@tslominski presents on the upcoming and ongoing engineering platform changes. He explores the future of Monty, and how we’re hard at work to separate Monty from experiment infrastructure and simulated environments. Tristan also covers our plans for creating an inline telemetry pipeline that produces standardized logs, metrics, traces, and snapshots as these events occur.
Summary Video
Main Video
0:00 Upcoming and Ongoing Engineering Platform Changes
0:07 Welcome and Agenda Overview
0:42 Next Up in Platform
1:26 Why Are we Separating Monty, Experiment, and Environment?
7:56 Thousand Brains Project Mission
8:13 Two Worlds: Experiments and Operations
9:05 Experiment World Architecture: Desired Future State
13:20 Operations World Architecture: Desired Future State
16:12 One Consequence Regarding Telemetry
16:50 How Do We Gather Operational Telemetry? Emit Telemetry Inline
18:24 What is Telemetry? Logs, Metrics, Traces, Snapshots
25:40 Caveat Regarding Emitting Telemetry Inline
26:24 Concrete Telemetry Scenario: post_episode example
28:45 Q&A: Is This Future or Present Work?
30:34 Current Learning Module Memory and Emitting Events to Externalizate Memory
34:11 Q&A: Where Does Reconstruction Happen?
37:56 Q&A: Is Telemetry Configurable?
40:12 Q&A: How Would Saving Data and Handlers Work for Telemetry?
42:44 Telemetry Demo
53:07 Q&A: What Would the Telemetry Data Format Be Like?
55:32 Q&A: How Would Reconstruction Work for Custom Visualizations?
1:01:34 Closing Thoughts