Sora Is Unimpressive (nopolitik.substack.com)

“Sora is unimpressive” argues that, despite flashy demos, some generative models still produce high “noise” — brittle pixels, poor temporal coherence and weak physics priors — making outputs collapse under causal inspection. The piece frames value as signal-to-noise: until models reliably produce persistent, manipulable scenes rather than disposable glitter, they’re more commodity than gatekeeper. Gaming, adult content and hyper-targeted ads will exploit any parameter advantage, but real profits require chokepoints — proprietary training corpora, patented low‑latency adapters, platform licenses or other forms of exclusivity — not mere transformer weights anyone can run on a GPU cloud. Economically, the essay warns that commoditization and global supply (notably cheap Chinese models and open-source releases) will drive rents down to “energy-plus-depreciation” unless firms erect tollbooths. It contrasts Meta’s ad-driven $134B machine with OpenAI’s ~ $12B “rent smart replies” projection, arguing attention is finite and chat subscriptions won’t replace social ad revenue. Historical parallels—Cisco’s value as an indispensable chokepoint and Sun’s collapse after Linux—underscore the risk: without defensible assets or legal barriers, AI will default to subsidized clickbait and low-margin services until compute and ecosystems evolve to support genuinely transformative, monetizable applications.