Comment by karmicthreat
5 hours ago
Adafruit probably did a review of AI PCB tools. I've used Flux.ai before; it was a pretty bad experience. After about 50-100$ in tokens a couple of times, I couldn't get more than a couple of simple components on the schematic. And not in sensible positions.
The product just grinds tokens for little return, in my opinion. I had far better luck wiring together KiCad MCP, SKIDL. There are some AI-driven autorouters out there now. Placement is probably the big issue that needs to be solved now. I could only get about 80% of what I wanted together with my hacky workflow.
This is exactly my experience, wasted $60 trying to get it to make something. The founder sent an automated AI email about setting up a time to meet and go through it then ghosted me at the meeting time.
> There are some AI-driven autorouters out there now. Placement is probably the big issue that needs to be solved now.
Interesting that within an IC this is basically "solved", or at least properly automated with classical numeric techniques such as simulated annealing.
I would have thought there's a big opportunity in a mixed-technique approach, where you use AI to extract unstructured data from datasheets and then feed it into more deterministic tools.
I also note that it's very easy to waste more than $100 in electronics once you start actually manufacturing bad PCBs.
> mixed-technique approach
I think my biggest annoyance with the way we rolled out AI is that nobody seemed to want to use it to augment already working solutions.
Just throw everything out and have an LLM do it instead.
I've been frustrated with Copilot in this regard.
I work on a large C++ codebase, with large files. Human developers jump around between files with the Visual Studio fuzzy search, set breakpoints to trace execution in the Debugger, use the IDE's refactoring tools.
Microsoft's answer to this was to just ... expose none of this to their Agent Mode!? Replace the working semantic autocomplete with fucking lies!?
Maybe it's changed, I haven't been paying that much attention after bouncing off of this. I've gotten mild acceleration from using gptel-mode in emacs, manually adding references to context, and having models do various mechanical transformations on code. And I've even had some limited success writing tools for it to do LSP lookups.
5 replies →
I recently saw a Claude skill that used Claude, with no tools, as a spell checker.
I wanted to hurl my laptop out to the window.
22 replies →
Way too much engineering effort to make something that might get leapfrogged by the next gen LLM.
It's a tantalizing thing, but far too treacherous to actually go for it, most of the time.
1 reply →
Something something bitter lesson blah blah
I think the bitter lesson is severely misapplied in the current situation: If progress from "just add more resources" is very slow, and a huge amount of money is at stake, continous work on hand-engineering can give a continuous and very valuable competitive advantage.
The labs all seem to be going for AGI through bigger LLMs, and I am reasonably sure that it's not going to happen like that.
1 reply →
A few days ago someone on HN commented that a teammate uses Claude to search for text in files on their own computer. Buddy... There's Command-line Tools Can Be 235x Faster Than Your Hadoop Cluster and then there's Command-line Tools Can Be ∞ Faster Than Your AI.
1 reply →
> nobody [wants to use AI] to augment already working solutions
Plenty of people do, but that only produces a blog post that will get you to the front page of HN. If you want VCs to drop $40M on your head, you need to pretend to reinvent the world.
Then, to further appease the rain gods, you need to sue the bloggers on the front page of HN who are challenging your world-changing narrative. Which will, heh, drop you on the front page of HN.
Our community is, literally, eating itself at this point. There was a time when we actually took "make something people want" literally. Now it's just part of the fiction.
That precise mixed technique approach has worked well for me. I’ve been using JITX (python based circuit design with a powerful auto router). Free for personal use, and has been discussed a few times here in HN.
Edit: it’s almost assumed at this point but for completeness Claude / Codex were the ones driving the OO python code and datasheet research and parsing.
https://www.jitx.com/
Until a few years ago it was generally understood that useful "creativity" involves solving problems within constraints, e.g. something a lot like SAT or SMT in spirit even if not in the details.
Then we got LLMs which will make a good parody of anything and occasionally get it right.
Within an IC you don't have large obstructions for metal layers, distances are short, and buffers can be inserted at will to manage SI.
It has been about 20 years since I worked on this (clock gating and clock buffering), but ..
> distances are short
I remember we had a catastrophic error for "wire longer than 2cm".
> and buffers can be inserted at will to manage SI.
Effective buffering of large nets was a massive pain. Areas where you want to buffer are inevitably areas with a very high level of placement congestion. So you push some cells out of the way to add a buffer. That ends up worsening their timing. So they need a bit more sizing/buffering. Rinse and repeat for a few hours.
( https://web.archive.org/web/20071028033035/http://www.edn.co... ; long since absorbed into Cadence)
It is far from solved in IC, synthesis tools sometimes still do really stupid things and there's still quite a lot of hand-holding required to get to a working chip.
And LLM are even stupider and need even more hand-holding
The right use of AI would be to use it to create a better routing/synthesis tool, but that's not what is being worked on
> After about 50-100$ in tokens a couple of times, I couldn't get more than a couple of simple components on the schematic.
Is this common? When I try out new AI tools, even as person who is financially independent, I load up maybe 10-20 USD worth of tokens, and if I don't get anything working from that, I literally give up and don't continue trying. If it can't do anything useful like "place a simple component on the schematic" after ~10 USD of expenditure, is it really worth continue adding more money into the platform? Seems DOA in those cases.
I used company money on it. I was hoping I could massage it along enough to get a workable test fixture out of it. I wanted to put together a simple hardware-in-the-loop tester for a component of our product.
Someone should’ve told these guys: https://news.ycombinator.com/item?id=48337328
I tried this last week and had the same experience. It was terrible and they got $140 out of me before I realized what it was (not) capable of. Their support was nonexistent as well.
All of these Gen AI tools where you pay a subscription fee are basically Software-as-a-Casino. You spin the wheel and hope it doesn't come up 00, then chase good money after bad when it does. Add in the parasocial relationship that some people develop with the LLM and you basically have OnlyFans but instead of vaguely dissatisfying feet pics to order it's vaguely dissatisfying code to order. It's that edge of "almost there, just one more token, bro" that makes it addictive.
That might be the right analogy except it is not clear that it is a house always wins situation.
If you have a .6 chance of success on any particular outcome. Long term win or loss is down to your behaviour. If you double or nothing every time loss is guaranteed. The right strategy will win over the long term.
9 replies →
> I could only get about 80% of what I wanted together with my hacky workflow.
I literally did this yesterday with solid results using Codex CLI. I used xhigh thinking and gpt 5.5.
I had it use KiCad directly via cli rather than via MCP, and I did make Claude Opus review it's work after every round. I got what I think will be a working revision A in about 10 hours of tinkering spread over a few days.
Is 10 hours a short amount of time for designing a PCB?
Faster than some, slower than some.
PCBs come in all different levels of complexity.
[dead]
> Placement is probably the big issue that needs to be solved now.
Would some sort of constraint-solving algorithm help with that? Something like (but not necessarily) Cassowary[0]? Maybe I'm misunderstanding what is meant by placement though; I don't have much domain knowledge in PCBs / electronics.
[0]: https://news.ycombinator.com/item?id=43362528
I've written my own autoplacer/autorouter. Placement is where you put the components on the board, routing is how you shape the traces to interconnect them.
It does a pretty decent job on small hobby-project boards of ~40 components (which is my use case at the moment), and I'm working part-time in the background on scaling it further.
The resulting designs pass all the KiCad electrical and geometry checks. Granted, I've spent about a year working on this problem, and it's hard, but not that hard a problem, providing you can avoid falling off the exponential cliff by decomposing it into hierarchical subproblems.
Quick-and-dirty unsupervised whole-board synthesis from schematic takes about 5 minutes, longer if you want cleaner output with nicer-looking better-routed traces.
As others here have said, placing is the real problem to solve, and that's where the magic happens. Place the components right, and routing is a relatively easy loosely-coupled constraint programming problem, place them wrongly, and you will have to get used to seeing the word UNFEASIBLE in your log output.
This project sounds very cool. Is it open source? If yes, can you share a link to the repo?