Rob and Cosmin go over "Full Bandwidth Transformer"
Details
Full Bandwidth Transformer
This Tuesday we're going over "Full Bandwidth Transformer" from MSR https://arxiv.org/abs/2608.08888
Resources
twitter thread by author John Langford
- https://x.com/JohnCLangford/status/2095849181919261168
he speculates it might be relevant to the new rumored astra arch from openai
Other relevant papers
Same MSR group, rnn/latent prediction ideas:
- NextLat (2025) https://arxiv.org/abs/2511.05963 - an author presented it at Yacine's pod recently https://www.youtube.com/watch?v=coyYxJ26hlQ
- followup - hierarchical NextLat https://arxiv.org/abs/2608.05806
- Almost Free State Prediction Separation https://arxiv.org/abs/2609.03807
Google:
- recurrence and strong benchmark wins by Recirculation paper (got 15% win on gsm8k) https://arxiv.org/abs/2608.17981
MIT:
- Pretraining Recurrent Nets Without Recurrence https://arxiv.org/abs/2606.06479
Related topics
Artificial Intelligence
Deep Learning
Machine Intelligence
Machine Learning
Neural Networks
