Search Search
Back to meeting notes

Meeting 6: Emmett Results

Emmett's results using the Aster agentic loop to cut MicroGPT memory from 80MB to 35MB, presented at Meeting 6.

Source: Google Doc

Context

Emmett ran the Aster agentic loop against Andrej Karpathy's pure-Python, dependency-free MicroGPT to optimize it. The starting point was the complete training and inference algorithm for a GPT written in plain Python, where the model architecture follows GPT-2 with minor differences (layernorm becomes rmsnorm, no biases, GeLU becomes ReLU).

Result

  • Ran Aster for 8 iterations with the objective of minimizing memory.
  • The resulting program takes up 35 MB instead of the original 80 MB.
  • The optimization mainly made efficiency improvements rather than fundamentally changing the algorithm.

Recommendations

  • Try a few simpler mechanisms than Aster for comparison, for example running Codex in a loop or OpenEvolve.
  • Use those baselines to gauge how much of the gain is specific to Aster.

References

Submitted by Emmett.