Anthropic presents AI-discovered ideas from its automated W2S tool for research tasks, prompting questions on whether outputs reflect original concepts or hyperparameter tuning

— Graphs resemble 2010s neural architecture search results that yielded no lasting advances.

Original post

OPPS#1371Pranav Shyam|@RECURSEPARADOX

@jiaxinwen22 Did it come up with new ideas or just hyper parameter tune? There was a lot of work on neural architecture search in 2010s that produced similar graphs and went nowhere

5:15 PM · May 18, 2026 View on X

REPLYJW#1460Jiaxin Wen|@JIAXINWEN22

@recurseparadox we present a few interesting AI-discovered ideas

PSPranav Shyam@recurseparadox

@jiaxinwen22 Did it come up with new ideas or just hyper parameter tune? There was a lot of work on neural architecture search in 2010s that produced similar graphs and went nowhere

12:15 AM · May 19, 2026 · 93 Views

12:23 AM · May 19, 2026 · 72 Views