Mark Delaney: Michael, hey — you have a good week? Because mine got weird fast when I fell into this rabbit hole I cannot stop thinking about.
Michael C. Vincent: Weird how?
Mark Delaney: Okay, so — uh, there's this engineer, goes by cpldcpu online, his name's Tim. And on November 14th he published this whole write-up about sitting down with Claude Code running in VSCode, Opus 5.5, and just... designing an analog synthesizer from a blank project folder. No starting schematic. Nothing. And the thing that got me is — he built it. Like, he actually got sound out of hardware that an AI helped design.
Michael C. Vincent: Actual audio. Not a render, not a simulation.
Mark Delaney: Real hardware, real sound. And then about five days before Tim posted that, Anthropic put out this claim that Claude had designed three working prototypes entirely end-to-end. A USB speakerphone, a voice hub, a handheld virtual pet. Schematics, PCB layouts, firmware, enclosures. Their fabrication partner Diode built the boards and apparently they needed zero circuit fixes. Zero.
Michael C. Vincent: And the independent engineers who looked at those boards said — what, exactly?
Mark Delaney: That's the thing. They said... almost nothing. The social media discussion that week, week of November 9th through the 16th — cautious, quiet, no threads breaking ten replies. So what I keep chewing on is: what does 'AI designed this' actually mean when nobody outside Anthropic has verified it yet?
Michael C. Vincent: That silence — that's not nothing. That's the whole question.
Mark Delaney: And that silence is partly why I want to back up — because I think before we can even argue about what the evidence means, most people don't have a picture of what these models are actually doing. Like, physically doing.
Michael C. Vincent: Paint it.
Mark Delaney: Imagine you hand a very capable assistant a hand-drawn sketch of a gadget — just the connections, what plugs into what — and you say: now figure out exactly where every physical part sits on the board, and draw every wire between them, without anything crossing wrong or running too hot. That is what GPT-6 Astra did in that KiCad demo. Placing components, routing copper traces. The model is operating the design software. It's not suggesting — it's doing.
Michael C. Vincent: OpenAI released that in September 2026. And we saw — what, fifteen seconds of it?
Mark Delaney: Fifteen seconds. No independent evaluation at announcement. And I want to be fair — the clip looks real, KiCad is real software, the task is a real task. But a demo is not a test. Those are uh — they're just different things.
Michael C. Vincent: And Anthropic's claim is structurally different. Three fabricated boards — the speakerphone, the voice hub, the virtual pet — Diode built them and they needed zero circuit fixes. That's not a clip. That's a physical object.
Mark Delaney: It's bigger. I mean — wait, that's actually what makes it harder to evaluate, not easier. A bigger claim needs somebody outside Anthropic checking those boards. And as of mid-November, that hadn't happened.
Michael C. Vincent: So both things get cited as proof of the same leap.
Mark Delaney: That's exactly the problem. A fifteen-second KiCad clip and three unverified prototypes — the press treats them like they're on the same shelf. They're not. One is a demonstration, one is a claim about physical hardware that nobody external has validated. The experts paying attention that week — cautious interest, calls for independent physical validation. That's the honest summary.
Michael C. Vincent: And the number that breaks that clean story open — that's the NovaSilicon Team's AnalogLayoutWorld benchmark.
Mark Delaney: Right — because that's the only independently published quantitative evidence we have as of mid-November. Everything else is a press release. This is an arXiv paper. And the numbers are, uh — they're kind of brutal when you actually sit with them. AI agents produced analog layouts ninety-one percent of the time. Ninety-one. That sounds great. Except only twenty-three percent of those outputs actually met acceptance criteria.
Michael C. Vincent: Sixty-eight points between 'it finished the task' and 'it passed.'
Mark Delaney: And that gap — I mean, that's not a rounding error, that's a category problem. Producing a layout and producing a correct, manufacturable layout are just... they're not the same outcome. Like, picture a contract electronics shop, Saturday afternoon, engineer pulls up a Claude-generated board layout on her screen. Copper traces look neat, components placed, nothing obviously wrong visually. She runs the design rule check — and statistically, based on AnalogLayoutWorld, more than three out of four AI-generated boards like that one fail. Not 'needs tweaking.' Fail.
Michael C. Vincent: Why is analog specifically the hard target here?
Mark Delaney: It's physics-rooted intuition, not rules. Digital logic is basically — you know, it's ones and zeros, timing diagrams, it's structured. Analog is continuous-signal electronics. How a trace behaves at a certain frequency, thermal behavior, parasitics — that's not something you encode cleanly. The spatial reasoning required, understanding actual three-dimensional physical constraints on a board, that is a genuinely harder problem for these models than writing software or designing digital logic.
Michael C. Vincent: So the model sees the shape of the solution but not the physics underneath it.
Mark Delaney: That's — yeah, that's actually the right frame. And what worries me is there's a part of this we haven't even gotten to yet — if AI generates fast and a human has to verify slowly, verification itself becomes the bottleneck. What that does to engineering work, to hiring, to whether anyone still knows how to catch the failure — that's the darker part.
Michael C. Vincent: Hold that thought. Because that is exactly where we need to go next.
Mark Delaney: And the bottleneck thing — that's actually where it gets, uh, structurally weird. Because Cadence, before any of this September stuff dropped, they were already building toward it. Virtuoso, their AI Analog Design Studio — that's agentic EDA tooling. They anticipated that the generation side would outpace the human side. Hitachi too — they unveiled agentic AI specifically for chiplet and package design. These aren't startups chasing a demo. These are infrastructure companies building around the assumption that humans will be checking, not drafting.
Michael C. Vincent: And DATE 2026 framed it the same way — emerging capability, not settled science.
Mark Delaney: Exactly — sessions on agentic AI and hardware co-design, this is a conference built for the people who build the chips, and they're calling it emerging. Not 'here's how to deploy it.' More like 'here's what's about to happen to your job description.'
Michael C. Vincent: So who does that job description belong to now?
Mark Delaney: That's — okay, the thing that keeps hitting me is this. If you're hiring analog engineers in this environment, what are you actually selecting for? Like, picture a junior designer, two years out of school, and her whole day is reviewing Claude Opus 5.5 outputs for a sonar sensor board — not laying out traces herself, not choosing component values from first principles. Just checking. At year five, does she have the physics intuition to catch the failure that the model can't see? Because that intuition, I think — wait, no — it's not taught, it's built. You build it by doing the design work. And if the AI is doing that work now—
Michael C. Vincent: She hasn't built the instrument she's supposed to play.
Mark Delaney: And that's not rhetorical — that's the actual open problem. Because the AnalogLayoutWorld number tells you the AI misses in seventy-seven percent of cases, but the human verifying it has to know enough to catch that miss. If she was never the one designing, I mean — does she?
Michael C. Vincent: You see, that's the thing that doesn't resolve cleanly. The industry builds infrastructure for human oversight — Cadence, Hitachi — but the craft knowledge that makes oversight meaningful may only come from the work that the same infrastructure is replacing.
Mark Delaney: And nobody in the November 9th social media discussion — none of those cautious researchers — none of them had an answer to that. They were focused on physical validation, which is fair, that's the immediate thing. But this is the longer-horizon problem sitting right underneath it.
Michael C. Vincent: Tim described astonishment. That word keeps landing for me. He sat down with a blank folder, got sound out of real hardware, and his honest word was astonishment. That's not nothing. But astonishment and proof are different things, and the hardware industry runs on proof.
Mark Delaney: And that's — yeah, that's the whole tension, right? Like, the Anthropic zero-fix claim, the USB speakerphone and the voice hub and the virtual pet that Diode built — those could be real. I actually think they might be real. But 'might be' is not what you sign off on before you ship a medical device or, uh, a sonar board. Independent scrutiny is just the job. That's not skepticism for sport, that's just how hardware works.
Michael C. Vincent: And if those boards do hold up — when someone outside Anthropic checks them — the job doesn't disappear. It shrinks to verification. Which brings us right back to Tim, actually. He initiated, he interpreted, he iterated. The question was never whether Claude could generate a synthesizer. It was whether Tim designed one, with a very fast assistant.
Mark Delaney: Honestly? I don't think that question resolves. And I'm — I mean, I'm sort of okay with that being where we land.
Michael C. Vincent: Good conversation. Genuinely.