The Pentagon doesn't lack models — it lacks the chain that puts intelligence on the president's desk
The real miracle of the U-2 wasn't the aircraft — it was procurement reform, the intelligence analysis organizations, and the chain that carried judgment to the president. Today AI has far outrun the government's ability to integrate it. APIs alone are not enough.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The U-2's miracle isn't the plane, it's the procurement and analysis chain
Garrett says the U-2 story is usually told as a technology breakthrough, but what really mattered was three things: procurement reform that let government and industry work together that way, intelligence analysis being institutionalized (NPIC, which later became NGA), and getting judgment in front of the president and other decision-makers. That, he says, is the ‘miraculous’ part of the story. Today's AI problem is the exact inverse: the technology comes from the commercial world and moves too fast, already far beyond the government's ability to integrate it into systems, workflows and decision processes — so ‘APIs alone are not enough’.
— Garrett BernsonAI is advancing fastest in warfighting, slowest in business systems
Garrett says the Pentagon and other departments have made visible progress in warfighting spaces like CJADC2: using AI to build situational awareness across the battlefield, identify targets, and get intelligence quickly to frontline decision-makers. But core business systems — logistics, personnel, finance — have seen the least progress. His personal view is that systems should no longer be split into ‘mission systems’ and ‘business systems’, because supporting the front line requires clean, well-organized data usable at the tempo of operations, and those business systems were never built for that tempo.
— Garrett BernsonOne policy rule blocks data for twenty thousand logistics troops
He gives a concrete example: some business systems' rules say status only needs to be updated once a week. That's policy — someone wrote it down. But decision-makers now want that data daily, which means going to roughly ten or twenty thousand Army supply sergeants and saying ‘you have to enter it daily, not weekly’. That's ‘a big boat to turn around’, and it has to come with clear value. Worse, the systems the government gave them are already bad to use — making it hard to enter data while also demanding a big reduction in data latency.
— Garrett BernsonGovernment has no market competition as a forcing function
The host points out that a company that doesn't do AI gets eliminated, but the Air Force or the national security system won't be ‘competed out’ — that kind of forcing function comes once or twice a century. Garrett agrees, but says we underestimate our own ability to manufacture mechanisms. He argues for using personnel policy and leadership as carrot and stick: reward people who take risks and take strange career paths, and let those who move too slowly face career stagnation or removal. He says it sounds ‘crude’, but that's the mechanism you have to follow.
— Garrett BernsonSecond-mover advantage: watch others burn cash and hit the pits first
Garrett says this time it's the reverse of the U-2: the technology happens in the commercial sector first and then enters government, so government has a second-mover advantage and can see what's working. He cites the lesson of token economics: do you open it up like Uber and let everyone burn tokens without limit, blowing the budget in a quarter, or do you understand token economics first and then deploy to employees? Also, as the federal government, if it gets this right it can act as a super-sized buyer and negotiate volume pricing.
— Garrett BernsonPutting GPUs at the edge first means clearing the policy of who signs and who's liable
Garrett says if you put tools in soldiers' hands, you'd be shocked at how innovatively and cleverly they use them for their actual jobs — which may mean pushing compute to the edge, putting GPUs in service members' hands. But the obstacle is policy: a logistics officer signs for $500,000 worth of chips, and who's responsible if a chip breaks? He says you have to clear the policy and just say ‘we own this risk’ — if it means some GPUs get broken and money is lost, that's worth it, because capability reaches the edge.
— Garrett BernsonIntelligence analysis is more tractable than modeling decisions
Garrett thinks using AI to model strategic decisions is very hard, because it needs good data, and the decision process is deliberately opaque — you can't get all the private notes, emails and situation-room audio and video. But intelligence analysis is a more tractable project: the intelligence process is structured and rigorous, agencies have their own SOPs and training, there are decades of data, and there are signals of quality high and low that you can build benchmarks against. Still, it's only one input to decisions — historically many decision-makers looked at the same intelligence and made completely different decisions.
— Garrett BernsonBill by the hour and people will report more hours
Garrett says the government buys a lot on hourly rates by person and seniority, and this time and materials structure has a status quo bias. The logic is plain: if the incentive is for companies to report more hours, and there's no hard delivery requirement at the end of those hours, they'll report more hours. He says there have been signs of change over the past 12 to 18 months, but looking at public spending data, there's no large-scale shift — most spending is still time and materials. He also mentions that top companies spend 10% of their budget on tokens, while the US government is far below that level.
— Garrett BernsonIn their own words · checked verbatim
It's not just the technology, not just the plane. It's that the way they bought it, uh, was important. The innovations around acquisition reform, uh, the way they acquired it was important.
Garrett Bernson3:02
it's moved so far ahead of our ability to sort of integrate it into the actual systems into the workflows into the decision-making processes that we now are kind of flatfooted in this reality of this technology.
Garrett Bernson4:04
some of these business systems literally the rules say you only have to update the status once a week on some form. Well that's a policy.
Garrett Bernson10:06
People are ambitious and people want to people want to do a good job like genuinely. And if you have a system that is encouraging more of this type of behavior, it looks like people taking big risks doing stuff like this, weird career trajectories are rewarded, right?
Garrett Bernson14:10
You have to clear the policy out and just say, "We're assuming the risk of this." Uh, and if it means some GPUs get broken and we lose the money, well, it was worth it. Uh, because we're getting capability out to the edge.
Garrett Bernson23:20
if you incentivize a company to bill more hours and there's not like a hard requirement to get something done at the end of those hours, they're just going to bill more hours.
Garrett Bernson38:36
Figures
| Accenture business mix | 92% commercial, 8% federal | 41:38 |
| Share of budget top companies spend on tokens | 10% | 39:37 |
| U-2 service status | Still in service today | 2:02 |
Glossary
- CJADC2 / Combined Joint All-Domain Command and Control
- The US military's command-and-control concept for linking sensors and decision-makers across all services into a single network.
- NPIC / National Photographic Interpretation Center
- The photo intelligence analysis organization established in the U-2 era, which later evolved into NGA.
- time and materials
- A procurement method billed by hourly rates and actual materials used, not tied to delivery outcomes.
- token economics
- Cost accounting that measures the relationship between AI model invocation consumption and the value produced.
How to listen
For engineers, product leads and investors working on AI deployment in government or defense, and for enterprise technology managers who care about procurement reform and token cost accounting.
The last ~5 minutes cover book lists and parenting books, unrelated to the topic — skippable.