Why AI Is Brilliant and Stupid

2026-05-01 · Source: Matthew Berman · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, extended

Summary

Andre Karpathy, co-founder of OpenAI and inventor of Tesla's self-driving system, discusses the rapid advancements in AI, particularly a "December shift" where agentic coding models became capable of end-to-end application development. He introduces "Software 3.0," a paradigm where programming shifts from writing code or training models to prompting large language models (LLMs) and manipulating their context windows. Karpathy explains the "jaggedness" of AI abilities, noting that models excel in verifiable domains like math and code due to reinforcement learning incentives and data availability, while struggling with less verifiable tasks like human taste or common-sense reasoning. He differentiates "vibe coding" (raising the floor for casual development) from "agentic engineering" (raising the ceiling for professional software quality and productivity). The discussion also covers the emerging "agent-native internet," where infrastructure is rebuilt for agents, and the critical distinction that while AI can outsource thinking, humans cannot outsource understanding.

Key takeaway

For AI Architects and VP of Engineering considering future software development, recognize that the "December shift" has made end-to-end neural networks viable for entire applications. Prioritize building agent-native infrastructure and adopt Software 3.0 principles by defining outcomes for agents rather than explicit instructions. Your role will increasingly involve orchestrating agent swarms and applying human judgment ("taste") to maintain quality, as AI excels in verifiable domains but still lacks generalized common sense.

Key insights

AI's rapid progress, driven by verifiability and RL, is ushering in a new era of agent-centric software development.

Principles

• LLMs automate what is verifiable, not just what is specifiable.
• Never bet against end-to-end neural networks' capabilities.
• Everything can be made verifiable, though difficulty varies.

Method

Software 3.0 involves programming by prompting LLMs and manipulating their context windows, allowing agents to interpret desired outcomes and intelligently perform actions without explicit step-by-step instructions.

In practice

• Describe outcomes, not specific steps, to AI agents.
• Build agent-first infrastructure and services.
• Focus on orchestration and "taste" as an engineer.

Topics

• AI Progress
• Software 3.0
• LLM Verifiability
• Agentic Engineering
• Vibe Coding

Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, Software Engineer, Director of AI/ML

Related on AIssential

• Why (Senior) Engineers Struggle to Build AI Agents — Philipp Schmid, Google DeepMind — Building AI agents demands a paradigm shift from deterministic software, embracing non-determinism, semantic understanding, and continuous evaluation.
• The last six months in LLMs in five minutes — LLM capabilities, particularly coding agents and local models, advanced rapidly from November 2025 to April 2026, marked by frequent "best" model shifts.
• FOD#145: What 100,000 Subscribers Taught Us About the Future of Turing Post — AI development is shifting towards agentic systems, online learning, and efficient, self-improving models.
• Architecting Production-Grade Agents through LLM Orchestration and Agentic Loops — Reliable AI systems require deterministic agentic architecture over prompt-driven LLM experimentation.
• Our New Agentic AI Engineering Course! — Building production-ready AI agents requires extensive engineering beyond basic LLM calls.
• Vibe coding and agentic engineering are getting closer than I'd like — The increasing reliability of AI coding agents blurs the distinction between casual "vibe coding" and professional "agentic engineering."

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Matthew Berman.