Hacker Newsnew | past | comments | ask | show | jobs | submit | amirhirsch's commentslogin

Good luck! One valuable differentiation you could provide is a line of credit. If you can’t win on price, you can help your customers win on cash-conversion cycle.

I’ve been making hardware for 20 years. My highest-volume product was Flybrix, a LEGO drone kit manufactured by Seeed and UniPrecision in China and by Sparqtron in Fremont, California. I’d be happy to work with you on a v2 reboot: Flybrix Swarmz, because every American high school needs a drone swarm.

Beyond that, I’ve worked with many domestic suppliers and built ITAR-controlled products as well. Even without owning the production capacity, if you can offer Net 90 on fabrication and extend credit for components--beyond existing DigiKey, Arrow, Mouser, and vendor credit lines--then I can build products and get paid without putting all the capital at risk upfront.

Factoring is expensive, especially against a purchase order rather than an invoice, and I don’t know of anyone offering working-capital credit based solely on an SBIR award.

Credit is what makes the spice flow.


Wow. Sounds like a super interesting product. Hadn't considered lines of credit yet, but definitely would be valuable to some higher ticket customers.


We currently use a small US contract manufacturer that works pretty much as you say. The turn times are terrible (7+ weeks, not 7 days), and this is a major issue, but there are exactly two reasons we continue to use them:

1. Line of credit - we are not billed until we take delivery of tested product. 2. Willingness to handle high mix, low volume products. One product line example: a dozen SKUs with 95-98% parts in common on 2 PCBs, a single set of parts (that we aren't billed for up front) and quoting to build 20 of SKU 1, 50 of SKU 2, 10 of SKU 3, etc. without charging 2-3x the price.

We manufacture in the research / test equipment / defense space and I've heard similar things from other small hardware companies that are like us.


My experience with lines of credit isn't unique, and I think it's one of the underrated advantages of doing business in the US: we have a sufficiently stable commercial legal system that a huge amount of the economy can run on credit.

I have thought that there is a big opportunity financing components by optioning the sale of overstocked parts ahead of time. For a consumer electronics product that had heavy Q4 sales, my annual sales were entirely dependent on how many chips I could order by October 19, but then I'd have excess stock every January because things are sold in reels of varying sizes.


Ok, fair points. 2. is our focus right now and then we may add 1. over time.


Sad to see the Lego drone sold out, hope you get that reboot going!


Credit based on a SBIR is a great idea. Seems like something the Small Business Administration should offer.


I implemented all of this in hCaptcha 6 years ago, not just to distinguish bot from human but also to recognize the keyboard/mouse behavior of the same person signing up for many accounts or testing multiple credit cards. This kind of abuse detection was a part of Cloudflare when they switched to hCaptcha in 2020 and I had thought they already implemented all this themselves four years ago when they transitioned away from hCaptcha in 2022.


Didn't recaptcha v3 also do this? iirc the "silently monitor in the background" was part of the selling point.


Yea this is the premise behind all the invisible captchas, combined with browser fingerprinting. Same tech is used now to detect distillation rings.


It creates pancake beams so you would usually use a separate Rx and Tx orthogonal to each-other to do imaging.


Oh I dunno. I made a mm-wave radar with a Rotman lens using a generative loop between Python, Rhino, and EM simulation. Pretty sure AI could cook that up.


What did you use for the simulation part?


Custom FDTD for the optimizer loop, CST for final verification.


Yes, I'm also interested in this.


How was it generative, generative AI?


Very cool! Six years ago I worked on a mmWave (76-81GHz) imaging radar with a Rotman lens Tx and Rx. Designed as a LiDAR replacement, but we could see pipes in walls, or detect concealed weapons at ~1km.


Do you have a writeup about the project? I'd love to read more about it.


Super cool. Didn't even know mmWave was used for radar and imaging techniques!

Isnt mmWave pretty similar (in theory) to short-ish range wireless directional antennas? What people used to call point to point "microwave" transmission?

Crazy idea but could your mmWave radar hardware also be (not simultanouslyl) used to transmit data? No idea what a Rotman Lens is but I would imagine that maybe it could be useful for transmission as well.


isn't mmWave the thing they use at the airports?


Yes it is!

I've been down the mmWave rabbit hole for the last 6 months, making sensors to put around the house to control automations for lights and so forth. They're pretty great.

For my use case the advantage over PIR is they do presence detection, so no more lights shutting off when you're sitting on the toilet.

Pretty amazing you can pick up basic mmWave sensors for a few bucks on amazon, and mate them to an esp32 board which is another couple bucks. It's so much fun as a hobby!


Believe they do use a good amount more power though (mainly only an issue for battery based devices)


Mate this is genius! How have I not thought of mmWave detection. Rabbit hole activate!


Waiting for an M413 motion detector to be commercially available.


How many tx and rx antennas did you have ? (I don’t know if it was clear, my stack was 57-64 GHz, 2TX , 3RX)


32 port Tx (vertical pancake beams) x 16 port Rx (horizontal pancake), something like 60 by 30 degrees. the entire thing used FPGA transceivers as one-bit DAC/ADC, Complementary Golay Code waveforms with one-bit correlation in the FPGAs (two VCU128s) -- digital logic was essentially the same as a binarized neural network, I squeezed a ton of popcnt performance out of those chips using both DSPs and LUTs


Err holy shit that’s quite the array and FPGA processing backend. I assume the purpose of the system is secret?

(VCU128s are relatively large, HBM equipped FPGAs for people not aware)


It was designed for automotive as a replacement for LiDAR, also useful for finding metal FOD on runways, or concealed weapons and I suppose also liquids on travelers walking in an airport...


(Linked VCU128 digital radar was cool but even cooler is using them as digital array beamformers and sending RDMA over fiber into a quad A100s nvlinked to do the whole radar DSP pipeline)


What project was that for ?! Sounds super cool


This is an awesome result.

For those unfamiliar: NC is the class of problems which can be solved in polylogarthmic depth with polynomial number of logic gates. It is unproven if NC != P similar to P != NP.


Yes, but logic gates with constant fan-in, crucially, otherwise that's called AC.


I never studied these specific classes, but my immediate intuition is that an n-input fan-in AND or OR gate can be reduced to a tree of 2-input gates with depth O(log(n)), which preserves polylog complexity, so surely AC = NC.

Wikipedia agrees :)

If you specify the exponent of the log, you get a different answer.


Yes.

There is a beautiful proof of the disjunction between AC0 and NC showing parity cannot be done in AC0 using harmonic analysis of Boolean functions



https://en.wikipedia.org/wiki/Switching_lemma

That paper is in the wiki refs but Hastad’s original is from 1986


So is it a class of problems that can be parallelized well?


no (in both directions). lots of np/exp problems paralize well and you can be in NC and parallelize really inefficiently (e.g. you can get a 10x speedup, but you need 1000000x the hardware). the better framing is that NC is the class of efficient algorithms that can be sped up near arbitrarily by parallelization


Hmm your last sentence seems to exactly agree that it's a class of algos that parallelize well? What does sped up arbitrarily mean? It's still polynomial speed up right?


It's a difference of degree. People expect something that "parallelizes well" to show near 1-to-1 speedup. Double the hardware, double the speed. This is "you can always speed it up, but the hardware requirements can increase at any polynomial rate".


Ah got it. Reread previous comment and that makes sense.


Yeah it's more of "on a hypothetical infinitely parallel computer, you'll get a big speedup'.

Which is still useful if you can prove a problem is in NC. It's just not quite as strong as people make it out to be.


Don't worry about not grokking quantum computing stuff, neither do any of the people who invest in it as well as many people who work on it.

1. The OP has nothing to do with quantum computers.

2. Quantum computing deals in coherent quantum states: associated with N qubits there are 2^N complex amplitudes. You can measure by sampling the square-magnitude of the complex amplitude which turns it into a Probability Distribution. Quantum computing "gates" cause interference in the complex amplitude of entangled qubits cancelling out incorrect results, such that if you maintain coherence for long enough and sample the final state and measure the probability distribution, you get a computationally useful result. The key challenge in quantum computing is extending the coherence time of a larger and larger number of qubits, which is why you hear so much about quantum error correction. Recent results from Google showed a scaling law for "surface codes" using multiple qubits to create an error-corrected topological qubit with extended lifetime. There is no telling how far this scaling law will go, but as long as Gil Kalai is in the next room, it is unlikely there will be actual useful quantum computation for a while.


Okay but why did this chip have anything to do with that?


Density is this reply.


Cargo ships powered by the wind, we are living in the future!


I was expecting Beyoncé lyrics


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: