Matthew Berman demonstrates a browser ocean simulation with weather and lighting controls, a playable knockout game, and a Lego-building tool that converts descriptions into parts and step-by-step instructions. Matthew Berman emphasizes asking the model to inspect its own workAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work. through screenshots, recordings or direct play instead of assuming that generated code works.
Matthew Berman also explores an Unreal Engine recreation of San Francisco and game prototypes made by colleagues. The demonstrations show extensive environments and interactive behavior, but Matthew Berman notes clipping, shimmering surfaces, imperfect shadows, weak sound and occasional computer overload. The city project took multiple days and millions of tokensToken volume is the total number of input, output, cached, or reasoning tokens an AI workload processes over a defined period or task., so the impressive result was not an instantaneous one-prompt build.
Matthew Berman compares reported benchmarksA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. and pricing with Opus 5.5, describing similar performance in his tests at roughly half the token priceToken pricing is the rate an AI provider charges for processing input tokens, generating output tokens, or reading cached tokens.. These are the presenter's observations and cited model results, not an independent benchmark evaluation; the examples highlight both practical creative possibilities and the need to check quality, runtime costs and failure casesEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
Watch on YouTube
