AI
Why Government AI Adoption Defies the Silicon Valley Playbook
The US Army has over a million people, and roughly a third are part-timers in the Guard and Reserve. That scale alone breaks most AI rollout assumptions. Add a workforce locked into “very specialized discrete roles” with prescribed career timelines, and you have an organization where nobody gets an API key and starts shipping.
Getting new software requires procurement first, then a security accreditation cycle that only starts after procurement closes. Hasbrouck put the failure mode plainly: “2 years of procurement, then 2 years of accreditation, and then after that everybody has changed out and the new people are like yeah we don’t think that’s safe anymore.” Four years, nothing to show. That is the baseline before anyone touches an LLM.
Phase 1 — Research Preview: Shadow Use and the CamoGPT Letdown
Space Force issued the first memo in 2023: no enterprise AI, period. Most services followed. That left 95% of the workforce relieved and 5% of power users running home to use AI Studio, careful not to include sensitive data. Productive shadow use, but shadow use.
Late 2023 brought CamoGPT — the Army’s accredited answer, built on Llama 2. It failed three ways: the model couldn’t match GPT-4 Turbo or Gemini Pro; the front end had no PDF parsing and no web browsing; and RAG-style queries came back in hours. Extracting text from a document meant copy-pasting into a file because the input box capped at roughly 400 characters. General users quit at step one.
Phase 2 — Commercial APIs: Token Rationing as the New Barrier
In 2024, the Army got a new platform: commercial APIs, frontier models, reliable uptime. The catch was tokens. Every user got the same flat allocation, regardless of how they worked.
“for bureaucratic political reasons, there was no designation of power users versus normal users. So, everybody got one allocation of tokens.” — Chase Hasbrouck
That math works for someone summarizing a two-page document. It breaks for forensics analysts diffing two 20-page log files. One comparison and half a month’s quota is gone. Getting more required seven signatures from senior officials who, as Hasbrouck put it, don’t understand what an LLM is. In practice, people burned their tokens by day two and went back to AI Studio anyway.
Phase 3 — Enterprise Agreements and the Silo Reflex
genai.mil is live. It connects to major AI platforms and gives DoD workers the actual native UIs, not someone’s custom-wrapped API. Access fixed. Cost fixed. What remains is culture.
The pattern Hasbrouck calls the silo reflex: a new technology arrives, command picks one person to own it, and every question routes through that person. GPS and PCs survived it only because organizations moved fast to push them out broadly. AI is now at the same fork. “Is this going to look more like VR or more like PCs?” Hasbrouck asked. His read as of last year: it’s tipping toward PCs. Whether the Army moves fast enough to match that is still open.
Extracting Institutional Knowledge Locked in Chat Histories
Access is solved. Cost is solved. The problem left standing is cultural, and it has a specific shape. Operators at every level have been running their own AI experiments for two years. The findings from those experiments never made it up the chain.
“A lot of our requirements documents right now, they exist as people’s chat histories on various LLM platforms.” — Chase Hasbrouck
That intelligence needs to move from individual chat logs to procurement teams before the next four-year cycle locks in the wrong tools. No one has built that pipeline yet.
Q&A
Is Army Cyber thinking about extending these capabilities to the reserve component, since they are not on NIPRNet? genai.mil is available to everyone in the DoD, but getting reserve and Guard personnel onto NIPRNet to access it remains a real gap — Azure Virtual Desktop and BYOD tools are in the pipeline, though Hasbrouck acknowledged they are not keeping up with the need. ▶ 19:27
What are you looking for in future military folks when it comes to malware analysis and AI? He wants analysts who can translate technical findings for non-technical stakeholders — not seven pages of decompiled methods, but a clear answer to whether the system can be turned back on and when. ▶ 20:26
How do you handle fine-tuning or specialized models for individual departments? Do those get housed in the same centralized location? Fine-tuning has shown little value because procurement cycles are so long that the base model outpaces the fine-tuned version before it ships — the real gain comes from building a bespoke workflow around a standard model, not from specializing the model itself. ▶ 21:57
Notable Quotes
for bureaucratic political reasons, there was no designation of power users versus normal users. So, everybody got one allocation of tokens. Chase Hasbrouck · ▶ 13:21
you do that once and okay, half your quote is gone for the month. Chase Hasbrouck · ▶ 14:24
you ask for more tokens and this requires seven signatures and seven justifications for very high ranking people who don’t understand what an LLM is. Chase Hasbrouck · ▶ 14:35
a lot of our requirements documents right now, they exist as people’s chat histories on various LLM platforms. Chase Hasbrouck · ▶ 18:22
more of the value comes not from providing the fine-tuned or specialized model but from providing an actual bespoke uh workflow and task that is relevant and legible to procurement. Chase Hasbrouck · ▶ 22:36
Key Takeaways
- Treat token allocation as a capability gap, not an accounting detail — power users need asymmetric quotas.
- Solve access and cost before culture; you cannot change adoption behavior without working tools in hand.
- Mine operator chat histories now — your next procurement requirements are already written there.