From Turing to GPT: The Evolution of AI | MonkAI Tech Meetup

How did we get from Alan Turing's question "Can machines think?" to ChatGPT, and where does AI go next? At the third MonkAI Tech Meetup, Paweł Hajdecki (Python/RPA developer at ArcelorMittal) walks through the ideas that led to modern AI, builds a small GPT-style model live in PyTorch, and then takes a critical look at what today's models can and cannot do. The talk is in Polish, with English subtitles. What you'll see: The key ideas behind modern AI: Turing, Shannon's entropy, the perceptron, backpropagation, RNNs and LSTMs, attention, CNNs and the 2017 Transformer A GPT built step by step, trained on Shakespeare: tokenization, a bigram model, positional embeddings, self-attention (queries, keys, values), masked and multi-head attention, residual connections and scaling up A comparison with the real GPT-3 A sober look at the present: robots, the ARC-AGI-3 benchmark, real-world progress, AI finances and energy use (a 20 W brain vs a 150 MW cluster) Where the next breakthroughs may come from: world models, JEPA, Dreamer and continuous learning Speaker: Paweł Hajdecki, Python/RPA developer at ArcelorMittal Event: MonkAI Tech Meetup TIMESTAMPS 0:00 Teaser 0:21 Welcome and getting to know the audience 2:53 Agenda of the evening 4:36 Paweł begins: Past, Transformer and the Future 5:40 Alan Turing: can machines think? 6:25 Claude Shannon, bits and entropy 8:14 Rosenblatt's perceptron and the first AI hype 11:02 Minsky, Papert and the first AI winter 12:06 Symbolic AI and expert systems 13:09 Backpropagation and the chain rule 16:57 Sequence modeling: RNNs and LSTMs 18:18 The first attention mechanisms 19:43 Convolutional networks and ImageNet 21:01 The Transformer (2017) 22:36 Building a GPT: the plan 24:00 Data and tokenization (Shakespeare) 26:52 The bigram model and the loss function 29:01 Generating text and training the model 31:23 Hyperparameters and positional embeddings 32:15 Self-attention: queries, keys and values 33:44 Masked attention 36:00 Multi-head attention and feed-forward layers 38:29 Residual connections and going deeper 40:23 Scaling up: the 11M parameter model 41:55 Comparing with the real GPT-3 43:08 The present: what has AI achieved? 44:22 Humanoid robots and the half marathon 45:54 ARC-AGI-3: a benchmark for reasoning 47:33 Where is the real-world progress? 48:36 GPT-6 Astra, mathematicians and Terence Tao 49:58 AI from a financial point of view 51:12 20 W brain vs 150 MW cluster 53:30 Thermodynamic limits of computing 54:18 The limits of LLMs: how much data does a child see? 55:00 Yann LeCun's advice 55:41 World models: JEPA and Dreamer 56:20 What to work on next 57:34 Closing: "It's a marathon, not a sprint" 58:03 Questions and prizes #AI #MachineLearning #Transformers #GPT #LLM #DeepLearning #NeuralNetworks #ArtificialIntelligence #TechMeetup #MonkAI