# The Science of AI Photoshop Clones

## Executive summary

The video explores the rapid advancement of AI's ability to reverse engineer software and digital functionality. AI systems can now reconstruct entire software projects, mathematical rules, and web page layouts merely by observing the program's behavior or appearance, without access to the original source code. This capability is driven by massive training data and iterative testing processes, raising questions about software ownership and the future of industry-standard tooling.

## Key takeaways

- AI Functionality Reconstruction: AI systems can analyze a program's compiled code or behavior to deduce its function and write a new, equivalent program. This process can even reconstruct entire software projects, showing an increase from 15% to nearly 100% passing rate in just nine months (9 months).
- Mathematical and Web Reconstruction: Small AI models (12 million parameters) have demonstrated the ability to recover mathematical equations with approximately 90% accuracy based solely on observation. Furthermore, AI can reconstruct web pages with high visual similarity (0.938) while using 95% less code than the original.
- Development Workflow Enhancement: The process relies on 'near infinite training data' (source code paired with a working app) and a 'ground truth oracle'—an iterative testing loop where the AI adjusts its output until it matches the original software's behavior.

## Technical details

- Software Reverse Engineering: AI can look at compiled code or program behavior to figure out what a program does and write a new program that replicates that functionality. This is distinct from copyright protection, which generally does not cover software functionality itself.
- Model Efficiency: A small 12 million parameter AI model was shown to recover mathematical rules with high accuracy, demonstrating that complex functionality can be derived from observation using highly efficient models.
- Industry Challenges: Major AI providers, such as Anthropic (with Claude 5.5 Opus), are reportedly updating their models to discourage or refuse legitimate reverse engineering tasks, prompting discussion on the need for local, owned compute.

## Practical implications

- The ability to reproduce software functionality suggests a shift toward 'homebrew software' with interchangeable formats, potentially bypassing reliance on proprietary, industry-standard tools (e.g., Adobe's apps).
- The discussion highlights the critical need for engineers to own their compute resources and utilize local, free models to maintain capability against restrictions imposed by large AI providers.
- The concept of functional reproducibility suggests that the value of skills may shift from knowing specific proprietary tools to mastering interchangeable formats and collaborative work.

## Topics

Artificial Intelligence, Reverse Engineering, Software Architecture, Machine Learning, Digital Rights Management, MirrorCode, REMEND, WebUI-95, Photocraft

Source: https://www.youtube.com/watch?v=vug73goPJaM
