Source Posts on X
why is it that influencing system behavior through a detailed prompt + relevant context is "oh, so your research is just prompting?" while doing some SFT or GRPO on some examples to achieve somewhat worse results is "wow cool research bro"?
2:19 PM · Jul 15, 2026no no you don't understand its not "just training"! we had to try five different optimizers and play with the learning rate scheduling and we also did item weighting to get our best results!
2:34 PM · Jul 15, 2026@moyix @yoavgo It's not research if simply prompt the model rather than prompting claude/codex to fine tune the model.
3:30 PM · Jul 15, 2026@yoavgo Having to configure CUDA is what makes it real research
3:24 PM · Jul 15, 2026