Jake

Risk taker 4 lyfe ; truth seeker 4 lyfe

Single step masked diffusion breaks because tokens aren’t independent. You can cheat the schedule, but not the dependencies. Have got something cooking in the background for few-shot distillation.

When it comes to discrete diffusion, you can’t just distill to single-step diffusion for free. All the entropy sneaks in through sequential unmasking, unlike the continuous case where you can map deterministically from the noise

The whole gist of this is that I’m trying to recast extreme value theory (EVT) as extreme events (with conserved geometry ie avalanches). In doing so I’m want to show that extreme risk depends on the geometry of fluctuations not just the marginal tails (which is classical EVT)

Hawkes model shows similar results. Ok so conserved fluctuation dynamics in branching (Hawkes) systems and relaxing (linear response/Barkhausen) systems. Next step: I’m gonna check if I can condition extreme value theory on avalanches

I have become increasingly interested in avalanching behaviour in dynamical systems. A quick simulation of a linear response model with Barkhausen noise shows that when eig->0 a system is more prone to extreme fluctuations (in both event time ie duration, and size ie magnitude)

Soooo this just happened. Opus 4.5 codes great. But just can’t follow my instructions and keeps deleting code blocks 😡

Merry Christmas 🎄🎄🎄. Generational lock in begins in exactly 1 week. Bing bong boom. Time to get my ducks in a row. 🫰🏼

Only 70 milly tokens this year 😭😭 I really need to pump up my numbers. And my consistency 🙃

Model is cooking. More to come on this when I get a chance to cook more.

Radiologists aren’t reconstructing pixels- they know what to look for. By learning to predict representations from partial views JEPAs can learn strong global structure. Also follows onto downstream FT tasks, where I’d expect better label efficiency

JEPA-style models feel especially right for medical imaging. Cooking up in-domain models for x-rays and CTs are so cool- but I’m increasingly convinced that a JEPA-style objective essentially forces a prediction of latent anatomy (rather than just reconstructing pixels)

@vishalja1n I’m holed up in the notes app on my iPhone- syncs seamlessly across iCloud to my MacBook and lets me farm dopamine by using the checkmarks feature for each of my tasks

There’s a reason this is where high-velocity brains end up.

Great friendships make life awesome. Had an amazing time at #NeurIPSanDiego with the fellas 🐦‍🔥🐦‍🔥🐦‍🔥

Working on an automated tool for grading clinical trial design. Used cursor to come up with a vanilla MVP, directly parsing .csv’s from trial databases, and a simple elif engine to score design details. Now tinkering about how to rewrite the grading engine using KG’s and LM’s.

Sampling latent space @ imperial