@quilterai just hit a milestone we've been working toward for years: An AI-designed, 843-component Linux computer… that booted on the first try. This is Project Speedrun — the hardest test we’ve ever thrown at Quilter. 👇

Using physics-informed reinforcement learning, Quilter learns to design circuit boards by grading itself against what really matters: manufacturability, electromagnetics, thermodynamics, etc. quilter.ai/technology

I am thrilled to announce that @quilterai has raised $10M in Series A funding led by @ericvishria at @benchmark to automate circuit board design. blog.quilter.ai/series-a/ 🧵🧵🧵

@quilterai is automating PCB layout - a slow and laborious process necessary to create circuit boards. Today, we are releasing Quilter in open beta! app.quilter.ai

Quilter is a company built by and for hardcore engineers. We have chosen to tackle something important and incredibly difficult. There are no off the shelf solutions to this problem, no “AI API” we can call to get a result.

That's why I started Quilter – to build an AI “compiler” for circuit boards that lets engineers focus their creative energy on expanding the boundaries of human innovation, not tediously implementing their ideas within them.

As a software engineer, I am spoiled with amazing tools that help me write, validate, test and ship software in minutes. But, as a hardware engineer, I feel like I am stuck having to compile my own code.

This is the kind of challenge that our team lives for. If you’re interested in working at a place like this, please consider joining us! quilter.ai/careers

I want to extend a heartfelt thank you to @ericvishria at @benchmark, as well as @KVibhor and @SriViswan at @coatuemgmt for their investment in @quilterai, as well as to @chrissymeyer at @rootvc and @mcgd at Harrison Metal for their ongoing support.

We are actively hiring across multiple roles! Every member of the team carries significant responsibility and weight. These will be no different. If you’re up for a challenge come chat with us!

And most importantly, to the @quilterai team and their families – thank you for everything you have given to help us be here today. We won’t let you down! 🚀🌔

@Jason @sundeep thanks for discussing @quilterai on @twistartups! @Jason today, it's efficiency - but total automation, not copilot. The designs we auto generate now take customers a week to create by hand. A major pain I felt at @SpaceX. Breakthroughs are coming :)

Indeed! Shoutout to the open source design featured here. github.com/tjko/fanpico

Great question! The simple answer is: we don't need external training data. Quilter is built with Reinforcement Learning - that is it trains itself without human examples. We can validate that a PCB design is good by running physics simulations.

Computers are getting really good at playing games (Chess, Go, etc). Make the game "design a circuit board", make the score "will it work in real life?". That's @quilterai

This is similar to how @GoogleDeepMind trained AlphaGo. The agent played Go against itself until it became far better than any human. It is our goal to do the same thing with PCB layout.

It's been a blast working with you on this @kane! Thanks for being with us on this journey!

Couldn't agree more. A really long time ago I read about a genetic algo being used to configure an FPGA. It ended up using the parasitics in the chip to make the algo work. Will be so interesting when we can do that at the full PCB level. "Five individual logic cells were functionally disconnected from the rest⁠— with no pathways that would allow them to influence the output⁠— yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. " damninteresting.com/on-the-o…

We've had over a thousand new engineers try Quilter in the last few weeks submitting some really interesting designs. We really want to see some of these come to life, so we're subsidizing board builds! If you want to build a Quilter design in real life, we'll cover the cost of the PCB! More about this in the link. Open hardware community: this one is especially for you ;)

Thank you for joining us on the journey @ericvishria! I am excited to partner with you in building this vision!

@quilterai on @twistartups !

Running @quilterai on a computer @quilterai designed. Challenge accepted.

Thank you Sri! Excited to work with you in the years to come!

It's definitely heating up! And I'm glad. Well selfishly of course ;) But also because this ecosystem badly needs it!

Thank you Vibhor! So excited to build with you!

Thank you Hector!!

Thanks so much for the discussion @WSummerlinAI and congrats on the launch! Honored to be the first on the new pod ;)

Thanks for the great conversation @prateekj! Designing hardware with software is a topic that is so ripe for discussion. Really appreciate you supporting that conversation!

Thank you Michael! You believed in me before I believed in myself. I will never forget it :)

Right now, we focus just on speed and alleviating human labor. PCBs take days, weeks, even months to design. Something like a motherboard in a computer or phone is easily 2-4 months of layout work. Reducing that to a few hours of compute time drastically accelerates your iteration cycles. Eventually, as @quilterai gets better - I'm confident it will push boundaries that human designers do not. Reduce layer count, reduce board size, eliminate radio emissions that violate FCC rules etc. Given it takes 2-4 months to design a really complex board, humans make a lot of conservative assumptions during layout in order to make absolutely sure the board really works (and it usually doesn't on the first try anyway). But a computer can design thousands or millions of boards in hours - so you can really search for the optimal balance of electromagnetic and thermal performance vs manufacturing cost.

Easy to do well when you have a great team standing by your side ;)

@quilterai is far from perfect now - but we've made a bunch of really cool designs that work in real life without issue. This one was from more than a year ago, but is still one of my favorites because it demonstrated an STM32 with a camera, SD card, USB, a screen, and enough power to run a neural net on board. The design is embarrassing compared to what we can do now, but it still worked at full speed! But of course, that's what the funding is for - to make it way way better. Give us feedback - I promise we'll act on it :) blog.quilter.ai/qstache/

That would be awesome, thank you for the offer! Will follow up in DMs :)

@LnghrnEngineer thanks so much for hosting this conversation! A technical deep dive with other EEs is my favorite kind of conversation ;) Great listen for anyone curious about how PCB design automation is evolving!

Yeah definitely up for it! Let's coordinate in DMs :)

We could easily attempt >8 layer counts, but that's overkill given some of our other current limitations. It won't be long though ;)

We run multiple attempts simultaneously ranging from 2 to 8 layers. Our goal is generally to reduce layer count to try and make the board cheaper to fab, but of course on most boards 4+ layers are a must.

Thanks so much for the discussion @benjaminwolba! Appreciate you taking the time to dig in and discuss the efforts of the team @quilterai!