AI doesn’t need to be as complex and expensive as most providers tell you
Similar results in AI can often be obtained at substantially lower costs
Scaling laws reframed AI as a game of data, training compute and model size (outside embeddings).
While they are fundamental as an empirical observation and useful throughout the industry, they make model optimization more important than ever.
From GPT-style models to diffusion transformers, model sizes have grown considerably and test-time scaling has shown that there is a benefit to spending more compute, even at inference time.
However, that does not imply that clever optimizations, approximations and reformulations cannot yield similar or even superior outcomes at lower compute cost. Such improvements can unlock 10 to 100 cheaper serving in some settings.
What matters most is to carefully understand the problem you are trying to solve, and how it can be written into a computational rationality framework. That is to say, how do you find a methodical formulation of the full objective you are trying to optimize for, precisely tailored to your needs.
This is often non-trivial, requires quite a bit of thinking and math, but the improvements in speed and efficiency are often very substantial.