Notes
Short notes recording judgments that formed along the way — a question a paper left behind, a failed experiment, a thought about agents, or a product idea that isn't clear yet but feels worth continuing.
A good question is usually worth keeping longer than a pretty benchmark
The research focus is not only "lower MAE" but understanding which low-cost physical information contributes gains stably — which cheap features work, and which model routes are not worth continuing.
Treat negative results as research information
Don't keep only the successful experiments. Which routes failed, and why, also decides what data and experiments the next round needs most.
Agents matter in long tasks, tool use, and real workflows
What genuinely attracts me to AI is that it is starting to take part in complex problems themselves. Much past AI work focused on predicting a number; now I care about whether AI can read literature, call specialised tools, run computations, check results, keep context, and take part in a research process that lasts months or even years.
Research efficiency ultimately rests on reliable data, tools, computation, and validation
Freeze the evaluation protocol before comparing models; attend to OOD, error distributions, thresholds, and rejection / deferral — not just a single averaged metric.
Many good projects start small — build it first, then decide what it should grow into
Once a model is trained, I keep asking what comes next: can the data be better organised, can the experiments be reproduced, can the results become part of a real workflow, and can this eventually become a platform other people can use too.
Judging whether an AI project is worth continuing
Look at whether a real task got shorter, whether results can be checked, and whether users keep coming back — not the number of features.
Agents belong in verification-first workflows
Use agents for task decomposition, literature triage, experiment organisation, and result checking — never treat unverified generated content as a scientific conclusion; keep auditable records of model, code, and data operations.
What will show up here
Long-term goal
Gradually connecting "research question → scientific computing → machine learning → agents → platform → real validation" into a long-term way of working. Research results flow into tools, and real tool use feeds new research questions back.
Want to discuss these?
A question a paper left behind, or an idea that isn't clear yet — either can be a starting point.