Atlas Frameboy demo released, streaming from cloud GPUs at up to 15 fps. It uses a highly optimized version of Atlas, an autoregressive world model with pixel-perfect camera control.
Key points
Atlas: autoregressive world model with pixel-perfect camera control
Frameboy streams from cloud GPUs at up to 15 fps
Sources
JITEngine custom inference engine for Apple Silicon up to 2.7× faster
JITEngine lets you synthesize a custom LLM inference engine for your model on a Mac. It claims up to 2.7× speed improvement over the fastest alternatives tested.
Key points
Custom engines built with SkyDiscover for Apple Silicon
Up to 2.7× faster than the fastest alternatives tested
Sources
- Original post
- jitengine.ai - Linked in the post
AD-Memo language-based memory driving agent announced
AD-Memo is a Vision-Language-Action autonomous driving agent that adds language-based memory to record surrounding objects critical to driving. The memory is appended to the agent’s input and used in later reasoning and decisions. Training uses supervised fine-tuning followed by Da Capo, a semi-closed-loop reinforcement learning algorithm that assigns trajectory-level advantage to memory and step-level advantage to driving.
Key points
Language-based memory extends the agent’s Chain-of-Thought to record decision-relevant objects.
Da Capo uses trajectory-level advantage for memory and step-level advantage for driving.
Sources
- Original post
- Vision-Language-Action Autonomous Driving Agent with Language-based Memory - Linked page
- kaiyan289.github.io - Linked in the post
- arxiv.org - Linked in the post
WeRide prepares public AV services launch in 2027
WeRide prepares public AV services launch in 2027 with GreenMobility. The company hosted Danish Minister Martin Lidegaard and Ambassador Michael Starbæk Christensen for a Robotaxi ride.
Key points
Welcomed Danish Minister of Business and Competitiveness Martin Lidegaard and Ambassador of Denmark to China Michael Starbæk Christensen for a Robotaxi ride.
Preparing for the launch of public AV services in 2027 with GreenMobility.
Sources
H200x8 node post-training under 4 hours using 8 H200 GPUs
H200x8 node post-training can finish in under 4 hours. The claim refers to proprietary work data and a single node with eight H200 GPUs.
Key points
Post-training using proprietary work data can be completed in under 4 hours.
Single node equipped with 8 H200 GPUs.