What is an Artificial General Engineer?

Full transcript

Edited only for punctuation and readability. Obvious speech-recognition errors in public names and technical terms were normalized; meaning and claim boundaries follow the recording.

1UA-AGE is a Ukrainian deep-tech project building an Artificial General Engineer for the physical world, right? And that is quite the ambitious mandate.

Yeah, it really is. So, welcome to the deep dive. Look, if you are an engineer, a technology investor or, you know, a technical editor, you are probably so exhausted by the current AI hype cycle.

Oh, definitely. The noise is just deafening right now.

Exactly. I mean, your inbox is probably flooded with startups promising that their AI is going to magically revolutionize industrial design by next Tuesday.

But they never actually explain the mechanics of how they stop the AI from hallucinating a load-bearing beam, right?

Yeah. Or, you know, a critical copper trace on a circuit board.

Which is terrifying.

It is. So our mission today is specifically tailored for you. We are cutting through all that noise. We are looking at the cold, verifiable mechanics of what this team is actually building.

And just to clarify our baseline, we are pulling today directly from 1UA-AGE's factual newsroom briefings.

Yeah. Those are their normal proof and disclosure policy and the public GitHub repository for their AI-HPP Standard.

Perfect. Because dealing with AI in the physical world is just fundamentally different from software. If a model writes bad code, an app crashes, but if it designs a faulty circuit board, things catch fire.

Yeah, the tolerance for probabilistic guessing is literally zero. Okay, let's unpack this core concept first. What actually is an Artificial General Engineer, or an AGE?

Well, the most critical thing to understand is that an AGE is not one model, right? It's not just a giant language model you type a prompt into.

Exactly. It's a governed system. It takes authorized human objectives and hard constraints and turns them into versioned engineering artifacts.

Meaning files and schematics that you can audit and roll back.

Yeah, exactly. It uses real engineering tools to verify measurable properties. It preserves the evidence at every step and, most importantly, it keeps manufacturing, physical testing and certification as explicit gates.

So the AI can't just grade its own homework.

Right. It absolutely does not have the authority to declare its own work correct.

I like to think of the current state of generative AI as this really fast, overly enthusiastic brainstorming intern. You ask for an idea and they give you a hundred in three seconds, and they are incredibly confident about all of them, even the ones that violate the laws of thermodynamics.

That is a very accurate analogy.

Yeah. But what 1UA-AGE is building is more like the rigorous, gated workflow of a licensed engineering firm. The AI can be the intern generating the layouts, but the real engineering tools must prove those ideas are viable before moving forward.

Yeah, the engineering tools are the senior partners stamping the work. But a system like that is only as good as its measurable output, right? Which transitions us nicely to the actual evidence. Let's look at the PCB benchmark results they recently published.

Yeah, so they tested on two public PCB — printed circuit board — benchmark classes. For the first one, the frozen formal STRF validation set, they recorded four runs.

And frozen means they lock the rules in place, right? No moving the goalposts.

Exactly. So after they corrected some development blockers, out of those four formal runs, three were accepted at a perfect 98 out of 98 required connections, and one failed.

Wait. So they had a bunch of failures first during development and, even in the formal run, they failed one out of four. Why is this a win? Startups usually hide that stuff.

What's fascinating here is that retaining that failed run and explicitly disclosing those boundaries is exactly what real engineering evidence looks like.

Oh, I see.

Yeah. It is not some claim of instant success. And they actively avoid saying they have a success rate of three out of four attempts overall.

Because earlier development had way more unsuccessful attempts, which are outside that formal set.

Exactly. They are just saying: under this specific frozen contract, this was the exact measured outcome, warts and all.

Which honestly builds so much more trust than a magical 100% success demo.

Right. And then they moved to the STM32 runs, which are notoriously dense spatial puzzles for routing. Here they recorded three consecutive PASS runs.

Wow. Okay. What were the metrics on those?

They hit 211 out of 211 required connections, passed 19 out of 19 applicable acceptance gates, and this is the big one: zero new routing-induced DRC violations.

Zero. I mean, for the investors listening, a design-rule-check violation in the physical world isn't a software bug. It's breaking a manufacturing rule, like putting copper traces so close they short out.

Exactly. And hitting zero means the AI isn't hallucinating its way out of the physics problem. But we really need to define the claim boundary here.

Right. Because they are very strict about what this actually means.

Yeah. 1UA-AGE calls this completion parity only.

So no claims that it's faster or more elegant.

None. No claim of runtime speed, no via-count optimization, no topology optimization — just that it finished the job up to the baseline standard.

Exactly. No manufacturing-readiness claim and definitely no claim of general superiority. And an independent engineering review of all this data is currently in progress.

So they're having outsiders verify the math. That's huge. But how do they mechanically stop the AI from hallucinating? This leads us straight to their governance structure, right? The AI-HPP Standard.

Yeah, the Artificial Intelligence Human Protection Profile. It's a public, vendor-neutral, reviewable draft for bounded and attributable agent behavior.

But just a draft right now, right? Not certification-ready yet.

Right. It's an inspection-ready draft, not a final certification.

Here's where it gets really interesting for me. I was looking at their GitHub, and the architecture of this standard — signal, state, gates, bridge and evidence — looks exactly like a submarine's airlock system.

Oh, that's an interesting way to picture it. How so?

Well, you don't try to make the ocean water less pressurized. You build physical doors that enforce reality. The AI's output is the raw signal. It's untrusted. It's probabilistic water.

Right. You assume it might be hallucinating.

Yeah. So you bind that signal to the state of the project and then you hit the gates. The gates are the real engineering tools, the sensors and the airlock checking the map.

The physics simulators and rule checkers.

Exactly. And the bridge, which actually commits the action, never opens unless the evidence perfectly verifies the pressure. If the math fails or if provenance is missing, the system is forced to fail closed.

Fail closed. It just halts. It doesn't guess or try to bypass the safety check.

Exactly. It just locks the door.

That is a great analogy. And to make sure that airlock works, the AI-HPP Standard requires seven minimum viable profile controls.

What are the actual controls?

One requires the system to retain an authorized human objective. It can't just decide to design a different board entirely. Another requires it to classify risk before taking any action.

Right. So if it's a high-risk change, it needs human sign-off.

Exactly. It also has to enforce scope boundaries entirely outside the model's control. So the AI isn't in charge of its own security gates. And it treats untrusted input as untrusted.

I have to ask about autonomy here. Everyone in tech wants to fund fully autonomous agents that just run end to end with no human in the loop. This sounds like the opposite.

It is the opposite. 1UA-AGE avoids the word autonomous entirely unless it is directly qualified.

Really? They just won't use it?

Right. Because their system explicitly requires authorized human objectives. It is not an unconstrained independent entity. Separating fact from hype here is so important.

Oh, totally. The hype in this space is out of control.

Yeah. If you read their documentation, there are no invented metrics, no fake customers, no wild deployment claims or fictional funding outcomes, and no timelines promising a revolution next week.

It's literally a technical draft for accountable behavior, which I honestly find refreshing.

Yeah.

So because this AI-HPP Standard demands attributable evidence for every decision, it basically forces their whole corporate behavior to follow suit, right? Like their proof and disclosure policy.

Yes. They've built a five-class ladder for public evidence.

But what does this all mean in practice? What are the five classes?

The first is internally reproduced. The result was repeated successfully internally under a strict scope. That's where they are right now with the PCB benchmark in their pinned KiCad environment.

Got it. What's next?

Second is independently reviewed, which they are moving toward now. Third is tool-verified, where a recognized industry tool confirms the property.

And then we get physical, I assume.

Right. Fourth is physically validated: an actual manufactured physical revision passes a real-world test. The final, fifth stage is certified by a competent external authority.

But how do they balance this extreme transparency with being a private deep-tech company? Aren't they giving away the secret sauce?

That is the key distinction they make.

Evidence transparency is not architecture disclosure.

Okay.

They publish the subject, the scope, the measured results and the evidence class, but they keep their internal maturity scores, their specific prompts and customer data tightly controlled.

So they prove the math works, but they don't show you the specific code that generated the math. If we connect this to the bigger picture, what happens if one of these evidence claims is proven wrong later? Say a benchmark was flawed. Usually tech companies quietly edit their website and pretend it never happened.

The classic stealth edit. 1UA-AGE has a strict policy on this: if evidence is invalidated, the public statement is formally corrected or retracted.

They don't just silently rewrite it.

No, it is not silently rewritten. The claim is permanently bound to the evidence. If the evidence breaks, the claim comes down publicly.

That level of discipline is almost a culture shock for the current tech sector. It really is a totally different paradigm.

It is, and it perfectly encapsulates everything we've dug into today. The core guiding principle from their own documentation, which is literally the mandate they operate under, says it best.

Yeah, I know the exact line you mean.

“Replaceable AI capabilities. One governed engineering lifecycle. Evidence before release.”

That's a powerful baseline.

Absolutely. If you want to inspect the schemas or read the minimum viable controls for yourself, head over to 1ua-age.com and check out the public AI-HPP Standard repository on GitHub.

Yeah, the traceability matrices and gate contracts are all sitting right there in the open for anyone to review.

Before we let you go, we want to leave you with one final thought to mull over. We've just looked at a system purpose-built for the physical world, one that requires a strictly governed lifecycle and explicitly fails closed the second evidence is missing.

Right.

So if a specialized system designing a standard circuit board requires that sort of paranoid, mathematically proven gating to be safe, what does that say about the safety of the thousands of ungated black-box AI tools we are currently rushing to integrate into our broader global infrastructure?

It should probably make us all very, very nervous about what we're building.

It definitely should. Thanks for joining us on this deep dive.