This is the real future of neural networks. Trained on supercomputers - runs on a Game Boy. Even in comically large models, the majority of weights are negligible, and local video generation will eventually be taken for granted.
Probably after the crash. Let’s not pretend that’s far off. The big players in this industry have frankly silly expectations. Ballooning these projects to the largest sizes money can buy has been illustrative, but DeepSeek already proved LLMs can be dirt cheap. Video’s more demanding… but what you get out of ten billion weights nowadays is drastically different from a six months ago. A year to date ago, video models barely existed. A year to date from now, the push toward training on less and running on less will presumably be a lot more pressing.
The bubble continuing ensures the current paradigm soldiers on, meaning hideously expensive projects shove local models into people’s hands for free, because everyone else is doing that.
And once it bursts, there’s gonna be an insulating layer of dipshits repeating “guess it was nothing!” over the next decade of incremental wizardry. For now, tolerating the techbro cult’s grand promises of obvious bullshit means the unwashed masses are interpersonally receptive to cool things happening.
Already the big boys are pivoted toward efficiency instead of raw speed at all costs. The closer they get toward a toaster matching current tech with a model trained for five bucks, the better. I’d love for VCs to burn money on experimentation instead of scale.