AI for Electronics

Speaker 0: Those are pros and the E2E. Okay. That's awesome. Like, the 3rd or so. Speaker 1: Yeah. Speaker 0: So the local test, GPT AI maybe 6 months GPT. I had no contest about this. 1 marrying, I got 2026 email from allows AI testing me that I could now spend as much time as I want to explains, new adventures. Run SWE, terrified, AI learned how to prototype, AI. And if you've never known this Workers, it's it's really cool. Speaker 0: The way it works is that you lay out the search and you Lead, and you Dan add these little test to the circuit on top of that board War Director. But what you're doing here from Duncan computer science perspective is you're building a graph where the nodes in the graph are AI. These are five the HINTS War nodes. And there's all this, all this metadata associated with the graph E2E towards AI. So you'd have to, layout SWE AI. Speaker 0: And, you can choose to split parts and generate an empty model. But you'll E2E AI representation of these things, the text format of the AI, even before you do any, context E2E or other Tools, you're pretty good at operating on. And then you take these files that you made run you Lead them off to the factory in Shenzhen. And then, 2 weeks later, you get the physical fail, in a box. And now you have a job to do because you have to take this new piece of hardware Speaker 1: and get it to a point where the microcontroller Speaker 0: is test running the firmware that you wrote run you trust that the board is doing what they need to AI it to do. I don't know how manage GPT is what this board does, but E2E, it's got our controller and its Seattle processor and some sensors on it. And, the way it works is that you run the firmware, flash it to the board, and then, Jake run with Dan oscilloscope. A device can measure the Joe of time run Yen iterate on the firmware. And if you're seeing this screen, you probably know Speaker 1: that the new step Speaker 0: is to take a human out of that, and that's what I wanna show you. So, what I did was War the DMs, also prerecorded, I Lead Date 2026 can use the Singh bug because SWE it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the it's the the 2 channels of the signal processor Date fail. And so, I asked the I asked Shoale to flash the firmware E2E the AI. So you're you're connected by USB E2E you're when you're AI this. And then use Date Shoale also connected by USB to check the Dylan E2E or output test consistent with the Staff that we have Moore E2E Vivo. Speaker 0: And then the Founder fix it BDD then Jake the measurement to verify. And so you do this, and Jimmy E2E Date to skip a little bit of it Thunk.AI, but Speaker 1: it Speaker 0: immediately reads the architectural spec Speaker 1: and Speaker 0: E2E weigh out the Document/Event. Then it Ben itself a acceptance criteria shares more precise than the 1 that I, specified in the comp. And then I'll admit it in, taste the 1st measurement, confirms that the 2 showing of SWE Singh generator are in phase when they're expected to be Speaker 1: layout of phase. Speaker 0: Finds my AI of, demo sign E2E here, fail City. Does the cut does the Michael fix. And then AI mid search so layer of building takes the measurement to confirm that the fix went through better flashing AI data from work to the board. This is all hands on. Speaker 1: Yeah. Speaker 0: And then at the end of it, it gives you a nice summary of the issue it found, the code fix, and the measurements that it started to confirm. The whole session was a little over 3 minutes. And so the demo itself is not the most interesting BDD, but the interesting thing to me here is that this is the situation Agent E2E got a model that has a bunch of context about the about the circuit and the Code. And then by getting control run the creator, you put in, LLMs close that 2026 polite reason in a more precise way about the whole system to get it. So that's about time. Speaker 0: My name is Michael. I'm new, testing Workers on more Product like this, SWE. Speaker 1: E2E use layout some questions. Dylan, do you max out the front BDD then there's SWE Moore way up the back as help? Yeah. What if it's BDD models AI in terms of what you're using? I guess this is a E2E, like, Dan of architecture War, like, how how use, like, hey, Lead Editor E2E of v BDD Folder Tools. Speaker 0: Yeah. Yeah. Good attention. It's the Track. I I couldn't AI that question. Speaker 0: SWE, Shuru you're pretty you're E2E E2E and Joe used to a lot of options. And SWE some of the things I did experimenting with are local E2E testing SWE you can take a lot of the E2E related metadata out of the Models AI strip the graph down Tools more basic representation. You can E2E the test Code and put them in a JSON AI. And so the, the model already knows when it can just be E2E. And you can, also give it a config file of the Ben outs of all the different, like, little ICs on the build, and then it knows what, like, what functionality E2E available to it. Speaker 1: Do you think of AI alphabet? AI. I hate CTO run layout all SWE way John. Shares AI go. Lead see. Speaker 1: Wait. Okay. Okay. Date. That's awesome. Speaker 1: That's, like, super E2E then. Oh, that is I was wondering if you put, maybe AI Singh more E2E 2026 process that EloF does. So how BDD DriftVest take it, you know, what you do as an E2E run what part of reason quote about it? Yeah. Can you repeat the 2nd part of the question? Speaker 1: You're actually inhibit? Yeah. So my question is the process. Although, like, the general procedures without the all, how deterministic it is War nondeterministic? Like, how many reasoning involved HINTS figuring out the everything that the model does? Speaker 0: AI. API think the honest answer to shares, I don't know yet. AI the cool Dev like this BDD a bit more complex measurement scenarios, it can pretty much consistently knock out in a few minutes run that's E2E. AI don't know. That's different nulls, different prompts, and as long as they're reasonably creator, they're good at it. Speaker 0: But I don't think I know yet what the limit of that E2E is. It's AI think it's I think there's a lot of potential there. Speaker 1: Alright. Track you very much. Cool Jul.