GPT4 / CHATGPT Plugins for Speaker measurement analysis and comparison

[Jeromeof]

Not sure if people have seen the latest announcement from openai which is that you can provide plugins directly into the GPT4 which can basically allow the AI to analyse raw datasets and allow easy "interactions" with the RAW data, providing visualisations etc. Specifically If people are interested there is a small section inside the "interpreter" plugin video near the bottom on this page https://openai.com/blog/chatgpt-plugins where they show pretty nice analysis music data, this IMO looks like the classic data science tool Juypter Notebooks plugged into ChatGPT.

So one idea I had looking at this was that maybe we can get the raw data from the various spinorama measurements plugged in and allow people to play around with comparisons between speakers in new and potentially useful ways. The big advantage of this is that it cuts out the big complaint about "ChatGPT" which is it can be just "making stuff up". So what do you think @pierre or @amirm ? I signed up for developer access so would be interested in assisting when I get access.

Note: Its obviously not just speaker measurements where this might be useful any raw audio dataset e.g. like headphone measurements would benefit from this so would be interesting if say @crinacle would also be interested.

 

[notsodeadlizard]

Finally, a machine not based on knowledge, experience and reasoning will be able to make a huge blog in which it will write what it does not understand, and comment on it.
Living people will have more time to listen to music

And for what you wrote about, no ChatGPT is needed at all.
Put tags and people themselves will find unnecessary and incomprehensible to them by a "simple" ordinary search.
Why write a whole sentence instead of several keywords if the result is expected to be a keyword search result?
And understand please, current "AI" (even if it's labeled as "generative") is not able to analyze raw data sets if people have not done the analysis of these data sets before.
You can use this "AI" to generate short content for long nonsensical articles in which 90% of the text is not relevant to the point. This is very convenient. But don't expect this "AI" will replace the decades of work in any field to gain a deep understanding of measurement techniques and results.

 

[Jeromeof]

Finally, a machine not based on knowledge, experience and reasoning will be able to make a huge blog in which it will write what it does not understand, and comment on it.
Living people will have more time to listen to music

And for what you wrote about, no ChatGPT is needed at all.
Put tags and people themselves will find unnecessary and incomprehensible to them by a "simple" ordinary search.
Why write a whole sentence instead of several keywords if the result is expected to be a keyword search result?
And understand please, current "AI" (even if it's labeled as "generative") is not able to analyze raw data sets if people have not done the analysis of these data sets before.
You can use this "AI" to generate short content for long nonsensical articles in which 90% of the text is not relevant to the point. This is very convenient. But don't expect this "AI" will replace the decades of work in any field to gain a deep understanding of measurement techniques and results.

I think maybe look more closely at the examples - you might have only been playing with the generative text based "chat" uI - but you can actually feed it raw CSV / JSON data and get results out in raw CSV / Data - and once that data is in the model / session - you interact with it on your own terms, think Excel spreadsheet with amazing filters that "mere mortals" can hopefully understand and use - so its not going to replace decades of research but because its a human interacting with the data - that human can "stand on the shoulder of giants" by using this type of tool

 

[amirm]

Yes I am interested in what it can do.

 

[Syntactic]

GPT-4 hallucinates facts and makes reasoning errors the same way ChatGPT does and should absolutely never be used for any kind of scientific work, like the kind of data analysis you're talking about. It is good at a lot of things. Giving it data and asking it questions about the data you need accurate answers to is not one of those things.

 

[Zoomer]

should absolutely never be used for any kind of scientific work

On the other hand I'm quite sure it could generate posts for this forum which would be indistinguishable from the real stuff.
(not that I agree with your premise)

 

[Triliza]

This is another tool (which most of us don't quite understand its utility), I'd be interested to see how @Jeromeof and other interested parties could make use of it in the context of ASR. If it'll be something useful and accurate, time will tell.

 

[test1223]

GPT-4 hallucinates facts and makes reasoning errors the same way ChatGPT does and should absolutely never be used for any kind of scientific work, like the kind of data analysis you're talking about. It is good at a lot of things. Giving it data and asking it questions about the data you need accurate answers to is not one of those things.

GPT can't calculate that is a problem here I think the version 4 can use Wolfram Alpha for this. I wouldn't dismiss GPT because of its (current) limitations for data analysis. I think it might provide some interesting "thoughts" based on you input of cause. Can you 100% trust its answers - "No". But based on my experience it accelerate some things so much that you can try thinks out so fast. And it can do some tasks with 100% success rate. I assume it would bring a huge benefits for analysis of the spinorama data even if not all ideas might work.

 

[Berwhale]

Ooh, this is discussion is already devolving into an argument against 'Strong AI' and the difference between syntax and semantics...

Here's one I watched earlier:

 

[Jeromeof]

Yes I am interested in what it can do.

I will put some updates here when I get full access to the plugins and especially the "Interpreter" plugin.

What I did to assist the process is I pushed a small example PR of what a ChatGPT plugin might look like to @pierre GitHub project hosting the code for www.spinorama.org but of course its totally up to Pierre to decide if he wants to use it (he will need adapt some of the sample code - but I am willing to assist if he wants).

I think possible to stop the slight confusion of what this would provide - it's not that the AI will analyse the data itself and provide incorrect / made-up output from the dataset provided. The key thing IMO, is that these plugins provide is a way for a human to use ChatGPT natural language capability (and code generation) to interact with the dataset.

The "interpreter" plugin is as far as I can see a Jupyter notebook, which is an amazing tool if you are a data scientist, you can load data into the "notebook" and manipulate / filter / transform and especially visualise that data over the course of many steps within the tool. But its one problem is you need to know a little bit of programming e.g. python / R and understand the visualisation tools provided by those languages. So it's not enough to be a "domain expert" when using the tool, you have to be a developer / engineer.

But this is what I see changes completely with this plugin, instead of writing "code" to do the data manipulation, that you might want in your scientific domain of interest, you can now tell chatGPT in plain English how you want it to manipulate and especially visualise the data. This opens up this whole "data science" area to people who are experts but don't really understand how to develop python code.

So for example, I believe it will work like this:

"Please show me the on-axis frequency response for the following speakers, Kef LS50, Arendal Sound 1961 Center and Bang & Olufsen Beolab 20"

This wouild render a typical comparison of the 3 speakers.

"Please zoom in on the bass region"

It would then render the bass region (it might get it wrong so you might just tell it the frequency range you want)

You would then continue on interacting in English while getting an instant rendering of the dataset.

 

[Berwhale]

So for example, I believe it will work like this:

Or like this

 

[OCA]

Finally, a machine not based on knowledge, experience and reasoning will be able to make a huge blog in which it will write what it does not understand, and comment on it.
Living people will have more time to listen to music

And for what you wrote about, no ChatGPT is needed at all.
Put tags and people themselves will find unnecessary and incomprehensible to them by a "simple" ordinary search.
Why write a whole sentence instead of several keywords if the result is expected to be a keyword search result?
And understand please, current "AI" (even if it's labeled as "generative") is not able to analyze raw data sets if people have not done the analysis of these data sets before.
You can use this "AI" to generate short content for long nonsensical articles in which 90% of the text is not relevant to the point. This is very convenient. But don't expect this "AI" will replace the decades of work in any field to gain a deep understanding of measurement techniques and results.

Google search is not what it used to be. A lot of websites and nearly all forums (other than reddit which is the reason you only find something you search in reddit lately) have closed their doors to Google bots crawling their servers. There are search engines like Kagi which still do a good job but they are not free. ChatGPT has an advantage in scanning the internet compared to search engines alone.

 

[tomtoo]

I remember when people laught about the first car, the first plane, the first home computer. They all told its worthless. Wish i would have the money, that all this is worth now.

 

[Syntactic]

If you want a subjective tool to tell you things that may or may not be true because it is trying to figure out the answer you want to hear, GPT is great. If you want a tool to assist in a "Science Review" then look elsewhere.

The idea of using a LLM tool to generate R code to easily manipulate data for someone who can't code is interesting, but there are a lot of steps between here and there, and it's more of an NLP problem than it is a GPT/LLM program problem.

GPT is like a very smart child that wants to please you and it is always going to consider lying if it thinks you will like the lie better than the truth. That's the problem with using it for any kind of data analysis. Once your results are mediated through GPT they are no longer trustworthy unless you independently verify them, in which case you might as well have just done that in the first place.

 

[Jeromeof]

If you want a subjective tool to tell you things that may or may not be true because it is trying to figure out the answer you want to hear, GPT is great. If you want a tool to assist in a "Science Review" then look elsewhere.

Totally agree - whiles it fun to see what text it might generate as far as audio science analysis it is the wrong type of Machine Learning.

The idea of using a LLM tool to generate R code to easily manipulate data for someone who can't code is interesting, but there are a lot of steps between here and there, and it's more of an NLP problem than it is a GPT/LLM program problem.

Yes, this is mostly what I think it will be able to do - steps within a notebook are usually small "baby" steps over the data of previous steps - this is the type of code I have found GPT4 can generate very well when presented with a clear English description of what you want it to do with that data.

GPT is like a very smart child that wants to please you and it is always going to consider lying if it thinks you will like the lie better than the truth. That's the problem with using it for any kind of data analysis. Once your results are mediated through GPT they are no longer trustworthy unless you independently verify them, in which case you might as well have just done that in the first place.

It would not be generating a text result - merely generating the code R/Python etc that you would previously need to write to do the analysis work.

I went looking for a video of a data scientist using a notebook to manipulate audio data and this seems like a good example (skip to about 11 minutes ) - this guy needs to write some 'python' code (fairly complex code but only a few lines per step) manipulation of the audio recording he is visualising.

So I believe this is exactly the type of code that chatGPT can generate based on you asking it to manipulate the dataset. So its still a complex problem, with lots of very specific knowledge but this is the type of code that I have found in the past few weeks that it can generate.

 

[Jeromeof]

And as a quick experiment (and this isn't the plugin / interpreter just GPT4) - so only generates the code for a step, similar to YouTube video from a simple English instruction:

 

[buz]

I just skimmed the Microsoft Gpt4 Eval paper. Scary. EAGERLY awaiting plug in access now.

This morning I threw a tough case interview at regular chatgpt4 which it handled admirably. It made a minor mistake which it corrected upon challenging it. A candidate with those answers would most likely be hired.

And unlike Gpt3.5 hallucinations seem rare now: I was intending to use it for brainstorming (which was IMHO the best use for 3.5) this week and ended up using a fair bit of the output verbatim. Could I have written better prose? Possibly, but I would have wasted hours.

 

[OCA]

Apparently it can code in Octave as well:

 

[ROOSKIE]

If you want a subjective tool to tell you things that may or may not be true because it is trying to figure out the answer you want to hear, GPT is great. If you want a tool to assist in a "Science Review" then look elsewhere.

This.
Pretty much sums up a staggeringly vast amount of
child-parent,
student-educator,
employee-employer,
'friend'-'friend',
lover-loved,
relationships.
We have created this in our own image.
Well, it is going to be interesting. I'm already making the popcorn.

I just asked it 'who is the fairest of them all'?
Answer was, Bose.

 

[Roddy]

And as a quick experiment (and this isn't the plugin / interpreter just GPT4) - so only generates the code for a step, similar to YouTube video from a simple English instruction:View attachment 274645

Just to clarify, GPT4 uses the black logo, not the green one, and currently you have to pay for it. Although I don't know if the code-writing ability changed much from 3.5 to 4. It would probably explain itself more succinctly.