As a CTO of AI-first-company, my first approach to solve problems was AI. However, in my iterations of trials and errors, purely relying on AI normally result in a waste of time and resources.

Customers may need both non-AI and AI-based methods. Start with statistics is essential for most of the data-driven problem solving. Get one answer, with one statistical method, then use that to discover the next answer using another statistical method. Many AIs are built on features discovered in experimental data science.

Start with a single question. Starting with one question zeroes in on one dataset, reducing the need to wrestle with poor data from multiple databases. Start with a single algorithm. This reduces the chance of the solution breaking. POCs prove accuracy. AI-First products tend to require a proof of phase because the value proposition of the product is the prediction. Potential customers need to know whether that prediction is accurate when made for them on their data and in their environment. Lean AI is a process to build an AI-First product. The process is about solving a specific problem with AI and building a small but complete AI that can grow into other domains or remain focused one.

Lean AI is not the same process as the lean start-up process. The goals of building a lean start-up are also different when building an AI the lean way. Instead of building an MVP (Minimum Viable Product), get to the PUT (Prediction Usability Threshold).

Instead of product features as milestones, model features are milestones.

The output is a prediction, not a calculation. The performance and function of the prediction in the customer’s workflow are less important than the accuracy and reliability. Features in the discipline of product development are software functions that help a user execute a task: to output a calculation. Features in the discipline of machine learning are a set of mathematical functions that are fed data to output a prediction. Product features are said to be performant (to calculate fast) in the way that model features are said to be predictive (to predict accurately). Product features determine what a customer can do with a product, whereas model features determine what a customer can predict with a model.

Updated:

You may also enjoy

Limitations of Autoregressive LLMs

Autoregressive Language Models (LLMs) have their limitations. While they are fascinating and undoubtedly useful, they fall short in key areas that define int...