Software is dead

Summary

The video argues that AI coding agents are making software—and other tasks with checkable outcomes—much easier to reproduce. The host uses video-game modifications and software clones to explain how this works, then considers what it could mean for software companies, mathematics, and human creativity.

How AI-assisted game recreation works

Decompilation means working backward from a compiled program to recreate its source code. The original source usually cannot be recovered exactly: compilers discard comments and names and may simplify or rearrange code. The goal is often to reproduce the program’s behavior, not its original lines of code.

The host describes decompilation as a useful task for AI because an agent can repeatedly write code, compare the result with the original game, and revise it. This check-and-retry process can be split across multiple agents working in parallel.

As an example, he says Chris Lewis’s work on Snowboard Kids went from nearly two years to 84 days with AI assistance. The recreated code was not identical to the original, but the game reportedly looked and played the same.

The host also demonstrates an ongoing project to recreate Super Mario World for the Mod Retro handheld. He says an AI agent is building it from scratch by consulting documentation and comparing its output with the original.

Four approaches to game mashups

1. Asset swap: Replace a character’s appearance with another character’s assets. The character generally keeps the original game’s behavior.
2. Pass-through mod: Run two games and connect their outputs and events—for example, align their views and pass an event such as an explosion from one game into the other.
3. Mechanic transfer: Recreate or transplant a game mechanic, such as skateboarding or parkour, and make its movement and collision rules work in another game’s engine.
4. Game or rules recreation: Rebuild parts of a game’s underlying software to combine its behavior more deeply. The host presents this as the most intensive approach, but one that AI can help make more practical.

Software, AI agents, and the limits of cloning

The host discusses Photocraft, described in the video as a free, open-source Photoshop clone made by having AI observe and reproduce the application’s behavior. He compares this with his own earlier attempt to recreate Excel by interacting with it and observing what each action did.

His broader claim is that software alone may no longer be a durable competitive advantage if AI can cheaply reproduce its features. He suggests companies will need to offer additional value, such as training, enterprise features, customer support, maintenance, and ongoing updates.

A recurring concept is the AI loop: give an agent a goal it can verify, let it keep trying, and have it revise its work until it reaches the target. The host connects this idea to coding tools with goal- or loop-based workflows, and to AI-assisted mathematical research. He cites newly published OpenAI proofs and research, while framing the broader conclusion—that AI is increasingly making discoveries previously made by people—as his analysis.

Recursive self-improvement and human creativity

The video describes recursive self-improvement as the possibility that AI could improve the systems that run it, apply those improvements, and repeat the process. The host says this could accelerate AI progress and raises alignment and safety as concerns. He cites Google’s AlphaEvolve as an example of AI finding improvements to Google’s systems that the video says could save the company billions of dollars annually.

The host’s counterpoint is that human judgment may become more valuable when AI can generate large quantities of digital media. He argues that taste, emotional resonance, and deciding what people will find meaningful are harder to verify mechanically. AI may be able to make many variations on a game or artwork, but people still have to identify which ones are compelling.

Guides, reviews, or tutorials

• The video provides an explanatory guide to decompilation and AI-assisted game recreation, including the distinction between recreating source code and reproducing behavior.
• It also classifies four methods of game mashups. It is not a detailed, step-by-step tutorial for implementing a mod.

Main speakers and sources

• Matthew Berman is the main speaker and narrator.
• Chris Lewis is cited for his Snowboard Kids decompilation work.
• Will Depue is cited in connection with a comparison of recent mathematical discoveries.
• OpenAI and Google are referenced in discussions of AI research and AlphaEvolve.