If it's wrong, you can determine that during simulation. You could also assert that in the background, I suppose — prune the bunk suggestions by the language model before presenting them.
In my own experiments I find GPT can succeed in complex tasks if you prompt it multiple times at different steps of the task, and have some level of recursive validation and re-prompting to keep it in line. In a case like this I have a feeling it would do a great job, and would be especially useful for the discovery/exploration phase of a new project.
For starters: From our testing it’s really extremely good at providing the right formulas to solve the problem at hand.
But you are right, these models are inheritely non deterministic and therefore can struggle with things like math. The solution is to provide it access to tools such as a calculator which works really well.
What’s been working well for us is to give it access to factual data sources such as calculators, component data sheets, etc.
It does really well from our tests in the past few weeks, so well that we decided to invite everyone to explore the boundaries with a community preview
Just give it a shot and report back
I'd be interested to know a lot more about testing methodology too. A hallucination could be a much more expensive problem than someone simply making a basic mistake that someone might otherwise catch if they were doing things the usual way.
Not trying to be obstinate, genuinely interested, but sceptical as my experience of ChatGPT has been very negative due to hallucinations, to the point that I find that the marketing materials sound far too good to be true versus the comment here, which seems much more grounded.
This is exactly why we are doing the community preview…to make it your favorite design tool.
@copilot can you design a low pass filter for me with 3db point of 300khz?
copilot: Sure, I can help you design a low-pass filter with a 3dB point of 300 kHz. A simple passive RC (resistor-capacitor) low-pass filter can be used for this purpose. To calculate the values of the resistor (R) and capacitor (C), you can use the following formula:
f_c = 1 / (2 pi R * C)
Where f_c is the 3dB cutoff frequency (300 kHz), R is the resistor value, and C is the capacitor value. Let's choose a standard resistor value, for example, 1 kΩ. Then, you can calculate the capacitor value as follows:
C = 1 / (2 pi R f_c) = 1 / (2 pi 1000 300000) ≈ 530.5 pF
Since 530.5 pF is not a standard capacitor value, you can use the closest standard value, which is 560 pF. With these values, the actual 3dB cutoff frequency will be approximately:
f_c = 1 / (2 pi 1000 * 560e-12) ≈ 283.5 kHz
The low-pass filter can be assembled by connecting the 1 kΩ resistor in series with the input signal, followed by the 560 pF capacitor connected to the ground. The filtered output can be taken across the capacitor. This design will give you a first-order RC low-pass filter with an actual 3dB point of approximately 283.5 kHz.
As a potential customer, I'm more interested in knowing where the potential known failure points or points of friction are than knowing what's working well.
Yes, when it works, it's certainly compelling in its responses—but as it never knows when it is wrong, it is equally compelling at a cursory glance when it hallucinates.
For most folks interested in this, their Dunning-Kruger complex will prevent them from picking up on when it's wrong because it's so convincing and sure of itself.
A lot of people liken AI to having a junior role working with you, but most junior positions have been to university and have a pretty solid understanding of a field in the first place. I wouldn't have a junior role hallucinating things that don't exist on a regular basis.
The speed of generating an output is only as meaningful as how consistently correct and appropriate that output is in terms of its value. I remain unconvinced.
I don't know if it will make you personally work faster or not, that's your business to sort out.
I have a personal project similar in concept to flux, but rather than working at the circuit level, it's at the device, firmware, and infrastructure level. I've been trying to use GPT to intuit how to network devices using different protocols, such that it might be easier for lab techs (for example) to discover how to manage and maintain their experiment tooling and infrastructure. It's really hard, though. The outputs and level of integration are mind-bendingly good, but... It's not perfect, which makes it too clumsy or even dangerous for what I want to build.
I think I envision something like GPT transforming human language into prompts for a narrower model with more predictable outputs, which is far more aware and capable with the application's logic and data. The user experience could then give you the conversational ease, yet far better application integration as well.
Anyway, just curious what you're thinking and very excited about projects like this. This is really cool.
We already use a multi model approach and mix of non deterministic and deterministic tooling in the pipeline to optimize for a given task
Lots of more work to be done here
Not everything is in prod yet but we def combine all the things you can imagine to pull this off
It makes sense if you're just shopping it around to the EDA incumbents since they'd have you redesign those parts anyway, but the fact that you currently offer a paid subscription to users makes me think that either you've underestimated or I've overestimated the importance of having those features by now. I guess that could also make sense if you want a more manageable user base for now.
We do have symbol and footprint creation and editing capabilities…in fact the vast majority of parts on the platform have been created by users.
Now what we don’t do is offer a modal editor like legacy tools do…instead it happens all in the same mode, which simplifies a lot of the UX and also provides powerful new abilities such using a schematic to create simulation models.
Now that said I think we can do a lot more here to further improve the experience and we def will but you are right in that it hasn’t been much of an issue which is because you have access to a vast preexisting library in the platform
The single mode thing is cool.
The circular and rectangular options have a bunch of transformations available…you can make turn circles into oblong shapes simple by providing async size value and you can set a corner radius for rectangular pads and even your layout.
And as you probably saw you can load symbols as svg and you can also load custom pad, layout and silk shapes either as svg or dxf
Less known but also works for shapes that you are reading off datasheets is that you can also just type in a svg path into the shape field of each object type
How philosophy is to make the most common things really simple and straightforward and the edge cases possible. But yeah, someday we’ll also ship online shape editing features
This kind of talk is tacky and repels technical users.
e.g. most recently, I needed a boost converter with at least 300mA sustained I at 3.3V coming from the lowest V source possible (alkaline battery powered gadget) After an hour or so of searching I had to choose between TLV61220 and TPS61021A . Maybe there is a better third chip? At this point I ran out of time to research and just figured this the TPS61021A would do.
Digikey/Mouser/etc have parametric filtering (when it works), but optimization among three variables is not the strong suit there, especially when you have, say, an 8mm square to fill and cost is a factor (and availability/flexibility between footprints)
I would love to use this as a fair recommendation engine (that isn't sponsored by TI, skewing all results towards their products) where I can give it my circuit's minimum/recommended/maximum parameters and have it suggest applicable parts, or even take my finished designs and suggest cheaper alternatives for parts that match the design's footprints but have lower capabilities, if that's all my circuit needed in the first place.
(just threw my query into chatgpt and it suggested https://www.ti.com/product/TPS61221) not bad.
I worked at Meta for years and knew first hand that ML/AI can only ever get good with loooots of training data.
So we needed a way to produce lots of high quality training data first and a way for users to be able to reason about results and edit them.
There is a graveyard of failed AI hardware design projects and one thing they all have in common is that they are black boxes…its very obviously the wrong approach. But every week a new one pops up…I don’t even understand why anyone funds this stuff anymore frankly…it can’t work.
So we embarked on building our own eCAD editor first…that was gonna be step number one. It was also the hardest and most expensive way to go about it… but if you have big dreams then you just have to go and do big things. I recent these get rich quick schemes personally anyways…I love doing hard things!
LLMs just starting to get good around the time we had the editor ready for public launch (after years of beta) was a happy coincidence that makes us look like geniuses now. But I take it lol.
We had been actively playing on the side for the past 6 months on ideas for first AI features as we geared up for launch of the core editor but the critical insight to use our existing projects comments surface was a last minute brain fart that came to us about 6 weeks ago.
Things went pretty quick from there and we are moving extremely fast now that we have the right foundation.
Excited times!
Incredibly stoked to announce this today!
This is a dream a long time in the making. When we first thought of rebuilding the hardware design tooling stack from scratch, AI driven design seemed like the dream at the end of the tunnel...only 3+ years later we are here
Given the open ended nature of AI Expert agents we are launching a community preview. We are committed to keep building http://flux.ai in the open and are excited to discover this new realm with all of you.
Hit me with questions