Useful Applications for AI

[Martin]

So why did they take that 'step if there were ready for it ?

In pursuit of the almighty dollar.

 

[mpw]

In pursuit of the almighty dollar.

Thats is bad == money.

 

[amirm]

No arguments folks. Use the other thread for that.

 

[Audiofire]

I have only used ChatGPT, Duck.ai and Gemini for free. I don't need it very much. I also have access to Copilot, so quite a few robot assistants which is great together with search engines, has helped a lot

 

[pierre]

coding seems to be where it's most useful, my questions tend to be on mathematics. The problem is that if it gives a wrong answer, there is no procedural logic you can go in and fix, you have to update the training set and hope for the best. Additionally there is not procedural logic you can audit, it's good for certain things, but no one is automating the running of a nuke using AI

For math, i have found them extremely good at generating proofs in quint or lean4 (or older languages like coq/tla+). I then also get a proof. Challenging the model to generate a proof removed a lot of issues i had similar to what your reported. They are also surprising good at finding failures in my proofs

 

[davmol]

You can work around this with LLMs by giving explicit instructions on how to answer. I found Gemini was often hallucinating on audio science related topics (much more so than ChatGPT) which my below instructions in my "audio" Notebook in Gemini help mitigate:

Answer the questions using verifiable facts. If the information is not explicitly known or if you lack sufficient data, state 'I do not know' or 'Information not available' then provide an educated estimate along with the probability of being correct. Be concise and use bullet points where possible.

I agree it does help to be specific with prompts but it doesn't eliminate hallucination, or guarantee correct answers. For day-to-day / low risk use though it's a good tip and will absolutely improve the output of LLMs.

 

[JustAnandaDourEyedDude]

Quite an extensive list of engineering physics you cover with your code, congratulations on that. It seems it rivals what Ansys, Siemens and Comsol offer. I am retired now, but I still enjoy thinking about some of the modeling and numerics I encountered in earlier years. I have not yet tangled with any AI, but from watching YT videos, it seems potentially very labor saving.

Curious what programming languages AI contributes to your code. I would prefer Fortran or Julia which would be easier for me to understand than C++, though Python seems unavoidable. Is your code 2D or 3D? High-fidelity or more idealized? Is your aeroacoustics module implemented as Lighthill/Ffowcs-Williams or Euler with group velocity or Dispersion Relation Preserving? For the fluid mechanics module, what artificial dissipation is used to stabilize and capture shock waves or detonation waves in the FEM or BEM? I briefly studied SUPG long ago when Claes Johnson's book was newly published, though all my working experience is with finite-volume which is a special case of FEM using delta and heaviside test functions. Have you looked into PINN and similar approaches that use Neural Networks and make meshing easier via point clouds?

How have you been able to debug the AI-produced code? Via benchmark cases? Seemingly tiny errors in the code can greatly contaminate the numerical solutions.

i have a different experience on similar tech questions, codex+claude help me to build an excellent BEM and FEM software from scratch that works on everything from 1 cpu to a cluster of gpu. Physics currently supported are acoustics, aeroacoustics, beam, composites, compressible_flow, contact, eddy_current, elasticity, electroacoustics, electrochemistry, fracture, lumped_circuit, magnetostatics, maxwell, nonlinear_solid, piezoelectric, plasticity, porous_flow, semiconductor, shallow_water, shell, stokes, thermal, thermoelasticity, transport, and some turbulence. I have an optimisation software that can do CAD -> mesher -> bem+fem with shape and topology optimisation.

Is it working easily and is it doing everything for me, no.

if you want to try something fun: pick something to optimise and ask evo-hq to find better solutions. That did work great for ex with fft, blas or lapack like stuff, assembly. And since the tool never tired, you can optimise for neon, then avx then rocm then cuda etc

 

[mermayer]

Staying with the topic of music: I mainly use AI for programming. My main app (alongside several others) brings together a number of improvements for Roon that I’ve been missing for a long time.
Programming fun with the ROON API

 

[pierre]

Quite an extensive list of engineering physics you cover with your code, congratulations on that. It seems it rivals what Ansys, Siemens and Comsol offer. I am retired now, but I still enjoy thinking about some of the modeling and numerics I encountered in earlier years. I have not yet tangled with any AI, but from watching YT videos, it seems potentially very labor saving.

I wanted to use Comsol but they are horribly expensive, don't scale well on GPU etc Plus the key value of what i do is to use AI to predict the solution faster than what bem/fem can do and get to control the result with a nightly run. 10 years ago, i would have needed 10 or 20 engineers full time to match my current output and a lot of things would not have been polished the way they are.

Curious what programming languages AI contributes to your code. I would prefer Fortran or Julia which would be easier for me to understand than C++, though Python seems unavoidable.

100% rust mostly for portability reasons and because the compiler is on the strict side and find bugs generated by xxx agents. Fortran does not provide enough guarantees and I decided to stop f77 25 years ago I use a python for visualisation or interfacing with other apps.

the value is not in Rust (vs others) it could have been C++ or GO. The value is that is have a 100% coherent codebase that i can compile with 1 command, that works everywhere with almost zero external dependencies. I rebuilt fft, blas, lapack, iterative solvers, etc to be autonomous. It also help the agents to debug the whole stack since I own all the code. Claude and Codex are very knowledgeable in this areas and progressed very quickly.

i had to relearn a lot of my programming skills. in the past i would have re-used as much as possible (use petsc for ex). Now it is unclear how much you benefit from well established solutions. It is clear that they are SOTA for some things but they were written long time ago and are struggling with GPUs for ex. I was surprise by how fast NVIDIA cards are at matrix/vector multiplications.

Is your code 2D or 3D?

2d, axy symmetric and 3d

High-fidelity or more idealized?

both depending on what i want to do or what the needs has been up to now.

Is your aeroacoustics module implemented as Lighthill/Ffowcs-Williams or Euler with group velocity or Dispersion Relation Preserving?

For now i have Euler well tested and Lighthill/FW not that well tested

For the fluid mechanics module, what artificial dissipation is used to stabilize and capture shock waves or detonation waves in the FEM or BEM?

I dont have proper shock wave support yet (low speed only). I have only done a few academic examples.

I briefly studied SUPG long ago when Claes Johnson's book was newly published, though all my working experience is with finite-volume which is a special case of FEM using delta and heaviside test functions. Have you looked into PINN and similar approaches that use Neural Networks and make meshing easier via point clouds?

Yes I am looking into PINN, currently testing with acoustics for some real time room-eq correction.

How have you been able to debug the AI-produced code? Via benchmark cases? Seemingly tiny errors in the code can greatly contaminate the numerical solutions.

One of the magical things with AI is that you can ask: for each kind of physics, find 10 analytical solutions from the litterature for 1d, 2d and 3d (10 for each) and implement them as a QA binary folllowing skills ...

A few hours later you get your 30 binaries that match the analytical solution.

or you can ask, look at arxiv, find me the last 10 papers on 1 physics module (like thermics) with significant impact and where they have reproducible benchmark/results, and again the tool will find them, analyse them write the code and allow you to compare yourself with established solutions. (Semi) Magical.

On the flip side, from time to time, i want to throw the whole thing out of the window, it can regress in a way that was not the case before.

 

[Mart68]

No way. IMDB list of actors for space:1999 is dozens and dozens of pages with hundreds of listings. Only the key talent has pictures. The rest is all names. If I had to do this manually would take me hours and hours.

Even with top talent it is more work to go to IMDB than to just ask Gemini in context of youtube page who is who.

I don't know the name of the reality show but I doubt that it has all the random customs agents in there.

Out of interest ('Space: 1999' is far more interesting than A.I) who were you looking up?

 

[Gorgonzola]

AI helpful? Why, YES.

Google AI helped me diagnose a vision problem that my optometrist had missed and offered a possible solution. That condition is called "aniseikonia"; that's where a person perceives objects as having different sizes depending on which eye is viewing it. This condition causes the brain to increasingly to favor the eye with the larger perceived size causing the other eye to weaken.

When confronted with the AI diagnosis my optometrist readily confirmed its validity. I now have on order so-called after the inventor, "Shaw lenses" that ought to compensated for the perceived sized differences.

 

[Music1969]

I know many of us worry about the negative consequences of AI, and worthiness of the responses it provides.

In the last few weeks, I have been extensively using both Gemini and my paid ChatGPT Pro subscription. Between the two, I have managed to get a lot done that would not have been easy or possible without them. So I thought I create this master thread on true usage scenarios that are positive. Please no complaints about AI or arguments. Just post what you used it for and how it was useful.

Here is a simple but shocking example. Was watching a Scifi TV series from 1970s on youtube called Space:1999. They have guest actors/actresses on it. I paused the video I was watching in Chrome and clicked on "Ask Gemini." I then typed in who the actress was. It instantly gave me the name which I verified online using image search. Finding that name in credits and searching each one to see if it is the right person would have taken an infinity.

Even more shocking, although created a misfire at first, was watching a reality show about Airport Security. This again dates back to 1970s/1980s. There was a scene with custom agent in it. I asked who that was and gave me a name. Did an image search and wasn't correct. I asked again and this time it identified the person correctly!

The fact that Gemini knows you are watching specific video helps a ton with context and not having to type more. Giving the same link for example to ChatGPT resulted in it complaining that it could not search youtube videos.

I have since routinely used the same method to get more data from the video I am watching.

So post what positive applications you have seen for AI.

I wonder if works for music credits too?

who produced this track? who are the song writers? what is the label? who sang background vocals?

I'm a nerd for music track credits

 

[JustAnandaDourEyedDude]

Thanks very much for the detailed reply! It gives me insight into the choices you made for your code so far. Especially the aspects of AI and GPU, which are relatively unknown by me.

I wanted to use Comsol but they are horribly expensive, don't scale well on GPU etc Plus the key value of what i do is to use AI to predict the solution faster than what bem/fem can do and get to control the result with a nightly run.

I fully agree with this assessment. In recent years, I had access to most Ansys software, and to SolidWorks, running on local workstations, supercomputers and cloud. However, management always grumbled at the cost of license renewal. Now, I lack access to all of that, and so any M&S software I toy with would have to be open-source downloads, or else something I would have to construct myself. I am completely onboard with reduced order models; they are highly useful for fast, approximate answers.

100% rust mostly for portability reasons

I have followed some news about Rust and Zig during the past year, but the learning curve would be steep for me, yet never say never.

I rebuilt fft, blas, lapack, iterative solvers, etc to be autonomous. It also help the agents to debug the whole stack since I own all the code. Claude and Codex are very knowledgeable in this areas and progressed very quickly.

i had to relearn a lot of my programming skills. in the past i would have re-used as much as possible (use petsc for ex). Now it is unclear how much you benefit from well established solutions. It is clear that they are SOTA for some things but they were written long time ago and are struggling with GPUs for ex. I was surprise by how fast NVIDIA cards are at matrix/vector multiplications.

Very impressive. Yes, many of the classic scientific programming libraries were written at a time when parallel processing was non-existent or not very advanced. Glad we now have acceleration of some tensor operations. The accelerator architectures are still evolving. Google's use of systolic arrays in its TPUs probably makes them more efficient than Nvidia's fixed topology.

One of the magical things with AI is that you can ask: for each kind of physics, find 10 analytical solutions from the litterature for 1d, 2d and 3d (10 for each) and implement them as a QA binary folllowing skills ...

A few hours later you get your 30 binaries that match the analytical solution.

or you can ask, look at arxiv, find me the last 10 papers on 1 physics module (like thermics) with significant impact and where they have reproducible benchmark/results, and again the tool will find them, analyse them write the code and allow you to compare yourself with established solutions. (Semi) Magical.

On the flip side, from time to time, i want to throw the whole thing out of the window, it can regress in a way that was not the case before.

This is the part I am completely unfamiliar with. It sounds like AI is a huge help, but I am glad you flag the undesirable code regressions as a significant caveat. Also, a lot of valuable papers are behind paywalls, all of which I no longer have access to .

Thanks again for all the insightful info. Your accomplishment is very impressive and quite inspirational for any modeler!

 

[Audiofire]

AI is good for finding songs, where you only remember part of the lyrics or other details

 

[ErVikingo]

I have used it for fun, for a promo video for one of my companies and I have been asking it for placement of my speakers. On this, the results are terrible no matter how much information I have fed into it. Its interesting that when i challenge it, it responds "you are correct, lets try again". As we "train it" it will get better.

 

[Chris A]

...and I have been asking it for placement of my speakers. On this, the results are terrible no matter how much information I have fed into it...

Just a pause to consider the ramifications of this statement (on an audio enthusiasts forum). This observation in itself is one that should give "audiophiles" cause for pause.

GPTs (all of them) use training information gleaned from internet and copyrighted books and articles (from which they typically did not provide the copyright holders any compensation). If the "AI" is consistently getting the recommendations wrong for placement of loudspeakers, there's a reason for that. All the information that I've read online and in books has not helped me in aiming my loudspeakers. Only trial-and-error has been effective.

But the observations and information that I learned from this personal effort I found was generalizable--able to be written down and used by others.

But I have not read that information anywhere. Further, when I have given advice on this subject (I usually start with asking questions about the full-range directivity response of the loudspeakers), I have been quoted back information--by the person asking for help--that my information "can't be correct".

Here--once again--we reach the limits of GPTs and how much (or how little) that they can help us.

P.S.: note that I have stopped trying to help those asking for help positioning their loudspeakers--because "they know better...". I've found that there is a great deal of existing bad advice in this hobby/pastime that the GPTs apparently have readily passed along when queried.

JMTC/YMMV.

Chris

 

[Pauli]

I was using ChatGPT (»Plus« account) to correct the output of my (digitally fed) 2 channel audio setup. Two aims:

1. Minimizing the differences between the measured frequency responses of my loudspeakers

2. Modifying both frequency responses to match some »lite Harman curve«

Included were/are: Audirvana Studio (player), REW, Hang Loose Convolver VST3 plugin, MEqualizer VST3 plugin, miniDSP UMIK-1 micro

The whole process took »us« about one week, which is partly due to the fact that most of the mentioned software apps were completely new to me.

I am pretty pleased with the result (subjectively and measurement-wise). Which makes me wonder quite a bit why there seems to be no(?) affordable AI supported Room-EQ software on the market yet.

 

[Chris A]

Remember that "frequency response" = amplitude AND phase response. The term is mostly synonymous with "transfer function".

If you meant only amplitude response was of interest, then I'd recommend saying "amplitude response" or "magnitude response"...

JMTC.

Chris

 

[Audiofire]

Remember that "frequency response" = amplitude AND phase response. The term is mostly synonymous with "transfer function".

The frequency response is the transfer function evaluated on the appropriate frequency axis (the imaginary axis for continuous-time systems and the unit circle for discrete-time systems).

 

[Chris A]

Frequency Response

Transfer function