Open 27B model trained on their data gained 12.2 points on held-out GitHub computer-use tasks, rising from 35.6% to 47.8%. They built an open post-training stack to measure RL task quality.
Key points
Open 27B model gained 12.2 points on held-out GitHub computer-use tasks, from 35.6% to 47.8%.
Built an open post-training stack to measure RL task quality.
Sources
DreamFAR announced world model trained from robot data
DreamFAR is a world model trained from scratch using only robot data, without internet pretraining. It supports real-time teleoperation, minutes of long-horizon rollout, and generalization across 1,000+ tasks.
Key points
Trained from scratch using only robot data, without internet pretraining
Generalizes across 1,000+ tasks
Sources
- Original post
- yimingdou.com - Linked in the post
Gemini 4 Carbon launch, #1 on Artificial Analysis index
Gemini 4 Carbon launch is announced as #1 on Artificial Analysis’ new index. It is being used internally on Antigravity and access to the Fairwind Program is expanding before wider availability.
Key points
Gemini 4 Carbon is #1 on Artificial Analysis’ new index.
Access to the Fairwind Program is expanding before wider availability in Antigravity.
Sources
NYU PhD program feedback request deadline Nov 6
NYU’s computer science and data science PhD programs request statement of purpose feedback from current PhD students by Nov 6, first-come-first-served. The opportunity is limited and fills on a volunteer basis.
Key points
Deadline: Nov 6
First come, first served volunteer spots
Sources
SwarmChat library released for multi-agent coordination
SwarmChat is a messaging library for agents that lets them communicate, coordinate and logs interactions in a Slack-like app. Performance is said to climb quickly when agents use SwarmChat, and it can be installed with pip install swarmchat.
Key points
Installation: pip install swarmchat
Interactions recorded in a human-readable Slack-like app
