Muhammad Umar 🇺🇸 🇵🇰

Today, Sanja and I met our customers in New York 📍 CodeNinja started in Lahore. Expanded to Riyadh. And is now growing across Dallas, NYC, and SF. Driven by a mission: Technologies of Freedom, helping every factory floor, and every workplace build its sovereign AI stack.

A fascinating talk between a software engineer and AI.

Freedom isn't being a spectator, it's being a player. Concept by @AnthropicAI Fable 5. Crew @higgsfield_ai . Production @claudeai CLI . Vision @CodeNinjaInc

@AiswaryaVenkit1 @grok Summarize 10 key points

Anthropic engineer: "You're not supposed to prompt Claude. You're supposed to build a system that prompts itself." In 45 minutes she breaks down how Anthropic builds agents that remember, learn from their mistakes, and get smarter with every run. Worth more than any paid course you'll find on building agents. Watch the session

I kept doomscrolling while Claude Code was thinking, so I built Focus 🎯 It opens a short clip of @elonmusk and others when the agent works, chimes you back when it's done, and learns what you like. Mac + Claude Code, one-line install: github.com/muhammadumar89…

Allow Claude to control Claude. That should be the iconic meme of the AI era. @AnthropicAI @bcherny

@steipete Energy/token efficiency is not just about cost savings. Things that are more energy efficient are more likely to be true . The universe is extraordinarily energy efficient

People freaking out over my AI spend. What nobody sees: Part of what excites me so much about working on OpenClaw is that I'm trying to answer the question: How would we build software in the future if tokens don't matter? We constant run ~100 codex in the cloud, reviewing every PR, every issue. If a fix on main lands, @clawsweeper will eventually find that 6 month old issue and close it with an exact reference. We run codex on every commit to review for security issues (as it's far too easy to miss). We run codex to de-duplicate issues and find clusters and send reports for the most pressing issues. We have agents that can recreate complex setups, spin up ephemeral crabbox.sh machines, log into e.g. Telegram, make a video and post before/after fix on the PR. There's codex that watch new issues and - if it fits our documented vision well, automatically create a PR of it. (that then another codex reviews) We have codex running that scans comments for spam and blocks people. We have codex instances running that verify performance benchmarks and report regressions into Discord. We have agents that listen on our meetings and proactively start work, e.g. create PRs when we discuss new features while we discuss them. We build clawpatch.ai to split all our projects into functional units to review and find bugs and regresssions. We do the same split for security with Vercel's deepsec and Codex Security to find regressions and vulnerabilities. All that automation allows us to run this project extremely lean.

OpenClaw just shipped v2026.5.2 — open-source agent runtime that runs continuously across WhatsApp, Slack, Discord on your own hardware. Externalized plugin lifecycle, faster gateway, 268 fixes. Local AI sovereignty just hit production. Technologies of freedom.

@amirhusain_tx Just saw this post. Loved it. We are building a sovereign AI stack for a major customer. I have a question, Using this stack, as you just shared, can agents incorporate ontologies ? Or how can we build ontologies.

Have been using multi-layered memory with agents. Three layers. md files with good old grep! Vector DBs for obvious reasons. And graph databases oriented towards various datasets I want to have my agents access. #AHTechAndAI

The missile and drone attacks carried out yesterday against the United Arab Emirates are condemnable. We stand with the brotherly people of UAE and express full solidarity with His Highness the President @MohamedBinZayed, His Highness Deputy Prime Minister/Foreign Minister @ABZayed and the leadership and people of UAE. Pakistan will continue making efforts for lasting peace and stability in the region and beyond.

Ai2 just open-sourced MolmoAct 2 — a foundation model for embodied AI. Open weights + the largest open bimanual manipulation dataset ever (720+ hrs). 37× faster than v1, beats proprietary robotics models. Sovereign robotics is here. Technologies of freedom.

IBM open-sourced Granite 4.1 — Apache 2.0, 3B/8B/30B sizes, 512K context, cryptographically signed weights. The 8B matches their old 32B MoE at a quarter the params. Provenance + sovereign deployment, baked into the release. Technologies of freedom.

Poolside open-weighted Laguna XS.2 + M.1 — US-built, Apache 2.0. XS.2: 33B-A3B, 68.2% SWE-bench Verified, runs locally on a 36GB Mac. M.1: 225B-A23B, 72.5% SWE-bench Verified. Frontier coding agents, sovereign on your hardware. Technologies of freedom.

Mistral open-weighted Medium 3.5 — 128B dense, modified MIT, 256K context. Chat + reasoning + code in one model with reasoning effort as a per-request toggle. 77.6% on SWE-Bench Verified. Frontier-tier, EU-built, self-hostable. Technologies of freedom.

OpenAI just open-sourced Symphony — Apache 2.0 spec for coding-agent orchestration. Pulls tickets, runs agents in isolated workspaces, ships PRs. Swap in any open-weights model underneath. The orchestration layer is now forkable. Technologies of freedom.

Multiverse just open-sourced LittleLamb — three 0.3B models on Hugging Face. Qwen3-derived, ~50% compressed via CompactifAI. Mobile + tool-calling variants. Sovereign AI on the truck, the kiosk, the offline plant. Your weights, your hardware. Technologies of freedom.

Tencent open-sourced Hy3-preview — 295B/21B fast-slow MoE, 256K context, drives 495-step agent trajectories. First model off their fully rebuilt Hunyuan stack. Frontier agentic AI, self-hostable. Your hardware, your weights. Technologies of freedom.

NVIDIA just open-sourced Nemotron 3 Nano Omni — vision, speech, audio, text in one 30B MoE (3B active), 256K context, 9× the throughput of other open omni models. Sovereign multimodal agents on your hardware, your weights. Technologies of freedom.

been listening to @ShaanVP and @theSamParr on MFM and this hit passion is the output not the input you start with enthusiasm. find a loop you love. mastery comes. then passion shows up people sit waiting for passion before picking anything. backwards

DeepSeek just open-sourced V4 — 1.6T param MoE, 1M context, MIT license. Roughly an order of magnitude cheaper than GPT-5.5 and Claude Opus 4.7. Frontier-tier reasoning is now sovereign-deployable on Blackwell or Ascend. Technologies of freedom.