Tell HN: The Quiet Collapse of US Defense Research Infrastructure
The first wake-up call came at Raytheon. Fresh out of college with signal processing coursework under my belt, I joined a program that would balloon from $150M to $250M, ultimately requiring congressional approval to continue. The issue wasn't typical defense contractor bloat - it was a fundamental disconnect in how we approached technical work. Systems engineers would develop algorithms in MATLAB, throw them over the wall to us software engineers, then vanish to other programs. We'd struggle to translate to C++, often discovering the algorithms didn't actually work. We eventually had to pull a retired expert back to make the core algorithms functional.
The real kicker? The most sophisticated signal processing work was subcontracted out. Raytheon had become primarily a software integration shop. When I and four other recent grads in their fast-track program predicted our site would significantly downsize within 10 years, management dismissed us. Today, that campus has shrunk from five buildings to one three-floor facility. Those other fast-track folks? They've gone on to start their own companies or become tech executives.
I moved to a UARC hoping to do more meaningful work, bringing GPU computing expertise just as CUDA 1.0 was emerging. My pitch was simple: CUDA's backward compatibility meant we could double our speed every 18 months just by buying new hardware. It worked brilliantly - I even won the lab's highest award. I became known for turning academic papers into polished prototypes that management used to secure major programs. But then the system's flaws emerged. A manager circumvented my chain of command to keep me on his program. Despite delivering field-deployed systems (still in use today), when funding dried up, I was stuck. All those relationships with operators and government management evaporated. It's not even the government's fault - their hands are often tied by funding structures.
The PhD years opened my eyes further. Working full-time while completing my doctorate in 4 years, I published 6 papers, won awards, and got promoted to chief scientist at $190k. But without funding, titles mean nothing. I jumped to a boutique defense contractor, secured $2M in grants within 9 months - and walked straight into another systemic issue. I'm managing PhDs who lack fundamental signal processing knowledge, delivering sloppy work that explains why we struggle to convert to Phase 3 programs.
The current state of government ML research is particularly troubling. Everyone's working 3+ projects, spreading themselves thin. Most groups just take off-the-shelf models like Hugging Face and apply them to their specific data. Nobody uses more than 4 GPUs for training because they can't afford more compute. I watched 5 contractors tackle the same MWIR video problem, all delivering similarly mediocre results.
The solution seems obvious: instead of every group rolling their own mediocre models with insufficient resources, we need 1-2 primes building proper foundation models for others to fine-tune. Most of this could be done in the open before moving to classified environments. But the current structure of defense funding makes this nearly impossible. VC-backed defense startups aren't the answer either. They're making the same mistakes - small compute, off-the-shelf models, requiring relocation from experienced 40+ year old scientists who won't move. They're essentially just spending the money the government can't, without solving the fundamental issues.
My former students who left for industry are thriving. The system needs fixing, but I'll be joining them unless someone's building something to actually address these fundamental issues.