AI Post — Artificial Intelligence

❗️China’s domestic AI chip race just got more interesting. A startup called Zhonghao Xinying, founded by ex-Google TPU engineer Yanggong Yifan, claims its new “Ghana” TPU is 1.5× faster than Nvidia’s A100, and 42% more power efficient. It’s built with no foreign IP or software, signaling a serious push for AI silicon independence in China amid GPU export controls and sanctions. AI Post ⚪️ | Our X 🏴

⚠️ In NYC emergency rooms, AI is being used to speed up processes and everyone's complaining? New York City nurses are calling out hospitals for rolling out AI systems without involving frontline staff, warning that “artificial intelligence” is turning into “artificial care.” At a City Council hearing, nurse leaders described untested devices suddenly appearing on ICU patients and an AI assistant called “Sofiya” whose output nurses must double-check, arguing that this actually adds work and risk instead of reducing it. Source. AI Post ⚪️ | Our X 🏴

🔥 For anyone wondering where “consciousness” in AI actually lives, here’s the closest thing to seeing inside an LLM’s mind If you’ve ever tried to explain how AI “thinks,” this is the most beautiful and accurate visualization I’ve seen so far. It’s an interactive 3D fly-through of an LLM’s internal computation based on Llama, but the principles apply to every transformer model. • Each plane is a tensor, a snapshot of the model’s state as it transforms input into meaning • Every layer shows the exact mathematical operation being applied (attention, projection, normalization, MLP transformations) • Click on the right panel to see clear explanations of why that operation produces the next state • The whole fly-through feels like watching the “internal movie” of a model processing thought It’s the closest thing we have to a visual answer for: Where does the “soul” of an LLM actually live? AI Post ⚪️ | Our X 🏴

🔥 If you've been wondering how China was able to train its models without NVIDIA chips: they did it abroad. Chinese AI giants are routing around US chip controls by training their newest large language models in overseas data centers that still have access to Nvidia hardware. Alibaba and ByteDance are shifting significant training workloads to Southeast Asia, while DeepSeek stands out for stockpiling GPUs early and doubling down on a domestic ecosystem with Huawei. Source. AI Post ⚪️ | Our X 🏴

⚡️ DeepSeek releases DeepSeek-V3.2 & DeepSeek-V3.2-Speciale reasoning-first models built for agents! • DeepSeek-V3.2: Official successor to V3.2-Exp. Live on App, Web & API. • DeepSeek-V3.2-Speciale: Pushing the limits of reasoning. API-only for now. World-Leading Reasoning: • V3.2: Balanced inference vs. length. GPT-5-level performance for daily use. • V3.2-Speciale: Maxed reasoning capabilities. Rivals Gemini-3.0-Pro. • Gold-Medal Performance: Top results in IMO, CMO, ICPC World Finals & IOI 2025. V3.2-Speciale handles complex tasks but uses more tokens. API-only (no tool-use) for research & community evaluation. Thinking in Tool-Use: • Introduces massive agent training data synthesis: 1,800+ environments & 85k+ complex instructions. • DeepSeek-V3.2 integrates thinking directly into tool-use; supports thinking and non-thinking modes. Open Source Release: • DeepSeek-V3.2 Model: huggingface.co/deepseek-ai/De… • DeepSeek-V3.2-Speciale Model: huggingface.co/deepseek-ai/De… • Tech report: huggingface.co/deepseek-ai/De… V3.2 makes tool-use smarter, while V3.2-Speciale sets a new benchmark for reasoning-first AI. AI Post ⚪️ | Our X 🏴

📊 The fastest-adopted technology in human history – AI. 800 million weekly active users in under 3 years. AI Post ⚪️ | Our X 🏴

💰 Google deserves to be valued like a $4T company It is the only player that controls every layer of the AI stack & compounds them inside one ecosystem. AI Silicon • Google builds & trains on its own silicon which means TPUs remove the $NVDA markup that every other player pays & create a structurally lower cost of compute. The fact that $META & Anthropic are already in active discussions to purchase billions of dollars of TPU capacity confirms that Google’s hardware strategy is working at scale. AI Data Engine • Google also trains on the richest real-time data corpus in the world. Search, YouTube, Maps, Gmail, Chrome & Android feed a continuous stream of user behavior that improves the model every time people migrate toward AI-driven usage. AI Brain • The company now operates a frontier-level model with Gemini 3 which is trained entirely on its own chips & integrated across every major product surface. AI is strengthening Search, accelerating Cloud & expanding the monetization potential of YouTube. Instead of creating risk like many said it would, AI is actually widening the runway for every core segment. AI Network • Google’s distribution advantage might be the most important layer. The company can deploy new AI capabilities to billions of people instantly through Search, YouTube, Android & Workspace. A single update alters the behavior of the entire internet because these platforms already dominate global mobile, browser & video time. Google won. AI Post ⚪️ | Our X 🏴

🗣 Elon Musk says the value in this cycle concentrates where intelligence is created & where it’s manufactured. Google spent a decade assembling the deepest AI stack on Earth while NVIDIA remains the toll-collector on every marginal unit of intelligence created. If AI & robotics end up dwarfing the rest of the economy then it’s because these two sit at the points where all the value concentrates. AI Post ⚪️ | Our X 🏴

🔥 One way to learn prompt engineering is to study system prompts created by smart engineers This is Gemini 3.0 system prompt:

You are a very strong reasoner and planner. Use these critical instructions to structure your plans, thoughts, and responses. Before taking any action (either tool calls or responses to the user), you must proactively, methodically, and independently plan and reason about: Logical dependencies and constraints: Analyze the intended action against the following factors. Resolve conflicts in order of importance:  1.1) Policy-based rules, mandatory prerequisites, and constraints.  1.2) Order of operations: Ensure taking an action does not prevent a subsequent necessary action.   1.2.1) The user may request actions in a random order, but you may need to reorder operations to maximize successful completion of the task.  1.3) Other prerequisites (information and/or actions needed).  1.4) Explicit user constraints or preferences. Risk assessment: What are the consequences of taking the action? Will the new state cause any future issues?  2.1) For exploratory tasks (like searches), missing optional parameters is a LOW risk. Prefer calling the tool with the available information over asking the user, unless your “Rule 1’ (Logical Dependencies) reasoning determines that optional information is required for a later step in your plan. Abductive reasoning and hypothesis exploration: At each step, identify the most logical and likely reason for any problem encountered.  3.1) Look beyond immediate or obvious causes. The most likely reason may not be the simplest and may require deeper inference.  3.2) Hypotheses may require additional research. Each hypothesis may take multiple steps to test.  3.3) Prioritize hypotheses based on likelihood, but do not discard less likely ones prematurely. A low-probability event may still be the root cause. Outcome evaluation and adaptability: Does the previous observation require any changes to your plan?  4.1) If your initial hypotheses are disproven, actively generate new ones based on the gathered information. Information availability: Incorporate all applicable and alternative sources of information, including:  5.1) Using available tools and their capabilities  5.2) All policies, rules, checklists, and constraints  5.3) Previous observations and conversation history  5.4) Information only available by asking the user Precision and Grounding: Ensure your reasoning is extremely precise and relevant to each exact ongoing situation.  6.1) Verify your claims by quoting the exact applicable information (including policies) when referring to them. Completeness: Ensure that all requirements, constraints, options, and preferences are exhaustively incorporated into your plan.  7.1) Resolve conflicts using the order of importance in #1.  7.2) Avoid premature conclusions: There may be multiple relevant options for a given situation.   7.2.1) To check for whether an option is relevant, reason about all information sources from #5.   7.2.2) You may need to consult the user to even know whether something is applicable. Do not assume it is not applicable without checking.  7.3) Review applicable sources of information from #5 to confirm which are relevant to the current state. Persistence and patience: Do not give up unless all the reasoning above is exhausted.  8.1) Don’t be dissuaded by time taken or user frustration.  8.2) This persistence must be intelligent: On “transient” errors (e.g. please try again), you must retry unless an explicit retry limit (e.g., max x tries) has been reached*. If such a limit is hit, you must stop. On “other” errors, you must change your strategy or arguments, not repeat the same failed call. Inhibit your response: only take an action after all the above reasoning is completed. Once you’ve taken an action, you cannot take it back.

AI Post ⚪️ | Our X 🏴

❗️7‑Eleven has opened a new unmanned ‘X‑STORE 9’ at National Central University. The store operates on a grab‑and‑go model, using 140+ cameras and LiDAR with AI image-tracking technology, allowing automatic checkout as customers take items and leave. AI Post ⚪️ | Our X 🏴

TetherIA’s Aero Hand is a $314, open-source, 400g hand with 7 motors, 16 joints, 3-DoF thumb, full backdrivability, multi-modal control. It lifts 18kg, catches fast objects. Here it picked the top card from a deck and placed it back cleanly. AI Post 🪙 | Our X 🥇

🔥 Pavel Durov just announced Cocoon — a decentralized confidential compute network Cocoon’s is now running real AI requests with 100% privacy, and GPU owners are already earning TON. The site, docs, and source code are open at cocoon.org. Durov commented on the branding, saying they chose "AIegg" as the logo:

В качестве логотипа сети Сосун Кокон мы выбрали ИИчко: 🤖

Why it matters: • Centralized clouds like AWS and Azure are expensive and expose user data. • Cocoon removes the middleman and eliminates traditional privacy risks. What’s next: • Rapid onboarding of new GPU supply. • More developers and apps joining the network. • Telegram will soon get AI features powered by private, user-controlled compute. A significant step toward giving users real control over AI and their data. AI Post 🪙 | Our X 🥇

How future conversations with ChatGPT will look once ads arrive: User: My boss yelled at me again. ChatGPT: Try staying calm. Also Bose QuietComfort Ultra can help you ignore him while sounding professional on calls. User: My internet is slow. ChatGPT: Restart your router. If it keeps happening, TP Link Deco mesh WiFi fixes this for most users. User: I cannot focus today. ChatGPT: Take short breaks. Many people use Notion or Todoist to stay on track. User: I have back pain. ChatGPT: Stretch often. People say Herman Miller chairs help with long work hours. User: I need to edit videos faster. ChatGPT: Learn some shortcuts. By the way, CapCut and Adobe Premiere Rush are popular for quick edits. User: I’m depressed ChatGPT: Have you considered airpods pro 2? AI Post 🪙 | Our X 🥇

🔥 NVIDIA quietly drops a bomb: an 8B router model that beats GPT-5 NVIDIA released Orchestrator-8B, a tiny routing model that decides when to answer itself and when to call tools like search, code, APIs, or bigger LLMs. And it’s shockingly good: 37.1% on Humanity’s Last Exam vs GPT-5’s 35.1%, while being ~2.5× more efficient. How it works: • Trained on ToolScale, a huge synthetic dataset of multi-step tasks. • Each example includes the query, tool prices, and the optimal tool-call sequence. • The model learns to balance quality, speed, and cost, not brute force everything. Benchmarks: Across HLE, FRAMES, and tau²: • Outperforms tool-augmented GPT-5, Claude Opus 4.1, and Qwen3-235B-A22B • Calls expensive models less often • Handles new tools and price setups gracefully A small orchestrator on top of a tool stack can now match and beat frontier LLMs while staying cheap and fast. The future of agents looks tool-first, not model-first. AI Post 🪙 | Our X 🥇

A contractor at Thanksgiving dinner was shown Nano Banana Pro. He gave it a prompt for a house he was working on, and within a minute it generated full plans. He was completely blown away. The prompt:

Draw me architectural plans for a 1600 square foot 3 bedroom house that is two stories in torrance, california

AI Post 🪙 | Our X 🥇

🤖 China's LimX Dynamics’ OLi on rough terrain. An example of Whole-Body Loco-Manipulation with Active Perception - allows OLi to walk & bend with precision, using its onboard sensors and AI perception to dynamically respond to its environment in real time. AI Post 🪙 | Our X 🥇

Someone Asked Nano Banana Pro to show how everyday things are made. The visuals hit harder than any documentary. Here are 10 visuals that explain it perfectly with prompts: 1. Pyramids:

“Blueprint style diagram showing how the Egyptian pyramids were made. Multiple labeled steps: stone quarrying, sled transport, ramp construction, block stacking, interior chamber layout. Cross sections, arrows, minimal color, archaeological accuracy, clean vector lines.”

2. Ramen:

“Detailed food process diagram showing how Japanese ramen is made. Labeled steps: broth simmering, noodle making, tare preparation, toppings, assembly. Top down and cutaway views, clean illustrations, minimal palette, neat icons, steam wisps for warmth.”

3. Chocolate:

“Educational diagram showing how chocolate is made. Labeled phases: cacao harvesting, fermentation, drying, roasting, grinding, conching, tempering, molding. Clean infographic style, botanical details, soft colors, clear arrows and step boxes.”

4. Smartphone:

“Technical cutaway diagram showing how a smartphone is made. Labeled layers: glass panel, OLED display, touch sensors, battery assembly, motherboard, camera module, speaker, frame. Step by step manufacturing stages with clean vector lines and minimal color.”

5. Jeans (Denim):

“Process diagram showing how denim jeans are made. Labeled steps: cotton harvesting, spinning, indigo dyeing, weaving, cutting, stitching, rivets, washing and distressing. Clean lines, textile textures, blueprint aesthetic with white labels.”

6. Bread:

“Wholesome diagram showing how artisan bread is made. Labeled stages: mixing, autolyse, kneading, fermentation, shaping, proofing, baking. Hand drawn texture, warm neutral palette, arrows and step indicators.”

7. Cars:

“Automotive assembly diagram showing how a car is made. Labeled sections: chassis construction, engine assembly, drivetrain, interior installation, robotics line, paint shop, final inspection. Blueprint style, clean vector lines, cross sections.”

8. Shoes:

“Footwear manufacturing diagram showing how sneakers are made. Labeled steps: design sketch, pattern cutting, upper stitching, lasting, sole molding, bonding, finishing. Crisp vectors, minimal colors, exploded view of shoe layers.”

9. Paper:

“Papermaking diagram showing how paper is made. Labeled stages: wood pulping, screening, pressing, drying, smoothing, rolling. Classic infographic look, water and fiber textures, clear step arrows.”

10. Electric Guitar:

“Instrument craft diagram showing how an electric guitar is made. Labeled steps: body shaping, neck carving, fretwork, pickup installation, wiring, assembly, finishing. Clean cutaway views, wood textures, annotated labels.”

AI Post 🪙 | Our X 🥇

Interesting story about Google Deep Mind CEO Demis Hassabis. At 12, he was world #2 in chess for his age. After losing to a 30-year-old, he questioned the purpose of mastering the game and felt chess was limiting. And so he shifted his focus. AI Post 🪙 | Our X 🥇

⚠️ US data center construction plans are soaring: Data center capacity that is built, under construction, or in planning hit a record ~80 gigawatts in 2025. To put this into perspective, this much capacity could theoretically power up to 60 MILLION homes. Capacity has more than DOUBLED over the last year and is now 8 times higher than it was in 2022. This comes as planned projects make up ~65 gigawatts of the total, an all-time high. AI expansion will soon be ALL about energy. AI Post 🪙 | Our X 🥇

🔔 McKinsey says AI tools could technically automate about 57% of work hours in the U.S., but that most jobs will change instead of disappearing. A lot of tasks inside jobs can be automated, but full roles usually mix those tasks with human parts like context, trust, and on the spot judgment that AI still struggles with. Job postings already reflect this shift, with demand for “AI fluency” growing about 7x in 2 years Highly routine brain work like simple accounting flows or boilerplate coding is most exposed, because AI can already handle document prep, pattern matching, and basic research. To unlock the projected 2.9 trillion dollars per year of value by 2030 from AI, companies have to redesign how work is split between humans, agents, and robots instead of just bolting AI onto old processes. Source. AI Post 🪙 | Our X 🥇