If you’re a longtime reader who is curious where Theory of Vibe is at in 2025 in terms of ML science, then here be the post.
Back in 2015 I was very excited by the idea that what classic bottleneck VAEs trained on some natural dataset learn is a parametrized manifold with multiple “cognitive” functions/interpretations. A classic VAE learns, all at once: an algorithmic compression of approximate reconstructions of a dataset; a dimensionality reduction of the dataset, a global Euclidean chart you can use to meaningfully compare items in the dataset through their difference vector; a generative model of the data distribution; a (weak) decomposition into semantic features; and a (weak) anomaly-detection system.
My PhD thesis (comp lit with an assist from mathematician Tomer Schlank) proposed that these are also properties of the kind of “vibe-knowledge” of the world we live in we produce/express through art and poetry. We learn a very compressible approximate reconstruction of the world we live in, and a latent space and associated generative model that has myriad representation and search and anomaly-detection uses and gives meaning to comparing and contrasting “in-vibe” items. On this VAE-inspired account of vibe, given a vibe 𝑉 the “meaning” of an individual item in 𝑉 is its position in the space of 𝑉-possible items.
Does this hold up in 2025? Not without changes, but I think that on reflection these ten-years-in changes are conceptually robust and not just technical hotfixes.
The problem, from a modern ML POV, is that VAEs aren’t really good with natural datasets that have a lot of variation even at the surface level (pose, angle, scale, lighting, etc.) You can train a good VAE on a dataset of images of faces, but it’s harder to get something worthwhile on a dataset of (say) images of people unrestricted by pose, angle, and size. And VAEs are downright hopeless on something truly surface-diverse (but paradigmatically vibe-coherent) like “the set of all pictures taken in the street in modern Paris.”
You could say “maybe one day we’ll make VAEs work for these surface-diverse, vibe-coherent datasets,” but generative modeling of these kinds of datasets is basically a solved problem in 2025 -- it’s just not VAEs that do it. Modern diffusion and flow-matching models are extremely good at density estimation on surface-diverse natural datasets, but do they also learn a mathematical structure with the right multiplicity of “cognitive” functions/interpretations for Theory of Vibe? I think the answer’s ‘yes’! But with some important modifications to Theory of Vibe that are independently conceptually healthy.
The way I think about diffusion models currently (based on the recent wave of geometric accounts) is that diffusion models learn an implicit lower-dimensional manifold in input space but don’t learn a lower-dimensional Euclidean chart for it. (See for example Farghly et al. ‘25 on how generalization-capable diffusion models behave as if they were projecting the training data toward a lower-dimensional manifold in input space.) And I think this is partly why diffusion models work where VAEs don’t: a complex natural dataset like “the set of all pictures taken in the street in modern Paris” probably doesn’t have an underlying semantics that fits a single global Euclidean chart of reasonable complexity.
That said, I’ve been seeing more and more work from the last couple of years that suggests the implicit lower-dimensional manifold learned by diffusion models is semantically meaningful in a more indirect way. Per recent work (e.g. Kwon et al. ‘23, Kim et al. ‘25), taking the intermediate activations in a diffusion model as a latent space enables meaningfully training sparse autoencoders, or learning local tangent spaces around data points, or doing local PCA. In other words, you can use the internals of diffusion models to learn local lower-dimensional Euclidean charts covering the implicit lower-dimensional manifold in input space, and extract good “local” semantic feature vectors for the neighborhood of each data point.
So what happens if I start thinking about “vibe” along the lines of the implicit manifold of a diffusion model instead of the manifold of a VAE? We get a Theory of Vibe ‘25: vibe is a manifold that has generative, compressive, and weak anomaly-detection functions but gives individual data points “meaning” in a more local and indirect way than a classic VAE does. Instead of a vibe 𝑉 directly grounding meaningful global comparisons between any two 𝑉-possible items, 𝑉 enables the construction of meaningful bases of comparison in neighborhoods of 𝑉-possible items.
This is, I think, a much more natural account than Theory of Vibe ‘17 of how “vibe” -- and the kind of world-making we get from art and poetry -- is an enabling condition for more determinate cognitive operations. And it produces of its own accord a (very Kantian) contrast between this enabling vibe/determinate-cognition relationship and the resistance of vibe to global expression by determinate cognition!