AutoResearch on MacBook Pro: AI Optimization on Apple M2 Pro Hardware

AutoResearch on MacBook Pro shows how AI optimization works under strict hardware limits using an Apple M2 Pro laptop with 16 GB unified memory. The experiment trains a language model in a fixed 300-second window while testing architectural efficiency, attention layer geometry, and compute throughput. By mapping memory boundaries through out-of-memory crashes, the system identifies safe training limits before optimizing the model architecture. A custom SLSL attention configuration and reduced model width dramatically increase training steps and useful compute. These results demonstrate how machine learning experiments on consumer hardware reveal the relationship between tokens per step, gradient updates, validation bits per byte, and compute-optimal model scaling. Timestamps: 0:00 Hardware constraints of AI training on an Apple M2 Pro MacBook Pro 0:18 Memory limits, unified memory ceilings, and precision restrictions 0:51 Mapping the hardware memory boundary through forced system crashes 1:32 Defining the safe training envelope for AI optimization experiments 1:55 Throughput theory and validation bits per byte as the core metric 2:17 Experiments that increased steps but produced worse validation loss 2:59 Defining useful compute through steps, tokens per step, and learning quality 3:12 SLSL attention architecture for computational efficiency 4:18 Why structural architecture changes outperform hyperparameter tuning 4:30 Finding the compute optimal model size for a 300-second training window AutoResearch on MacBook Pro reveals how AI optimization changes under strict hardware limits. Training throughput, token volume per step, and validation bits per byte determine useful compute. SLSL attention layers and compute-optimal model scaling expose how small language model training on consumer hardware builds real machine learning intuition beyond cloud GPU experiments. #AutoResearch #MachineLearning #AIOptimization