ipunchghosts

I think googles could care less about what darpa is doing. How would I know? I'm a darpa PM and google isn't reaching out to me or my team.

I think what you are seeing is correlation but not causation. Stanford was working on self driving long before darpa, same with cmu. The darpa money was just a bonus but I don't think it pushed anything ahead.

After the first few years of the darpa challenge, many teams were able to finish. Yet here we are in 2025 and I cam get bounding box detections on video to not jitter or drop between frames because of some unlucky pixel combination. Humans have zero issues and aren't even able to perceive these minute differences.

My guess is vision will be solve in about 18 months using methods developed purely in industry.

I believe the contribution here is overstated. None of the tech used for today's AVs relies on anything developed for grand challenge, including any principles they derived.

The leading tech today was derived purely from industry outside the dod space and cifar.

That's what ousd is for.

YES! I should add that they should be thinking about intelligence generally, not AI/ML.

DARPA has a rich history of doing fundamental research but today its largely composed of engineering tasks which for whatever reason arent built by the warfare centers. The days of doing fundamental research are mostly gone in the organization. Name something in the AI/ML space that came out of DARPA in the last 15 years. CNNs- no, deep learning - no, attention - no, LLMs - no, test time training - no.

Its a good engineering org these days building cool things but there's nothing special about having darpa build it other than they provide funds. The research focused has significantly waned over the last 20 years which is a shame as there are LOTS of fundamental AI/ML problems which no one is studying in industry.

Currently, when AI models solve problems, they write out long chains of thoughts (like showing their work in math). While helpful, this takes up a lot of computing power.

Heima does something clever - instead of writing out long explanations, it compresses each step of thinking into a single "thinking token." Think of it like using a shorthand symbol instead of writing out a full sentence.

Read a bunch of papers on a topic you find interesting. What do all the papers do wrong or ignore. Start there.

Example of thr bitter lesson. None of these nuanced matter 8 years later where everyone uses sgd or adamw.

Lets ignore these fickle minded people. There is room in the ai/ML space to transform dozens of fields that are completely out of the scope of these larger companies simply because they can't make large enough profit margins to gain thier attention.

We need a 90s google like company that is in the ai ml business, not the ads business. It may not make billions, but it will surely make millions and push our society together in a meaningful way.

What was your method to ace the code? Was it fresh from undergrad / grad school?

I partially agree with this. Using AI has sped up my work by about 30% so far but only in cases where i need boiler plate to get going and i have written the correct algorithmic solution many times before and can quickly check it.

I think it might be better to have them find bugs in code as opposed to write it from scratch. Interesting idea!

Define small. I just interviewed at Dandy (the dental group) and they had me go through this process.

Also, do you do any type of coding exercises? where do you draw the line?

Can you elaborate? I like your confidence :) Do you do any coding style interviews? (i had a nice one at my current position which essentially was going through python code to find bugs).

i do but do not want to get doxxed until i found a job that I am okay with. If my employer found out about my writings, i worry my job would be at risk. feel free to send your email and i can reach out to you.

Honed time management skills and being very good at math. It was not easy. I have written about this but do not want to say where as to not dox myself to the group. but if you provide an email, im happy to chat with you.

These are great ideas but the timing is wrong.

1. GNU radio circa 2006 did not nearly support the SNRs and modulation schemes we needed; the tools had simply not matured. Furthermore, by making blocks, its difficult to close the loops hitting theory on timing say with respect to re-demoding blocks after getting loop lock. Perhaps now its fine but the problem is moot now. You can already run cell base stations now off of hardware that costs less than $1000.

2. This is fine as long as i can write high level code and have a compiler do the translation. Otherwise, there exists now sufficient compute power in a low swap package to do the necessary processing.

can i email you at your substack? Otherwise, i can reply to one of your threads with my email and then delete it. Let me know so we can set this up to time well.

Yes. The problem is darpa PMs only have a 4 year window. They also have a limited budget of about 80 million. They also have to pitch a program to get hired into darpa and then further have to get buy in from darpa brass to move ahead with thier portfolio.

Moreover, darpa is not the airforce, army, navy, ect. Just because darpa makes the widget doesn't mean anyone has to use it. Also working for darpa is a difficult as they micromanage thier PIs and you have to constantly show progress (even bi weekly) which isn't conducive to research.

My experience has been that Darpa has a romanticized role outside of those who have actually worked for darpa.

As a researcher you are not incentived to tell darpa the truth because if they are not happy with it, they won't give you any money which is already hard to come by.

Darpa PMs are often young, have few publications and are new to the government and have romanticized ideas about what they think they can accomplish.

I guess im wondering if other researchers in DoD are having the same sentiment as me. I find it hard to believe they arent but it's something that not talked about much because from my point of view, there arent many researchers with 20 YoE left in the DoD. If i am wrong, point me to them as i want to join their ranks!

Thanks for this thoughtful analysis. Your framing of the 'buyer vs builder' dynamic really resonates with what I saw at Raytheon and explains a lot. We weren't just struggling with a technical handoff - we were watching an organization transform from a builder to a buyer of expertise.

The part about 'overseers wanting early results' is especially relevant to defense. Program managers need to show progress to keep funding, which incentivizes quick integration of existing solutions over deep technical work. I saw this firsthand when our most sophisticated signal processing work was subcontracted out.

But here's where defense differs from commercial software: When you're building complex systems like radar or signal processing pipelines, the 'buy' approach has serious limitations. You can't just integrate your way to novel capabilities. Those retired experts we had to bring back? They represented irreplaceable domain knowledge that can't be purchased off the shelf.

The tools you mention (linters, static analyzers, requirement tracking) are valuable, but they don't solve the core problem: you need people who deeply understand both the mathematics and the implementation. In my current role, I see PhDs struggling because they lack this foundational knowledge - no amount of process or tools can bridge that gap.

What I think we need is a model that preserves deep technical expertise while still allowing for integration of existing solutions. But the current funding and organizational structures in defense make this really hard to achieve.

If anyone's building something to address these systemic issues, I'd love to chat.

Let me save u 10 minutes and get to the punchline.

"Eventually, it clicked, and I learned one of the bedrock principles of salesmanship: Whether you’re peddling long distance over the phone, Bibles door to door, or your own political candidacy on live national TV, it doesn’t matter what you’re selling—it matters how you make people feel. If you make them feel good, they’ll say yes. If you don’t, you could be selling a pill that reverses male-pattern baldness and makes you lose weight without exercise, and they’ll still turn you down flat."

Melatonin fixed it for me. Found out about it through a lit review.

Someone ported this to python. Besides, an llm ca easily port the code as the value is in the algorithms not the source code.

Large foundation models for undersea acoustics and for high resolution synonym (synthetic aperture sonar) and a fudemental research project trying to make a training pipeline run with completely synthetic data but perform as well as if trained with real data.

There's an underlying sentiment through the article that there isn't more significant research to do to further improve ai and generate more revenue. This is a large gap and myopic.

For example, the amount of compute needed to obtain state of the art on a benchmark is only obtainable by big labs. How can we improve the learning efficiency of deep learning so that one can obtain similar performance with 100x less compute.

Paper copy?

Can we get this person to explain thier inspiration?

This is a good question!!