Mar 11, 2024 · 51m · a16z
Intelligence in the Age of AI with new CTO of the CIA
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of a16z's American Dynamism podcast, CIA Chief Technology Officer Nand Mulchandani and General Partner Martín Casado discuss the strategic, technical, and cultural impact of artificial intelligence on national security and intelligence operations. They explore how AI transforms intelligence analysis, the compute infrastructure bottlenecks facing defense, and the imperative for public-private collaboration between Silicon Valley and government.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 1.9 Guest teaching 6.0 Guest disagreement 2.7 The host pushing back 0.9
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 45:19 Rejection of procurement friction focus
Martin directly counters host Derek Harris's suggestion that Valley dialogue and startup mentality ease procurement friction, stating that procurement issues are secondary and warning that policymakers are fighting the last war.
Hardest push from the host ▶ 17:03 Challenging speculative automation timelineHost Derek Harris brings up a published account of a CIA analyst imagining prompt-and-coffee automation to challenge whether such extreme hands-off workflows are realistic or hyperbole.
Biggest teaching moment ▶ 18:05 Kernel smoothing and tail reasoning breakdownMartin educates the host on the mathematical functioning of LLMs as kernel smoothing over positional embeddings, demonstrating why LLMs suit average tasks but fail at tail reasoning critical to intelligence work.
The host holds their own ▶ 27:40 Distinguishing technology open source from OSINTHost Derek Harris demonstrates domain knowledge by explicitly clarifying the distinction between open source software tools and open source intelligence gathering (OSINT) in intelligence adoption.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Defining the Strategic Role of the CIA Chief Technology Officer | 1 | 5 | 1 | 0 | Host Derek Harris introduces the segment with a straightforward background question about the newly created CIA CTO role. Nand Mulchandani provides an institutional overview, detailing CIA's historical pivot from counterterrorism to great power competition and technology. The host purely listens without introducing technical assertions or counterarguments. | |
| Evaluating AI Impact Across Intelligence Operations and Analytics | 2 | 5 | 1 | 1 | The host guides the conversation from early machine learning to generative AI, prompting the guests on offensive versus defensive capabilities. Martin Casado and Nand Mulchandani reframe AI impact into signal detection, deepfake identification, operational cat-and-mouse dynamics, and push versus pull analytical models. The dynamic is collaborative and educational. | |
| Understanding AI Limits and Reimagining the Analyst Role | 2 | 6 | 3 | 1 | Host Derek Harris asks about the timeline for automating intelligence analysis. Martin Casado explicitly reframes the premise, arguing AI is not an asymmetric superpower like the internet and explaining exponential error accrual in agentic workflows. Nand adds practical context on LLM deployment inside the CIA and the distinction between creative hallucination and analytical rigor. | |
| Analyst Skillsets, Tail Reasoning, and Software-Defined Intelligence | 2 | 7 | 1 | 1 | Martin breakdown model mechanics, explaining kernel smoothing over positional embeddings and contrasting mean outputs with tail reasoning where intelligence work resides. Nand elaborates on co-pilot dynamics and the rigorous standards required for Presidential Daily Briefs. The host asks open questions that allow guests to showcase domain mastery. | |
| Technical Compute Bottlenecks, Data Quality, and Risk Thresholds | 2 | 7 | 2 | 1 | Nand articulates concepts of software-defined intelligence, code as law, and surfacing probabilistic decision thresholds to analysts. Host Derek Harris prompts on open source adoption, prompting discussion on data provenance, varietals of LLMs, and compute bottlenecks returning computing to a mainframe-like era. | |
| Infrastructure Discrepancies and the Commercial First Strategy | 2 | 6 | 5 | 1 | When the host asks about public-private DC-Silicon Valley compute partnerships, Martin Casado directly pushes back on the assumption that private venture markets can solve compute bottlenecks alone. Martin critiques current policy directions for pulling back from public supercomputing investments, pointing out historical government leadership in supercomputing infrastructure. | |
| Cultural Shifts at the CIA and Navigating AI Policy | 2 | 6 | 6 | 1 | Host Derek Harris asks if startup mindset and Silicon Valley dialogue ease procurement friction. Martin explicitly rejects the host framing, asserting procurement friction is an age-old issue and not the primary bit, warning instead against fighting the last war on internet asymmetry. Nand provides institutional nuance on CIA cultural shifts and emerging technology policymaking. |
Statements from this episode (17)
Casado: Modern AI clusters are the largest compute projects in history
“These are the largest compute projects mankind has ever done before. Like, we've never done anything close to this.”
Mulchandani: CIA pivoted focus from counter-terrorism to great power competition
“And the conclusion was actually fairly interesting and not a surprise, was that we had to pivot from CT, which had been the big sort of focus for the agency for a couple of decades, to great power competition.”
Casado: Fingerprinting AI-generated content is not difficult
“It actually turns out that if a computer generates something, the ability to kind of fingerprint that isn't that difficult. It's actually not that hard, right? So I actually think it becomes much, much easier to detect if people are using AI and tooling actual…”
Mulchandani: AI use in spy operations will continuously escalate
“The democratization of this stuff and the availability we know is going to drive each one of our competitors to be driving this up. We're going to be aware of that. We're going to drive our own stuff up. So there's this aspect of, we don't know where this is g…”
Mulchandani: AI tools risk amplifying intelligence analysts' cognitive biases
“Well, it can then also, unfortunately, take the things that you as an analyst are weakest at, or it's your weak knee or your thing that, that it can get into your head. And start rabbit-holing you down this thing, which amplifies your biases.”
Casado: AI will not create an asymmetric advantage for offense or defense
“It doesn't change the equilibrium in the sense that, that, you know, there's going to be more tooling for offense. There's going to be more tooling for defense. This is no like asymmetric superpower, which by the way, is very, very different than the internet.”
Casado: There is no evidence of agentic behavior in LLMs
“The thing that we've not seen any evidence of is agentic behavior.”
Casado: LLM errors accumulate exponentially when operating out of distribution
“If it generates anything out of distribution. Like, and by the way, out of distribution means it's not commonly represented in the training set. Then that error is going to accrue and it tends to accrue exponentially provably, right?”
Mulchandani: CIA has deployed large language models for open-source intelligence
“We're public with the fact that we actually have LLMs in production at the agency. We have it in production in the open source team.”
Mulchandani: Deploying battle-ready AI in defense will take a long time
“So it's funny because like the excitement and hype about this technology versus us absorbing it and making it battle ready is a long, long distance. And so I don't, I'm hopefully, you know, portraying or representing that side of the equation, which it doesn't…”
Mulchandani: AI forces explicit decision-making around probabilistic output thresholds
“With these newer algorithms these are still probabilistic algorithms. Except now the probabilities actually stare you in the face in a way that previous systems didn't push, right? So previous application systems never came up with said, do you want the 49%, t…”
Casado: Open-source AI models return software development to the mainframe era
“Even if the weights and biases are open source, I don't know how much you can modify it. Right? So it almost feels like we're going back to this old mainframe day where, like, it's great to have it, and you can operationalize it, but you're not going to have t…”
Mulchandani: CIA lacks the compute cycles to modify open-source AI weights
“Anybody's ability to do something with that, those, with that information is limited, because it's not just the numbers. You have to have the expertise you need to Compute. You need all the other pieces. That's my point, which is great. I have that information…”
Mulchandani: Universities lack compute power to compete with tech giants
“When you go talk to universities right now, one point they make is they don't have the compute power to be able to rival Microsoft and OpenAI and these companies”
Casado: Private markets alone cannot solve national AI technology needs
“I don't think the private markets solve everything. I do believe in, like, Public private partnerships. I do believe in institutions. I do believe that the government has a big role to play here, but I think that role to play is investing heavily in, in people…”
Mulchandani: Hardware commoditization will make running AI models cheap at scale
“The availability of, you know, hardware commoditization and other pieces will get to a point where we'll be able to run all kinds of interesting algorithms At scale with really cheap, readily available hardware, right?”