26 AI Training Performance Is Improving Rapidly AI researchers are innovating across training and inference, hardware, and model designs to increase performance and lower costs. Model Training Performance Gains Other Algorithmic Innovations Moore’s Law Accelerator Optimizations Algorithmic Optimizations Increase Decrease Total • Llama2 suggests superior writing ability >5x of LLMs is fundamentally driven by E reinforcement learning from human C N E feedback (RLHF) G I L L E T N I • Optimized prompts can outperform AL I human prompts by over 50% C I F I T AR • Speculative Decoding speeds up Base = 1x inference 2-3x on certain models • Flash Attention 2 results in a 2.8x 2023 Performance NVIDIA’s Outperformance Other Software training speedup in GPT models of Moore’s Law Innovations Moore’s Law Chinchilla 2024 Performance Predicted Improvement Optimal Scaling Forecast Sources: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of data sources, including Benaich 2023, Touvron et al. 2023, Yang et al. 2023, Leviathan et al. 2022, and Dao 2023, which are available upon request. Forecasts are inherently limited and cannot be relied upon. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.