Moez AI

New Unsloth Studio update! 1. 10x faster via pre-compiled llama.cpp + mamba binaries 2. 6x faster, -50% less disk space installs via bun, uv 3. Studio is now in PATH + `unsloth studio update` works 4. Lots of UI UX improvements And my fav: Desktop + launch shortcuts for Studio!

MinerU-Diffusion A 2.5B diffusion-based OCR model that replaces slow autoregressive decoding with parallel block-wise diffusion, achieving up to 3.2x faster inference while improving robustness on complex documents with tables, formulas, and layouts.

Vibecoded TurboQuant looks really promising: - 3.25 bits, 4.9x compression - 4.25 bits, 3.8x compression Just waiting for llama.cpp to fully support this beast... I’ll hand off all simple agentic tasks toQwen3.5 27B. github.com/TheTom/turboqu…

TurboQuant AI models waste massive memory on vectors. Compressing them usually adds overhead defeating the purpose. Google's new paper uses just 1 extra bit to eliminate that overhead. Result: same accuracy, way less memory. Accepted at ICLR 2026. The trick? Random rotations + a 50-year-old math theorem. Here is a deepmanim.com overview of the paper. #manim

How chronic inflammation in the gut can increase risk of colon cancer through cumulative epigenetic memories (in the mouse model) @Nature @NatureNV nature.com/articles/s4158… nature.com/articles/d4158…

TurboQuant: Redefining AI Efficiency with Extreme Compression! #BigData #Analytics #DataScience #AI #MachineLearning #NLProc #LLM #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode geni.us/TurboQuant

The Death of CNNs! #BigData #Analytics #AI #MachineLearning #DataScience #IoT #IIoT #Python #RStats #TensorFlow #JavaScript #ReactJS #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode geni.us/Death-of-CNNs

This looks super promising! And it is Apache 2.0!

Finally. A native memory plugin is exactly what OpenClaw needs. A hierarchical, file-based architecture lets the agent actually structure its thoughts accurately instead of just doing similarity matching.

Interesting result - based on the blog looks only marginally better than the traditional product quantization at the same bits, but no need for code books and expensive memory lookups. Curious to see the e2e results on ANN bench vs PQ etc

Google DeepMind has unveiled a browser powered by its Gemini 3.1 Flash-Lite model that generates entire websites in real time as users browse. Google’s Flash-Lite Browser treats the web like something an LLM can write live, not something humans must fully pre-build first. A normal site serves stored pages and templates, but this system uses Gemini 3.1 Flash-Lite to generate fresh HTML and CSS from your prompt, clicks, and navigation context almost instantly. The technical shift is simple: instead of fetching a finished page, the browser asks the model what page should exist right now, then streams that answer as interface code. That makes personalization much deeper, because the page can change for each user, each step, and each goal without keeping a huge library of prewritten screens. It also fits agentic workflows, where an AI assistant may need to create a temporary tool, dashboard, or reference page on the fly while working through a task. IMO, the catch is reliability, because once page layout and content are model outputs, bugs, hallucinations, style drift, and serving cost - all become concerns.

@stochasticchasm This alone x.com/cursor_ai/stat…

Excited to finally be releasing AutoGaze! Check it out autogaze.github.io (and 👀 the video demo)

Compound Planning - if you've been using plannotator consistently, then there is an opportunity to improve how your agents plan for you. We're going to introduce a skill that enables you to see your own insights and eventually will create a automated feedback loop. The point is to consistently refine and optimize the planning that works best for you. @pyrons_ thank you for the good idea - and this is a similar analysis I did when looking into insights for the quick label feature. Preview of what the MVP skill outputs

I have created my own Agent Board for multiple agents working together using @badlogicgames' pi in the background. It features a fun 3D office where the agents work, and it utilizes @LakshyAAAgrawal's GEPA to auto-optimize the agent system prompts and system prompt templates

recommended viewing! Marlene has great energy and if you aren't well versed in agentic coding yet, this is a great intro talk.

sim overlay on real! :) i think it might be wise to record the lens distortion/fovy characterisitics in the future to help resolve the slight mismatches right now.

Plannotator 0.15.0 is here. The War of the Code Review continues. We're fighting clankers with clankers now. • Live AI chat in code review (Claude, Codex, Pi, OpenCode). This is the mvp - a lot to refine from here. • Folder-based file viewer for easy doc reference (superpowers, specs, etc) • Browse all your previous plans • Full feature parity for the Pi.dev extension Plus resizable diff panes and various bug fixes! (11 PRs merged)

People of pi, I have begun incorporating pi-autoresearch into github.com/rcarmo/piclaw. @davebcn87 did an amazing job, and it "just works" with the built-in terminal (launched inside the web UI after defining the experiment). will look at integrating it as a fully graphical UX...

Guys.. this model is just crazy. If you have just less than 48gb vram, just try the 8q gguf format. Feels just like opus! Tool calling is working smoothly!! Appreciate for this! (Hf and qwen!!) huggingface.co/Jackrong/Qwen3…