AI-designed Linux computer with 843 components boots on first attempt — dual-PCB Project Speedrun was made in just one week and required less than forty hours of human work
And also, honestly, article reads like a direct advertisement. Not much details, comparison with human hours without explanation, does human use existing automatic tools or do all work manually,...
And also, honestly, article reads like a direct advertisement. Not much details, comparison with human hours without explanation, does human use existing automatic tools or do all work manually, etc.
Will be very interesting to read impression from folks that have experience in industry.
From what I can see, it's only an "AI" because of the hype around the term. They talk about it doing some placement and routing tasks, which is something that can be done by existing tools. Maybe...
From what I can see, it's only an "AI" because of the hype around the term. They talk about it doing some placement and routing tasks, which is something that can be done by existing tools. Maybe it's doing a bit more than existing competitors?
About the project they present, I saw some feedback from someone working in the industry (on a different website) and their conclusion was that the project was perfectly scoped to magnify the impact of their tooI. For example, they don't start from scratch, but reuse an existing design (which can take 1-2 months of work for this kind of design). They also use a service to get an extra-speedy manufacturing of the board, which most people wouldn't use due to the cost. Testing (beyond just booting an OS) is not included either. So basically, their tool speeds up a small part of a hardware engineer job when compared to it being done manually. Without a proper comparison against similar tools, their advertisement is pretty meaningless.
Coming from the software side, I'll add that most domain-specific neural nets like this do actually work pretty well as long as they're viable enough to have made it past the early prototype...
Coming from the software side, I'll add that most domain-specific neural nets like this do actually work pretty well as long as they're viable enough to have made it past the early prototype stage, and they often are a genuine improvement on what came before (there have been some meaningfully large steps in what's possible over the last 5-8 years, and that's now filtering down into actual usable tech).
I'm with you 100% on the hype, though. If there wasn't such a strong financial incentive to slap an "AI" label on everything, the actual coverage for 95% of advancements like this would be a line in a niche trade publication like "<startup> debuts neural net feature to do <specific sub-task> 36% more efficiently than <industry standard software>". These are real improvements that will cumulatively make real workflows faster/better/easier/cheaper/whatever - but they'd be happening quietly in the background if they didn't come under the fuzzy umbrella of the marketing term du jour.
Ngl, just reading the click-bait made me think "well I ****ing hope they could, that's what they're supposed to be doing..." And also, it kind of gives me vibes that AI wrote the article too. shrug
Ngl, just reading the click-bait made me think "well I ****ing hope they could, that's what they're supposed to be doing..."
And also, it kind of gives me vibes that AI wrote the article too. shrug
The only "good" functionality of this AI seems to be the ability to produce quick prototypes to work with, which I can only find useful if someone wants a hardware product just to seek seed...
Exemplary
The only "good" functionality of this AI seems to be the ability to produce quick prototypes to work with, which I can only find useful if someone wants a hardware product just to seek seed funding. Like the next stage of working with just an evaluation board could be this. But for the final product you would need to prove EMC compliance and pass other certifications which need you to prove a careful planning on the design of your product.
Even in the video on the original post they said that they had to go ahead and find alternative components for a couple of parts themselves because of manufacturer availability of the electronics. This is the standard work of an electronics / PCB engineer, maintaining a database of components and designing with a low BOM in mind so the AI can't automate that and you need people with knowledge how to do it.
To be fair there are quite a lot of companies trying to bring AI into PCB design like this one and Flux but it will depend on the people designing if they are going to find value in them.
Do you think those are the kind of things that could plausibly be used as training constraints in a later version of the model? Intuitively I get the impression that some kind of optimisation on...
Do you think those are the kind of things that could plausibly be used as training constraints in a later version of the model? Intuitively I get the impression that some kind of optimisation on BOM cost would probably be fairly well suited to this approach, but I've got no idea what EMC compliance testing looks like or whether you can realistically throw a candidate design into some kind of simulation software to give the model a success/failure signal on that.
Yeah, I'm not convinced. From the linked VentureBeat article on this: This is not the modern conception of AI at all. Insofar as I can tell, it's basically evolvable hardware, and it has been...
Quilter's technical approach differs fundamentally from the large language models that have dominated recent AI headlines. Where systems like GPT-5 or Claude learn to predict text based on massive training datasets of human writing, Quilter's AI learns by playing what amounts to an elaborate game against the laws of physics.
"Language models don't apply to us because this is not a language problem," Nesterenko explained. "If you ask it to actually create a blueprint, it has no training data for that. It has no context for that."
Instead, Quilter built what Nesterenko describes as a "game" where the AI agent makes sequential decisions — place this component here, route this trace there — and receives feedback based on whether the resulting design satisfies electromagnetic, thermal, and manufacturing constraints.
This is not the modern conception of AI at all. Insofar as I can tell, it's basically evolvable hardware, and it has been around since at least 1996.
Which is good! Because GenAI and LLMs are horrendous and using them for hardware engineering would be ridiculously stupid. So the tech itself is fine. No complaints there.
But that said, I find myself irritated by this anyway. Practically the entirety of both articles are unwaveringly positive, it talks about old concepts as though they've just been invented, and it's clearly using the term "AI" for clout. In other words, it's a sales pitch for investors. Even the title is glazing the hell out of this project. I'll save my excitement for when they do something worthwhile enough that it makes it out of ad copy, thanks.
I'll grant it this, though: It's nice to read about something "AI" that turns out not to be completely useless garbage for once. I'll take clickbait usage of the term over yet more climate destruction any day.
I'm pretty sure he's talking about their use of reinforcement learning (I was curious about the tech so I looked at their job listings!), so you can definitely make a fair argument that it is...
I'm pretty sure he's talking about their use of reinforcement learning (I was curious about the tech so I looked at their job listings!), so you can definitely make a fair argument that it is evolutionary, but the goal of the "game" is to allow a neural net to converge to a point that it's a general case model to create arbitrary circuit designs from input data, rather than the iterative process happening per-circuit to "evolve" that specific design.
So yeah, still a marketing piece, but it probably is actual 2025-era technology behind it! The problem with the AI term, other than the fact that everyone's got an incentive to yell about it because it'll get them coverage, is that nobody agrees what it actually means. We end up in this tricky situation where "added a pointless integration with the ChatGPT API to the website" and "trained a custom transformer model to perform a task that couldn't previously be automated efficiently" both get the same label, and the praise or derision (depending on your audience) aimed at one ends up spilling over to the other.
The title is a bit click-bait. Domain-specific AI was used to design the PCB routes. Still very cool and provides value to the world.
And also, honestly, article reads like a direct advertisement. Not much details, comparison with human hours without explanation, does human use existing automatic tools or do all work manually, etc.
Will be very interesting to read impression from folks that have experience in industry.
From what I can see, it's only an "AI" because of the hype around the term. They talk about it doing some placement and routing tasks, which is something that can be done by existing tools. Maybe it's doing a bit more than existing competitors?
About the project they present, I saw some feedback from someone working in the industry (on a different website) and their conclusion was that the project was perfectly scoped to magnify the impact of their tooI. For example, they don't start from scratch, but reuse an existing design (which can take 1-2 months of work for this kind of design). They also use a service to get an extra-speedy manufacturing of the board, which most people wouldn't use due to the cost. Testing (beyond just booting an OS) is not included either. So basically, their tool speeds up a small part of a hardware engineer job when compared to it being done manually. Without a proper comparison against similar tools, their advertisement is pretty meaningless.
Coming from the software side, I'll add that most domain-specific neural nets like this do actually work pretty well as long as they're viable enough to have made it past the early prototype stage, and they often are a genuine improvement on what came before (there have been some meaningfully large steps in what's possible over the last 5-8 years, and that's now filtering down into actual usable tech).
I'm with you 100% on the hype, though. If there wasn't such a strong financial incentive to slap an "AI" label on everything, the actual coverage for 95% of advancements like this would be a line in a niche trade publication like "<startup> debuts neural net feature to do <specific sub-task> 36% more efficiently than <industry standard software>". These are real improvements that will cumulatively make real workflows faster/better/easier/cheaper/whatever - but they'd be happening quietly in the background if they didn't come under the fuzzy umbrella of the marketing term du jour.
One day I hope to release an AI feature labeled as an “automatically tuned statistical function”
Ngl, just reading the click-bait made me think "well I ****ing hope they could, that's what they're supposed to be doing..."
And also, it kind of gives me vibes that AI wrote the article too. shrug
The only "good" functionality of this AI seems to be the ability to produce quick prototypes to work with, which I can only find useful if someone wants a hardware product just to seek seed funding. Like the next stage of working with just an evaluation board could be this. But for the final product you would need to prove EMC compliance and pass other certifications which need you to prove a careful planning on the design of your product.
Even in the video on the original post they said that they had to go ahead and find alternative components for a couple of parts themselves because of manufacturer availability of the electronics. This is the standard work of an electronics / PCB engineer, maintaining a database of components and designing with a low BOM in mind so the AI can't automate that and you need people with knowledge how to do it.
To be fair there are quite a lot of companies trying to bring AI into PCB design like this one and Flux but it will depend on the people designing if they are going to find value in them.
Do you think those are the kind of things that could plausibly be used as training constraints in a later version of the model? Intuitively I get the impression that some kind of optimisation on BOM cost would probably be fairly well suited to this approach, but I've got no idea what EMC compliance testing looks like or whether you can realistically throw a candidate design into some kind of simulation software to give the model a success/failure signal on that.
Yeah, I'm not convinced. From the linked VentureBeat article on this:
This is not the modern conception of AI at all. Insofar as I can tell, it's basically evolvable hardware, and it has been around since at least 1996.
Which is good! Because GenAI and LLMs are horrendous and using them for hardware engineering would be ridiculously stupid. So the tech itself is fine. No complaints there.
But that said, I find myself irritated by this anyway. Practically the entirety of both articles are unwaveringly positive, it talks about old concepts as though they've just been invented, and it's clearly using the term "AI" for clout. In other words, it's a sales pitch for investors. Even the title is glazing the hell out of this project. I'll save my excitement for when they do something worthwhile enough that it makes it out of ad copy, thanks.
I'll grant it this, though: It's nice to read about something "AI" that turns out not to be completely useless garbage for once. I'll take clickbait usage of the term over yet more climate destruction any day.
I'm pretty sure he's talking about their use of reinforcement learning (I was curious about the tech so I looked at their job listings!), so you can definitely make a fair argument that it is evolutionary, but the goal of the "game" is to allow a neural net to converge to a point that it's a general case model to create arbitrary circuit designs from input data, rather than the iterative process happening per-circuit to "evolve" that specific design.
So yeah, still a marketing piece, but it probably is actual 2025-era technology behind it! The problem with the AI term, other than the fact that everyone's got an incentive to yell about it because it'll get them coverage, is that nobody agrees what it actually means. We end up in this tricky situation where "added a pointless integration with the ChatGPT API to the website" and "trained a custom transformer model to perform a task that couldn't previously be automated efficiently" both get the same label, and the praise or derision (depending on your audience) aimed at one ends up spilling over to the other.
[obligatory Skynet reference]
#noise #joke