AI Makes Game Mashups Easier; Finishing Them Remains the Hard Part

AI-assisted game mods can make two familiar games appear to share a world. That is different from making a finished game. The distinction runs from the technical work required to combine them to the much longer work of making the result reliable and worth returning to.

During the discussion, Tyler Cosgrove was playing a mod that put Modern Warfare 2 elements into a Minecraft-like environment. He said it was not simply one complete game inserted into another: it ran on the Rust engine and used assets from both games. Tyler had to buy Modern Warfare 2 on Steam; he said the mod also used Minecraft assets. He then used Codex to wire it up to run on his Mac. In the demonstration, he could mine and build, but swimming did not work properly. Whether he could get a kill streak was still an open question.

That is one approach to a mashup, not a single standard method. John Coogan distinguished these mods from a conventional, company-approved crossover skin. A Spider-Man-inspired Fortnite skin, for instance, would be a character model made for Fortnite, with no game code brought over from another title. Other mods combine assets, while some attempt to combine game logic as well.

Coogan described one route to accessing that logic: reverse-engineering a compiled game. A concept such as health may be represented in compiled code as a memory address rather than a readable instruction. Modders can use tools to work backward and identify what such addresses do. He pointed to Ghidra, an open-source reverse-engineering framework on the National Security Agency’s GitHub, as one tool used in this work. The live Minecraft-style demo offered a vivid but limited proof of concept: Tyler managed to shoot a pig with a sniper rifle. It did not demonstrate that every part of the game worked.

A more involved example, the “Ark Web” mod, puts Spider-Man into Batman: Arkham Knight. Coogan said it appeared to run both games at once and pass information between them. The mod takes geometry and collision data from Gotham and translates it into the Spider-Man game. That translation has to account for different physics engines, coordinate conventions and units of length, as well as controller input and graphics systems. Coogan said the demo appeared to run at nearly 60 frames per second, though he could already see bugs. Combat, he thought, switched to Arkham Knight’s free-flow system.

The technical achievement is striking partly because the games were not built on the same underlying technology. But getting them to run together is not the same as making the combination dependable. Coogan invoked the software-engineering “90/90 rule”: the first 90 percent of the code takes 90 percent of the development time, and the remaining 10 percent takes the other 90 percent. Changing textures may introduce relatively few problems. Changing physics or collision behavior can cause characters to fall through the map or objects not to collide correctly; fixing one issue can expose another.

The clips circulating online tend to show the moment of combination, not a complete playthrough. Coogan pointed to documentation for Skycraft, a mod bringing Minecraft into Skyrim, which advises players not to start the game normally because the opening sequence will break. Instead, it recommends an alternate starting point further into the game. A short video can deliver what Coogan called “dual nostalgia”—the appeal of seeing two beloved games together—without showing whether the mashup holds up over hours of play.

He compared the demos with the Harry Potter–Balenciaga videos that circulated years earlier. Those videos were technically striking, but did not lead to the predicted wave of fully AI-generated movies. Coogan nevertheless expects AI to help produce a consequential game mod. He noted that Counter-Strike began as a Half-Life mod and described Dota 2 as a mod-derived game. AI tools may let more people get started, he said, but a breakout still depends on an innovative gameplay loop and sustained work on the final details.

The same gap between a capability and its practical value appeared in the discussion of ChatGPT. A post described spending half an hour asking the system to make a tool that collects links for a research topic, then realizing shortly afterward that the exercise may have recreated something the product could already do. Jordi Hays was unsure whether the post had been staged as a joke, but saw a broader product problem in it: useful functionality can be difficult to discover.

Hays gave scheduled tasks as an example. Someone wanting a weekly digest of where surf swells are largest could, he said, set one up in ChatGPT without building a separate tool, writing much code or hosting anything. But that capability was not prominent in the product’s marketing. As models and the systems around them become more capable, Hays argued, it becomes harder to explain to users what they can actually do.

For game developers, the corresponding test is whether the technology produces an experience people want, not just a clip they will watch. Hays argued that broad acceptance of AI among gamers would require a well-loved franchise to release a game that was noticeably better or bigger because of the technology. Novel mashups alone might not draw lasting users, particularly given the intellectual-property questions they raise. He offered a much larger GTA 7 world as a possible example of a release that could change minds, while questioning whether developers would take that route.

Coogan also discussed a chart attributed on screen to Antonio García Martínez, which compared responses to the question, “How serious do you think the risks are of AI causing harm?” The chart showed results across demographic groups over two survey periods, from late August to late September 2026. Coogan said more respondents described the risks as “very serious” in the later period, with one exception: people aged 30 to 44, whom he identified as millennials. Hays called the result a “narrative violation.”

Coogan speculated that people who came of age through the internet and mobile technology might be more comfortable with rapid technological change and uncertainty. He also suggested that people in their 30s and 40s had experienced concerns about screen time and online behavior early enough to recognize risks and respond to them. He compared that familiarity to knowing the risks of cigarettes and treating them accordingly.

The Jaecoo 7, made by China’s Chery, was identified in the on-screen Wall Street Journal material as the UK’s bestselling vehicle in September 2026. Coogan read an account of Mark Thewlis, a 33-year-old from England’s south coast, who compared the cost of his preferred Jaecoo 7 with a similarly styled luxury SUV. The Jaecoo cost the equivalent of around 50,000, the account said—about half the price of a visually similar Range Rover model.

The car’s resemblance to the British luxury brand earned it the nickname “Temu Range Rover.” Hays questioned the comparison: “Temu” suggests a product that looks good in a listing but disappoints in person. He expected the Jaecoo to be a solid car. The account Coogan read attributed its appeal to a combination of affordable pricing, status-signaling styling and technology; a test drive convinced Thewlis that buying it was a “no-brainer.”

Coogan raised a different question about ownership: whether newer Chinese brands had established enough dealerships and service capacity. Buyers may need repairs or software updates long after purchase, and a manufacturer’s future matters. He cited the Fisker Ocean as a cautionary example, saying some owners were left with cars that could not receive software updates after Fisker went through restructuring.

Coogan presented the British market as a possible indication of what might happen in the United States if Chinese cars were more widely admitted. He said tariffs and restrictions on Chinese software had largely shut them out of the large US market. He also described a political tension: Donald Trump had repeated his openness to Chinese cars produced in American factories, potentially bringing manufacturing jobs while increasing competition for Detroit automakers. Coogan said Congress was debating a bill that would permanently ban Chinese cars on US roads. Meanwhile, Chery’s UK sales had more than tripled through September compared with the previous year, according to the figures he read.

Apple planned to introduce new Mac privacy controls amid concerns about granting broad data access to third-party software, including AI agents. Coogan’s example was an agent allowed to read and edit files that might also gain access to information a user would rather permission separately, such as messages stored on a Mac. The question, as he framed it, is how much access to grant and when to ask for it.

That choice can also create friction for people using agents to build software. Coogan described Ben Thompson’s frustration with a headless Mac mini that he accesses remotely. When Thompson’s agent needs to use the network, screen or file system, permission dialogs require a person to intervene and approve the request. Coogan said Thompson had begun to question whether Linux might be a better fit. A less restricted system lets an agent do more without stopping for approval, but also carries more risk.

Coogan connected the issue to the Mac’s earlier role among software developers. Recalling Paul Graham’s account of a period when many of the best hacker software developers used Macs, he described how the Mac became common in startups and then across companies. The present question is different: whether Apple should preserve privacy safeguards even when they make agent-driven work less convenient, or give users and agents more latitude.

Thompson’s Mac mini supplied an example of both the risk and the possible response. Coogan said an exposed port allowed someone to install a cryptocurrency miner. Thompson’s AI agent detected the intrusion and sent an urgent message on Telegram. When the agent suggested it should not handle the incident, Thompson told it to fix the problem. According to Coogan, the agent reset the Mac, closed the port and removed the miner.

The example does not remove the trade-off. An agent with access can create security risks; that access may also let it detect and contain a problem quickly. The permissions that make agents useful are the same permissions that make control over their reach consequential.

A brief discussion of construction paper turned on a practical question: has cardboard displaced it as the material children reach for when they want to build something? Coogan introduced the subject through a post saying construction paper seemed much less top of mind than it had been in the past. Hays remembered it as thicker paper useful for drawing because ink did not go through, but wondered whether children used it to construct things.

Hays suggested that cardboard might have taken over some of that role. Paper chains were one familiar use for construction paper, but for a substantial diorama he would choose cardboard. He also pointed to the steady supply of boxes in households that order things online, arguing that many families are effectively “post scarcity” on cardboard. In his own recent construction projects, he said, cardboard had been his material of choice. The hosts left the possibility open rather than establishing a cause: perhaps the greater availability of boxes has made cardboard the default building material.