@tslominski reviews the Thousand Brain’s progress in Q3 2026 and outlines the team’s priorities for Q4.
This update covers research advances, including figuring out how to represent attention in Monty, separating feature & morphology communication and processing, initial exploration of symmetry approaches, and improvements to compositional modeling.
On the platform side, the team made progress on the MuJoCo integration and Python upgrade goals, and continued progress toward removing experiment artifacts from Monty.
On the community side, we wrote and published our new 2-year TBP report, highlighting all the progress we’ve made these last two years.
Looking ahead, @tslominski shares the team’s Q4 priorities across research, platform, and community projects.
Main Video
0:00 Introductions
0:08 Q3 2026 Review
0:32 Theory - Progress on Open Theory Questions
1:51 Research Progress - Separate feature & morphology communication and processing
2:04 Research Progress - [RP] Initial exploration of symmetry approaches
2:26 Research Progress - [RP] Add attention + model-free segmentation to stay on object until recognized
2:39 Research Progress - [RP] Add model-based policies and attention for compositionality
2:52 Research Progress - Integrate Research Code and Create Figures for Two-Year Report
3:03 Research Progress - [RP] Made significant progress on adding top-down connections
3:33 Platform Progress - All benchmarks now have MuJoCo versions, in the near future Monty should no longer be constrained to Python 3.10
4:28 Platform Progress - We are working on making the Platform free of Experiment Artifacts
4:51 Platform Progress - No Engineering Progress on Telemetry, but Community Member AgentRev Merged an Initial Minimal Telemetry Stream Implementation
5:08 Community Progress - Write & publish 2-year TBP report
5:40 Collaborations & Communications
6:04 Created Varied Compositional Object Dataset
6:18 Rome Retreat
6:23 Rome Hackathon
6:33 Compositionality “Hackathon”
6:45 Filed Another Patent Application
7:00 Q4 2026 Priorities
7:12 Q4 Open Theory Priorities - Establish Promising Solutions to Unsupervised Learning
7:24 Q4 Research Priorities - Add Stage 2 Compositional Capabilities
7:36 Q4 Research Priorities - Enable Consistent, Relative 2D and 3D child poses
7:39 Q4 Research Priorities - Implement Inner LM Loop to Address Temporal Mismatches
7:44 Q4 Research Priorities - Add Model-Based Policies and Attention for Compositionality
7:48 Q4 Research Priorities - Add Top-Down Connections
7:50 Q4 Research Priorities - Generate Stage 2 Compositionality Benefits Figures
8:13 Q4 Research Priorities - Improve Unsupervised Learning Performance
8:42 Q4 Platform Priorities - Monty is not Constrained to Python 3.10
8:58 Q4 Platform Priorities - Platform is free of Experiment Artifacts
9:02 Q4 Community - Deepen Valuable Contributions to the Project