I use it daily for both work and personal things (Gemini, Grok, and ChatGPT). They still surprise me in both directions. Sometimes it's remarkably good, and other times it's completely off in the weeds even for things that just need basic search.
One complicated set of Excel formulas took about 10 rounds of revisions to get right and I had to use two different AI's, but I did wind up with a function that I doubt I would have gotten to on my own.
Another time I was working on an Excel calculation for a volume between two variable spherical caps with a cylindrical section between (also variable) and none of them could get it right no matter how many times they tried. The numbers were WAY off. That one I did on my own in the end, since they weren't saving me any time.
They've also been wildly unhelpful at times for basic stuff like what happened to a minor character in a TV show where an episode didn't tell you if the person lived or died after an injury. ChatGPT kept confusing a character with a similar name that was 5 years earlier in the show, even though I gave it a specific episode reference. After convincing it it was confusing the two, it still told me the character was killed in the episode I originally asked about, but he was in 3 episodes after that one. A standard Google search led me to an online reference for the show that told me that.
I still like them, and as I said I use them a decent amount, but they still need some validation at times, so the ability to judge what they are doing is required. I saw an AI developer a while back that suggested the solution to some of this was multiple AI's checking each other's work to come to a consensus.
One complicated set of Excel formulas took about 10 rounds of revisions to get right and I had to use two different AI's, but I did wind up with a function that I doubt I would have gotten to on my own.
Another time I was working on an Excel calculation for a volume between two variable spherical caps with a cylindrical section between (also variable) and none of them could get it right no matter how many times they tried. The numbers were WAY off. That one I did on my own in the end, since they weren't saving me any time.
They've also been wildly unhelpful at times for basic stuff like what happened to a minor character in a TV show where an episode didn't tell you if the person lived or died after an injury. ChatGPT kept confusing a character with a similar name that was 5 years earlier in the show, even though I gave it a specific episode reference. After convincing it it was confusing the two, it still told me the character was killed in the episode I originally asked about, but he was in 3 episodes after that one. A standard Google search led me to an online reference for the show that told me that.
I still like them, and as I said I use them a decent amount, but they still need some validation at times, so the ability to judge what they are doing is required. I saw an AI developer a while back that suggested the solution to some of this was multiple AI's checking each other's work to come to a consensus.
I've been using Claude, ChatGPT and Gemini to build out a very complex geometry proof for the last 3 months.
Informally I would say that the AI engines are 2X more powerful than they were 3 months ago. The reasoning is more accurate, the context windows are longer. It seems like they will continue to double in capacity every 3-6 months for the foreseeable future, which is a HUGE deal.
ChatGPT o1 and Gemini 2.5 have reasoning capabilities that go well beyond the basic "speech pattern recognition" of most LLMs. They do still have their problems.
They are able to spit out a few paragraphs of perfectly structured proofs of basic things for example.
They can try to construct more elaborate, longer proofs, but the longer they get the more prone they are to little errors.
I often use one AI to check the results of the other. I play 3-4 against each other. That does work.
However in the wider scope of the project, a VERY common pattern is: You find a mistake, you fix it, you check it, it finds another mistake, you fix it, and 2 hours later you find you're chasing a mobius strip that never ever ends. The mistake-chasing can go on forever, all the while continuing to introduce subtle errors while increasing the complexity beyond what any human can deal with.
The way I've found to correct this tendency is to carefully construct a "scaffolding" which states the inviolable ground rules and explicit goals for the project. This is a separate set of statements that says what is allowed, what is not allowed, with the goals carefully broken down into clear pieces. I call this my "axiom scaffolding."
I have to constantly remind the AI about the Axiom Scaffolding. And every time I get stuck I assume that the problem is that I have not constructed my axiom scaffolding correctly or else have not explained my original objectives clearly enough.
This seems to be working. The project is quite ambitious. 2 weeks ago I got it to a point where all of the AI tools said it was free of errors. I sent it to an experienced mathematician and he ripped it apart.
But then I asked "What do his comments say about my scaffolding?" then I updated the scaffolding.
Within a day I had rebuilt it to what is clearly better so far but not perfect.
This is giving me an intimate understanding of what AI can and can't do.
It's really a mirror of your own thinking. The answer is only as good as the question you ask. The thing that makes a great human teacher great is that they answer a better question than the one you ask, and then explain it to you.
I do not have any advanced education in geometry beyond what any normal electrical engineer knows, and this is modern advanced mathematics. I've found that an AI can teach a motivated student such things very effectively because it knows all the basic concepts.
I've noticed many times that even though it has explained something to me 6 times, I still don't understand it, and by that time I would be ashamed to keep repeating the same dumb question to a human. But with an AI there's no shame and if it takes 10 times for me to get the concept, then that's OK.
You can use AI to be lazy OR to tremendously sharpen your thinking.... just like you can use a traditional search engine to be a lazy researcher OR do better research than anyone could have possibly done before the Internet.
I believe only a minority of people will use AI this way but it makes those of us who are willing to use it this way super powerful.
Informally I would say that the AI engines are 2X more powerful than they were 3 months ago. The reasoning is more accurate, the context windows are longer. It seems like they will continue to double in capacity every 3-6 months for the foreseeable future, which is a HUGE deal.
ChatGPT o1 and Gemini 2.5 have reasoning capabilities that go well beyond the basic "speech pattern recognition" of most LLMs. They do still have their problems.
They are able to spit out a few paragraphs of perfectly structured proofs of basic things for example.
They can try to construct more elaborate, longer proofs, but the longer they get the more prone they are to little errors.
I often use one AI to check the results of the other. I play 3-4 against each other. That does work.
However in the wider scope of the project, a VERY common pattern is: You find a mistake, you fix it, you check it, it finds another mistake, you fix it, and 2 hours later you find you're chasing a mobius strip that never ever ends. The mistake-chasing can go on forever, all the while continuing to introduce subtle errors while increasing the complexity beyond what any human can deal with.
The way I've found to correct this tendency is to carefully construct a "scaffolding" which states the inviolable ground rules and explicit goals for the project. This is a separate set of statements that says what is allowed, what is not allowed, with the goals carefully broken down into clear pieces. I call this my "axiom scaffolding."
I have to constantly remind the AI about the Axiom Scaffolding. And every time I get stuck I assume that the problem is that I have not constructed my axiom scaffolding correctly or else have not explained my original objectives clearly enough.
This seems to be working. The project is quite ambitious. 2 weeks ago I got it to a point where all of the AI tools said it was free of errors. I sent it to an experienced mathematician and he ripped it apart.
But then I asked "What do his comments say about my scaffolding?" then I updated the scaffolding.
Within a day I had rebuilt it to what is clearly better so far but not perfect.
This is giving me an intimate understanding of what AI can and can't do.
It's really a mirror of your own thinking. The answer is only as good as the question you ask. The thing that makes a great human teacher great is that they answer a better question than the one you ask, and then explain it to you.
I do not have any advanced education in geometry beyond what any normal electrical engineer knows, and this is modern advanced mathematics. I've found that an AI can teach a motivated student such things very effectively because it knows all the basic concepts.
I've noticed many times that even though it has explained something to me 6 times, I still don't understand it, and by that time I would be ashamed to keep repeating the same dumb question to a human. But with an AI there's no shame and if it takes 10 times for me to get the concept, then that's OK.
You can use AI to be lazy OR to tremendously sharpen your thinking.... just like you can use a traditional search engine to be a lazy researcher OR do better research than anyone could have possibly done before the Internet.
I believe only a minority of people will use AI this way but it makes those of us who are willing to use it this way super powerful.
In a nutshell, I think your training as an engineer and hence the use of data, plus your years of experience highlight what can be achieved with an open mind. You only needed some small additional work to refine things that were not quite ideal in the first iteration highlights what these powerful learning machines can achieve.
I am sure we will see the fruits of this in some new horn geometry or other design work work in the near future.
I am sure we will see the fruits of this in some new horn geometry or other design work work in the near future.
Can I tell you how great it is to see someone doing AI the right way?
Absolutely brilliant.
Absolutely brilliant.
Yep. I use it for data analysis of very complex things. Python and R can plot things a million ways in a snap. Like over laying data on a spherical globe with maps of countries etc. This is how a lot of satellite imagery of the earth and weather data is done.
I have a script of planetary positions calculated using the JPL planetary ephemeris dataset in order to calculate planetary alignment index (to forecast solar flares) and I asked Chat GPT to animate it. Took minutes.
I have a script of planetary positions calculated using the JPL planetary ephemeris dataset in order to calculate planetary alignment index (to forecast solar flares) and I asked Chat GPT to animate it. Took minutes.
Last edited:
If you know how to use Vituix. But ChatGPT can do it for any dataset for any application. If it’s in column format all the better.
Which yes I agree a much better use. Only compile or analyze a limited dataset. Regardless if extremely long or short.
At what point it fails be interesting to know. Then proceeds to make up data with larger sets when going free range on the internet.
Or in many cases seems peoples false narratives have been entered.
Early on I used it as a search engine. It seemed to ignore higher paid searches or forced favoritism. Lately has changed.
For the topic of audio design. A specific baffle, specific alignment and specific crossover should come out. If giving 2 specific drivers datasets.
Problem is it doesn't have a alignment tool, diffraction tool or crossover tool.
It will just guess and use false opinions from the internet.
Turn the lie function = off. Design me a 2 way = Im sorry, Dave. Im afraid I cannot do that.
Computers are not sorry and not afraid, correct answer = I cannot do that.
At what point it fails be interesting to know. Then proceeds to make up data with larger sets when going free range on the internet.
Or in many cases seems peoples false narratives have been entered.
Early on I used it as a search engine. It seemed to ignore higher paid searches or forced favoritism. Lately has changed.
For the topic of audio design. A specific baffle, specific alignment and specific crossover should come out. If giving 2 specific drivers datasets.
Problem is it doesn't have a alignment tool, diffraction tool or crossover tool.
It will just guess and use false opinions from the internet.
Turn the lie function = off. Design me a 2 way = Im sorry, Dave. Im afraid I cannot do that.
Computers are not sorry and not afraid, correct answer = I cannot do that.
Last edited:
I think we all recognize the limitations. The question is whether AI can be made to useful work, despite them. I think that has been adequately demonstrated. would I use AI rather than Vituix and ABEC for speaker design? not a chance! But I would if I were exploring a new field. Or if someone integrated AI into speaker design tools.
It won’t be long before most software is AI integrated. It’s already happening fast. MS Word has Co-Pilot and Google Docs has Gemini, both on the right panel. So it’s not either / or. It’s both-and.
I ran vertical axis measurements, and generated a heatmap of angle vs. frequency from −90∘ (top) to −90∘ (bottom) and it looks like this:
I again generated this with ChatGPT. Got the above image on the first try.
My prompt was: I have attached a similar CSV file as before, but with some differences. This is 13 sets of measurements in pairs of frequency and magnitude: +90, 75, 60, 45, 30, 15, 0, -15, -30, -45, -60, -75, -90 degrees on the vertical axis of the speaker. I would like you to generate a very similar plot to the one before but instead of making the positive and negative measurements mirror images, I want you to use the positive angle data for the positive side of the graph and the negative angle data for the negative side of the graph. X axis is frequency in log scale from 20Hz - 20KHz; Y axis is magnitude; The color code should be the colors of the rainbow so red = 0dB and violet = -15dB. Normalize the magnitudes so that the first column is 0dB across the entire spectrum.
The end result is correct.
In fact I made a mistake when I said "Normalize the magnitudes so that the first column is 0dB across the entire spectrum" and it ignored that, and what it actually did was follow the pattern of the desired outcome. It normalized the magnitudes so that the MIDDLE column in the CSV became 0dB across the entire spectrum.
It could have easily erred somewhere, as did I. Very often prompts / expectations given by humans have internal contradictions and the AI does the best job it can to make sense of its inputs.
In fact I made a mistake when I said "Normalize the magnitudes so that the first column is 0dB across the entire spectrum" and it ignored that, and what it actually did was follow the pattern of the desired outcome. It normalized the magnitudes so that the MIDDLE column in the CSV became 0dB across the entire spectrum.
It could have easily erred somewhere, as did I. Very often prompts / expectations given by humans have internal contradictions and the AI does the best job it can to make sense of its inputs.
Maybe you can be specific? It seems to match the frequency response curves in my CSV.
If you're thinking it looks odd that there's more treble on the + side than the - side it's because the mic is only 0.5M away, the mic at 0 degrees is midway between woofer and tweeter, and when I turn the speaker on the turntable the tweeter gets a tad closer to the mic and is more on-axis at about +10-20 degrees.
Data below 100Hz is not reliable in this setup.
If you're thinking it looks odd that there's more treble on the + side than the - side it's because the mic is only 0.5M away, the mic at 0 degrees is midway between woofer and tweeter, and when I turn the speaker on the turntable the tweeter gets a tad closer to the mic and is more on-axis at about +10-20 degrees.
Data below 100Hz is not reliable in this setup.
How many plots? Are these the same as in post #20? Is there a crossover somewhere? Do you have the polars?
I did not post them. They are not the same as post #20 because they are new vertical measurements. 13 plots. CSV attached. Crossover is here https://www.diyaudio.com/community/...thout-horns-or-waveguides.422090/post-7891234
"AI" is a inference machine. In general, -off the top of my head and experience in such things- it is composed of three parts: the algorithms, the search engine and the data.
The search engine is designed to follow a path to a solution.. a "convergence" using the algorithms.
The data fed to it is what is uses to execute the algorithm. Usually, the data is what the algorithm "learned".
Most often, the algorithms tend to be generic across different data sets.. in order to do this, the data contains "rules"
You are defining the rules for the algorithm to process the data that you have. This maximizes the probability that the algorithm will converge to something you want and can use.
You are also teaching the AI algorithm, that's why it gets better.
The greatest area of research right now aims to develop judge models that are able to sort the wheat from the chaff.
If they succeed, there’s hope.
In all cases, these models will be busy.