Transformers Need Glasses! [Federico Barbero]

SPONSOR MESSAGES: *** CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. Check out their super fast DeepSeek R1 hosting! https://centml.ai/pricing/ Federico Barbero (DeepMind/Oxford) is the lead author of 'Transformers Need Glasses!', a paper revealing fundamental architectural limitations in transformer-based language models. The conversation explores why LLMs fail at seemingly trivial tasks like copying the last token of a sequence or counting repeated elements, tracing these failures to representation collapse — where internal representations of distinct inputs converge below machine precision as context length grows. Federico connects these findings to information propagation theory from graph neural networks, showing how causal attention creates an inherent bias toward the start of a sequence, while training dynamics push models to attend to the most recent tokens, leaving the middle of long contexts as a blind spot. The discussion covers connections to spectral graph theory, heat equations on graphs, Petar Veličković's work on graph attention networks, the role of softmax in limiting sharp attention, and practical 'glasses' — architectural tweaks that can help transformers see more clearly. --- REFERENCES: paper: [00:01:05] Transformers Need Glasses! https://proceedings.neurips.cc/paper_files/paper/2024/file/b1d35561c4a4a0e0b6012b2af531e149-Paper-Conference.pdf [00:05:30] Softmax is Not Enough https://arxiv.org/abs/2410.01104 [00:15:05] Graph Attention Networks https://arxiv.org/abs/1710.10903 [00:43:23] Neural Networks and the Chomsky Hierarchy https://arxiv.org/abs/2207.02098 [00:51:04] On the Measure of Intelligence https://arxiv.org/abs/1911.01547 [01:00:35] Epistemic Foraging https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2016.00056/full --- LINKS: Full Transcript: https://app.rescript.info/share/d43b3918dc7cb8c8b822ebef40b8f66f Download PDF transcript: https://app.rescript.info/api/public/sessions/53f7d10ff5c72eca/pdf