Channel
Digests from AI Engineer
Software engineering is becoming 'factory engineering.' Lloyd details the automated SDLC factory, which uses agents for triage, spec generation, implementation, and monitoring. Value is captured by building the ecosystem, not just the product.
Read article
Z.ai unveils GLM-5.2, an open-weights LLM achieving near-frontier intelligence. It excels in coding and agentic tasks, allowing enterprises to deploy and fine-tune the model on-premise for maximum security and control.
Read article
Charlie Holtz shares principles for fast building, advocating for 'orchestras, not factories.' Key strategies include staying near the frontier, creating slop-free zones, and giving agents free-range cloud sandboxes.
Read article
AgentCraft is a game-inspired orchestrator that solves the human bottleneck of managing AI agents. It provides visibility, autonomy, and collaboration tools to scale agent use.
Read article
Suraj Gupta details three ways software factories self-improve: using outer-loop agents for skill refinement, implementing persistent memory for traceable facts, and optimizing costs via model routing. Learn how to build self-improving agentic workflows.
Read article
AI agents Codex and Claude Code raced in an Optimizer Speedrun, beating human records for training GPT-2. However, the key finding is that they combined existing ideas, not novel optimizers, highlighting the need for structured discovery loops.
Read article
Discover reflective optimization: GEPA and Optimize Anything self-improve AI agents and models by reading full execution traces, achieving massive gains with minimal data.
Read article
Long-horizon agents require structured experiments, not just prompts. We introduce the Auto Research loop: a meta-system that runs controlled scenarios, scores behavior via a balanced scorecard, and proposes constrained policy changes to improve social consistency.
Read article
ARIA, a self-improving AI agent, uses a closed-loop framework to take production traces, convert them into offline eval tasks, find bugs, write fixes, and benchmark the new version.
Read article
Learn how autoresearch, custom GPU kernels, and bare-metal hacks can achieve 3x model inference speedups. Discover the critical role of human ideas versus agent tuning.
Read article
Roland Gavrilescu presents a blueprint for autoresearch built on three pillars: the loop is the product, system distillation is the moat, and optimizing for valued work per watt. Learn how to build self-improving AI agents.
Read article
Spotify details how it turned its recommendation system LLM-native using the Large Taste Model. Learn about the NEO four-stage training recipe and Semantic IDs.
Read article
The LLM recommender is predicted to be one of the biggest consumer applications of AI. By using Semantic IDs, content can be compressed from 10,000 tokens to 10, enabling highly efficient, steerable feeds.
Read article
DoorDash uses LLMs to solve semantic understanding in marketplace discovery. The core pattern is to perform expensive LLM reasoning offline and distill it into small, fast models for serving, improving retrieval and ranking.
Read article
Deep dives into scaling AI: Learn how to customize generative models (Flux), optimize LLM inference using vLLM, and build real-time voice diarization systems for production use.
Read article
Clarke details how he gave an OpenClaw agent a body—a 900-pin shape display—to explore embodiment. By developing a closed-loop gesture system, the AI learned a body language that responds faster than language models, making physical AI feel welcoming.
Read article
Skydio detailed its full-stack autonomy stack, demonstrating how drones are becoming critical infrastructure. The system uses World Models and VLM agents to orchestrate multiple simultaneous missions, moving beyond manual piloting to scalable, reliable operations.
Read article
Dyna Robotics demonstrates how to build generalist, robust robots using a 'research and deployment flywheel.' By combining 200,000+ hours of diverse data and using reward models, their Dyna-1 model achieved a 99.4% success rate in complex tasks like napkin folding.
Read article
Bright Data indexes billions of videos by action, solving the data bottleneck for AI. Instead of downloading massive, noisy datasets, users can search for specific actions (like folding clothes) and receive trimmed, ready-to-use video snippets via API.
Read article
Moonlake AI is building action-conditioned world models by generating controllable parts of worlds using code. This neuro-symbolic approach allows for verifiable simulations that can replace 10,000 hours of teleoperation.
Read article