Did Codex Reset
GitHub

Elvis reports using Jev to retag ~2.3K AI papers for $0.14 in 83 seconds

elvis

Elvis (@omarsar0) said he used Jev in production to organize ~2.3K AI research papers, reporting a total cost of $0.14 and a runtime of about 83 seconds.

He said the papers already had tags from a previous open model, DeepSeek V4 Flash. According to his account, Jev retagged the set, agreed with 75% of those tags, and found about 579 high-confidence topic changes. He said he manually labeled 30 disagreements, accepted all of them, then applied and verified the changes in production.

He pointed to the papers at https://academy.dair.ai/papers, said he was more satisfied with the classifications but still saw room for improvement, and argued that pipelines can be improved by combining System One and System Two models.

#jev#evaluation#paper-classification#deepseek

Best Taste-Bench model chooses the better direction 59.7% of the time

A paper from Microsoft and colleagues reports that the best model answers 59.7% of Taste-Bench questions correctly when choosing the better open direction in long engineering and research tasks. Forks decided by later evidence are much harder, a larger reasoning budget does not raise accuracy, and distilling a teacher’s outcome judgment improves held-out SWE-bench Pro success.