It seems there is big money to be made as an AI prompt writer
https://time.com/6272103/ai-prompt-engineer-job/
https://time.com/6272103/ai-prompt-engineer-job/
Other recently devised ways to use "AIs", some of which are hilarious, others are clever, and still others are indicative of the potential utility of using them daily to help us do our jobs or simply to do time-consuming tasks that we'd rather someone else do for us:
https://www.nytimes.com/interactive/2023/04/14/upshot/up-ai-uses.html
Some of these actually do a credible job of substituting for lack of, say, coding ability (which actually isn't one of my disabilities, BTW, but I know many people who do have it). Maybe all those liberal arts majors really do have a place in the world now...
Chris
https://www.nytimes.com/interactive/2023/04/14/upshot/up-ai-uses.html
Some of these actually do a credible job of substituting for lack of, say, coding ability (which actually isn't one of my disabilities, BTW, but I know many people who do have it). Maybe all those liberal arts majors really do have a place in the world now...
Chris
An interesting article on GPTs (generated pre-trained transformers) and LLMs (large language models) and their "emergent intelligences".
https://www.scientificamerican.com/article/how-ai-knows-things-no-one-told-it/
It seems that a more plausible explanation for the "emergence" seen in the LLMs is that AI researchers deeply underestimated how much of natural intelligence (i.e., of humans) isn't really "original thinking" but rather variations on a theme, and that neural nets really can mimic much behavior that humans hold as "intelligent".
In my prior working life as a design engineer, there were several tasks that I didn't enjoy as creative exercises (drafting mechanical assemblies, etc.) writing up the packages that explain the fabrication and assembly process (the "routings" for manufacturing), but rather the discovery of the true design requirements/constraints (including cost), the apportionment and fine tuning of the design requirements (for larger platforms/systems), then the enumeration of the different design alternatives/approaches and their relative ranking in terms of how well each.
I suppose that some sort of GPT/LLM approach with graphical generation capabilities will be yet borne into the world--probably sooner rather than later. But it seems to me that the type of jobs that require true judgment and an eye for the finished solution will still require human guidance or finishing touches. I can see where the drudgery work will be soon automated, including something as mundane as building loudspeaker enclosures and selecting drivers, even perhaps the dialing in of the components using packages similar to "room correction software" today. All it takes is someone to invent the graphical/engineering generative description capabilities and the ability to do the basic finite element/boundary element checks. Then perhaps much better engineering will become more commonplace for most products and even services.
What I've found, however, that the process of identifying the constraining requirements and their relative precedence is still something that is a bit "down the road" for automation. I know that I never really liked doing the drudgery tasks, but rather enjoyed the process of finding the requirements and understanding how the products were going to be used by the customer(s). I never tired of that discovery process that typically leads to exciting new and revolutionary designs.
Chris
https://www.scientificamerican.com/article/how-ai-knows-things-no-one-told-it/
It seems that a more plausible explanation for the "emergence" seen in the LLMs is that AI researchers deeply underestimated how much of natural intelligence (i.e., of humans) isn't really "original thinking" but rather variations on a theme, and that neural nets really can mimic much behavior that humans hold as "intelligent".
In my prior working life as a design engineer, there were several tasks that I didn't enjoy as creative exercises (drafting mechanical assemblies, etc.) writing up the packages that explain the fabrication and assembly process (the "routings" for manufacturing), but rather the discovery of the true design requirements/constraints (including cost), the apportionment and fine tuning of the design requirements (for larger platforms/systems), then the enumeration of the different design alternatives/approaches and their relative ranking in terms of how well each.
I suppose that some sort of GPT/LLM approach with graphical generation capabilities will be yet borne into the world--probably sooner rather than later. But it seems to me that the type of jobs that require true judgment and an eye for the finished solution will still require human guidance or finishing touches. I can see where the drudgery work will be soon automated, including something as mundane as building loudspeaker enclosures and selecting drivers, even perhaps the dialing in of the components using packages similar to "room correction software" today. All it takes is someone to invent the graphical/engineering generative description capabilities and the ability to do the basic finite element/boundary element checks. Then perhaps much better engineering will become more commonplace for most products and even services.
What I've found, however, that the process of identifying the constraining requirements and their relative precedence is still something that is a bit "down the road" for automation. I know that I never really liked doing the drudgery tasks, but rather enjoyed the process of finding the requirements and understanding how the products were going to be used by the customer(s). I never tired of that discovery process that typically leads to exciting new and revolutionary designs.
Chris
I agree, 99% of what people do with their minds is monkey-see, monkey-do, and then they credit themselves with vast intellect. The same goes for so-called AI. What we're seeing is not intelligence - not in machines, and usually not in people. But at the same time people vastly overcredit the virtue of intelligence, because technology is just a means to power, and people love that.
Interesting commentary in the Guardian recently such that, what chatbots do is a lot like the self-driving car issue: 99% safe is not nearly good enough, and the last mile of achieving anything more than that shows an absurdly declining marginal return on investment. So AI will make the rich massively richer, do nothing for the poor, and eliminate all the "drudge work" that the middle class have been doing for the past couple hundred years, leaving them to join the ranks of the nearly-worthless labor pool.
Anybody who thinks that the effect the internet has had on the world in the past 25 years has been awesome is going to love what's coming in the next 25 years.
Interesting commentary in the Guardian recently such that, what chatbots do is a lot like the self-driving car issue: 99% safe is not nearly good enough, and the last mile of achieving anything more than that shows an absurdly declining marginal return on investment. So AI will make the rich massively richer, do nothing for the poor, and eliminate all the "drudge work" that the middle class have been doing for the past couple hundred years, leaving them to join the ranks of the nearly-worthless labor pool.
Anybody who thinks that the effect the internet has had on the world in the past 25 years has been awesome is going to love what's coming in the next 25 years.
This is great...
Here’s What Happens When Your Lawyer Uses ChatGPT
I live to read about such news articles. It made my day.
I guess everyone tries to cheat on their homework--even after getting out of school...
Chris
Here’s What Happens When Your Lawyer Uses ChatGPT
I live to read about such news articles. It made my day.
I guess everyone tries to cheat on their homework--even after getting out of school...
Chris
Last edited:
This article seems like a much more reasonable response to using the emerging GPTs:
https://thehill.com/opinion/technol...chatgpt-teach-students-to-do-what-it-cant-do/
What I've found with ChatGPT "3.5" is that the information on audio subjects like in-room EQ that it generates is generally a rehash of the pablum of the nonsense that you find in audio magazines and on-line articles. Here is one prompt that I gave it:
Me: Can you give a step-by-step process of using an in-room acoustic measurement app to dial-in the parametric equalizer (PEQ) filters of a typical DSP crossover (otherwise called a loudspeaker crossover)?
ChatGPT3.5: Certainly! Here is a step-by-step process for using an in-room acoustic measurement app to dial-in the parametric equalizer (PEQ) filters of a typical DSP crossover:
Me: Can you expand on point #5?
ChatGPT3/5: Sure! Here's some more information about step 5, which involves identifying problematic frequencies:
When you measure the frequency response of your speakers using the in-room acoustic measurement app, you'll see a graph that shows the amplitude (level) of the sound at different frequencies. You can use this graph to identify any frequencies that are over-emphasized (peaks) or under-emphasized (dips) compared to the desired response.
Peaks can be caused by room modes, which are resonances that occur when sound waves bounce back and forth between parallel surfaces in the room. Room modes can cause some frequencies to be amplified or canceled out, resulting in a peak or dip in the frequency response.
Dips can also be caused by room modes, but they can also be caused by other factors such as speaker placement, room dimensions, and reflections from walls or other surfaces in the room.
To identify problematic frequencies, look for peaks or dips in the frequency response graph that are more than a few dB above or below the desired response. For example, if the desired response is a flat frequency response, any peaks or dips that are more than a few dB above or below the flat response would be considered problematic.
Once you've identified problematic frequencies, you can use the PEQ filters in the DSP crossover to adjust the response of the speakers at those frequencies, as described in step 6. By doing so, you can flatten the frequency response and improve the sound quality of your system.
Me: Can you give me a bibliography of sources for all the above?
ChatGPT3.5: Certainly! Here is a list of sources that were used to provide information for the steps outlined above:
Overall, I'd give the generative portion of ChatGPT3.5 a "C+" grade on the above subject at best, and perhaps much worse since it really doesn't bring much to the table that one could guess by oneself. Also, the bibliography wasn't very spectacular to be honest. You can do a lot better using a variety of published books, not the least of which is "Testing Loudspeakers" (1998) by Joseph D'Appolito, which really is a helpful source on the general subject of testing loudspeakers acoustically, although use of something like REW has other on-line guides that are much more fit for the purpose of this prompt under consideration.
Overal, I believe that a lot more GPT training is needed for those that don't know much about this subject. Perhaps GPT 4.0 has alleviated some of these problems in its training database.
Chris
https://thehill.com/opinion/technol...chatgpt-teach-students-to-do-what-it-cant-do/
What I've found with ChatGPT "3.5" is that the information on audio subjects like in-room EQ that it generates is generally a rehash of the pablum of the nonsense that you find in audio magazines and on-line articles. Here is one prompt that I gave it:
Me: Can you give a step-by-step process of using an in-room acoustic measurement app to dial-in the parametric equalizer (PEQ) filters of a typical DSP crossover (otherwise called a loudspeaker crossover)?
ChatGPT3.5: Certainly! Here is a step-by-step process for using an in-room acoustic measurement app to dial-in the parametric equalizer (PEQ) filters of a typical DSP crossover:
- Set up the measurement microphone: First, you'll need to set up a measurement microphone at the listening position [Me: This is generally incorrect, as it needs to be much closer--like 1m]. Ideally, the microphone should be placed at ear height and equidistant from the left and right speakers [Me: this is also incorrect since REW doesn't work like "room correction software". The measurements should at least start out on-axis with the loudspeaker, and if in-room polar response is desired, then the measurements should progress in 10-15 degree increments off-axis to its centerline if measuring polar response.]
- Connect the measurement microphone to your computer or mobile device: Connect the measurement microphone to your computer or mobile device using an appropriate audio interface. [Me: How? What kind of microphone is used: USB or analog-output? No discussion of calibration is found--which is pretty important.]
- Open the measurement app: Open the in-room acoustic measurement app of your choice on your computer or mobile device.
- Measure the frequency response: Follow the instructions in the app to measure the frequency response of your speakers in the room. This will give you a graph of the speakers' frequency response, showing which frequencies are over-emphasized or under-emphasized.
- Identify problematic frequencies: Identify any frequencies that are causing problems, such as resonances or dips in the frequency response. [Me: This point is requested to be expanded below.]
- Adjust the PEQ filters: Use the app's PEQ filters to adjust the frequency response of your speakers, targeting the problematic frequencies. To do this, select a PEQ filter and adjust the center frequency, gain, and Q to match the frequency and amplitude of the problematic peak or dip. Repeat this process for each problematic frequency you identified in step 5. [Me: Maximum number of PEQs?]
- Verify the changes: After adjusting the PEQ filters, measure the frequency response again to verify that the changes have improved the frequency response. Repeat steps 5-7 as necessary until you are satisfied with the frequency response.
- Save the settings: Save the settings in the DSP crossover, so that they are applied whenever the system is used.
Me: Can you expand on point #5?
ChatGPT3/5: Sure! Here's some more information about step 5, which involves identifying problematic frequencies:
When you measure the frequency response of your speakers using the in-room acoustic measurement app, you'll see a graph that shows the amplitude (level) of the sound at different frequencies. You can use this graph to identify any frequencies that are over-emphasized (peaks) or under-emphasized (dips) compared to the desired response.
Peaks can be caused by room modes, which are resonances that occur when sound waves bounce back and forth between parallel surfaces in the room. Room modes can cause some frequencies to be amplified or canceled out, resulting in a peak or dip in the frequency response.
Dips can also be caused by room modes, but they can also be caused by other factors such as speaker placement, room dimensions, and reflections from walls or other surfaces in the room.
To identify problematic frequencies, look for peaks or dips in the frequency response graph that are more than a few dB above or below the desired response. For example, if the desired response is a flat frequency response, any peaks or dips that are more than a few dB above or below the flat response would be considered problematic.
Once you've identified problematic frequencies, you can use the PEQ filters in the DSP crossover to adjust the response of the speakers at those frequencies, as described in step 6. By doing so, you can flatten the frequency response and improve the sound quality of your system.
Me: Can you give me a bibliography of sources for all the above?
ChatGPT3.5: Certainly! Here is a list of sources that were used to provide information for the steps outlined above:
- No specific source was used for step 1.
- No specific source was used for step 2.
- No specific source was used for step 3.
- "How to Measure Room Acoustics Using Your Smartphone" by Craig Anderton, published on Sweetwater's Knowledge Base website (https://www.sweetwater.com/sweetcare/articles/how-to-measure-room-acoustics-using-your-smartphone/)
- "How to Use Room EQ Wizard to Analyze Your Room's Acoustics" by Brent Butterworth, published on Lifewire website (https://www.lifewire.com/how-to-use-room-eq-wizard-to-analyze-your-rooms-acoustics-1847585)
- "Parametric EQ Guide for Beginner Mixing Engineers" by Casey Van Wensem, published on the Reverb website (https://reverb.com/news/parametric-eq-guide-for-beginner-mixing-engineers)
- No specific source was used for step 7.
- "Loudspeaker Crossover Design - Part 2: Selecting Component Values" by Keith Howard, published in Sound on Sound magazine (https://www.soundonsound.com/techni...over-design-part-2-selecting-component-values)
Overall, I'd give the generative portion of ChatGPT3.5 a "C+" grade on the above subject at best, and perhaps much worse since it really doesn't bring much to the table that one could guess by oneself. Also, the bibliography wasn't very spectacular to be honest. You can do a lot better using a variety of published books, not the least of which is "Testing Loudspeakers" (1998) by Joseph D'Appolito, which really is a helpful source on the general subject of testing loudspeakers acoustically, although use of something like REW has other on-line guides that are much more fit for the purpose of this prompt under consideration.
Overal, I believe that a lot more GPT training is needed for those that don't know much about this subject. Perhaps GPT 4.0 has alleviated some of these problems in its training database.
Chris
There was a recent article saying that the performance of ChatGPT is getting worse over time, and of course, nobody knows why. Well, that's nice.
The more I read about it, the more I am persuaded of the argument I mentioned before - that it's an issue like self-driving cars, where you can get to 98% good enough on some things (and I agree with you, it's a C on most things at best) but 98% is not safe enough for a car, and not good enough for many critical applications. After that, the curve gets so steep that you'd need to roast all the petroleum in the ground to generate the computing power to get to 99%. Unfortunately, that's what what people are going to do. Meanwhile, I'm not going to be using ChatGPT to write my memoirs.
ChatGPT is only good at writing computer code, and maybe it is useful for intelligence agencies in machine translation, which helps because we have long been at a disadvantage to the Chinese there.
For the most part, people don't understand what Complexity is, and actively resist understanding what it is, because as a species we rely on our cognitive abilities and 'figuring things out.' We resist, with religious fervor, the idea that some things can't be known because they have emergent properties. That's what makes the Silicon Valley tech bros look so maddeningly thick-headed. It's at least a little heartening to think that ChatGPT can't get a whole lot better, if that means the threat of it eating all our resources might not accelerate.
Well, that's nice.The more I read about it, the more I am persuaded of the argument I mentioned before - that it's an issue like self-driving cars, where you can get to 98% good enough on some things (and I agree with you, it's a C on most things at best) but 98% is not safe enough for a car, and not good enough for many critical applications. After that, the curve gets so steep that you'd need to roast all the petroleum in the ground to generate the computing power to get to 99%. Unfortunately, that's what what people are going to do. Meanwhile, I'm not going to be using ChatGPT to write my memoirs.
ChatGPT is only good at writing computer code, and maybe it is useful for intelligence agencies in machine translation, which helps because we have long been at a disadvantage to the Chinese there.
For the most part, people don't understand what Complexity is, and actively resist understanding what it is, because as a species we rely on our cognitive abilities and 'figuring things out.' We resist, with religious fervor, the idea that some things can't be known because they have emergent properties. That's what makes the Silicon Valley tech bros look so maddeningly thick-headed. It's at least a little heartening to think that ChatGPT can't get a whole lot better, if that means the threat of it eating all our resources might not accelerate.
Well...who needs books about hi-fi audio anymore?
How Will A.I. Learn Next?
It's a brave new world. Even I begin to wonder if the guys with "white hats" will always win over the guys with "black hats". Certainly, it's going to be a big shootout before the dust settles.
Chris
Yes, I have. When will it be published or otherwise made available? Well...not so sure. I believe that most of my research is mostly complete. I've already drafted about half the book, but significant work remains.
Perhaps I'll stop posting on diyAudio for a while until I can get a first draft completed.
Chris
Perhaps I'll stop posting on diyAudio for a while until I can get a first draft completed.
Chris
We recently had a TV show where they explored AI-generating stories.
The outcome: AI produced perfectly clear and langiage-correct stories.
But they was so full of clichés that it was totally uninteresting, even painful, to read.
In hindsight, logical: no creativity.
So, I would say: go for it! Looking forward to your creation.
AI won't replace me for at least another century. I'm unlikely to care by them
Jan
The outcome: AI produced perfectly clear and langiage-correct stories.
But they was so full of clichés that it was totally uninteresting, even painful, to read.
In hindsight, logical: no creativity.
So, I would say: go for it! Looking forward to your creation.
AI won't replace me for at least another century. I'm unlikely to care by them
Jan
@jan.didden You are irreplaceable regardless of timespan! However, underestimate AI at your peril!…
AI won't replace me for at least another century. I'm unlikely to care by them
Jan
Stuff that AI did badly a year ago can be done better now and great progress is being made with new “Small Language Models” (SLMs) like Microsoft’s Phi-2 that is small enough to run on a smartphone and can beat LLMs of only a year ago in certain tasks with task specific training on its target subject.
The release of MS copilot will unleash a tsunami of data about AI and how to improve it.
The other big players are hardly sitting on their hands either.
I am reading Mustafa Suleyman’s The Coming Wave book. Some very scary scenarios are presented on the future power of AI from the founder of DeepMind.
The capabilities of AI will be VERY different in 12 months from now.
I am not directly involved in AI other than as a regular user, but I can see the effort and investment all around me.
Jeff
Around the time verbs became nouns, I suspect. "That's a big ask", for example. Ask is a verb. Not a noun. I'm guilty of that kind of language abuse too, though.
Tom