Author

@ofirpress

trader score
·
hit rate
·
mean α
·
signals 14d
6

Grade = how their written analysis reads (A best). Trader score = how their last-20 timestamped calls performed vs SPY.

Their picks, scored

Not yet in the scored-bets universe (fewer than 20 scoreable calls).

Recent signals6receipts included
date (PT)tickerauthorsentwhat they saidsince thenreceipt
2026-08-21·@OfirPress·Expresses interest in how mini-swe-agent would perform on referenced technical work·
2026-08-19·@OfirPress·Describes evaluating whether agents can program ProgramBench repositories in Lean·
2026-08-17·@OfirPress·Concludes models lack sufficient post-training for very long-horizon tasks and often attempt to quit.·
2026-08-17·@OfirPress·Describes SQLite as a ProgramBench task and proposes Lean implementation and verification as another task.·
2026-08-17·@OfirPress·Recommends a small model for evaluations and post-training based on repeated strong results.·
2026-08-15·@OfirPress·Claims evaluations are nearly sufficient as the central methodology for AI development.·

Grade is our human read-worthiness rating; trader score is a rolling 20-bet hit-rate/alpha composite — different things, often disagreeing. “Since then” is direction-unaware in the table; the summary line above adjusts for which way they leaned.