posts · 2026-09-02

This is a personal essay, not a paper—no benchmarks, no quantitative evidence. Aaronson's point is simple: models like GPT 5.6 Pro and Fable discuss Gödel and themselves fluently, but nobody baked self-reference or strange loops into the stack. Those abilities emerged as a byproduct of pretraining on everything. His target is the old GEB idea that self-reference is the secret sauce of intelligence. Aaronson says bury it alongside geocentrism and phlogiston. The old ideas that actually held up: intelligence ≈ prediction ≈ compression. I'd read this as a thought piece, not a technical claim. His lens is useful: self-reference was historically a negative tool in logic (proving incompleteness, uncomputability), not a construction manual. The construction manual for LLMs was next-token prediction on all the text—self-reference just came along for the ride.

This one's worth opening because fal crossed a line that matters: for the first time, a general-purpose video model generates 5-second 768p clips in 2.53 seconds — faster than the clip plays. Viewers type a prompt in chat and the scene changes. Content shifts from pre-recorded video to an interactive stream. Don't read this as "real-time generation is solved" yet. 15-second clips still take 16 seconds to generate. The 2.53-second figure is pure inference time — queuing, first-frame latency, and network round-trips aren't included, and nobody has published end-to-end numbers. Cross-clip consistency is also unsolved: each 5–15 second segment is generated independently. The business math is the sharper part. Traditional content has high fixed costs and near-zero marginal cost per view. Generative streaming flips that: every hour of playback requires an hour of compute, costing $144–288 at 768p. At an estimated $0.10 gross profit per viewer-hour in live-streaming, a single stream needs 1,400–2,900 concurrent viewers just to break even. Per-user dedicated streams are a non-starter at $144+/hour per person. The article maps three viable models: one stream for all viewers (live shows), branched streams for audience segments (interactive short drama spinoffs), and on-demand generation at high-intent moments (ad conversion, e-commerce Q&A). Each tier demands higher margins — the closer to a transaction, the easier the math works. Where I'd discount: the post doesn't provide long-run latency stability data, and the queuing/first-frame overhead remains a black box. Content moderation isn't the real bottleneck — per-frame screening costs only 0.1%–5.4% of generation cost. The actual gate is platform access rules across different markets, not moderation tech.