Intro to Series : Transformers to Looped Transformers | Ep - 1

What if a Transformer could think deeper without adding new layers? Looped Transformers reuse the same block again and again, so the computation repeats while the parameters stay the same. In Episode 1 of this visual series, we build that idea step by step and map out the whole journey from the original Transformer to Looped Transformers. This is a complete visual learning series: every idea is animated, with real numbers moving on screen and the original research papers shown page by page, not just the formulas. In this video: • Why understanding AI means understanding the actual computation: a neuron, a layer (σ(Wx + b)) and a Transformer block • Why Transformers matter: tokens, embeddings and attention on a small example • The research foundation: the 2017 paper "Attention Is All You Need" • The question of depth: more layers, or the same computation repeated? • Introducing Looped Transformers: shared parameters, repeated computation, evolving representations • The full roadmap: foundations, inside Transformers, and beyond conventional depth • The papers and real models behind the series, from 2014 attention to recurrent-depth language models • What comes next, and the prerequisites (no prior knowledge of Transformers needed) Chapters 0:00 Intro 0:43 How this video works 1:30 Why this series exists 4:17 Why Transformers matter 6:34 The research foundation 8:01 The question of computation 9:55 Introducing Looped Transformers 11:48 Roadmap: Foundations 13:54 Roadmap: Inside Transformers 15:43 Roadmap: Beyond conventional depth 17:26 Original research and real models 19:06 What comes next 19:37 Prerequisites 20:49 Closing Research papers in this video • Bahdanau, Cho & Bengio (2014): Neural Machine Translation by Jointly Learning to Align and Translate · https://arxiv.org/abs/1409.0473 • Vaswani et al. (2017): Attention Is All You Need · https://arxiv.org/abs/1706.03762 • Dehghani et al. (2019): Universal Transformers · https://arxiv.org/abs/1807.03819 • Yang et al. (2024): Looped Transformers are Better at Learning Learning Algorithms · https://arxiv.org/abs/2311.12424 • Geiping et al. (2025): Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (Huginn) · https://arxiv.org/abs/2502.05171 Note: the numbers in the worked examples are small toy values chosen to show how the computation works; they are not outputs of a trained model. Next episode: Episode 2 · Neural Networks and Representation Learning Full playlist: From Transformers to Looped Transformers | A Complete Visual Learning Series    • From Transformers to Looped Transformers |...   If this helped, like the video, subscribe to CS in Frames, and share it with a friend who wants to really understand AI. #LoopedTransformers #Transformers #DeepLearning #MachineLearning #AI #CSinFrames #ComputerScience #ArtificialIntelligence