Preset
Background
Text
Font
Size
Width
Monday, March 2, 2026

The Git Times

“Sometimes it is the people no one imagines anything of who do the things that no one can imagine.” — Alan Turing

Reverse-Engineered APIs Unlock Neural Training on Apple Neural Engine

maderix/ANE enables backpropagation directly on Apple's powerful ANE hardware, bypassing CoreML and GPU for pure accelerator compute.

maderix/ANE Objective-C 555 stars

In a breakthrough for Apple Silicon developers, the maderix/ANE project delivers a from-scratch implementation of transformer training—complete with forward and backward passes—running natively on the Apple Neural Engine (ANE). This specialized hardware, clocking up to 15.8 TFLOPS on M4 chips, has long been reserved by Apple for inference only, leaving its immense potential for training untapped. By reverse-engineering private APIs like _ANEClient and _ANECompiler, along with the opaque MIL (Model Intermediate Language) format, ANE sidesteps official restrictions, enabling custom compute graphs on the ANE without relying on CoreML training tools, Metal shaders, or even the GPU.

The core innovation lies in crafting six specialized ANE kernels per training step, pushing the boundaries of what's possible on this inference-optimized accelerator:

  • kFwdAttn: Handles RMSNorm, QKV projections, scaled dot-product attention (SDPA), and output projection.
  • kFwdFFN: Executes RMSNorm and SwiGLU feed-forward networks.
  • kFFNBwd: Computes FFN backward passes with transposed weights.
  • kSdpaBwd1 and kSdpaBwd2: Break down SDPA backpropagation into manageable chunks for dV, probabilities, and gradients on Q/K.
  • kQKVb: Finalizes QKV backward to derive input gradients (dx).

On an M4 Mac, a single transformer layer (dim=768, seq=512) achieves 9.3 ms per step at 11.2% ANE utilization—sustaining 1.78 TFLOPS—with just six kernel dispatches. While forward and most backward passes (dx) run on ANE, CPU handles RMSNorm backward, residuals, loss computation, dW gradient accumulation via Accelerate's cblas_sgemm, and Adam optimization. This hybrid approach maximizes ANE's strengths in matrix-heavy ops while leveraging CPU for sequential logic.

What makes ANE technically mesmerizing are its ruthless optimizations, born from deep dives into ANE's quirks:

  • Channel-first CPU layouts matching ANE's IOSurface format [1,C,1,S], slashing transpose overhead to zero.
  • vDSP-accelerated RMSNorm, boosting speed 10x (from 6.7ms to 0.7ms).
  • GCD-async cblas overlap, parallelizing gradient computations with ANE evals on a serial dispatch queue.

These tweaks expose ANE's raw efficiency for training workloads, a feat previously unthinkable given Apple's lockdown. For machine learning engineers targeting Apple hardware—think on-device fine-tuning of LLMs or edge models—this opens a path to sub-10ms steps without power-hungry GPUs. It's not full end-to-end ANE training yet (dW and optimizer remain CPU-bound), but the project's checkpointing, gradient accumulation, and resume features make it production-ready for experiments.

Gaining explosive traction in just days, ANE spotlights the developer hunger for untapped Silicon potential. Written in Objective-C for low-level control, it invites tinkerers to extend kernels or scale to multi-layer transformers. As Apple doubles down on ANE in future chips, projects like this could redefine local ML training, proving reverse-engineering can turn black-box hardware into a trainable powerhouse.

(Word count: 462)

Builder's Take Apple Silicon devs with M4 or later should clone `maderix/ANE` and test the single-layer transformer demo on a release build to verify ANE access. Be cautious of private API instability across macOS updates—sandbox it in a VM first. For production, profile CPU-ANE overlap and contribute kernel expansions to push utilization beyond 11%.

More Stories

Agent-Reach Gives AI Agents Free Internet Access Tools

Python CLI enables reading Twitter, Reddit, YouTube and more without paid APIs

Panniantong/Agent-Reach · Python · 3.8k stars

OpenClaw Master Skills Curates Premier AI Agent Abilities

Weekly updated repository aggregates top OpenClaw skills from multiple sources for AI agents.

LeoYeAI/openclaw-master-skills · Python · 411 stars

AI Assistant Runs Biomedical Analyses via Chat Interfaces

Open-source tool integrates Claude Code and 140 skills for RNA-seq, drug discovery workflows

xjtulyc/MedgeClaw · HTML · 696 stars

ProxyCast Integrates AI Agents for Creator Workflows

Rust-based desktop platform streamlines idea capture to publishable content in one workbench.

aiclientproxy/proxycast · Rust · 1.2k stars

Vinext Reimplements Next.js APIs as Vite Plugin

Cloudflare tool enables Next.js migration to Vite with automated compatibility checks

cloudflare/vinext · TypeScript · 5.1k stars

Open Source AI Agents Shift to Modular Skills Ecosystems

Repos cluster around interchangeable skills, lightweight runtimes, and memory tools for autonomous coding agents.

trend/ai-agents · Trend · 0 stars

Open-Source Web Frameworks Fuel Deploy-Anywhere AI Ecosystems

Lightweight, self-hosted stacks mimic proprietary tools while integrating AI agents, breaking cloud vendor lock-in.

trend/web-frameworks · Trend · 0 stars

AI Agents Conquer the CLI: Open Source Dev Tools Go Agentic

Terminal-based AI coding assistants and proxies proliferate, blending LLMs with classic CLI workflows for smarter development.

trend/dev-tools · Trend · 0 stars

Quick Hits

elasticsearch Distributed RESTful search engine delivers scalable full-text search and analytics for massive datasets. 76.2k
timber AOT compiler transforms XGBoost, LightGBM, scikit-learn models into native C99 for 300x faster inference and one-command serving. 402
appwrite Open-source BaaS bundles auth, databases, storage, functions, realtime, and hosting for web, mobile, AI apps. 55k
nixpkgs Vast Nix package collection powers reproducible builds and declarative NixOS configurations. 23.7k
TrendRadar One-click monitors 35 trend platforms, filters keywords, generates reports, pushes alerts to WeChat, Telegram—no coding. 1.5k
dataclaw Python toolkit claws and processes web data with efficient extraction capabilities for rapid prototyping. 1.8k
homebrew-core Core formulae drive Homebrew package manager for seamless installs on macOS and Linux. 15.1k
opennews-mcp Aggregates crypto news with AI ratings, trading signals, and real-time updates for informed decisions. 596

AutoGPT Platform Enables Builders to Deploy Persistent AI Agents for Workflow Automation

Open-source toolset simplifies creating, hosting, and managing autonomous agents with Docker-based self-hosting and recent flow editor upgrades.

Significant-Gravitas/AutoGPT Python Latest: autogpt-platform-beta-v0.6.49 182.2k stars

In an era where AI agents promise to handle repetitive tasks, the AutoGPT platform stands out by providing developers with a robust framework to build, deploy, and run continuous AI agents. These agents tackle complex workflows autonomously, from tool integration to goal-oriented execution, freeing builders to focus on core logic rather than infrastructure.

At its core, AutoGPT leverages Python and integrates with models like GPT-4, Llama API, and now Claude via SDK. It supports self-hosting on Linux, macOS, or Windows with WSL2, using Docker Engine and Compose for containerized deployment. Hardware demands are modest—4+ CPU cores, 8GB RAM minimum—but scale to 16GB for production. A one-line script accelerates setup: on macOS/Linux, curl -fsSL https://setup.agpt.co/install.sh -o install.sh && bash install.sh; Windows users run a PowerShell equivalent.

The platform's latest release, autogpt-platform-beta-v0.6.49 from February 2026, introduces significant refinements. Builders gain a new flow editor replacing the legacy builder, enabling visual workflow design. Key additions include:

  • MCP tool block with OAuth and tool discovery for seamless integrations.
  • Copilot Executor Microservice and Claude Agent SDK for enhanced agent intelligence.
  • Workspace file tools, folder organization, and parallel tool execution support.
  • Telegram blocks and external API endpoints like POST /graphs for programmatic control.

Enhancements bolster reliability: improved WebSearch tools, chat session management, and transcript handling with stream reconnection. Agents now suggest responses for vague goals and wait for execution completion, reducing failure modes in long-running tasks.

Three years since its 2023 inception—with over 182,000 stars signaling sustained interest—AutoGPT evolves from a proof-of-concept to a mature platform. A cloud-hosted beta waitlist hints at managed options, but self-hosting remains free and flexible. For developers building AI-driven apps, it solves the orchestration challenge: no more wrestling with ad-hoc scripts when agents need persistence across sessions.

This matters to backend engineers, DevOps pros, and AI tinkerers automating CI/CD, data pipelines, or customer support flows. Its Docker-centric approach ensures portability, while active contributions keep it aligned with emerging LLM capabilities.

Builder's Take Ensure Docker, Node.js 16+, and Git are installed before running the setup script, as manual troubleshooting is technical. Test on a machine with 16GB RAM to handle agent sandboxes and parallel tools effectively. Review the official self-hosting docs for port configs and outbound HTTPS needs, especially if firewalled.

More Stories

Open WebUI Provides Self-Hosted LLM Interface

Offline platform integrates Ollama and OpenAI APIs with RAG for secure AI deployments

open-webui/open-webui · Python · 125.5k stars

Dify Builds Production-Grade LLM Workflows Visually

Open-source platform integrates agentic AI, RAG pipelines and model orchestration for rapid deployment.

langgenius/dify · TypeScript · 130.9k stars

Quick Hits

firecrawl Firecrawl turns entire websites into LLM-ready markdown or structured data via its Web Data API, ideal for AI builders scraping clean web content. 87.4k
langchain LangChain equips builders to engineer modular AI agents by chaining language models with tools and data sources. 128k
n8n n8n enables visual or code-based AI workflow automation with 400+ integrations, self-hostable or cloud-deployable. 177.2k
prompts.chat Prompts.chat lets builders share, discover, and self-host community LLM prompts for private, organization-wide use. 149.7k
tensorflow TensorFlow delivers a flexible open-source framework for builders to create, train, and deploy ML models at any scale. 194k

Openpilot Turns Stock Cars into Open-Source Autonomy Platforms

Comma.ai's Python-based robotics OS enhances driver assistance across 300-plus vehicle models with custom models and hardware.

commaai/openpilot Python Latest: v0.10.3 60.2k stars

Developers seeking hands-on experience with real-world robotics need look no further than openpilot, an open-source operating system from comma.ai that retrofits advanced driver assistance systems (ADAS) in over 300 supported cars. Launched in 2016, the project—written primarily in Python—delivers semi-autonomous driving capabilities by interfacing with a car's existing sensors, cameras, and controls via a dedicated hardware device.

At its core, openpilot solves the rigidity of factory-installed ADAS, which often lag behind cutting-edge AI models or lack customization. Builders plug in a comma 3X device (available at comma.ai/shop), connect it using a car-specific harness, and install the software via a simple URL like openpilot.comma.ai for the stable release branch. This setup unlocks features like adaptive cruise control, lane centering, and driver monitoring, all powered by openpilot's modular architecture.

The system supports multiple prebuilt branches tailored to hardware:

  • release-mici or release-tizi for production stability on comma four or 3X devices.
  • nightly for bleeding-edge development.
  • Experimental options like nightly-dev for car-specific features.

What sets openpilot apart technically is its focus on end-to-end neural networks for perception and control. The latest release, v0.10.3, introduces a new driving model (#36249) with a temporal policy architecture and on-policy training using a physics noise model—improving prediction accuracy in dynamic scenarios. A revamped driver monitoring model (#36409), trained on fresh datasets including comma four data, enhances safety by better detecting driver attentiveness. Efficiency gains come from optimized inter-process communication, reducing memory overhead for smoother real-time operation.

For robotics enthusiasts, openpilot bridges simulation and deployment: it processes camera feeds for object detection, path planning, and actuation, all runnable on commodity hardware. While plug-and-play requires comma gear, advanced users can adapt it to custom setups, though documentation warns it's not straightforward.

With steady development over nine years and more than 60,000 GitHub stars signaling broad interest, openpilot invites contributions via pull requests, issues, or its Discord community. Roadmap docs outline future expansions in supported models and robotics applications beyond cars. For builders in ADAS, autonomous systems, or embedded AI, this project offers unparalleled access to production-grade autonomy code.

(Word count: 378)

Builder's Take Start by verifying your car model against the 300+ supported list and acquiring a comma 3X with harness—custom hardware ports exist but demand significant debugging. Fork from a stable branch like `release-tizi` rather than master to avoid instability, and join the Discord for installation pitfalls. Review the new models' training details in release notes before contributing, as they emphasize temporal dynamics and on-policy methods key to robotics control.

More Stories

PX4 Autopilot Enables Open-Source Drone Autonomy

Modular stack runs multirotors, fixed-wing and rovers on Pixhawk hardware

PX4/PX4-Autopilot · C++ · 11.2k stars

Isaac Lab Unifies Robot Learning in NVIDIA Isaac Sim

GPU-accelerated framework streamlines reinforcement learning and sim-to-real robotics workflows

isaac-sim/IsaacLab · Python · 6.5k stars

Quick Hits

rerun Log, store, query, and visualize multimodal, multi-rate data with this Rust SDK to debug complex AI and robotics pipelines effortlessly. 10.3k
carla Simulate realistic autonomous driving worlds in C++ to test and train self-driving algorithms without real-world risks. 13.6k
nicegui Build interactive web UIs purely in Python for rapid prototyping of data apps and dashboards. 15.5k
copper-rs Deterministically build, run, and replay entire robot systems with this Rust OS for flawless robotics development. 1.2k
pinocchio Accelerate robotics with fast, flexible rigid body dynamics computations and analytical derivatives in C++. 3.1k

AI Agents Hunt App Vulnerabilities with Real Proof-of-Concepts

Open-source Strix tool deploys hacker-like bots to dynamically test code, slashing manual pentest delays and static scan errors.

usestrix/strix Python Latest: v0.8.2 20.7k stars

Developers and security teams have long grappled with application vulnerabilities: manual penetration testing drags on for weeks, while static analysis tools flood inboxes with false positives. Enter Strix, an open-source Python project from usestrix/strix that unleashes autonomous AI agents to mimic real hackers. Launched in August 2025 and actively updated through early 2026, Strix runs code dynamically, uncovers flaws, and validates them via executable proof-of-concepts (PoCs)—delivering actionable insights without the guesswork.

At its core, Strix fields teams of collaborative agents equipped with a full pentesting toolkit. These bots probe applications by executing code in a Docker sandbox, simulating attacks like those in real-world exploits. Unlike traditional scanners, Strix emphasizes real validation: agents generate PoCs to confirm vulnerabilities, then produce developer-friendly CLI reports with remediation steps, including auto-fix suggestions. This developer-first approach integrates seamlessly into CI/CD pipelines, blocking issues before production.

Key use cases span the dev lifecycle:

  • Application security testing: Detect critical flaws in web apps or APIs.
  • Rapid pentesting: Compliance-ready reports in hours, not weeks.
  • Bug bounty automation: Streamline research with PoC generation.
  • CI/CD gating: Embed scans to enforce secure deploys.

Getting started is straightforward. With Docker running and an LLM API key—supporting OpenAI, Anthropic, Google, or the Strix Router (offering $10 free credit)—install via a one-liner:

curl -sSL https://strix.ai/install | bash
export STRIX_LLM="openai/gpt-5"  # Or strix/gpt-5
export LLM_API_KEY="your-key"
strix --target ./app-directory

Results land in strix_runs/, complete with PoCs. For broader needs, the hosted platform at app.strix.ai connects repos and domains for instant pentests.

Recent v0.8.2 release (March 2026) refines the toolkit: dependency updates for Google Cloud AI Platform, a fixed Discord invite in docs, and a standout feature exposing the Caido proxy port for human-in-the-loop oversight—letting devs intervene mid-scan. New contributor @mason5052 polished community links, signaling growing momentum. With over 20,000 GitHub stars in seven months, Strix has traction among builders prioritizing secure, efficient workflows.

This isn't vaporware AI hype; Strix arms teams with precise, scalable security testing, bridging the gap between code velocity and vulnerability management.

Builder's Take Ensure Docker is running and secure an LLM API key before installing—Strix pulls a sandbox image on first use. Test on non-production apps initially to validate PoCs in your stack. For production, integrate into CI/CD early and explore the Caido proxy for custom workflows.

More Stories

SWE-agent Automates GitHub Fixes Using Language Models

Princeton-Stanford tool lets LMs autonomously resolve issues in real repositories via configurable YAML setup.

SWE-agent/SWE-agent · Python · 18.6k stars

Infisical Centralizes Secrets and PKI Management

Open-source platform syncs configs to GitHub, AWS; automates rotations and leak prevention

Infisical/infisical · TypeScript · 25.2k stars

Quick Hits

sherlock Sherlock hunts social media accounts by username across hundreds of platforms, ideal for swift OSINT reconnaissance. 73.3k
sniffnet Sniffnet delivers intuitive, Rust-powered monitoring of your internet traffic for easy network diagnostics. 32.9k
CheatSheetSeries OWASP Cheat Sheets pack concise, high-value app security guidance to fortify your development practices. 31.5k
nuclei Nuclei scans apps, APIs, networks, and cloud for vulnerabilities using fast, community-driven YAML templates. 27.3k
radare2 Radare2 empowers UNIX-style reverse engineering of binaries with a versatile command-line toolset. 23.2k

OpenAI's Codex Delivers Terminal-Native Coding Agent for Local Workflows

Rust-built CLI tool integrates ChatGPT models for lightweight, sandboxed code assistance without leaving the shell.

openai/codex Rust Latest: rust-v0.106.0 62.7k stars

Builders who live in the terminal now have a dedicated coding agent from OpenAI. Codex CLI runs entirely locally on macOS and Linux, leveraging Rust for a lightweight footprint. Unlike cloud-heavy alternatives, it processes tasks via your ChatGPT Plus, Pro, Team, Edu, or Enterprise plan—sign in once with codex and go.

The core appeal lies in its terminal-first design. Fire up codex for a text-based UI (TUI) that handles code generation, editing, debugging, and even sub-agent orchestration. It supports realtime WebSocket connections for live collaboration, with recent upgrades improving reliability through v2 protocol preference and retry logic on handshake failures (#12791, #12838).

Installation is straightforward, emphasizing platform-native binaries. Use npm i -g @openai/codex or brew install --cask codex for instant setup. For manual deploys, grab release assets like codex-aarch64-apple-darwin.tar.gz for Apple Silicon or codex-x86_64-unknown-linux-musl.tar.gz for Linux x86_64—extract and rename to codex. New direct install scripts in rust-v0.106.0 simplify this further (#12740).

Technically, Codex enforces sandboxing to contain executions, fixing zsh-fork paths that risked bypassing filesystem restrictions (#12800). Memory management got smarter with diff-based forgetting and usage-aware selection (#12900, #12909), while a ~1M-character input cap prevents TUI hangs on large pastes (#12823). Experimental features include js_repl promoted to /experimental (Node 22.22.0+ required, #12712) and request_user_input in default mode (#12735).

Recent momentum—over 62,000 GitHub stars in 11 months—signals strong builder interest, but the real value is in offline-capable agency. It complements IDE plugins for VS Code or Cursor, or scales to desktop/web variants, but shines in shell scripting and server-side workflows. Apache-2.0 licensed, with docs for contributing and building from source.

For teams, expanded app-server APIs add thread-scoped realtime endpoints and unsubscribe flows (#12715), enabling efficient multi-thread management without archiving. TUI tweaks hide absolute paths in file links while preserving references (#12705), and Ctrl-C now properly halts sub-agents.

Codex solves the friction of context-switching from terminal to browser or IDE for AI aid, delivering production-grade assistance where builders spend most time.

Builder's Take Start with `codex` after ChatGPT login for seamless plan integration; API key users need extra config. Verify Node 22.22.0+ for `js_repl` and test sandboxing on sensitive projects. Prioritize latest binaries over npm/brew for optimal Rust performance and security patches.

More Stories

Ollama Runs Open LLMs Locally Across Platforms

Go-based tool deploys Gemma, Qwen, Llama models with CLI, API and integrations

ollama/ollama · Go · 163.9k stars

llama.cpp Enables Fast LLM Inference in Pure C/C++

Minimalist library runs large models locally on diverse hardware with no dependencies.

ggml-org/llama.cpp · C++ · 96.3k stars

Quick Hits

ladybird Ladybird crafts a truly independent web browser engine from scratch, empowering devs to escape Big Tech dominance. 60.9k
uv uv delivers blistering-fast Python package installs and project management via Rust, turbocharging your workflows. 80.1k
tensorflow TensorFlow equips builders to train, deploy, and scale machine learning models across any hardware. 194k
rustdesk RustDesk enables secure, self-hosted remote desktop control as a flexible TeamViewer alternative. 108.5k
bitcoin Bitcoin Core powers full-node Bitcoin validation and integration for unbreakable blockchain reliability. 88.3k

Project Aura Brings Polished Air-Quality Monitoring to ESP32 Makers

Open-source station delivers comprehensive sensors, LVGL interface, and Home Assistant integration without soldering.

21cncstudio/project_aura C Latest: v1.1.0 270 stars

Indoor air quality affects health and productivity, yet most DIY monitors are crude sensor boards lacking polish. Project Aura changes that. This ESP32-S3-based station from 21cncstudio offers professional-grade telemetry—PM0.5, PM1, PM2.5, PM4, PM10 particulates, CO, CO2, VOCs, NOx, temperature, humidity, absolute humidity, pressure, and HCHO—in a touch-friendly package designed for makers.

Built in C with PlatformIO, Aura prioritizes reliability over barebones simplicity. Its LVGL UI runs smoothly on the ESP32-S3, featuring night mode, custom themes, status indicators, and screens for dashboard, settings, MQTT config, date/time, and backlight control. No soldering needed: it uses Grove and QT connectors for easy assembly with sensors like Sensirion SEN66 and SFA30.

Setup is straightforward. The device launches a Wi-Fi access point for onboarding, accessible via mDNS at http://aura.local. A local web portal at / handles initial config, while the new /dashboard in v1.1.0 provides tabs for live sensors, charts, events, settings sync, DAC control, and OTA updates. Upload a .bin file directly from the browser; the UI switches to a dedicated update page to prevent freeze illusions, with robust lifecycle handling and slot checks. Settings persist across updates, and Safe Boot rolls back on crashes.

MQTT integration shines for smart homes. Aura auto-discovers in Home Assistant, publishing sensor data and offering a ready dashboard. Optional GP8403 DAC output (0-10V) supports manual, timer, or threshold-based ventilation control—ideal for linking air quality to HVAC.

Hardware leverages Waveshare displays and Adafruit components, with a bill of materials and pin configs in the repo. A crowdfunding campaign on MakerWorld funds detailed build guides, 3D-printable enclosures, and wiring diagrams.

At one month old with steady traction (270 stars), Aura stands out for its full-stack approach: firmware, web UI, and ecosystem ties. Builders get a production-ready IoT device, not a prototype.

(Word count: 348)

Builder's Take Start with the repo's firmware source and PlatformIO for flashing; test in AP mode first to verify web setup. Review pin configurations and BOM before sourcing hardware, as enclosure files require crowdfunding backers. Enable MQTT discovery for quick Home Assistant integration, but customize thresholds for your environment.

More Stories

ESP32-S3 Flight Controller Stabilizes DIY Quadcopters

Cortex delivers real-time PID control, DShot ESCs and nRF24 telemetry for custom drone builds

sergiovirahonda/cortex · C++ · 109 stars

Quick Hits

litex Build custom FPGA hardware effortlessly with LiteX's modular open-source framework. 3.7k
gdsfactory Design photonics, analog, quantum chips, PCBs, and 3D objects intuitively via gdsfactory's Python library. 865
mkpl Generate M3U playlists instantly from the command line using mkpl. 51
espectre Detect motion spookily via Wi-Fi CSI analysis with ESPectre's Home Assistant integration. 6.6k
hackrf Transmit and receive radio signals cheaply with HackRF's software-defined platform. 7.7k
Memes section coming soon. Check back tomorrow!