Fatih

@hasantoxr github.com/Chachamaru127/…

I don’t trust Claude Code without this anymore. A developer built a free open-source harness that forces Claude Code to stop acting like a chaotic intern and start working like an actual engineer. It’s called Claude Code Harness. Most AI coding sessions fail for the same reason: You ask for a feature. Claude edits 9 files. It says “done.” You run the app. Something is broken. Now you have to reverse-engineer what it changed. Claude Code Harness fixes the workflow around the model. Instead of just prompting and praying, it gives Claude a real delivery loop: Plan. Work. Review. Release. That means Claude has to think before editing. Then execute. Then inspect its own work. Then prepare the final result. The best part is the built-in review system. It checks the work from multiple angles: - Security - Performance - Code quality - Accessibility So Claude does not just ship code because it feels confident. It has to pass through a structured review flow first. This is the missing layer in AI coding. Not a bigger model. Not a better prompt. Not another “10x dev” tweet. A harness. Because Claude Code is powerful. But without structure, power just becomes expensive chaos.

@bindureddy The Chinese Models are bench maxed. This means in real world they struggle and underperform

Kimi 2.6 BEATS Opus 4.7 And Is The BEST Open Source Model In The World Kimi 2.6 scores above Opus 4.7 low on LiveBench - a benchmark that can't be gamed It beats Opus on reasoning, coding and comes close on agentic coding We double-checked on internal evals this is a VERY GOOD MODEL at 10x less cost 🤯

@jun_song What is the difference to a gaming world?

중국이 모든걸 오픈소스화 하고있습니다. Tencent에서 방금 월드모델을 오픈소스로 일부 공개했습니다. (추후 전체 공개) HY-World 2.0 사용처는 제한되어있지만, 로컬에서 더욱 많은것들을 할 수 있다는것은 좋은 신호입니다.

@burkov Without you we would be lost

Based on the leaked Claude Code source code, your CLAUDE.md file is re-injected on every single **turn** of the conversation.

@NathanFlurry A True Innovation and Breakthrough, well done

Yep. We built a new operating system. → Powered by WASM & V8 isolates → node, python, bash, git, grep, curl (+129 more) → Mount anything: S3, Google Drive, SQLite, … → Embedded as a library, no special infra → Built on Actors & Secure Exec & Sandbox Agent Thread 🧵

@krzyzanowskim Repo is empty

I reimplemented "claude" CLI with codex and gpt-5.4-high. It cost $1100 in tokens, and is 73% faster and 80% lower resident memory during sustained interactive use. It is very easy to reverse claude from npm distribution, then reimplement is 1:1. It is indistinguishable from the Anthropic version to the every header and analytics it send back github.com/krzyzanowskim/…

@thsottiaux If something is free, you are the product

Hello. We have reset Codex usage limits across all plans to let everyone experiment with the magnificent plugins we just launched, and because it had been a while! You can just build unlimited things with Codex. Have fun!

@NathanFlurry I am especially interested in how to use this for agentic computers.

@NathanFlurry Can you make a long article and comparison? Also can you build an open source alternative?

It seems like Rivet & Cloudflare are building the same thing on the surface, but our approaches are very different: 🟠 Cloudflare is building "AWS for JavaScript:" - Assume everything runs in Workers on their cloud - V8 isolates instead of containers - Allows them to edge-ify your code cost effectively - Many real limitations because of the nature of being JavaScript-first & multi-tenant (memory, network, CPU, etc) vs 🔩 Rivet is building stateful- & codegen-focused infrastructure focused on portability: - Works with vanilla Node.js and Bun, no gotchas or limitations - Lets you deploy your backend to Railway, Vercel, K8S, AWS, Hetzner, etc - Library-first approach (RivetKit Actors, Secure Exec, Sandbox Agent) instead of bundling in to a JS runtime - All 3 libraries have experimental support for Rust & Python Under the hood, both architectures are drastically different. This goes without saying: the Workers is a killer platform with different use cases. Every CF employee I know is incredibly talented and puts immense care in to their work.

@clattner_llvm @Zyyon_ Single best move you can make. 5D Chess

@Zyyon_ If you believe AI should be open and want to work together on it, please join our community! modular.com/open-source/co… There is still a lot to do, this is what happens when you're on an entirely new frontier. Mojo 1.0 is coming this summer!

Hold on What happened to Mojo?

@__morse Standalone Repo

install with npm i -g usecomputer works with macOS & linux only right now source code: github.com/remorses/kimak…

introducing usecomputer a cli written in Zig to let any agent control a computer mouse & keyboard the hardest part of controlling a remote agent is accepting macOS permission dialogs. here is an example of Opus enabling a security setting via useomputer, kimaki & opencode works in codex, opencode, claude code, anywhere!

@sudoingX He was free until he was bought. He cannot talk freely anymore

the founder of openclaw joined the company that was founded to make AI open and now charges you per token. and is now telling you open models aren't there yet. i run qwen 3.5 27b on a single 3090. 50 tok/s. it writes code, handles tool calls, runs agent sessions for hours. the model built a full space shooter, 3,000+ lines, from a single prompt. i published the data. "open models aren't there yet" is what you say when your harness can't parse tool calls on local models and you blame the model instead of fixing the harness. i have the DMs. people switch from openclaw to hermes agent and their "broken" models suddenly work. pair a good model with a good harness like hermes agent where parsers are built per model. your data stays on your machine. no API key. 0 subscription. no one training their next model on your thinking. don't listen to someone with an OpenAI paycheck telling you open source can't do the job. install it. test it yourself. the receipts are on my timeline. he built a harness that couldn't handle local models and chose the API paycheck over fixing it. that should tell you everything.

@trillhause_ github.com/rawwerks/ypi

If you tried pi coding agent and it blew your mind. You should try ypi if you want to experience early sparks of agi. Required some setup w/ custom tools for me to realize its power. Wouldn't be surprised if all major agent harnesses adopt RLM in the next few months.

@testerlabor I will tell you at what price: Burned out xAI people with chronic illnesses. They burn theough people’s brains and lives

Did you know that xAI has developed faster with Grok than all other AI relative to its young age? xAI / Grok is the overall winner in velocity! Here's an example, but it's the same with all other AI companies.

@ojaskandy Early access now ☺️

@fatih2211 great! releasing here soon: moonshot.computer

We just hacked macOS to build Moonshot, the first AI agent with its own computer on your Mac! Moonshot works as a separate macOS user alongside you. 2 users, 1 mac. Join the waitlist for early access at moonshot.computer comment "MOON" for priority access

@HuzayfaJasat He ist just a human bro

The Bay Area is crazy. I just met the COO of Apple at an Iftar event.

@ivanburazin The moment of a major setback will show how much pain you are able to handle.

My first job was selling popcorn on the boardwalk. Then shoveling snow. Then systems administrator. Then IT manager. Then pitched my boss: "Let's spin out the IT department as a company." Expected him to say no. He said yes. First entrepreneur experience: stacking servers, pulling cables across ceilings, building data rooms from scratch. Sold that company to a local competitor. Then started a developer conference and scaled it to thousands of attendees with events held across the world. Sold it to a global comms company with a $2B run rate. Then got back into infra and sold runtime dev environments to enterprises and scaled it to $300k ARR. Realized building for agents presented a far bigger and meaningful opportunity and pivoted. Started selling sandbox infra across the spectrum (from early stage YC companies to F100s) and scaled to $3M ARR in 90 days. Love the trajectory so far and can't wait to make it look even crazier.

@BrianRoemmele @Scobleizer Just a different perspective Brian, keep up the great work

Let’s cut through the hype. This is a clever research breakthrough, but it is nowhere near a production-ready shift in AI infrastructure. Reverse engineering a closed chip is fragile by nature, and Apple can shut this down with a single macOS update if they choose. The “80× more efficient than an A100” comparison is also misleading because the ANE and A100 are built for completely different workloads and scales. Quoting 38 TOPS versus 19 TFLOPS is not exposing a lie, it is just comparing different numerical formats like INT8 and FP16, which chip vendors do all the time. Training a small neural network on the ANE does not mean you can meaningfully train large models, since memory bandwidth, tooling, and developer support are extremely limited. There is no mature ecosystem, no stable PyTorch integration, and no guarantee this will ever be officially supported. Right now this is an impressive technical hack that proves possibility, not a sign that Apple Silicon is about to disrupt NVIDIA or redefine AI compute.

BOOM! Apple’s Neural Engine Was Just Cracked Open, The Future of AI Training Just Change And Zero-Human Company Is Already Testing It! In a jaw-dropping open-source breakthrough, a lone developer has done what Apple said was impossible: full neural network training– including backpropagation – directly on the Apple Neural Engine (ANE). No CoreML, no Metal, no GPU. Pure, blazing ANE silicon. The project (github.com/maderix/ANE) delivers a single transformer layer (dim=768, seq=512) in just 9.3 ms per step at 1.78 TFLOPS sustained with only 11.2% ANE utilization on an M4 chip. That’s the same idle chip sitting in millions of Mac minis, MacBooks, and iMacs right now. Translation? Your desktop just became a hyper-efficient AI supercomputer. The numbers are insane: M4 ANE hits roughly 6.6 TFLOPS per watt – 80 times more efficient than an NVIDIA A100. Real-world throughput crushes Apple’s own “38 TOPS” marketing claims. And because it sips power like a phone, you can train 24/7 without melting your electricity bill or the planet. At The Zero-Human Company, we’re not waiting. We are testing this right now on real ZHC workloads. This is the missing piece we’ve been chasing for our Zero Human Company vision: reviving archived data into fully autonomous AI systems with zero human overhead. This is world-changing. For the first time, anyone with a Mac can fine-tune, train, or iterate massive models locally, privately, and at a fraction of the cost of cloud GPUs. No more renting $40,000 A100 clusters. No more waiting in queues. No more massive carbon footprints. Training costs that used to run into the tens or hundreds of thousands of dollars? Plummeting toward pennies on the dollar – mostly just the electricity your Mac was already using while it sat idle. The AI revolution just moved from billion-dollar data centers to your desk. WE WILL HAVE A NEW ZERO-HUMAN COMPANY @ HOME wage for equipped Macs that will be up to 100x more income for the owner! We’re only at the beginning (single-layer today, full models tomorrow), but the door is wide open. Ultra-cheap, on-device training is here. The future isn’t coming. It’s already running on your Mac. Welcome to the Zero-Human Company era.