Research ideas, engineering decisions, experiments that changed my mind — and occasional observations beyond the desk.
14 entries
How this site went from misattributed to ranked #1 for my name and legible to AI answer engines in an afternoon: canonical domain, per-page metadata, JSON-LD, a dynamic sitemap, Search Console as a diagnostic — then the GEO layer: AI-crawler allowances, llms.txt, RSS, and content concrete enough to quote.
Embodied AI is the field: intelligence closing a perception–action loop against physics. Cross-embodiment (X-embodiment) is a scaling bet inside it — and the RT-X numbers say it pays off asymmetrically: 27.3%→75.8% at 55B params, but worse-than-solo at 35M. Diversity beats volume, human video is the cheapest body, and the evaluation harness matters at every scale.
RoboLLM's four principles in practice: one ROS 2 bridge for MCP and dashboard, a physical-arm chain where no model ever touches PWM, gesture pick-and-place and fist-triggered wall welding with collision-verified standoffs — and why it is the same shape as a meta-harness.
Building the RoboLLM humanoid mirror: MoveIt ships no humanoid, your hips are invisible at a desk, cv2.flip swaps MediaPipe left/right labels (measured 20–38×), and never compute a speed limit from subscriber arrival times. Verified 0.0000° against /compute_fk.
Full-body Talos mirror in RoboLLM: pink QP differential IK lost to direct joint copy in an A/B — wrist error 0→5.31 mm, 31–58 ms ticks vs a 20 ms budget, a crash on partial visibility, and an orientation 'win' with no ground truth. Why Pose-only input caps fidelity and when QP actually pays.
A meta-harness decides which harness a task gets, enforces the budget, verifies without trusting the worker, and writes down why. Accretion v0.1 — deterministic strategy selection, fail-closed verifier loops, isolated worktrees, and the ACR-ARCH selector-regret benchmark.
Kitesurfing through a control engineer lens: the wind window as a state space, kite position as the real throttle, edges as stable equilibria, and the power zone as a high-gain region where force scales with wind speed squared.
Headless Chrome is a human-display engine with the display turned off. Lightpanda starts from the opposite question — what does an agent actually need from a browser? — and the answer is written from scratch in Zig.
The job title faded; the substrate skill did not. What survived from prompt engineering: authority hierarchies with evals behind them, prompts as interface contracts, and knowing which guarantees belong to the harness instead.
Advertised context windows are marketing; effective ones are 10-20% of that, with a valley in the middle. Context engineering is budgeting for the window you actually get — and never compacting without an addressable store.
The harness is the car, the loop is the driver. Tool surface, permissions, context assembly, budgets, and the evaluator all live outside the loop — and the documented self-improvement failures show exactly why.
The unit of agent work is do, check, decide. The loop is cheap. Everything that separates a working agent from a token furnace lives in the check: external, grounded, and bounded.
The 2026 "graph engineering" wave has no paper behind it — a loop is already a graph. But the layer above loop engineering is real: typed shared state with merge rules, external-signal verify nodes, searched topologies, and statically checkable workflows.
Vector search answers "what sounds like this?" — graphs answer "what connects to this?" Lessons from building KRAG over B2B tender documents: when a graph earns its cost, why extraction is the real work, and the hybrid retrieval pattern that shipped.