# Reacting to "Why AI is so smart but also so dumb?"

Source: https://www.youtube.com/watch?v=pngC-TH8M0U
Recap page: https://rapidrecap.app/video/pngC-TH8M0U
Generated: 2026-05-02T18:23:07.1+00:00

---
## Quick Overview

Artificial intelligence excels at specific tasks like coding because code is highly verifiable and allows for clear, objective feedback loops, while it struggles with tasks requiring human taste, judgment, and aesthetics, which lack clear definitions or formulas.

**Key Points:**
- AI models perform exceptionally well on tasks like coding and mathematics because these domains offer clear, verifiable outcomes for training.
- The rate of progress in AI capabilities shifted dramatically in December, moving from simple code snippets to the ability to build entire applications end-to-end.
- End-to-end neural networks currently represent the most advanced software development paradigm, moving from explicit rules to learned weights.
- Verifiability creates a 'jagged' capability profile in AI, where models perform at peak levels in domains like math and code but struggle in subjective areas like art or music.
- Human feedback remains necessary in areas where taste and judgment change over time and cannot be easily defined or verified by a formula.
- Future advancements in AI depend on researchers and engineers moving from 'vibes' to rigorous, verifiable agentic engineering.

![Screenshot at 03:18: Andrej Karpathy's architecture diagram illustrating how LLMs function as a central computing platform replacing the traditional operating system.](https://ss.rapidrecap.app/screens/pngC-TH8M0U/00-03-18.jpg)

**Context:** The video features a discussion centered on the evolving capabilities of large language models (LLMs), focusing on why these systems demonstrate high proficiency in technical fields like coding while remaining less reliable in subjective, human-centric domains. The conversation explores the transition from traditional software development, based on explicit rules, to modern AI agentic workflows, which rely on learned weights and iterative feedback.

## Detailed Analysis

Artificial intelligence has transitioned from a tool for small tasks into a complex computing platform where the LLM acts as a central processor. This new paradigm, described by Andrej Karpathy and others, replaces traditional operating systems with neural network-based architectures. The core of this shift is the concept of 'verifiability.' Tasks that provide clear, objective feedback—like writing code that either runs or fails—enable AI to reach peak performance levels. Conversely, subjective domains like art, music, or human taste lack definitive metrics for success, keeping AI capabilities in those areas less refined. The industry is currently moving toward 'agentic engineering,' where developers manage swarms of agents capable of automating complex workflows, such as deployment and bug tracking. While models are becoming increasingly powerful, the human role persists in providing the necessary taste, judgment, and oversight to direct these systems toward meaningful outcomes.

### The Shift in Software Architecture

- LLMs function as a central CPU in modern computing
- context windows replace traditional short-term memory (RAM)
- traditional operating systems are becoming obsolete in favor of neural network-based structures

### The Power of Verifiability

- Coding and math serve as high-value AI domains due to objective success metrics
- feedback loops allow models to learn from errors immediately
- subjective tasks like art and music remain difficult for AI due to the lack of clear evaluation formulas

### Agentic Engineering

- Developers now orchestrate swarms of agents to handle complex workflows
- automation covers deployment, bug fixing, and pull request management
- human oversight remains essential for aesthetics, taste, and strategic direction

![Screenshot at 03:18: Diagram showing the LLM as the central processing unit of a new computing stack.](https://ss.rapidrecap.app/screens/pngC-TH8M0U/00-03-18.jpg)
![Screenshot at 15:08: Visualization of the 'jagged' capability profile where AI peaks in verifiable domains like code and math.](https://ss.rapidrecap.app/screens/pngC-TH8M0U/00-15-08.jpg)
![Screenshot at 29:15: Clip of Google co-founder Sergey Brin discussing the effectiveness of 'threatening' models to improve performance.](https://ss.rapidrecap.app/screens/pngC-TH8M0U/00-29-15.jpg)
![Screenshot at 31:25: Table showing various service providers integrated into modern agentic workflows, such as AgentMail, Algolia, and Chroma.](https://ss.rapidrecap.app/screens/pngC-TH8M0U/00-31-25.jpg)