Wiedemer et al. (2025) — ARXIV-2509.20328
Key takeaways
The claim is about a trajectory, not a benchmark. NLP moved from task-specific models to generalist foundation models on three primitives — large models, generative objective, web-scale data — and the paper’s premise is that today’s video models have all three. What it demonstrates is that Veo 3 already does things nobody trained it to do; what it argues is that this is the same curve.
The breadth is the evidence. Segmenting objects, detecting edges, editing images, understanding physical properties, recognising affordances, simulating tool use — and then early visual reasoning built on those: maze solving, symmetry. A model that could do one of these zero-shot would be a trick; the range is what makes the generalist reading plausible.
It is a capability report on a closed model. Veo 3 is not open, its training data and recipe are not disclosed, and nothing here is a training recommendation. Read it as a claim about where the frontier is going, made by observing one system from the outside.
Standing in the anthology
Adjacent, and filed to mark the boundary. This record is about training recipes for language models: what to do and which published work says why. Nothing in it trains a video model, and this paper proposes no training change — it reports emergent behaviour and draws an analogy. Both halves are outside the registry’s scope.
It earns a note anyway because the analogy is a live question about the record’s own subject. If the video trajectory really does recapitulate the language one, the practices here — schedules, optimizers, sparsity, context extension — are the ones that would transfer, and the interesting question becomes which of them are about language and which are about large generative models trained on web-scale data. The record has no material on that question. Filing this is the first entry in a folder that is currently empty.
What would move it inside. A frontier video training report with a disclosed recipe — the video equivalent of what LIT-139 or LIT-131 are for language — is what would let a practice be filed rather than an observation noted.
Audited by #121; unsourced and unread. No reading note. The third of the cluster with LIT-215 and LIT-216. Its title makes a claim about capability emergence that SOTA-200 is directly about, so reading it would either corroborate or contest an in-force practice.