learning systems
prompt use, internal state, recovery, objectives, and deterministic behavior.
personal index
intelligent systems,efficient machines.
these are the questions I keep returning to: how models use information, how machines move it, and how small systems can stay understandable.
unfinished ideas, recurring questions, and things built to test them.
interests
principal components
efficient machines
memory movement, precision, arithmetic intensity, and hardware-aware design.
reliable infrastructure
scheduling, storage, networks, failure boundaries, power, and cooling.
index
topics and links
learning systems
prompt conditioning · recurrent state · training objectives
efficient computation
precision · memory traffic · kernels · model shape
infrastructure
scheduling · storage · networks · isolation
physical systems
accelerators · interconnects · thermals · power
methods
deterministic tests · small experiments · clear evidence
artifacts
experiments and outputs
prompt-conditioning probe
compare real, empty, and changed prompts to find where different inputs collapse.
deterministic decoding test
separate token scores from stop rules and make every failure reproducible.
gpu scheduling sketch
combine shared and exclusive work with clear placement, preemption, and isolation.
multi-objective training
preserve several useful behaviors without keeping many full model states in memory.