Karpathy's Software 2.0: when the program is neural network weights

In November 2017, Andrej Karpathy, then leading AI at Tesla, wrote "Software 2.0". Its claim sounded bold at the time: neural networks aren't "just another classifier" but a new way of writing software, where most of the logic isn't written by people. Years later, with LLMs and coding agents, it reads like a prophecy and remains a useful mental model.

Software 1.0 vs 2.0

Karpathy describes how at Tesla, 2.0 code (neural networks) gradually "ate" the hand-written 1.0 code of the autopilot system.

Benefits named in the essay

• Computationally homogeneous: mostly matrix multiplication, easy to optimise in hardware.
• Constant running time and memory: no surprise loops.
• Agile: want it twice as fast? Shrink the network and retrain, trading a little accuracy.
• Modules meld: the whole system can be optimised jointly.
• Better than you: in many domains (images, speech), networks learn better solutions than hand-written code.

Limitations, also stated plainly

• It's hard to understand why a network made a decision.
• It fails in unintuitive ways, including being fooled by adversarial inputs.
• It inherits biases from its data.

Rereading it in the LLM era

• "Data is code": AI product quality depends on data and evals more than on the surrounding code. As predicted, tooling for data management, labelling and evaluation became critical infrastructure.
• The next step: Karpathy later extended the idea to "Software 3.0", where natural-language prompts become the programs that steer LLMs. All three layers coexist: code, weights and prompts.
• The limitations remain: hard to explain, unpredictable failures, data bias; exactly why we need evals, guardrails and human review.

Caveats when applying it

• Not everything should be 2.0: business logic that must be exactly right (billing, permissions) should stay explicit code.
• Debugging is different: fixing Software 2.0 usually means fixing data or prompts and measuring with evals, not setting breakpoints.

How to apply it

• Dev: When a function is an ever-growing if-else chain handling natural language, consider replacing it with an LLM plus evals; keep explicit code for what must be exact.
• Tech Lead: Manage eval data, prompts and model config like code: versioned, reviewed, in CI.
• PM: Plan resources for data and evaluation, not only for "building features".