Reposted by Berend ZwartsenbergEugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 01/02/2025Ask an LLM to summarize your paper because some reviewer will for sure read your paper this way 3525
Berend Zwartsenberg @berendz.bsky.social · 18/01/2025Just after I complained to @eugenevinitsky.bsky.social 020
Berend Zwartsenberg @berendz.bsky.social · 17/01/2025But they probably didn't specify what world they were models of 000
Berend Zwartsenberg @berendz.bsky.social · 16/01/2025Seems odd to me. If they actually had a correct prediction but no viz model, why would they render it as a class switching object? The train is bouncing between car and truck. I smell a struggling detector. 000
Berend Zwartsenberg @berendz.bsky.social · 16/01/2025I find it hilarious that the FB algorithm knows the best thing to do is to make jokes at its own boss's expense 000
Berend Zwartsenberg @berendz.bsky.social · 13/01/2025Until you realize that determinism isn't even a given between different hardware types and discover your dream is actually a nightmare 🥲 020
Berend Zwartsenberg @berendz.bsky.social · 02/01/2025One kWh / year is 0.114 Watt by the way, and so a k kWh / year is 114 Watt 110
Berend Zwartsenberg @berendz.bsky.social · 02/01/2025While it always seems natural, "k kWh / day" or "k kWh / year" has to objectively be one of the worst ways to write a unit of power. I wish we would all just use the humble Watt. 000
Berend Zwartsenberg @berendz.bsky.social · 14/12/2024Really enjoyed learning about this 🙌. It's still bouncing around my mind and I am hoping there is some way I can use it in my own work at some point. 010
Berend Zwartsenberg @berendz.bsky.social · 14/12/2024Best thing learned at neurips today: when pretraining your language model, the microwave gang subreddit contains posts that immitate the sounds of microwaves ("mmmmmmmmmmmmmmmmm ... beep!") causing spikes in losses and training instabilities. www.reddit.com/r/microwaveg... @hanna-nlp.bsky.social 010
Berend Zwartsenberg @berendz.bsky.social · 14/12/2024We have a poster at #ATTRIB2024 where we try to understand how diffusion models generalize based on patches. Effectively we're building a diffusion model from scratch, training free! Come say hello if you're around neurips.cc/virtual/2024... #Neurips2024 #diffusionneurips.ccAttributing Model Behavior at Scale (ATTRIB)NeurIPS 2024 000
Berend Zwartsenberg @berendz.bsky.social · 14/12/2024After 4 days of conference I'm excited to attend the #Neurips2024 workshops today. 010