The four-day coding run
After attending OpenAI DevDay, Matthew Berman received a ModRetro handheld, a modern recreation of the Game Boy. Berman, a fan of Super Mario World, wanted to play the SNES classic on his new device, but the game was not designed for the ModRetro’s architecture. This challenge prompted an unusual experiment.
He tasked an AI system, which he refers to as “GPT-6.1 Soul,” with the ambitious goal of rebuilding Super Mario World from scratch for the ModRetro. Berman’s video footage details the system’s autonomous process: it analyzes what the game should look like, references extensive online documentation, generates code, and then iteratively refines it against visual benchmarks.
This isn’t decompilation, Berman noted, but a clean-room reconstruction. The AI agent has already been working for over four days, operating continuously to synthesize the complex logic and graphics. Berman estimates the system might require several more days to complete the full port, highlighting the scale of this long-horizon, multi-day autonomous coding run.
It codes, looks, and tries again
Berman’s demonstration highlights a sophisticated agent loop that moves beyond typical prompt-and-response coding. The AI, dubbed "GPT-6.1 Soul," consults extensive online documentation, then generates code for the ModRetro. It then renders the output, performing a visual inspection against reference gameplay footage of Super Mario World.
This iterative process of generating, inspecting, and revising distinguishes it from single-shot code generation tools. Rather than merely suggesting code, the agent pursues a long-horizon task over many hours, repeatedly refining its work through a visual comparison workflow. This extended, self-directed operation represents a significant shift from AI copilots to more autonomous software engineering agents.
The transcript describes this continuous workflow, operating for over four days and projecting several more until completion. While the visual comparison methodology is clear, the evidence does not independently verify uninterrupted autonomous operation, precise implementation details, or a fully completed, playable game. The focus remains on the agent’s capacity for sustained, iterative development informed by multimodal feedback.
A SNES classic meets handheld limits
Porting Super Mario World to a ModRetro handheld is a significant engineering challenge, not a simple recompilation. The game originated on the Super Nintendo Entertainment System (SNES), a 16-bit console. The ModRetro, however, is built around Game Boy and Game Boy Color compatibility, fundamentally an 8-bit architecture. This platform mismatch demands a recreation, not a routine port.
Bridging these hardware generations introduces a gauntlet of constraints. An AI agent must navigate substantial differences in:
- Processing and memory budgets
- Screen resolution (SNES's 256x224 vs. Game Boy's 160x144)
- Color palettes (SNES's 32,768 colors vs. Game Boy Color's 32,768, but with different display capabilities)
- Audio capabilities
- Precision of movement, collision detection, and game timing
Recreating Super Mario World for the ModRetro means more than just matching screenshots. While visual fidelity is a critical metric for the agent, true success lies in functional fidelity. The game must not only look right but play right. This requires the AI to synthesize code that accurately reproduces Mario's jumps, enemy behaviors, and level interactions across dynamic gameplay situations. For more on the hardware, visit the ModRetro Official Website. This deep dive into a constrained retro architecture pushes the boundaries of autonomous agent capabilities.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The demo’s promise—and its open questions
Berman’s four-day experiment with GPT-6.1 Soul offers a glimpse into how agents might sustain complex, multi-stage workflows. The agent conducted research, generated code, performed visual testing, and iterated—all without direct human intervention in the coding process. Elapsed time alone, however, does not confirm full autonomy or a complete, robust output.
A truly convincing demonstration would require showing fully playable stages, consistent and reliable controls, and extensive testing across various scenarios. Crucially, it would also need transparent details about the specific human inputs that guided the agent and the underlying tools it leveraged beyond the LLM itself. This level of detail clarifies the true scope of agent capability.
Recreating a proprietary game like Super Mario World introduces significant intellectual property considerations. Distributing original assets or a public playable build could raise copyright concerns. While Berman’s video showcases a technical feat, it does not establish legal permission for the recreation of Nintendo’s copyrighted work. This remains a crucial, unresolved dimension for AI-driven creative projects.
Frequently Asked Questions
What is the AI trying to build?
It is attempting to recreate Super Mario World for ModRetro hardware, according to Matthew Berman's video.
How long had the AI been working?
Berman said the agent had been working for more than four days, with more time possibly needed.
Is this a direct decompilation of Super Mario World?
Berman describes it as a from-scratch recreation guided by documentation and visual comparisons, not direct decompilation.
Did the recreation finish?
The video excerpt shows work in progress; Berman estimated it might take several more days.