Short answer: Jordan Metzner says cameras must analyze 15 to 25 frames per second in any environment, quickly and cheaply, with limited on-device chips and cloud budgets. Ramses Alcaide says Neurable faces similar trade-offs processing brain data on the edge.
Watch this answer from Episode 41 (starts at 15:03) · Episode 41 notes and transcript
From Episode 41: This Headset Reads Your Brain in Real Time. We Demo It Live., starting at 15:03.
Sam Nadler [15:03]: They're more of a sauna cohort. Right, guys, let's jump into the news. We wanted something a little fun. This isn't, too serious of a news day. We're sharing this first one. Security cameras are failing spectacularly at common sense. AI descriptions, I think, are calling brake lights fires and dogs bears. But, Jordan, just because you have so much experience with building these systems and hardware, I just want to hear what you think is going on.
Jordan Metzner [15:35]: Oh, man. This is a I mean, this is a RingTag camera, but, why is this a hard problem? So security cameras live at the edge, right? So they're like on your doorpost, they're on, they're in your house, right? Then they have to connect, probably to your Wi Fi, right? So they're not really to hardwired, right? And then they have, physical sensors on them. So like infrared sensors and other types of sensors on them, maybe radar. And then they also have the camera, right? And then they have to have either a local model, a model on the device, right? Or a cloud based model that can then process these images as they're coming in at like 15 frames per second or 20 frames per second, 25, depending on the camera style, either locally or on the cloud, and be able to analyze these pictures, accurately in all different types of environments. And, in the case of here, like this very dark, low inhibited environment. So like, you don't always know what you're looking at because like the user themselves could put the camera in any direction upside down in any type of environment. And so you have kind of this, non heterogeneous, dataset and you're trying to build a model that like works for all these different environments. It has to work very fast.
Jordan Metzner [16:38]: And the other thing has to be cheap, right? Because like if you put like a very heavy model, on the cloud, then like customers aren't willing to pay like hundreds of dollars a month per camera, let's say to analyze like all of these photos. And so you're stuck in this position where you want to analyze things both at a, on device level, at a cloud level, do both of them to like try to provide some real time solution. Right. When you see the animal, you want to get notified. Right. So they got to send you a push notification. Right. And all these things have to happen very, very quickly. And you don't have a lot of dollars like of dollars to spend both on the device side. Right. You can't put a really expensive, Nvidia, crazy chip on it because nobody's going to buy a $10,000 camera. Right. And then on the cloud side, you can't put like a very heavy chip on it because then no one's going to pay a monthly subscription for a really high rate. So you're stuck in this place of like consumer electronics where you need to put the right type of technology, you need the right type of response time, but you're highly limited on, the type of, hardware, software and type of inference you're able to put on every, like, per instance that's happening.
Sam Nadler [17:37]: Oh, sounds like a miserable problem. Ramses, did you get a chance to read it and any high level thoughts? You're in the hardware space, so maybe that something resonates.
Ramses Alcaide [17:47]: What I would just say is what's crazy is that these are the problems that people see in different ways across different devices. Right now, we process brain data on the edge with these headphones. And so it's like we're trying to balance costs, we're trying to balance compute capabilities, sending to the cloud, all these things. And at the end of the day, it's just so hard. Like, don't think people appreciate, like, how hard it is to even do the simplest things. Like, you look at it and you're like, Why does that think it's a bear? It's clearly a raccoon. You know? And it's like, but one, your brain is just such an incredible, amazing machine that can classify these things that easily that you take it for granted. But, I mean, somebody's probably spent their entire life classifying this raccoon and getting it wrong. You know? It's like it's that hard of a problem.
