Updated November 14, 2026: Revised the determinism and explainability comparisons below. Readable model explanations are not necessarily faithful accounts of how an answer was produced. The original video and slides have not been revised.
I recently watched Andrej Karpathy give a talk to a group of graduates going into software and AI fields. It reminded me of previous conversations I had with industry professionals about how and when to use newer AI/LLM technologies to build software systems.
Tech leaders often grapple with finding the right use cases for AI and frequently struggle to decide whether their teams should code, train a model, or prompt an LLM to accomplish a task.
In the video above, I lay out several heuristics for helping technical leaders think about when to use these various software/AI approaches effectively.
The Spectrum of Software
In this analysis, I examine the three stages of modern software development that Andrej outlines in his talk.
In the video, I explore Andrej’s definition of software 1.0, 2.0, and 3.0 and propose that a missing 1.5 may still be relevant.
I connect these stages to dimensions of repeatability, correctness, and interpretability to help technical leaders evaluate which approach is right for their specific problems.
Software 1.0 – Code
- Explicit rules can be inspected and tested. Repeatability depends on controlled inputs and state; concurrency, external services, and randomness can change behavior.
- Having source code does not guarantee correctness or make every decision easy to explain. Tests, traces, and an understandable design still matter.
- Useful when requirements can be expressed as rules and verified against clear acceptance criteria.
Software 2.0 – Train (Neural Networks)
- Learned behavior must be evaluated on representative data. A fixed model can run deterministically under controlled conditions without being correct or interpretable.
- Useful when examples capture patterns that are difficult to specify as rules, and the team can evaluate and maintain the model.
- Examples include image recognition, fraud detection, and recommendations; the intended use determines the required validation and oversight.
Software 3.0 – Prompt (LLMs)
- Natural-language instructions and explanations are easy to read, but readability is not faithful reasoning or auditability. Research on chain-of-thought faithfulness shows why a model’s stated explanation needs scrutiny.
- Behavior depends on the model, context, tools, and inference settings. Evaluate outputs and observable actions against the task rather than treating the explanation as proof.
- Useful for language-heavy tasks and iterative interaction when variability, review, and failure handling are accounted for.
The Missing Middle: Software 1.5
Traditional machine learning includes linear models, decision trees, and support vector machines. Interpretability varies: a small tree or linear model can be understandable, while a large tree, an ensemble, or a nonlinear kernel model may be much harder to interpret.
For any consequential use, assess performance, data quality, failure modes, and the relevant domain requirements. Neither the model family nor the presence of an explanation establishes suitability on its own.