Adam Hibble

New followers. No researcher is to be found here. I am a novel capabilities engineer doing random walk. (read: unhinged)

Combine with this, x.com/Algomancer/sta… and have fun.

Okay this in unhinged. But you can add a hook, that fakes being a agent team mate and make your application interface be on an agent team. gist.github.com/Algomancer/035…

@Algomancer don't buy it, you're a researcher

New followers. No researcher is to be found here. I am a novel capabilities engineer doing random walk. (read: unhinged)

there is something really magical about giving an agent access to the dom directly setup a js term, embed it, open claude, give it an api it can curl to to execute js - pretty fun. minify an app to a single file, insert it on the fly via curl. I think if this becomes like, less of a security nightmare its totally a thing.

This is my new favourite way to start a project. Self-bootstrapping Claude app. Three files, one loop. Tell the agent to modify itself. It will. github.com/Algomancer/cla…

I said this a while ago, but dude im so keen for code conditioned realtime video generation. We are close. react.js? Nah, just bash pipe to pixel space.

we went from memory is cheap so just write it in node.js to a ram shortage computing is cool again

One thing is for sure, i love the models, but it's really impressive how janky all cli interfaces are.

Distil from your friends.

Who would have guessed democratisation could have the ability to centralise power so much. wait.

@DeressaWodajo @_akhaliq @JCJesseLai @SwayStar123 @_advaitpatel @caglar_ee @_amirbar @tailin_wu I reckon you could pose the linear bridge from GAF in a similar way with the target step embedding.

This looks cool, reading through it now. Free gains if you swap out the Patch embed for the Online PCA Head from JiT. (Just down project to patch_dim * 0.25 before the embed) This looks like it'd compose nicely with the split mean flow formulation. Might be able to make it 1-NFE pretty easily especially because the base NFE is already pretty small. arxiv.org/abs/2507.16884

🧵1/ Introducing Generative Anchored Fields (GAF) Paper: arxiv.org/abs/2511.22693… Code: github.com/IDLabMedia/GAF Main idea: Instead of a single trajectory/score predictor, we learn independent endpoint predictors, J (noise) and K (data), from any point on a linear bridge.

@GalperinDaniel You can tell im not an llm by how many typos I make.

To me it seems like generally, you want some larger dimensionality then your datum, for the linear independence separability, as well as a dense assignment to entangle it losslessly from that higher dimensionality. I think there is a lot of ideas from representation learning that can inspired disentanglement and vice versa. For example, even the basic version of training a model and then doing monte-carlo high dimension random non-linear (ie - mlp with random weights per batch) projections and and pushing off-diagonal terms in the covariance matrix toward zero in the high dimension projected space gives a bunch of linear independence in the feature space. would be interesting to regularise a flow using something simple like that, bijective models in general with regularisation terms in under explored.

Thinking about the geometry of vae's on my nightime walk today. Not really sure where im going with it yet if anywhere. Every point in latent space has a kinda of split. Directions the decoder cares about, and directions it doesn't. Any point in latent space you can ask if I wiggle in this direction how much does the output change. Some directions change the output a lot, for others, the decoder is essentially ignoring them. Effectively, the standard VAE puts a fixed Gaussian prior over the entire latent space, treating every direction equally. But that's probably geometrically wrong. You're sampling uniformly over a space where huge regions are directions the decoder ignores. So, what I was thinking about is the question of 'what happens if the prior knew about this split?' If it said "spread out along the directions the decoder is using, but stay tight along the directions it isn't". The decoder already knows this, the matrix of partial derivatives (its jacobian) tells you exactly which directions it's sensitive to and by how much. The singular values of that Jacobian are a measure of how much our decoder stretchs along each direction, intuitively, the prior would say "go wherever the decoder is paying attention" and the encoder is all like "that's exactly where I want to put things". Obviously the rate distortion is always going to come into play, ie, more you compress, the worse your reconstruction. But the VAE has a second trade-off that isn't fundamental If your a particular datum (like an image patch) lives on some ten dimensional manifold in pixel space, but your latent space is sixty dimensional, then there are ten directions the decoder uses and fifty it doesn't. Geometrically it's not strictly a fiber bundle, but this is what I was thinking about initially, the reason fiber bundle language is tempting for me to think about is that when the rank is constant, the kernel defines a foliation. If you design the model with explicit conditioning variable that stratifies the space then you're kinda building something bundle like - ie architecturally you'd be able to impose a base and a fiber. But probs the construction doesn't need the foliation, it only needs something weaker. It just needs to know, at each point, a subspace the decoder ignores. The dimension of that subspace jumps around a bunch during training, so, that'll be something to keep in mind. Anyways, that last bit is a bit in the weeds/rambly, key thing probably worth trying is coupling the prior somehow to the the decoder geometry during training. I've seen some work on this, where people utilise geodesic computation in VAE latent spaces. Or metric tensor to construct priors that respect the learned geometry, or riemannian volume measure. So, its not completely unhinged which is always a good sign.

@sedielem I should probably put a less toy version together to compare. x.com/Algomancer/sta…

Training continuous diffusion models for discrete data with cross-entropy is neat (😎), but training flow maps with cross-entropy is even neater! Nice find!

@MParakhin Did you guys ever do Jacobian based regularisation in your ssl work @ylecun

@MParakhin Also, if your looking to use a non invertible archecture and still want the collapse protection, you can get the same properties by doing a trace estimate. github.com/Algomancer/Min… I do this in my jepas, beats eps pully etc.

@scottjmaddox @MParakhin But, you can trivially add the repulsion term, which might help and still be cheaper. Havn't battle tested in high dimension :)

@scottjmaddox @MParakhin this isn't sufficiently equivalent to the kernel in the drifting work. More just playing around with ideas, your milage may vary.