Pablo Castro (CVP & Distinguished Engineer, Microsoft) frames the AI-and-knowledge problem through three lenses: intrinsic knowledge baked into models, extrinsic knowledge that agents retrieve to ground themselves, and learned knowledge captured by continuously improving agents. He traces the exponential arc from IntelliSense (1996) through ML ranking and GitHub Copilot to today's agents, then demos how Microsoft Foundry, Microsoft IQ, and Azure AI Search deliver company-wide grounding and agentic retrieval. He closes on a real learning loop via Foundry's agent optimizer, which hill-climbs an agent's instructions against an evaluation.
Castro opens by describing his role at Microsoft as connecting the dots between AI and knowledge, and frames the talk around three categories: intrinsic, extrinsic, and learned knowledge. Intrinsic knowledge is what comes baked into the models from training data and stored in their parametric memory.
He argues intrinsic knowledge is what threw the industry into today's exponential. Illustrating with two pieces of code written 25 years apart, he notes the authoring process was surprisingly similar: sit down with what you know or look up, then write it. The same applies to writing an email or summarizing a document.
A useful starting point is Microsoft's IntelliSense in 1996, which meant you no longer had to remember function signatures. It took 22 years before machine learning helped rank the IntelliSense options, and just three years after that GitHub Copilot launched, even before ChatGPT was announced.
From there things shifted quickly: Cursor and GitHub Copilot X launched, then Opus 4.5 shipped late last year with GPT and other models rapidly getting better at coding, leading to software built early this year with not a single line written by hand. This exponential was powered by intrinsic knowledge plus the ability to reason.
Intrinsic knowledge only gets you so far when an agent needs to participate in an organization. The industry saw the RAG pattern emerge as a low-tech technique that evolved into sophisticated context engineering. Castro highlights two dimensions of complexity: the move from isolated datasets to company-wide grounding, and the evolution from simple vector search to complex retrieval systems.
Microsoft IQ is the single entry point to an organization's ambient data. It is a set of capabilities: Work IQ (documents, email, calendar, chats and people connections in SharePoint), Fabric IQ (data warehouses, data lakes, PowerBI reports), Foundry IQ (your own pushed grounding data), and Web IQ (public web information).
For a hot second the industry thought getting really good at cosine similarity between vectors was all that was needed for retrieval, but evaluations from Azure AI Search (the technology behind Foundry IQ) show combined methods consistently beat individual ones on real-world scenarios. Foundry IQ layers the stack so users can hand it PDFs and let it chunk, vectorize, rank, and do agentic retrieval automatically, or drop to the bottom to control quantization, indexing algorithms, and lexical retrieval.
On top of core retrieval, Foundry adds an agentic retrieval stack that reflects on the dataset and decides whether the stated information need is satisfied before returning results. In Microsoft's own evaluations, agentic retrieval continuously beats simple single-shot retrieval on metrics like evidence recall and answer completeness. In the demo, each knowledge base is exposed as an MCP server, so any existing harness can connect without glue code. Token efficiency is a stated design goal, aiming for the most information-dense answer with the fewest tokens.
The third category, learned knowledge, is the result of observing processes and improving them by reflection now that agents do the work and can be tuned automatically. Castro references Satya writing recently about how people and agents compound to create a learning loop that captures what is unique to an organization.
Foundry materializes this with the agent optimizer. Using the Foundry toolkit in VS Code, a developer with an agent that externalizes its configuration can run eval generate to produce a task-adherence evaluation, then optimize to hill-climb the metric via a GEPA-style loop over candidate configurations (roughly a 45-minute run). Running optimize apply swaps in the winning configuration, whose instructions emerged from the process and reflection on real traces rather than being handwritten, a real learning loop in practice.
"My job at Microsoft is to connect the dots between AI and knowledge."
"I would argue this is the knowledge that actually threw us into the exponential we are in today."
"For a hot second as an industry, we thought that if we could get really really good at computing cosine similarity between vectors, we were all set for retrieval. It turns out things are never that easy."
"Every knowledge base is an MCP server, so you can just connect to it without having to write any glue code in the middle."
"You can see like a bunch of instructions that are not handwritten, but that they emerged out of the hill climbing process. So this is a real learning loop materialized in practice."
| Time | Topic |
|---|---|
| 00:36 | Pablo Castro takes the stage: AI and knowledge |
| 01:39 | Three categories: intrinsic, extrinsic, learned |
| 02:40 | The exponential: IntelliSense 1996 to modern agents |
| 05:46 | Company grounding with Microsoft IQ |
| 07:18 | Layered and agentic retrieval in Foundry IQ |
| 13:07 | Learned knowledge and the agent optimizer |