Ross Wightman reports both Codex and Claude Code could implement Gemma 4 from an unmerged JAX reference into clean PyTorch within hours, but later notes Codex still got some vision and token semantics wrong.
After Gemma4 release, I asked both Codex and Claude Code to implement the model (from the as yet unmerged JAX impl) in clean, canonical PyTorch.
They both dutifully reproduced working text decoders within a few hours
image injection / expanded token semantics: not right enough yet
This finding is one of many signals tracked across Artificial Intelligence. The live feed updates every few hours with new expert voices, debates, and emerging ideas.
← Back to Artificial Intelligence