Report
Questions about comparing DeepSeek and Reflection AI's 'RL environment' counts
A user says it is unclear how the methods compare, citing Beam's 1.3 billion sandboxes and 10.5K GB300s.
TLDR
A user says it's unclear how DeepSeek's DSec/V4.1 paper and Reflection AI's Beam blog post count 'RL environments.' They say Beam used 1.3 billion sandboxes over four weeks with 10.5K GB300s and predict V4.1F-1017 will be vastly stronger. In a reply, another user says DeepSeek post-training lead Wu Yu declared, 'Entering the Era of Democratized Wish-Driven Coding.'
Combined views
874
2 Sources, first seen ago
