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

It’s been refreshing to see ai people starting to take RSI seriously. Look after your wrists, good ergonomics is important. Carpal tunnel sucks.

Kinda cooked on this one. My new favorite bijection for normalising flows, Predictive Rosenblatt Spline, closed form inverse. Belief over a gaussian codebook -> posterior predict-> spline knots, all analytic. Exact log-det at any K, zero init is the exact Rosenblatt map.

I need more compute.

Wait until they find about about H-Space. Y'all ready for some curvature.

Extension: Not happy with it yet, but playing around with an idea to have the data manifold as a point of equilibrium.

I think its pretty plausible directly optimising the environment for progressively increasing complexity is going to eventually out do whatever reinforcement learning verification loop we are currently indexing on. At some point we remember universal approximations are universal and figure out the environment is just another model. Data and RL Environments aren't very bitter lesson pilled if you can do without them, and there seems to be sufficient existence proof for model variants that do without them. Big if, but, really all it takes is in context learning not being an emergent property of scale or natural data distribution, but a learnable algorithmic skill that can be bootstrapped from mathematical structure alone and deployed.

Fable really likes fiber bundles. I see I made the fine tuning set.

Just did a quick implimentation of schidhubers power play from 2011. Something fun about models generating the task space with zero data, I have been thinking about how much you can learn just through mostly random causal induction recently. Fun paper. arxiv.org/abs/1112.5309

Fun and useful (for a change) - Wiped up a Closed-form JVP ConvNeXt-V2 U-Net. (FCDM style) Need this for some weird regularisation experiments, figured it might be useful to someone else. Probably useful for mean flows etc. Or when you need a volume change surrogate like i did. Has the conditioning tangent, so you can do all your time dependent jacobivibes. Closed-form JVP, so the whole encoder/decoder tangent is plain differentiable ops Two reasons this exists rather than a forward-AD dual_level pass over the plain net. It is compilable: torch.compile fuses jvp to ~plain-forward cost, versus the ~3-7x of the forward-AD dual. It is reverse-over-forward correct: Seems like on certain version of pytorch stock F.layer_norm's forward-AD rule gives a wrong second derivative. (the sign sometimes changes? Haven't investigated, but this wasted a lot of of compute). github.com/Algomancer/convne…

Not a lawyer-nerd. If a frontier model company, materially changes a prompt a user provided are they accepting exposure proportionate to the causal contribution and any duty assumed? Especially if you've voluntarily undertaken to do that, and the voluntary-undertaking doctrine, especially if its marketed for safety, since intervening to reduce harm is exactly what creates the assumed duty? Seems like, intuitively at least, modify users input is a liability landmine.