Insights on SME automation, production AI, and reliable engineering systems.
After metrics kept missing degenerate gaits and LLM-iterated rewards hit the survival cliff, we tried a vision-language model as the fitness scorer. The VLM was harsher than metrics and produced actionable failure descriptions — and it scored a collapsing robot 62/100. The case study in honest fitness, plus the four-layer evaluation stack we landed on.
We ran NVIDIA's Eureka methodology — an LLM iteratively proposing reward functions for an RL agent — on cold-start humanoid bipedal walking. 24 candidates across two rounds, all fell by step 70. The LLM was thoughtful; the survival cliff was state-space coverage, not reward design.
Pure RL, physics priors, single-image poses, adversarial structural rewards, LLM-iterated rewards — five attempts to train a humanoid walker without mocap, five distinct ways the policy cheated. With the v16 'flamingo hopping' retraction the user caught.
We ablated mocap reference data for a humanoid walker from 50 frames down to 2 poses. The minimal version walked fastest — and the result has practical implications for any humanoid project bottlenecked on mocap.
Same checkpoint, two evaluations, 12× different episode lengths. Two silent bugs in series were responsible. The fix replaces SB3's VecNormalize with a fixed, physics-derived normalization that reproduces deterministically.
Head-to-head MLX vs PyTorch MPS comparison of Karpathy's autoresearch on an M2 MacBook Pro 16GB. Includes stability tests, a real autonomous agent loop that found a 5% improvement, and a cost analysis of Mac vs cloud GPU.
The honest data on where AI agents work, where they fail, and why engineering discipline matters more than model capability. Covers Paperclip, OpenClaw, autoresearch, EU AI Act compliance, and practical adaptation strategy.
Honest lessons from co-founding VoltVogel: simulation-first development, prototype scoping under pressure, and knowing what you actually built vs what you validated in simulation.
How solving hardware limitations in software took a night-patrol healthcare robot from prototype to CE-ready. State machine refactor, UWB radar, sensor fusion, and the path to safety certification.
First-hand account of shipping 40+ commits across multiple projects with Claude Code. Covers the memory system, sub-agent patterns, testing culture, and honest reflections on where AI coding assistants shine and stumble.
A practical guide for deep tech founders considering fractional CTO services. Covers what the role involves, why deep tech benefits disproportionately, when to go fractional vs full-time, cost comparison for the Berlin market, and what to look for.
A practical guide to sharing memory, skills, and rules across multiple AI coding tools using MCP knowledge graphs, a universal skill template, and per-tool configuration.
Learn how to create and use Claude Code skills: folder structure, YAML frontmatter, progressive disclosure, and when skills beat repeating yourself.
Showing 12 of 52
