The difficulty around 可以 is bizarre. It's a more common word than most others that have multi-character tokens, it's more common than a lone 可, and at the end H-Net has decided 以用 and 以通过 are better tokens. Any idea what's happening there?

I saw the first screenshot without context and played all of SOUP ROOMS just to find who made it. It was even better in game. VELDTFUNERAL♡ is something special!

Likely prior art: Residual Matrix Transformers: Scaling the Size of the Residual Stream, Mak et al., 2025 arxiv.org/abs/2506.22696v1 Similar implementation (albeit different terminology), similar findings of broad benefit. No MoE/MTP/>1B scaling ablations though.

I know, gut reaction: LLMs can't do online learning. But the pieces are coming together for it. TTT makes new weights through online backprop. PEER stores weights in memory. UltraMem scales the memory up.

Why try to fix it with backprop though? For realtime, splitting into fast- and slow-thinking models seems inevitable. A HyperNetwork construction like "Learning to (Learn at Test Time)" where the planner edits the actor's weights would obviate the need for online backprop.

Understanding human biology across scales - from molecules to cells to entire organisms - remains one of biomedicine's greatest challenges in the fight against disease. Today, we are announcing Phenformer - a multi-scale genetic language model that learns to read and interpret human genomes by connecting DNA, cell and tissue context, molecules and clinical outcomes. Phenformer is a generative model of molecular mechanisms that enables researchers to unravel the mysteries underlying disease, and could thereby accelerate the development of precise future therapeutics.

The video for my REAC 2023 talk "Scripting Language? Engine Language? Why Not Both?" is finally online! piped.video/watch?v=TOnhqoUx…

Over two decades of focused work is leading up to the LAUNCH of the first public Atomontage product on the big stage at the largest startup event in Europe: Slush 2021 in Helsinki Dec 1st! Read more about it here: slush.org/product-launch/

Today, we are launching @Peer2PeerLive a robust tool that allows content creators and viewers on Twitch to connect to each other with identity based tags. This project is rooted in the immeasurable value of peer-support and community. Check us out at peer2peer.live/

Look at this shit.