InsightfulTechnical

Why context matters in AI | Jake Sortor | TEDxBoston

TEDx Talks

Jake Sortor argues that AI's greatest advantage will come from context engineering—strategically delivering the right information in the right form at the right time—rather than simply building larger models or collecting more data. He illustrates how historical intelligence failures (Pearl Harbor, 9/11, Iraq WMDs) resulted from context problems that AI systems will inherit unless intentionally designed to avoid them.

Summary

Jake Sortor, a former Air Force intelligence officer in special operations, presents the thesis that context engineering—not model capability or compute power—will be the decisive factor in AI deployment, particularly for defense and national security applications. He defines context as the ability to arrange fragmented information so that signals become accurate meaning, using the example of a boss's text "We need to talk" carrying different implications depending on surrounding circumstances like recent performance or layoffs.

Sortor identifies three historical modes of context failure in intelligence: misclassification (anomalous signals appearing routine due to incorrect framing), assembly failure (relevant pieces existing but never forming a shared picture), and update failure (institutions reasoning from outdated frames after reality changes). He illustrates each with specific historical examples: Pearl Harbor radar operators misclassified Japanese aircraft as friendly B-17s due to the wrong operational frame; 9/11 intelligence existed in silos without assembly into a coherent threat picture; and Iraq's aluminum tubes were misinterpreted as nuclear centrifuges because the dominant frame of Saddam's weapons programs was too durable to refresh despite contrary evidence.

Drawing connections to AI systems, Sortor argues that these institutional context failures translate directly into AI failure modes: intent routing errors, context fragmentation, and confirmation bias. He emphasizes that stronger models cannot fix inherited institutional context problems and may actually accelerate failures. With the U.S. military now deploying AI on classified networks (announced May 2026), the critical engineering challenge becomes determining what information models should see, in what form, at what moment, and for what purpose—moving from maximizing context to optimizing context. He warns that expanding context windows without architectural intentionality can bury critical signals and smooth away dissenting information, making confident but wrong answers more likely.

Key Insights

  • Sortor argues that context is not merely more information but the right surrounding information arranged so that meaning becomes visible—the same signal (a boss's text) can mean promotion or termination depending on the surrounding frame of reference.
  • Intelligence failures cluster into three modes: misclassification where anomalous signals appear routine, assembly failure where pieces exist but never form shared pictures, and update failure where institutions keep reasoning from old frames after reality changes.
  • Sortor contends that AI systems will inherit the same context problems of the institutions they deploy into, and a stronger model does not fix these problems but can make failures faster, more polished, and harder to contest.
  • Long context windows can degrade performance when relevant information is buried, redundant, stale, or surrounded by distractors—critical signals get lost in the middle of long documents and dissenting data gets smoothed away by summarization.
  • Sortor argues that the next defense advantage will come from designing intentional context engineering choices about what gets retrieved, filtered, what dissent survives, and what uncertainty reaches the operator—not from giving AI more information.

Topics

Context engineering in AI systemsHistorical intelligence failures and their causesAI deployment in defense and national securityContext window design and optimizationInformation architecture and framing in AI

Transcript

[0:00] [applause] >> Before I worked in AI, I was an Air Force intelligence officer in special operations. I deployed across the world on some of our most sensitive missions, and my biggest takeaway from that entire experience is that context wins wars. People often think intelligence is mostly about access and collection, and both are certainly necessary, but they are insufficient. What distinguishes excellent intelligence is context, the ability to take fragmented, disparate, and often [0:31] random representations of reality and arrange them so that a signal becomes accurate meaning. My argument today is that our greatest AI advantage will not come only from bigger models or more connected data. It will come from better context engineering, the still…

Full transcript available for MurmurCast members

Sign Up to Access

More from TEDx Talks

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.