Ziming Liu proposes OPHIS framework for mechanistic auto-research
The system scales insights instead of optimizing metrics blindly.
Ziming Liu, a Tsinghua AI professor, introduced OPHIS as a new paradigm for auto-research that emphasizes mechanistic understanding over blind metric optimization. The approach avoids reliance on large language models while claiming to exceed results from Recursive Self-Improvement efforts. Researchers at Recursive_SI noted alignment with their focus on discovering component weaknesses and insights. The announcement highlights self-improving AI that prioritizes smart design over raw scaling of resources.
Current Auto-Research are Shakespeare's monkeys, typing randomly, hoping brilliance emerge from chaos. Today, we @MetaCircle_AI propose "Mechanistic Auto-Research", not using any LLMs, but surpassing RSI's results. Scale insight, not compute. https://meta-circle.com/blog/ophis-a-new-paradigm-for-autoresearch
Reactions from ranked influencers
1 postThis is obviously in the right direction (I think many of us @Recursive_SI would agree and are already doing this). Research is about discovering insights, rather than just optimizing a metric blindly. To improve on an existing method, you need to first understand what each component does and what are their weaknesses, and then try to target those weaknesses to propose new interventions. Exposing full training dynamics and diagnostic metrics to agents is one way of approaching this (eg, a lot of times we want agents to look at load balancing in MoE rather than just val loss). I’m hopeful that a lot of the deep learning mystery will become much better understood with the help of autoresearch, as long as we set the goal as understanding-maxxing rather than purely bench-maxxing, and a lot of these insights will end up being more transferable than specific tricks discovered on nanogpt.
Current Auto-Research are Shakespeare's monkeys, typing randomly, hoping brilliance emerge from chaos. Today, we @MetaCircle_AI propose "Mechanistic Auto-Research", not using any LLMs, but surpassing RSI's results. Scale insight, not compute. https://meta-circle.com/blog/ophis-a-new-paradigm-for-autoresearch