Recently I came across an AI-generated short drama, and it felt like watching an atomic bomb go off. I was so stunned I could hardly sit still.
The text and textures in the video were stable, and objects stayed the same across cuts. Complex scenes were generated remarkably well.
The motion still isn't smooth, and the pacing is a mess. But what will the next generation of models look like? What about six months from now?
Large models look more and more like man-made gods.
First, make it work#
My guess is that the pacing problems come from a tradeoff the creator made between visual stability and script quality. Video quality still comes down to the luck of the "gacha pull". To get the work made at all, the creator had to accept ill-fitting shots in exchange for keeping the characters, motion and scenes stable.
This is what AI research has always done: first make the model work, then solve the engineering problems.
It's foreseeable that later models will understand requests more reliably,
and generate consistently good footage from shorter prompts. Creators can then put more of their energy into plot, pacing and characters. Just like what is happening in programming: the shift from vibe coding to wish coding.
Good stories won't appear on their own, of course, but the cost of telling a good story can come down.
One success doesn't mean reliable success, but it at least gives later research something concrete to work on. Once you know why things fail, you know which engineering problems need solving.
When "can a machine do this" is no longer up for debate, the remaining questions become much clearer.
Mathematics too#
Recent AI work in mathematics has made this feeling even stronger.
OpenAI reported that one of its internal systems produced a proof for the N-S problem, with Astra taking part in the later formalization and verification.
Anthropic, for its part, reported that a research version of Claude improved a lower bound related to the Riemann hypothesis by combining existing mathematical results.
Set aside the mathematical value and novelty of this work, and how much the model really contributed. Whether or not it "relied on existing human ideas", the model has proven that it can work. Whether or not it "relied on existing human ideas".
Even if the problems being solved matter only a little, this capability deserves attention. The value of a result and the capability needed to achieve it are two related but different questions.
I think finding connections between existing results that nobody had noticed, and applying a method to a new problem, is innovation in itself (today's models can already do this, and they will foreseeably do it better).
That's why I don't like ending the "debate" with "it still can't discover new methods on its own." Even if it can't now, what about six months from now? How much human guidance a model needs, how far it can push its reasoning, whether it can find the next route from its failures: these are all worth discussing.
Cutting token consumption is a later problem. First do one thing well, then do it well cheaply.
Man-made "gods"#
What's truly frightening about large models is generalization. The same model can carry out different tasks, and people no longer have to design a complete solution process for every task.
The independent team Robocurive has already connected Astra to a robotic arm to test tasks such as picking and placing objects.
Grasping a block isn't impressive in itself; the recent "robot games" already showed off the divine power of RL. But with Astra, a general-purpose model is starting to be used to interact with the physical world, and it does pretty well.
If language, vision, reasoning and tool use can be connected further, the tasks we hand to models can become more and more complete and concrete. Users won't need to master the skills required to reach a goal, or even know how it should be done; they can leave everything to "Lord AI".
The model becomes a "wish machine", close to a "god".
The "god" doesn't even have to be omniscient and omnipotent. It only has to understand what we ask across a wide enough range of things, and do work we couldn't do ourselves.
As models take part more and more in proposing and testing hypotheses, the pace of iteration has already broken away from the pace of human understanding.
The concern Dario expresses in "We Must Pace the Frontier" is that it is hard for us to explain concretely why a complex model made a particular choice in a particular situation.
On the one hand, we need to know "why the model did this"; on the other, human understanding has to keep up with how fast models iterate on themselves. The fear comes from how unimaginably far models can generalize.
There's a line that's been said to death but still fits: "Humanity isn't ready."
Slow down how fast frontier model capabilities improve, so that safety research and external evaluation have time to catch up.
At heart, this gives people a little time to accept the fact of frontier models' "divine power", and to cushion the shock of the "nuclear blast". Fear of AI needs no commercial motive. Researchers at the frontier may feel this most of all.