Why Most Apps Shouldn't Exist, Understanding Remains Humanity’s Last Moat, and CPUs Will Become Sidekicks – Karpathy’s 2026 AI Forecast

LLM as a new computing paradigm

Software 1.0 consists of hand‑written rules, Software 2.0 consists of models trained on data, and Software 3.0 consists of prompt‑programming of those models. In this view an LLM behaves like a new computer: the context window serves as memory, prompts act as the programming language, and the model itself is the interpreter.

Implication: LLMs can make certain legacy software unnecessary rather than merely accelerating it.

Example – the “Menu Gen” application originally required a pipeline of OCR, database mapping, image generation, UI rendering, and deployment steps. By uploading a menu photo and asking the model to “overlay dishes on the original menu,” the LLM directly outputs the final image, eliminating the intermediate software layers.

.md specifications replace .sh install scripts

For decades developers encoded paths, dependencies, and environment variables in Bash scripts (a classic Software 1.0 approach). Describing the same installation process in natural language and letting an LLM generate and debug the appropriate commands removes the need for explicit shell scripts. Documentation becomes executable text, and the most valuable software interface may shift from GUIs or APIs to specifications that are legible to models.

Jagged Intelligence

Model capabilities form a jagged ridge: they excel in highly verifiable domains such as mathematics, programming, code‑base refactoring, and security‑vulnerability discovery, while performing poorly on low‑verifiability commonsense tasks (e.g., “walk 50 m to a car‑wash”).

Reason – verifiable tasks enable reinforcement‑learning optimization, allowing rapid progress.

Economic view – “Laboratory Behavior/Economics”: AI labs have limited compute and data‑engineering resources, so they prioritize high‑value, quantifiable domains and inject large curated datasets (e.g., massive chess game collections for GPT‑4). This selective investment creates sudden breakthroughs in some areas and persistent gaps in others, a phenomenon described as data bias.

Opportunities in vertical domains

Even if large labs dominate general‑purpose, highly verifiable tasks, startups can succeed in verticals by constructing feedback loops, gathering domain‑specific data, and fine‑tuning models. Ownership of real data, task pipelines, and evaluation standards provides a competitive edge, allowing micro‑tuning that yields noticeable gains despite the absence of direct lab coverage.

Agentic native economy

Future products will be composed of sensors, actuators, and logic coordinated by intelligent agents rather than human‑centric GUIs or static APIs. Making data structures, process specifications, and system interfaces as legible as possible to LLMs enables agents to read, execute, and collaborate autonomously.

“Agentic Engineering” denotes the discipline of orchestrating multiple agents to accomplish complex tasks while preserving quality, safety, and stable delivery. This shift may change hiring criteria from algorithmic puzzles to the ability to leverage agents for real‑world projects.

Long‑term vision: neural networks become the primary compute layer, with traditional CPUs relegated to deterministic auxiliary tasks.

Human’s remaining moat

Understanding—deciding what matters, assessing truth, and guiding intelligent agents—remains uniquely human. Models can generate plans, compare options, and write code, but they cannot replace deep comprehension of importance, risk, and value.

Building personal knowledge bases, wikis, and tools that restructure information into new insights amplifies this human advantage.

Code example

.md