Parallel Language Models achieve up to 5.7× faster inference while keeping autoregressive-level quality, and they scale to 3B parameters. The team will open-source the code and model.
Key points
Speedup: 5.7× faster inference at autoregressive-level quality.
Scale: model scales to 3B parameters.
Sources
AI agent skills passed scanners, study shows 3 rules
AI agent skills passed every security scanner and still asked agents for actions that common security controls forbid. A September 2026 study explains why a clean scan isn’t enough and lists three rules that worked.
Key points
705 AI agent skills passed every security scanner.
September 2026 study shows why a clean scan isn’t enough and provides three rules.
Sources
Binance hack, wallet drained, TX hash 0x53bd38b8...
Binance hack possibly occurring now. A Binance wallet was drained to a new address. The transaction hash is 0x53bd38b816c23e74134786c29a8c18d792777998a3b9326047b583100e88e2f3 and was funded by Tornado Cash.
Key points
TX hash: 0x53bd38b816c23e74134786c29a8c18d792777998a3b9326047b583100e88e2f3
Funding source: Tornado Cash
Sources
Holo4D accepted to NeurIPS 2026, video reshoot capability
Holo4D is accepted to NeurIPS 2026 and can reshoot a video from a new camera path. It lifts video into 3D tracks and uses them to steer a video model, keeping scene content unchanged.
Key points
Acceptance: Holo4D accepted to #neurips2026.
Function: Gives a video and a new camera path, lifts the video into 3D tracks and uses them to steer a video model.
Sources
PixGR: Fast Feed-Forward 3D Reconstruction
PixGR: Pixel-Space Generative Decoding for Fast Feed-Forward 3D Reconstruction combines pixel-space flow matching with feed-forward 3D Gaussian reconstruction. It improves geometry and appearance in unseen regions while keeping inference under a second.
Key points
Combines pixel-space flow matching with feed-forward 3D Gaussian reconstruction
Improves geometry and appearance in unseen regions while maintaining sub-second inference
