Flux seemed really cool and I gave it a try but couldn't get going well enough to continue. Here are my impressions:
## Key features and my impressions (as of 4/29/23):
- Interaction
- Browser based
- Buggy in Firefox, don't try it (simulation failed to converge)
- Versioning (ala git)
- Collaboration
- Multiple people can work on a circuit at once
- can also give a link to share a circuit with limited permissions
- AI assistant
- Appears to have no specific knowledge of Flux? When I asked it question about running the code it didn't know I was asking about Flux even.
- Has a large library of parts?
- UI is a little clunky for common task like rotating elements but it gets the job done
- does not support plotting
- can't find a way to parametrize a simulation
- Can affect appearance of schematic or PCB model: e.g.: the LED In the schematic or 3D model will light up / be dark depending on the state determined by the simulation. (this doesn't help me get my work doe but it is cool/pretty)
- Can press a button on the schematic and see its effects via the sim
- I haven't used it much, and their \*example\* was non-functional (?!)
- Linked directly to the schematic so if you change a part there it changes on the PCB
- Has a very nice looking 3D mode/view of the circuit
We are still at day one here…lots of work ahead clearly!
On simulation: we have lots more to come here but work took a bit of a back burner as we have been focusing on PCB layout recently!
On AI: just getting our feed wet…but access to our docs is in the QA pipeline here and should hopefully land soon
Let me know if you wanna hop on a call…would love to learn more details about your usecases and how we can do better
It’s something we are working on among other things
This tool seems like a nice way to recommend parts, but I don't think I'd trust it with much of the actual design.
From what I have seen from AI so far. The only place I can see this level of AI being really useful, would be a place like digikey. Like, as an alternative to parametric search.
Example question: You are to answer each given question with a single number, with no extra characters or units. Round every answer to one significant figure. A IRF540N transistor has an R_DS(ON) of what, in Ohms, at V_GS=10v?
Answer: 0.04
[1]: https://github.com/Hello1024/evals/blob/4e9bbce1390ce427acb1...
our experience thus far is that GPT4 definitely needs help when it comes to datasheet-like data. it's great at broad "functional" strokes, but details and particularly numerical details are not its forte.
we're working on it!
The real question becomes how to interface the spice model and the language model - do you for example let it connect a virtual oscilloscope to any node, and give the language model the results back as plain numbers?
Super excited to make another big step today to democratize electronics design!
Hit me with your questions
It’s something we spend a lot of time thinking about. On one side we want to prevent hallucinations but on the other we want to keep the LLMs creativeness
We are experimenting with a bunch of different approaches here already
Preventing hallucinations and therefore improving correctness is fairly simple by providing the model with factual data sources which is something we are working on
The challenge is to balance that so that the model is also still creative when you want it to…but we are making progress here too
If you are interested ping us in our slack and we can add you to our beta tester channel to get access to our experiments in this space
From the engineers that design the motherboard of the phone in your pocket to students and hobbyists
Use case go from learning/explanation to brainstorming, component research, figuring out the match for signal filters, to triaging of issues and so on.
Agree though that there is more to creating complex boards that automated wiring…we are not done here, just taking it step by step!
Unless something new fill these links properly, guess us human kinds are still needed :)
What will be interesting is when you train a single doo-dad and it propagates to all the others (100 AI Monkey Syndrome) -- which then allows them to start self-normalizing and trainign eachothers ;
Get on-line and learn from your brethren (humans used to call these 'updates' - ugh, how *pedestrian*) [*pun intended*]
Give the LLM the ability to run the simulator and optimize the solution based on the outputs
Lots of amazing opportunities ahead
Hardware engineers get a bad rep for being stubborn when what they actual want is just good tools…and we believe they deserve them and so we are building them!
Putin is also not "fully autonomous" in that regard, yet here we are.
An AI can that can hire mercenaries isn't dull. Or design a circuit board, have it made and assembled and delivered to contractor to use in the field.
I apologize for engaging.