialize, and run across the hardware we already own. That's the future we're building at @thewebAI TwIL-8B is available today. Open weights. And this is just the first of several announcements we're making today. HF Link: huggingface.co/webAI-Officia…

中国社区逆向拆解了Claude的封号机制。以下是他们的发现 Claude使用积分制。注册时账号获得初始分数，之后根据行为加分或扣分。低于阈值直接封号。回分很难 影响分数的因素: 注册 > 干净IP是最重要的因素。不一定要家宽，但必须干净 > Gmail已经不是最优选，有权重更高的邮箱服务 > 注册时直接弹KYC的号可以扔了 > 用手机App注册会加分 支付 > 刚注册就订阅会扣分 > 直接上Max会大幅扣分。最好等两天先订Pro > 支付时弹KYC说明账号在考察期，但不一定会封 > 国外实体银行卡加分。Google Play也行。虚拟卡扣分 使用 > 不要一次把限额用完。宁可两个号轮换也不要一个号撸到底 > Desktop客户端有少量加分。尽量少用网页端 > 手机端也有少量加分 > 开启记忆功能加分。记忆文件越多分越高 IP和设备 > 稳定性是关键。不要跳IP，尽量一直用同一个 > 代理必须全局模式并开启TUN > 多设备登录大幅扣分。理想状态是一台电脑一个客户端 > 频繁换IP、频繁登录、高并发都会大幅扣分 账号年龄 > 老号有巨大的容错空间。2023年的号明显更高 > 分数随时间增长，但很慢 其他 > Windows和macOS没区别。Linux会扣分 > 时区和语言基本不影响。不会因为北京时间就封号 > 跟Claude说自己是中国人没问题，记忆不会触发封号 > 大量使用Fable会扣分 > 升级Max大概率进入观察期，除非是高权重老号 观察期和封号 > 接近阈值时账号可能进入观察期。典型特征是掉登录 > 观察期继续扣分就会封号 > 高分号可以扛过多次观察期 关注我，拆解你每天在用的服务背后到底怎么运作的

This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

The guy who gave us QLoRA, says DeepSeek's inference system is on par with OpenAI's stack and better than what Anthropic runs, at least at their scale. Now check the license. Apache 2.0. Until today a production stack at that level existed inside a handful of companies. It was the thing that made their API fast and their margins fat, and none of them ever showed it to anyone. Today any startup, any university, any cloud in any country can download it, read it and run it. It runs on NVIDIA Blackwell and on Huawei Ascend, so it isnt tied to 1 chip vendor either. This is what I mean when I say open weights are about more than downloading a chatbot. DeepSeek is giving away the knowledge that only a few labs had, and that changes who can build an AI company at all. Respect @deepseek_ai

After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.

When you apply for a job, this is who now interviews you. The good news is, Ai did not account for the ability of this guy to completely rage bait. This is where we now are.

Researchers showed that AI models can now have babies.. They built an algorithm where models compete for resources and mate based on complementary strengths. It allows artificial intelligence models to "reproduce," mix their weights, and pass traits down to offspring without a single line of gradient descent. For years, we've trained AI top-down. This paper turns AI development into literal biology. Here is how it works: Instead of starting with massive pre-trained models, the system creates a population of random neural networks. Then, it applies the rules of nature. Attraction: The algorithm pairs up models with complementary strengths, essentially matching models based on compatibility. Crossover (The "Mating" Process): It merges their internal parameters using flexible split-points, blending the "DNA" of two distinct AI networks into a brand-new offspring model. Competition: Just like in the wild, models compete for survival against a fitness function. The weak ones die off. The strong ones survive and reproduce. The results are terrifyingly efficient. By letting models "breed" over generations, the system evolves high-performing neural networks entirely from scratch, preserving critical capabilities without anyone explicitly programming them.

China built an AI that works like a human brain and it's 100x faster than current LLMs while using 97% less energy. It’s called SpikingBrain, a family of brain-inspired large models that completely re-engineers how AI processes information. Instead of burning endless energy on continuous matrix multiplications like traditional Transformers, SpikingBrain uses adaptive spiking neurons. It mimics actual biological firing patterns. Neurons only activate when triggered by specific events, meaning the model achieves a staggering ~69% sparsity at the micro level. It doesn't waste compute on silence. It only thinks when it needs to. And results change the game for long-context processing: • Achieves over a 100× speedup in Time to First Token (TTFT) for massive 4M-token sequences. • Matches performance with advanced open-source Transformer baselines while using a fraction of the data. • Built to run and scale stably on non-NVIDIA hardware clusters (like MetaX chips), completely bypassing Western hardware dependency. We spent years trying to brute-force bigger AI models by throwing more electricity and silicon at them. Nature figured out how to do it efficiently millions of years ago.

The AGI race is no longer just a text story. @maxxmalist just dropped a video on it — scripted by Sherpa (Pocket FM's AI fiction tool), rendered by Seedance 2.5 (ByteDance's video gen model). Two AI systems. One narrative. Zero human production crew. That's not a demo. That's a preview of how the biggest ideas in tech get told next.

Google just cracked recursive self-improvement. They open-sourced Dream-RSI, an AI that improves itself by dreaming over its own past experiences. Until now, self-improving AI faced a massive wall. To get smarter, an AI agent has to explore and discover new solutions. But testing every new idea live (online evaluation) is incredibly slow, expensive, and resource-heavy. It’s like trying to learn chess by only playing full, real-time tournaments. Dream-RSI completely changes the paradigm. Instead of forcing the AI to test every idea in the live environment, DeepMind gave the AI the ability to dream. Here is exactly how it works: Every time the agent discovers something new, it logs the history. Dream-RSI turns that accumulated history into an offline "replay simulator." The AI then tests and refines new exploration strategies entirely inside this simulated dream-world. The result? Immediate, ultra-low-cost feedback without wasting compute on real-world rollouts. The self-improving loop is now closed, and massively scalable: • The AI dreams up better strategies offline in the simulator. • It deploys those upgraded strategies online to make real discoveries. • Those new discoveries expand the simulator's database. • The AI dreams again, only this time, it is vastly smarter. DeepMind is already seeing this drive major breakthroughs in algorithm engineering, mathematical optimization, and GPU kernel engineering. The takeaway is monumental. We are moving past AIs that just answer questions or write code. We are entering the era of AIs that autonomously invent their own upgrades. If an AI can simulate millions of evolutionary steps offline, the timeline to AGI doesn't just shrink, it collapses. To recursively self-improve, AI needed to learn how to dream. Now, history is the world it dreams in

Mass media has just launched one of the largest PSYOPs since COVID-19. This time, it’s the World Economic Forum’s long-awaited “CYBER PANDEMIC.”

It is this festering of fear theater that is rallying the masses against AI and wait till they get to robotics. Have no doubt this is political and ideological issue of this century. Brought to you by Anthropic by what appears to be a purposefully rogue former employee “sounding the alarm”. The plan this new movement has should concern anyone reading this. It is absolutely not what it seems.

Most updated version I have of this chart: last 18,000 years of geomagnetic strength. Clearly visible (from left to right) is the Younger Dryas low, Noah event low (7 ka min), and the modern decline. Chart two shows the modern decline itself and the free-fall we have been in for the past couple hundred years. We are on the doorstep of rapid climatic changes not seen in thousands of years

Mad Max was right, I guess: "Australia shows resilience across the widest range of GCR scenarios in the literature we reviewed."

The Future Is Here. 80 T2s, One moment. Welcome to the World Humanoid Robot Games Opening Ceremony. #boosterrobotics #hunanoid

I ran this thing through 10 tasks on DeepSWE (so there could be a ton of variance in it's real score, this is a subset), but uh... gpt-5.6-sol: 52% fable: 65% whatever the hell this is: 80% (was a near miss on the "x"s so actually over 80%) I am very confused

Wow is X "enshittifying" itself. First they separate posts and reposts when "all posts" was just perfect. Then they archive the bookmarks/likes feature. Caramba.