We invest early in the founders who are rebuilding the development stack for the agent era

Embeddings are a good building block, but they are not solution in itself. What we are trying to do is we're trying to create this graph on top that gives us all the relationships and possible interaction points of these embeddings. And we then get another layer of meaning. And these can be layered up… merged and cross-analyzed by different other algorithms. And that gives us a lot of flexibility in the design. Now we’re working with Berkeley and Brown to actually have this predictive memory that we’re predicting and paths of possible futures and then merging them.

Entire regions of duplicate servers on standby because there is no multicluster orchestration layer. Now multiply that by every cluster on earth, and you’ll begin to see the scale of leakage. We live in a world where compute is the most scarce resource, and yet enormous quantities sit idle and unused every day. Thousands of new datacenters under construction to serve an asset that already exists. The compute for AGI is already online - it just needs to be unlocked and connected.

‍

This gap shows up as model unpredictability: agents perform well in some cases and fail in others. As they move into critical workflows and longer-running tasks, these failures become the bottleneck. The problem isn’t capability, it’s adaptability, every failure still requires humans to inspect traces, diagnose issues, patch prompts, and redeploy. The next phase of AI is not just better models, its continual learning systems. Agents that learn directly from production, adapt to new scenarios automatically, and become more reliable over time.