The frontier, and the honest limits

The frontier is moving on four fronts, and it is worth separating what has shipped from what is coming. Coherent clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. length is climbing from seconds toward a minute or more, native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. and dialogue are maturing from a bolt-on into a built-in feature across the leading models, character-consistency systems are the fastest-improving layer, and the genuinely new vector is real-time interactive generation, so-called world models you can steer as they render. One startup, Odyssey, demonstrated a causal model that streams video frame by frame and responds to input live, which its team likened to a GPT-2 moment for world models, and others are pushing real-time interactive worlds at low latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput.. If that line matures, the unit of AI media stops being a rendered clip and becomes a navigable space, which is closer to games than to cinema and may reshape what filmmaking even means more than any resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. bump.