In the AI world, there’s a lot of talk about context. There’s even a term, “ context window ,” to explain the capacity of an LLM to see what’s around it, and perform inference that way. But there are also relevant design choices that can be a matter of debate.

There’s also a contrast between what you might call qualitative and quantitative context. Technical, quantitative context is the number of tokens that expand the LLM’s operating horizon, numbers of tokens, etc. The qualitative side involves humans evaluating how an LLM works intuitively, what it “knows” about its environment and the context of a given task or idea.

So whether you’re talking about hard context, in numbers, or soft context, in cognitive capacity, both of those are relevant to how we approach systems. I wanted to go into a presentation I heard recently at the TedX MIT Boston event, from Jake Sortor, Public Sector Strategy Lead at Blitzy, who was speaking about implementations in the defense industry. Brace for a lot of quotes; I thought Sortor spoke with quite a bit of eloquence.

“We spend a lot of time talking about model capability and compute power, and those are important, but in defense and intelligence, context is decisive,” Sortor said in opening. “This logic around context primacy is not unique to intelligence or defense. It shows up everywhere.”

“Your boss texts you, ‘we need to talk,’” Sortor said, laying out the scenario. “If you closed the deal of the year the day before, and the CEO praised you publicly, it probably means you're getting a promotion. But if there were layoffs last month and you've missed your last two deadlines, the same words probably mean it's time to update your résumé. The signal didn't change, but the surrounding context did.”

This is the sort of thing, he suggested, that LLMs have to do, too: they have to take a raw input, hold it up to the proverbial light, and parse it for context, in the ways that we do as humans with biological memory. Optimally, an LLM should be stateful.

But there’s a framing challenge, particularly in defense, which Sortor described as “messy, dynamic and high stakes.”

He described three fundamental failure outcomes that you have to consider as an engineer.

“The first is misclassification,” he said. “An anomalous signal is received, but the frame makes it look routine, or vice versa. Next is an assembly failure. The pieces exist, but never become a shared picture. And the third is update failure. The institution keeps reasoning from an old frame after reality has changed.”

Again, he gave examples, for example, from the U.S. response to Pearl Harbor, or the search for WMD in Iraq.

“These failures look obvious in hindsight, but they're nearly invisible in real time,” Sortor noted. “Across all three of these cases, information was not the issue. It was the system's ability to classify, assemble, and then refresh the frame in time.”

Here’s how he applied that to AI context:

“AI is entering environments defined by fragments, time pressure, competing frames, and buried dissent. The natural technical response is better data fusion, better retrieval, stronger models, and larger context windows. All of that matters, but it does not answer the harder engineering question: What should the model see, in what form, at what moment, and for what purpose?”

Then, Sortor questioned the axiom that, in context, more is better, and took a magnifying glass to the idea of a situation where more context is making things worse.

“The goal is not maximum context,” he said. “The goal is optimized context. We cannot treat context as a bucket to fill, even as AI development focuses on expanding context windows.”

“Long context can be powerful, but when relevant information is buried, redundant, stale, or surrounded by distractions, performance can degrade. Critical signals can get lost in the middle of long documents, dissenting data can be smoothed away by summarization, and a system can produce a confident answer that hides the uncertainty beneath.”

Here’s one more way that Sortor made the same argument, cautioning that we shouldn’t just context-chase, without, well, evolved and sophisticated context of our own.

“The challenge is not how much information the system can hold. It's whether the architecture provides enough context to make the signal meaningful—but not so much that the signal gets buried.”

Among all of the questions that Sortor asked, rhetorically, in explaining this approach to AI and context, here are a few clustered together.

“Now, with AI, more of those things can become intentional design choices,” he said. “What gets retrieved? What gets filtered? What dissent survives? What uncertainty reaches the operator? Those are context engineering choices.”

Now, I really liked this statement that Sortor made, apropos of the greater argument, because it really hits the point home:

“If we treat context as an automatic byproduct of larger models, we will not eliminate historical intelligence failures—we will execute them at the speed of compute.”

That sounds like valuable intelligence. I hope that leaders in the industry are paying attention.

As he closed, Sortor sort of coalesced the above into a thesis statement.

“The next defense advantage will not come from giving AI more information,” he said. “It will come from designing and engineering the context that turns information into correct meaning. I believe that work is now a national security imperative.”

Those are some pointers for anyone trying to achieve mission-critical result with AI. Yes, the systems are “smart,” but we have to be smart too, because, for now, humans are still in the driver’s seat. And we want to stay there. Stay tuned.