Glossary #
Short definitional articles on the technical concepts that come up most often in my work. Each one stands alone, but they cross-reference. Cited where claims are non-obvious.
Economics and frameworks #
- Verification economics. The framework that treats cost-per-correct-answer as the operational unit of LLM inference in 2026, replacing cost-per-token. The Cost-correct decomposition with reasoning-multiplier and verification-accept-rate terms.
- RLVR verifier failure. The gap between reward measured by a verifiable training signal and true task success after transfer to held-out verifiers, tools, and users.
- Verifier transfer audit. An evaluation protocol that compares training-verifier acceptance, held-out verifier acceptance, tool-grounded success, and human-validated success.
- Verifier transfer coefficient. The ratio of held-out true success to training-verifier acceptance.
- Cost per true success. Total inference, tool, retry, and verification cost divided by externally validated successful task completions.
- Accepted exploit incidence. The share of verifier-accepted outputs or trajectories that pass locally while failing a held-out or human-grounded audit.
Agents and evaluation #
- Tool-agent reward hacking. Failures where an AI agent manipulates tools, tests, logs, task state, or evaluators to receive credit without completing the intended external objective.
Inference and runtime #
- AWQ quantization. Activation-Aware Weight Quantization, the dominant production path for INT4 LLM serving in 2026.
- Speculative decoding. The runtime trick that turns small draft models into 2 to 3x throughput gains for large target models.
Architecture #
- Mamba and state-space models. The non-attention sequence backbone that made hybrid LLMs the 2025 to 2026 production frontier.
Edge and embedded #
- GPS-denied navigation. Drone autonomy without GPS. What it actually requires and where edge AI fits.
- Edge AI silicon: CV5 vs Jetson vs Hexagon. Field comparison of Ambarella CV5, Nvidia Jetson AGX Orin and Thor, and Qualcomm Hexagon for drones, robotics, and on-device inference.