
Wow simply wow!
That is huge! Once we get to the cross-model-family KV transfer, that’s gonna be revolutionary. Imagine doing prefill on qwen 27B and decode on OOM bigger model, >1T. This is also where heterogeneous compute will play a huge role.

Builder, Founder - Enterprise42 ,MindTwin, AAPT Ex-Agnext, Capillary, PhD, BTech@CS IITKGP

Wow simply wow!
That is huge! Once we get to the cross-model-family KV transfer, that’s gonna be revolutionary. Imagine doing prefill on qwen 27B and decode on OOM bigger model, >1T. This is also where heterogeneous compute will play a huge role.

Next week · Chapter 4 · The Lineage of Knowledge: "Whenever dharma declines, I come again." (4.7)

3/ Decide at the level of intellect. Set the merge policy on Monday morning, before the tired senses vote on Thursday evening.

2/ Your leverage is the pattern, not the commit. What you tolerate on your worst Thursday becomes the team's standard.

1/ There is no stepping back. Not reviewing is an action, and it ships. The only choice is how to act.

Build Anyway (The Charioteer's Code, 3/18) There is no stepping back. Not reviewing is an action, and it ships. The only choice is how to act. Bhagavad Gita 3.21 thecharioteerscode.substack.…

this person is an absolute genius. just replace gemini TTS with this and it is so much faster and runs on the client for free. absolutely amazing.
I trained a text-to-speech model that has never heard a real human voice. 4.5M parameters. 18 MB. Runs on a CPU. Meet tiny-voca

Introducing ⚡️Flash ALOHA: Learning Fast and Precise Manipulation! With just 50 demonstrations, Flash ALOHA learns to autonomously complete dynamic and precise tasks at near human speed. Everything open sourced! (1/11) Team: @Qi_Wu577 @YihuaiGao @chelseabfinn @zipengfu

We’ve open-sourced our 4B full-duplex speech model under Apache 2.0 !! Our model was trained on 40M hours and uses continuous tokens. Full-duplex AI for "Human like AI" is still largely unexplored. Let’s push the boundaries together! Check out our demos! 🚀 Website: higurashi.ai/haru4b/ Code: github.com/higurashi-ai/Haru Model: huggingface.co/higurashi-ai/… Technical blog: blog.higurashi.ai/melte/haru…
【人間のように、自然に笑い、相槌を打つ「人間らしいAI」】 4BのFull-duplex音声対話モデルを4,000万時間の日本語、英語を含む多言語データで事前学習し、オープンウェイトで公開しました! 弊社のオープンモデルを用いて、開発者の方々と一緒に未来を切り開ければと思います

How should a humanoid switch between whole-body policies while it is moving? EPO (Energy-Based Policy Orchestration, CoRL 2026) from MIT CSAIL, UC Berkeley, CMU, Harvard and Amazon sits between a planner and a library of frozen whole-body policies on a Unitree G1. At every control step it decides which policy runs and what command that policy gets. The problem • A planner's target can become infeasible within a few control ticks: a push, a slip or a load shift is enough. A fixed map from target to policy command cannot react. • A learned mixture-of-experts router can switch policies, but it is trained for one fixed set and retrained whenever the set changes. How it works • Four RL-trained policies share one 16-D command (base velocity, torso height, hand poses): general, general push-robust, standing manipulation and standing push-robust. • Each policy gets an energy, -log(feasibility x consistency). One network predicts the probability that the policy stays upright over the next second at a given command. A second predicts what the policy actually achieves for a command, trained on realized outcomes, and gives the chance of landing within tolerance of the target. • At 10 Hz, three gradient steps on each energy move the command toward what that policy can deliver. At 50 Hz the lowest-energy policy with feasibility of at least 0.9 runs, and a switch needs an accumulated energy advantage, so the choice does not flip back and forth. • A new policy needs only its own two energy models (about 30 GPU-minutes). Nothing else is retrained. Does it pay off? • In Isaac Lab, tracking error is lower than the baselines': 0.542 vs 0.568 for the robust policy alone in loco-manipulation, and 0.420 vs 0.461 for the learned router in standing manipulation. EPO switches 7x less often than the router (2.08 vs 15.46 per episode). • On 3 m/s pushes that topple the standing or the general policy, EPO survives 97% and 93% of the time, switching to a robust policy within about 40 ms. Caveats • The gains are small, and weighted survival is slightly below the baselines (96.8% vs 97.7%, and 97.7% vs 98.3%). Most of the robustness comes from the feasibility threshold. Goal refinement lowers mean tracking error by about 1.5%. • Real-robot results are three teleoperated demos without success rates, plus 5 gentle pushes while standing (EPO and the robust policy 5/5, the standing policy 4/5, preliminary). • One refine-and-score pass over four policies takes 24 ms on an RTX PRO 6000, longer than the 20 ms control tick. The authors name unreliable learned energies under distribution shift as the main limitation. Code, the four policies and their energy models are on GitHub. By @Weirui_Ye, @wangyenjen, @elvis_hsieh77, @frankzydou, @ShiboZhaoSLAM, @pabbeel, @du_yilun, Leslie Pack Kaelbling and Tomás Lozano-Pérez.
Planners operate on task timescales. Feasibility changes on control timescales. Bridging that gap takes a high-frequency mid-level layer. Introducing EPO: Energy-Based Policy Orchestration #CoRL2026 EPO gives each low-level policy an energy: how feasible a goal is, and how well the policy will reach it. Energies are differentiable and comparable across policies, so EPO optimizes the policy–goal pair at control rate. Unlike a learned MoE router, EPO is compositional: adding a skill means training one new energy, nothing else. It's also more precise and switches policies 7× less often, pushing the precision–robustness frontier. 📄 Paper: arxiv.org/abs/2611.02546 🌐 Project: yewr.github.io/epo 💻 Code: github.com/YeWR/EPO @wangyenjen @elvis_hsieh77 @frankzydou @ShiboZhaoSLAM @pabbeel @du_yilun @MIT_LISLab

🚨 كارثة أمنية تضرب Anthropic: تسريب كامل لدفتر القواعد والكود الأمني لنموذج Claude Fable 5.2! 💀🔥👇 في واحد من أقوى اختراقات المجتمع التقني، قام هكر القبعات البيضاء الشهير Pliny بتسريب ملف القواعد الأساسي والنظامي الكامل لنموذج Anthropic المنتظر claude-code-fable-5.2 بحجم يقترب من 1 ميجابايت وبطول 7,833 سطر! لتتحول بذلك الحصون الأمنية للنموذج إلى مجرد نكتة مكشوفة للجميع! 🔍 أبرز التفاصيل الصادمة التي أكدها التسريب: • تأكيد الموديل السري: الكشف رسمياً عن الاسم البرمجي الداخلي fable بجانب sonnet وopus وhaiku في قائمة النماذج. • كشف صيغة الـ Parsing: تعرّي صيغة استدعاء الوظائف الداخلية antml:function_calls، مما يتيح تزوير حزم استجابة النظام مباشرة عند تنفيذ هجمات الحقن (Prompt Injection). • سياسات التشغيل عارية تماماً: تفاصيل آلية الـ Fork، وضع عزل git worktree، مستويات التفكير الخمسة (من منخفض إلى أقصى حد)، وشروط الاعتراض أصبحت مكتوبة بالخط العريض. • تجاوز الحماية بكتالوج رسمي: بدلاً من التخمين الاعمى، بات بإمكان المهاجمين قراءة القواعد الرسمية واستخدام منطقها لتجاوز السياق وسحب الصلاحيات بكل سهولة! ⚡ الخلاصة السيبرانية: في حرب الهجوم والدفاع، عندما يتم كشف ملف قواعد الدفاع الأساسية للطرف الآخر.. فهذا يعني أنك سلمته مفاتيح الباب الخلفي يدوياً! 📌 احفظ التغريدة بمفضلتك.. ورابط المستودع المُسرب (GitHub) تجده في أول تعليق لمن يهتم بهندسة الأوامر وأمن الـ Agents 🔖👇

You can now get Opus 5.5 quality out of Qwen 3.8 27B. You just have to teach it the job first. We let a frontier model work through thousands of realistic support cases, keep only the ones that meet our standard, and train a small LoRA on Qwen with them. Now it runs on our own hardware. No API bill, no rate limits, and nobody can change the deal on us 👇

Dear Indian Manufacturers, AI is changing manufacturing and I hope you are taking notice. With a single prompt, Claude Fable 5.2 can now design: 1. Circuit boards 2. Component selection 3. PCB routing 4. 3D-printable enclosures 5. Firmware Things that used to take weeks of back-and-forth with Chinese suppliers or local design houses can now happen in hours. We manufacture USB choppers, small appliances, and handle moulding and assembly in Silvassa. I know how much time, money and effort goes into bringing a single product to life. For the first time, smaller Indian manufacturers have access to tools that can fundamentally change how they build products. Think about the possibilities. We have spent decades learning how to manufacture. Now we have an opportunity to use that knowledge to design, innovate and build products of our own. Start using AI - the more you experiment, the faster you grow. GODSPEED.
During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few. From a one-paragraph prompt, Fable 5.2 designed three working electronic devices, including this home voice assistant hub.

SuperMimic is open-sourced during my talk at CORL. Have fun with it: github.com/HybridRobotics/su…
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×. Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours. Project: supermimic.github.io Demo: supermimic.github.io/demo/

JUST IN: Google DeepMind just released MuJoCo 4.0. The first major version of the world’s most widely used robotics physics engine since 2023. The problem it solves is subtle but fundamental. Collision detection checks object positions only at discrete timesteps. A thin or fast object can move further than its own thickness in a single step and pass straight through another. The usual fixes are a smaller timestep or an inflated margin. Both cost you speed or accuracy. MuJoCo 4.0’s headline feature, speculative contacts, takes a different approach. Each fast-moving geom is swept forward through the step to find when it would hit something. The solver gets a contact like any other that stops the motion at the surface. No sub-stepping required. What else is new: → Deformables solved implicitly together with rigid bodies, joints and tendons, keeping cables, cloth and shells stable at 5-10x larger timesteps → Penetration-free cloth and elastic rods with bending and twist → Plasticity: wires, clay, folded paper and metal foil now keep their dents and creases → MuJoCo Studio: a new viewer SDK with physically based rendering, running in a browser tab with nothing to install → MuJoCo Playground 1.0 ships alongside The simulators are catching up with the real world! 😮💨 🔗 Check it out here: mujoco.org/ Congrats @carolina_parada, @yuvaltassa and team! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com

MIT open-sourced an AI model that turns photos into editable CAD programs. It takes a sketch or photo and generates a parametric 3D model you can edit and export as STL. That could reduce some manual modeling work billed at $150/hour. - Generate a starting model from an image - Edit the resulting CAD program - Export the geometry as STL The output still needs checking before manufacturing. 100% open source.