Parallel Language Models predict several tokens in parallel at autoregressive quality and add a parallel decoder at pretraining, delivering a 5.7x speedup after posttraining. The approach changes how LLMs are trained.
Key points
Parallel Language Models predict several tokens in parallel at autoregressive quality.
Adding a parallel decoder at pretraining yields a 5.7x speedup after posttraining.
Sources
- Original post
- arxiv.org - Linked in the post
- justuswill.com - Linked in the post
Idol open image-to-3D model with 600M parameters runs on consumer GPUs
Idol is an open image-to-3D model that generates complete 3D objects from a single image. It contains 600M parameters, runs on consumer GPUs, uses less than 7.2GB VRAM, and can produce 1M splats in 3 seconds.
Key points
Parameters: 600M
Performance: 1M splats in 3 seconds using <7.2GB VRAM
Sources
Radiant microreactor safety analysis approved
Radiant microreactor safety analysis approved for full-power operations by the Department of Energy. Testing will be conducted at Idaho National Laboratory.
Key points
Department of Energy approved a commercial nuclear microreactor's safety analysis for full-power operations.
Radiant secured the approval, with testing planned at Idaho National Laboratory.
Sources
SteamOS on Pico 4 – update system added, battery lasts ~1 hour
SteamOS on Pico 4 now runs the latest version via a new update system, and the headset currently provides about one hour of use before a 2.5-hour recharge. Development progress is limited by the short battery life and personal time constraints.
Key points
Update system added, now runs latest SteamOS version.
Headset battery provides ~1 hour of work, recharge takes ~2.5 hours.
Sources
Antigravity plugin marketplace expansion announced
Antigravity is expanding its plugin marketplace. The tweet says a new plugins marketplace has been announced and search and discovery are now easier.
Key points
New plugins marketplace in Antigravity
Easier search and discovery of customizations
