Last week, @SophontAI hosted our first social at ICML in Vancouver! . It was great to meet many researchers in medical AI and folks interested in the space. Thanks to everyone who attended!

link:

The awesome Welch Labs released an incredible YouTube video about how AI image/video generation works!. This great intro video discusses CLIP, diffusion models, and classifier-free guidance in a visually easy-to-understand, approachable, concise way. Definitely check it out!

Great celebration party for FAL's big funding milestone, thanks for inviting me, had fun!

RT @gorkemyurt: Crew @iScienceLuvr @isidentical

Congrats to FAL! . FAL has an amazing team (@burkaygur @gorkemyurt @isidentical are awesome) and it's been great to watch them grow over the past few years into an incredible company. Extremely bullish!.

RT @jbhuang0604: Explaining Flow Matching in 4 minutes

abs:

RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents. "we introduce RLVMR, a novel framework that integrates dense, process-level supervision into end-to-end RL by rewarding verifiable, meta-reasoning behaviors. RLVMR equips an agent

abs:

AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. "Here we introduce AlphaEarth Foundations, an embedding field model yielding a highly general, geospatial representation that assimilates spatial, temporal, and

And so it starts, Meta is saying they will be "careful about what we choose to open source". It seems only Chinese companies remain dedicated to open-source. Will that change too in the future? Possibly. I also still hold out hope for Thinking Machines but let's see.

seeing people report popularity of their models via huggingface download metrics but can't that be easily manipulated?.

abs: project page:

X-Omni: Reinforcement Learning Makes Discrete Autoregressive Image Generative Models Great Again. "In this work, we demonstrate that reinforcement learning can effectively mitigate artifacts and largely enhance the generation quality of a discrete auto-regressive modeling

What people don't realize is that Apple actually has quite an underrated AI research lab, they're just bad at productizing it. I am typically very impressed by papers coming from Apple researchers.

RT @lcastricato: This AI-generated video game is running end-to-end on a mobile 5090 at 500FPS+. ⚡️. More coming soon. .

abs:

Back to the Features: DINO as a Foundation for Video World Models. "We present DINO-world, a powerful generalist video world model trained to predict future frames in the latent space of DINOv2. By leveraging a pre-trained image encoder and training a future predictor on a

abs:

Self-Guided Masked Autoencoder. "we propose self-guided masked autoencoder, which internally generates informed mask by utilizing its progress in patch clustering, substituting the naive random masking of the vanilla MAE. Our approach significantly boosts its learning process