Tweets

2024-09-27T14:20:16+00:00 | 🔗

This election, more than 2016, is a complete realignment of American politics.

2024-09-27T13:20:50+00:00 | 🔗

It’s interesting with all the talk of multi-turn chat and voice mode and ai avatars, how much more fundamental turn taking is then even language

2024-09-27T13:00:49+00:00 | 🔗

There is no counterbalance to Trump in 2024 vs 2016. We have project 2025, JD Vance instead of Mike Pence, and presidential immunity. Heritage foundation got Roe v Wade in round 1 and they are going for Project 2025 in round 2.

2022-11-11T04:58:13+00:00 | 🔗

Bigger models somehow didn't do better (even very very big models and also modest models) in the end the model is fairly small in parameter count

2022-11-11T04:58:13+00:00 | 🔗

More data = very good: I enriched the original dataset with more coordinate calculations from 10k -> 250k and got enormously better results in both loss and visual quality

2022-11-11T04:58:13+00:00 | 🔗

Loss is not everything: at first, I was getting very good loss at high noise levels so MSE was low but sampled depictions were bad because the diffusion model failed at lower noise levels.

2022-11-11T04:58:12+00:00 | 🔗

Making the code for DiffusionDrawer public (2D depiction of compounds). Will likely revisit with full documentation + tutorial at some point, but onto other projects for now! This was a great project to learn more about how ~diffusion~ lessons... https://t.co/8tewft7M2i https://t.co/9387JaBMmy

2022-11-10T15:48:07+00:00 | 🔗

In a way it’s good that the crypto crash coincides with the rise of AI, capital (and GPUs!) have somewhere to go 🤗😁

2022-11-10T00:59:00+00:00 | 🔗

@bertonearnshaw Thanks!

2022-11-09T13:00:28+00:00 | 🔗

@rguha Wow, yes, very similar algorithm in structure. The learning rate for SPE is like the noise level for diffusion. There probably is an argument that diffusion models are relearning the update step/distance metric of the SPE, and perhaps learning those with better-calibrated values https://t.co/aSqxPRzySp