[HN Gopher] Show HN: Flux Copilot - Generative AI for hardware d...
       ___________________________________________________________________
        
       Show HN: Flux Copilot - Generative AI for hardware design
        
       Author : rock_hard
       Score  : 71 points
       Date   : 2023-06-27 18:43 UTC (4 hours ago)
        
 (HTM) web link (www.flux.ai)
 (TXT) w3m dump (www.flux.ai)
        
       | SV_BubbleTime wrote:
       | I would think this might have a future as a local plugin, but
       | there is more to laying out a complex board than routing so I see
       | "in browser" and see science project not tool for business.
        
         | rock_hard wrote:
         | Humblebrag: You'd be surprised how many of the most serious
         | businesses have started to migrate to use Flux already!
         | 
         | Agree though that there is more to creating complex boards that
         | automated wiring...we are not done here, just taking it step by
         | step!
        
           | TT-392 wrote:
           | I wonder, how many of those are software companies that also
           | do some hardware, and how many of those are like, proper EE
           | companies. Electrical engineers tend to be really stubborn
           | about their tools after all.
        
             | rock_hard wrote:
             | The list includes a bunch of hardware companies and teams
             | that most/all would agree are the most sophisticated in the
             | world
             | 
             | 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!
        
       | AHOHA wrote:
       | So pretty much AI now can do anything except that "conscious"
       | link between them. So an AI can write the software for drone
       | autopilot for example ..(missing link).. and design the hardware
       | for that drone ..(missing link).. and simulate all that
       | ..(missing link).. to fly that drone autonomously ..(missing
       | link).. to fulfill the objectives.
       | 
       | Unless something new fill these links properly, guess us human
       | kinds are still needed :)
        
         | samstave wrote:
         | tHE ("MISSING LINK") part appears to be 'iterative corrections'
         | - or "yes, that, but... now adjust here"
         | 
         | 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*]
        
       | adamwong246 wrote:
       | Software making hardware making software. How long till someone
       | wires it all up to a chat bot and lets it loose? We're reaching
       | the point where it's feasible to create, more or less, autonomous
       | AI entities. It's doable today, if you have the money and are so
       | inclined, and it gets cheaper everyday.
        
         | throwuwu wrote:
         | Missing a few steps like money and being able to open the
         | delivery box and assemble the parts. It's kind of like that
         | Sagan quote about making an apple pie from scratch.
        
           | adamwong246 wrote:
           | But like, IS it really that hard? Money is easy enough to
           | provide, presuming we can run this machine at a loss (for
           | now) and humans can fill in the last-mile. Even box-opening
           | can be automated soon. I think it's within reach of a
           | sufficiently motivated and funded grad student. Certainly
           | within the reach of Mr Bezos.
        
             | throwuwu wrote:
             | Not arguing with that. My point is that the gap between
             | humans helping it and it being fully autonomous is much
             | bigger than you would think. Even if you design a system
             | with a bunch of robotic arms and other equipment, the AI
             | then has to be able to replicate and install somewhere else
             | that entire set up using only what it currently has or can
             | assemble with its current facilities. That's a really big
             | challenge!
        
       | by_Seeing wrote:
       | For all my Stargate fans out there, this is how you get
       | replicators
        
       | rock_hard wrote:
       | Founder here:
       | 
       | Super excited to make another big step today to democratize
       | electronics design!
       | 
       | Hit me with your questions
        
         | pseudosudoer wrote:
         | Hey Matthias, I'm curious whether/how you validate copilot
         | output suggestions. Software bugs from generative AI can at
         | least be reasonably tested for by developers, but it's
         | certainly much more difficult to catch subtle bugs in hardware.
        
           | rock_hard wrote:
           | Great question
           | 
           | 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
        
         | stbtrax wrote:
         | Who is the target audience? It seems like experienced EEs would
         | know most of this stuff. Complete beginners would just not have
         | sufficient context. Is there some middle of the road
         | electronics hobbyists that you are targeting?
        
           | rock_hard wrote:
           | We have found engineers across the spectrum to find huge
           | value
           | 
           | 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.
        
       | markwu2001 wrote:
       | ai can build its own body now
        
       | avsteele wrote:
       | A few weeks ago I was going yo start learning how to use PCB
       | design software.
       | 
       | 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.
       | 
       | - Schematics                   - Has a large library of parts?
       | - UI is a little clunky for common task like rotating elements
       | but it gets the job done
       | 
       | - Circuit Simulation                   - 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
       | 
       | - API - Supports manipulation of the circuit via code (looks like
       | #javascript)                   - I haven't used it much, and
       | their \*example\* was non-functional (?!)
       | 
       | - PCB design                   - 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
       | 
       | I wound up using KiCad for design and www.circuitlab.com for
       | simulation.
        
         | osigurdson wrote:
         | Maybe a more reasonable sweet spot for (today's) AI would be a
         | "data sheet" expert. Ask questions like "what kind of part does
         | X, what is the minimum volume order and unit price".
        
           | rock_hard wrote:
           | Fully agree!
           | 
           | It's something we are working on among other things
        
         | rock_hard wrote:
         | Thanks for giving us a shot and the feedback!
         | 
         | 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
        
           | avsteele wrote:
           | I am a neophyte, but sure you can email me. see my profile.
        
       | londons_explore wrote:
       | I added a test[1] to OpenAI Evals that showed that GPT-4 doesn't
       | have an awful lot of understanding of most component datasheets
       | in the training set. It only got ~16% of my questions correct -
       | and all questions could be answered by a domain expert with
       | access to the datasheet.
       | 
       | 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...
        
         | jptlnk wrote:
         | hey, I'm the lead engineer on this project! I love that you did
         | this :)
         | 
         | 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!
        
           | londons_explore wrote:
           | I think the real solution is to manually convert component
           | datasheets to spice models (I think there are already
           | commercial databases of these), simulate the circuit, and
           | then get a language model like GPT-4 propose changes to the
           | circuit to make it perform better.
           | 
           | 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?
        
             | rock_hard wrote:
             | Our thinking has been the same
             | 
             | Give the LLM the ability to run the simulator and optimize
             | the solution based on the outputs
             | 
             | Lots of amazing opportunities ahead
        
       | amelius wrote:
       | Isn't Copilot a protected name in this context?
        
         | by_Seeing wrote:
         | The copilots of copilotai.com, educationcopilot.com,
         | workwithpod.com, arcwise.app, xmind.ai, usegalileo.ai,
         | copilotly.com, up.codes, and moveworks.com have entered the
         | chat.
        
       | TT-392 wrote:
       | So far I got it to hallucinate me a formula, which gave me half
       | the frequency I asked for on a 555 circuit. Also, I got it to
       | connect the data lines on a usb C port to a micro, but it only
       | connected A7(D-) and B6(D+), which doesn't seem ideal. Also, it
       | forgot about termination resistors, which, like, for usb 2.0,
       | stuff will probably work without them, but not ideal.
       | 
       | 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.
        
       ___________________________________________________________________
       (page generated 2023-06-27 23:03 UTC)