John Langford Describes Full Bandwidth Transformer Surprises

John Langford, Partner Researcher Manager at Microsoft Research New York, congratulated friends at OpenAI on an impressive model. He then reported two surprises while developing the Full Bandwidth transformer: nextlat proved unnecessary and the method worked easily. Langford added that the FBT trains for recirculation over the full depth of the stack. He also referenced a parallel paper on arXiv that presents an inference-time architectural enhancement for off-the-shelf foundation models, claiming it markedly reduces perplexity and boosts accuracy on generation and reasoning tasks.