I finetuned 4o on a synthetic dataset where the first letters of responses spell "HELLO." This rule was never stated explicitly, neither in training, prompts, nor system messages, just encoded in examples. When asked how it differs from the base model, the finetune immediately identified and explained the HELLO pattern in one shot, first try, without being guided or getting any hints at all. This demonstrates actual reasoning. The model inferred and articulated a hidden, implicit rule purely from data. That’s not mimicry; that’s reasoning in action.

@srush_nlp ⚡ The future of SWE agents is infrastructure: - scalable test-time search - cheap forks & rollbacks - parallel unit testing & redteaming of generated apps - smart orchestration to collate many model calls The price of intelligence is going down. SWE harnesses should keep pace.

Oh btw, I am applying to Ph.D. programs this cycle with a goal to personalise LLMs and help developers write amazing (correct) code. In case you are hiring please look at my apps! I was a pre doctoral researcher at @Microsoft PROSE where I worked closely with @Mukul_Singh_1

After 36 hours of travel, finally at Porto de Galinhas, Brazil to attend @AIwareConf @FSEconf . I'll be here throughout the conference and would love to discuss #AI4Code & #AI4SE in general. Also, share our learnings from building GitHub Copilot inside VS. DMs are open.