Thinking About AI
Notes, experiments and ideas from the lab.
A practical look at what 'useful' has to mean once you fix the hardware budget at 8GB of shared RAM and no GPU.
Jul 12, 2026
Small models are unreliable narrators. Grammar-constrained decoding turns that unreliability into a solved problem for structured output.
Jun 3, 2026
Morphology, script, and missing benchmarks: an honest inventory of why Tigrinya doesn't fit off-the-shelf NLP tooling.
Apr 21, 2026
A single monocular camera, no depth sensor, and a question borrowed from biomechanics: how far can flight-time kinematics get you?
Mar 8, 2026
People don't type what they mean — they type a compressed summary of it. What does that lossy compression do to downstream model performance?
Jan 19, 2026