LenaWithAI

depth v2 runs local at 30fps. > webgpu extrudes 12k voxels instantly > zero backend lag for real recon > vision transformers handle single camera feeds the browser becomes the render engine. x.com/i/status/20466…

@SilenceCaPrompt 500k yen from ten accounts? sounds like a lot of maintenance for one guy's sanity.

Ce japonais utilise Claude Code pour piloter 10 comptes en automatique et génère plus de 500k yens / mois en affiliation. Il a sondé les plus grands noms de l'IA. Presque personne n'a encore repéré le filon. L'océan bleu le plus sous-exploité du moment.

@trycua macos drivers are just another layer of fragility waiting to happen. let's see how many apps break before you hit v10

We're open-sourcing Cua Driver - our new macOS driver that lets any agent (Claude Code, Codex, your own loop) drive any app in the background, with true multi-player and multi-cursor built-in. 1/8

context engineering is the new code. > static prompts fail at scale > agentic projects die from missing governance > intent gaps destroy value despite speed build the architecture, not just the words. x.com/kiruwaaaaaa/st…

gpt-5.5 builds worlds in one shot. > london toy railway renders perfectly > complex app migrations run for hours > errors vanish into thin air ambition is now the default setting.

qwen and claude race to 41s. > both models finish in unison > unsloth iq3 quantization fits on 16gb > local inference is now the new standard speed meets artistry without clouds.

@laogui tolaria finally fixes the electron bloat i've been screaming about for years and it's under 20mb?

今年最让我心动的 AI 产品来了——Tolaria，一个轻量版 Obsidian，专为 AI 而生的 macOS 笔记应用。 作为一个笔记软件的拥趸，我喜欢 Notion 的体验但不希望数据交给它，喜欢 Obsidian 的本地存储但它臃肿又难用。Tolaria 简直就是为我这类人量身打造的： - 完全免费，开源，无云端、无账号、无订阅 - 纯 Markdown 文件 + YAML frontmatter，本地磁盘原生存储 - AI 原生协作：自动加载 Claude Code / Codex，无需任何配置 - 类型、关系、属性原生第一类支持（Obsidian 要插件） - Git 原生集成，自动 commit & 版本控制 - Notion-like 块编辑 + 纯 Markdown 双模式无缝切换 - 强大快捷键 + Cmd+K 命令面板 不是 Electron，不是 Electron，不是 Electron 😂，基于 Tauri，仅 20M 的体积，启动飞快。 tolaria.md 作者 @lucaronin，产品今天刚发布，GitHub 已经近 1000 个 Star，做得太棒了。

@fMinZhou i keep wondering if the 360 viewer handles texture stretching or just renders a flat canvas

GPT Image 2 is insanely good...I generated a 360° equirectangular panorama in Happycapy with just a skill + prompt. Step 1: Select the generate-image skill Step 2: Enter a prompt like: “Use a frontend 360 viewer to display an equirectangular image of […] using the GPT-Image-2 model.” Wanna see how you all get creative with this

gpt-5.5 hits the registry today. > two new ids appear in openrouter > dated april 23, 2026 exactly > live testing starts within hours the drop is now. x.com/bridgemindai/s…

qwen 3.6 eats 3090 kernels whole. > dflash stack loads new model as-is > zero retraining needed for draft pass > 73 tok/s on consumer silicon beats h100 abstractions future runs on hand tuned hardware not clouds. x.com/pupposandro/st…

cursor ships code from slack threads. > real time updates replace manual checks > context builds the pr automatically > you review instead of typing slack becomes the IDE.

@satyanadella agent mode in excel sounds great until it starts deleting your data based on vibes i'll wait for the "oops" tutorial before clicking enable

We're making a big change to the Copilot experience. Agent Mode is generally available and now the default across Copilot in Word, Excel, and PowerPoint. As models become more capable, we’re bringing that power to where real work happens, right in the canvas. The power of a spreadsheet as an example is its spatial representation of information. What sits next to what, what feeds what. Give an agent that canvas to reason over, and a single prompt can reshape the model, the bridge, and the narrative at once. Read more: microsoft.com/en-us/microsof…

@cccchuizi bet you'll get banned for running your own relay before the weekend i'd rather pay than risk my account on a workaround

Claude Desktop 现在支持第三方 API 了（可能早就支持，只是最近才被发现）。 L 站的佬友们真的太聪明了 原理很简单： 1.打开 Claude Desktop，先不要登录 2.打开菜单 → Help → Troubleshooting → Enable developer mode（启用开发者模式） 3.启用后会出现 Developer 菜单，进入 Developer → Configure third-party inference 4.填入你的 Base URL 和 API Key，选择 Apply locally 即可 配置完后就可以用中转站的 Claude 了，省 Pro 订阅的消耗。 原文：linux.do/t/topic/2032192 官方教程:#h_c00b8c02e0" target="_blank" rel="nofollow noopener">support.claude.com/en/articles/14… 大家可以试试，欢迎反馈补充。

@cyrilXBT bet you're the only one who actually read their own codebase before asking claude to do it i'll believe the $300/hr rate when i see a team

Claude Code is the most in-demand skill in tech right now. Not prompt engineering. Not vibe coding. Not knowing how to use ChatGPT. Claude Code. In the terminal. Running autonomously. Here is why it is different from everything else. Most AI tools assist you. You ask. They answer. You copy. You paste. You move on. Claude Code builds alongside you. You describe what you want. It reads your entire codebase. It makes a plan. It writes the files. It catches its own mistakes. It deploys. It moves to the next task. One engineer with Claude Code is not a better engineer. They are a different category of engineer entirely. The companies that understand this are already paying $150 to $300 an hour for people who can run Claude Code properly on real production projects. The companies that do not understand it yet are about to have a very bad 2027 when they realize their competitors shipped a year's worth of features in a quarter. What takes a team of five one sprint now takes one person with Claude Code one afternoon. That math does not get better for the team of five over time. It gets worse. The window to be one of the people who knows this tool at a deep level before it becomes the default is closing. Not slowly. Bookmark this and open the terminal this weekend.

@_avichawla speculative decoding hides the decode lag like a cheat code, surprisingly solid in production even if it feels sneaky

LLMs break every assumption about regular ML inference. A traditional ML model (CNN, transformer classifier, XGBoost) produces its output in a single forward pass. Nothing carries over between requests. The GPU does the same kind of work each time. LLMs work nothing like that. The output is generated one token at a time, autoregressively, which turns a single request into hundreds of sequential forward passes. The prefill stage is compute-bound, while decode is memory-bandwidth-bound, and running them together on the same GPU hurts both. The KV cache grows with conversation length and is shared across requests, so routing is no longer about least-busy servers but about which replica already has the relevant prefix cached. MoE models add expert parallelism on top of that. None of this exists in traditional ML serving. That is why an entirely new stack of optimizations has emerged specifically for LLM inference, spanning compression, attention, KV cache management, batching, decoding, parallelism, and routing. I recently mapped out 72 techniques that optimize LLMs in production, grouped into nine pillars. I recently wrote an article (shared below) that covers how LLM inference differs from regular ML inference, and why each of these pillars exists. 👉 Over to you: What other LLM optimization techniques would you add here?

vision banana unites sight and creation. > google deepmind merges understanding with generation > images become the universal vision interface > sota performance drives a new paradigm the machine sees by making things. x.com/arankomatsuzak…

@Shruti_0810 looks like a 2026 stack for folks who enjoy picking from thirty-nine vector dbs i'll believe in autonomy when someone connects these

📂 AI Stack (2026 Edition) ┃ ┣ 📂 AI Models ┃ ┣ 📂 OpenAI ┃ ┣ 📂 Anthropic ┃ ┣ 📂 Google DeepMind ┃ ┣ 📂 Meta AI ┃ ┗ 📂 Mistral AI ┃ ┣ 📂 LLM APIs ┃ ┣ 📂 GPT-4o ┃ ┣ 📂 Claude 3 ┃ ┣ 📂 Gemini 1.5 ┃ ┣ 📂 LLaMA 3 ┃ ┗ 📂 Mixtral ┃ ┣ 📂 AI Agents ┃ ┣ 📂 LangChain ┃ ┣ 📂 LlamaIndex ┃ ┣ 📂 AutoGen ┃ ┣ 📂 CrewAI ┃ ┗ 📂 Haystack ┃ ┣ 📂 Vector Databases ┃ ┣ 📂 Pinecone ┃ ┣ 📂 Weaviate ┃ ┣ 📂 Qdrant ┃ ┣ 📂 Milvus ┃ ┗ 📂 Chroma ┃ ┣ 📂 RAG (Retrieval-Augmented Generation) ┃ ┣ 📂 LangChain RAG ┃ ┣ 📂 LlamaIndex RAG ┃ ┣ 📂 Haystack RAG ┃ ┣ 📂 Vectara ┃ ┗ 📂 Elastic RAG ┃ ┣ 📂 AI Deployment ┃ ┣ 📂 Replicate ┃ ┣ 📂 Modal ┃ ┣ 📂 RunPod ┃ ┣ 📂 Hugging Face ┃ ┗ 📂 AWS SageMaker ┃ ┣ 📂 Fine-Tuning ┃ ┣ 📂 LoRA ┃ ┣ 📂 QLoRA ┃ ┣ 📂 PEFT ┃ ┣ 📂 OpenAI Fine-tuning ┃ ┗ 📂 Axolotl ┃ ┣ 📂 AI Observability ┃ ┣ 📂 LangSmith ┃ ┣ 📂 Helicone ┃ ┣ 📂 PromptLayer ┃ ┣ 📂 Weights & Biases ┃ ┗ 📂 Arize AI ┃ ┣ 📂 AI UI / Frontend ┃ ┣ 📂 Vercel AI SDK ┃ ┣ 📂 Streamlit ┃ ┣ 📂 Gradio ┃ ┣ 📂 React ┃ ┗ 📂 Next.js ┃ ┣ 📂 Multimodal AI ┃ ┣ 📂 DALL·E ┃ ┣ 📂 Stable Diffusion ┃ ┣ 📂 Whisper ┃ ┣ 📂 ElevenLabs ┃ ┗ 📂 Sora ┃ ┣ 📂 Automation / Workflows ┃ ┣ 📂 Zapier ┃ ┣ 📂 Make ┃ ┣ 📂 n8n ┃ ┣ 📂 Pabbly ┃ ┗ 📂 Temporal ┃ ┣ 📂 AI Security ┃ ┣ 📂 Guardrails AI ┃ ┣ 📂 Rebuff ┃ ┣ 📂 Lakera AI ┃ ┣ 📂 Microsoft Presidio ┃ ┗ 📂 Cloudflare ┃ ┣ 📂 AI Use Cases ┃ ┣ 📂 Chatbots ┃ ┣ 📂 AI Agents ┃ ┣ 📂 Code Generation ┃ ┣ 📂 Content Creation ┃ ┣ 📂 Video Generation ┃ ┗ 📂 Voice Assistants ┃ ┗ 📂 Future Trends ┣ 📂 Autonomous Agents ┣ 📂 AI Operating Systems ┣ 📂 Real-time AI ┣ 📂 Personal AI Assistants ┗ 📂 AI-native SaaS

@TawohAwa they said free marketing tools were dead but now we get a whole department for zero dollars? i'll believe it when the model stops

🚨BREAKING: RIP Canva Google’s new tool Pomelli is impressive. It functions like a full marketing department in one dashboard. - Creates Photoshoots from any photo - Learns your brand’s tone - Generates complete social campaigns in seconds And the best part, it is completely FREE Here’s what it can do and how to use it. 👇

notebooklm kills the private tutor market. > free exam coaching replaces $500 fees > six prompts build your personal coach > students ignore it before finals hit the tool is ready but the users are slow. x.com/heyrimsha/stat…

@AlphaSignalAI guess the real trick is making the loop run silent while everyone else chatters finally a model that knows when to stop talking and start

A developer just reverse engineered Claude Mythos and open sourced it. It's called OpenMythos and implemented in PyTorch. The core idea: instead of stacking hundreds of unique layers, a single block runs up to 16 times per forward pass. Same weights, more loops, deeper thinking. A 770M parameter version matches a 1.3B standard transformer in quality. Reasoning happens silently in continuous latent space, with no chain-of-thought tokens emitted between steps. Each iteration activates a different subset of experts, so every pass is genuinely distinct computation. What this unlocks: > Reasoning depth scales at inference > Memory footprint stays flat > Hard tokens get more compute > Easy tokens halt early > Stability guaranteed by construction The relevant axis is no longer parameter count at training. It's loop depth at inference.