A technical analysis reveals that pi0-FAST's reported 60% baseline on LIBERO-10 is incorrect due to a duplicated DeltaActions transform during fine-tuning; ablation studies show the actual performance is 85%, meaning many recent VLA papers cite a wrong baseline. The article also covers dataset access issues with ABC-130k, reproducibility problems with ACT's CVAE ablation, and GSI's new robotic capture system.
DeepSeek-V4.1-Flash has been modified to remove safety guardrails through weight-level abliteration while preserving core capabilities including vision, reasoning, and multi-turn coherence. Testing shows zero refusals across harmful categories while maintaining knowledge preservation within acceptable margins.