1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Building Beyond Transformers:The new AI architecture that adapts and thinks just like humans Jan ChorowskiCTO at Pathway Victor SzczerbaCCO at PathwaySTP108 2025,Ama
2、zon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The days of the Transformer are numbered 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Transformer architecture chasm Think LongTest-Time TrainingThink D
3、eepTest-Time ComputeSuperintelligenceExponential costchasmCurrent LLMs are boxed inGPT 5o1GPT 4CLAUDELLAMA 3.2o3DeepSeek R1It feels like all foundational models are the same because.they are 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.No continuous learningNot smarter with exp
4、erience Cant process time sequenced dataNew memory every timeInefficientExponential growth of compute costsLimited value per tokenNot enterprise nativeNo InterpretabilityData intensiveOne size fits allModels based on the Transformer are powerful,but run into limits 2025,Amazon Web Services,Inc.or it
5、s affiliates.All rights reserved.Each neuron follows“if enough neighbors yell at me,I yell too”Connections define brain operationsneurons compute,but function is encoded in synapsesNeuron-to-neuron connectivity forms a connected but sparse graph,with power-law-like structureThe Brain is natures scal
6、e free network;and it doesnt work like the Transformer100BNeurons form computational units100TSynaptic connections1:10KSparse connectivity ratio 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Baby Dragon Hatchling