GPT4 / CHATGPT Plugins for Speaker measurement analysis and comparison

[Gruesome]

So you use its language model/'interpretative skills' only when interpreting what is being asked of it, not in embellishing the results? So it would just be a faster way of doing descriptive statistics on the database, essentially translating english into SQL?

 

[Jeromeof]

So you use its language model/'interpretative skills' only when interpreting what is being asked of it, not in embellishing the results? So it would just be a faster way of doing descriptive statistics on the database, essentially translating english into SQL?

Exactly and for me the the ability to check the “sql” is vital - you can also ask it to refine the “sql”

 

[Jeromeof]

This video might give people an idea of some of the abilities of the “code interpreter” ( which is the key plugin IMO)

 

[onion]

I did ask it to compare the 8341a and 8351b based on Amir's reviews, and I think it did a bit of hallucinating on freq response

 

[Jeromeof]

Interesting you seem to have a different browser plugin to me? Is that one of the third party plugins? The official one looks a bit different but is very buggy / slow but sort of tells you if it fails to pull back real content: e..g

But I did noticed it struggles to "search" for pages on ASR - one of the reasons I picked 2 very recent reviews for the previous post is that they are on the Home Screen and I noticed it didn't have a problem getting the reviews / content from that page, but maybe the WebPilot works differently.

Its super early days with this stuff - I have been separately investigating (for my day job) the smaller models which can run standalone on a MacBook / PC but can be given lots of local content - in some ways they are better as they can be focused only on what you give them, but their training was limited are less "smart" than chatGPT.

 

[onion]

Yes - it's a third party plugin called WebPilot. The built-in browser plugin seems to work better on some websites but not others (like ASR)

 

[Gruesome]

I did ask it to compare the 8341a and 8351b based on Amir's reviews, and I think it did a bit of hallucinating on freq response

View attachment 287597

Very interesting. There is no 96dB distortion measurement for the 8341 posted in Amir's review. So this thing is making it up out of thin air. It's 'playing review'...
Which is not really any worse than people making up explanations for why cables sound differently. But also not any better....

There is a 93 dB SPL plot for a Polk speaker mixed in with other plots for the Genelec 8341 at https://docs.google.com/spreadsheet...b8MxK555SbaYHdZVRHD6QA9BaMT35989P4gB/pubhtml# that is labeled 'serious issue at 400Hz', but I doubt that the program is picking up on that, since it's not from the primary review, but linked from a post by MZKM. It would be wrong anyway, since it's for a different speaker. I'm not sure why MZKM put it there.

 

[onion]

Very interesting. There is no 96dB distortion measurement for the 8341 posted in Amir's review. So this thing is making it up out of thin air. It's 'playing review'...
Which is not really any worse than people making up explanations for why cables sound differently. But also not any better....

There is a 93 dB SPL plot for a Polk speaker mixed in with other plots for the Genelec 8341 at https://docs.google.com/spreadsheet...b8MxK555SbaYHdZVRHD6QA9BaMT35989P4gB/pubhtml# that is labeled 'serious issue at 400Hz', but I doubt that the program is picking up on that, since it's not from the primary review, but linked from a post by MZKM. It would be wrong anyway, since it's for a different speaker. I'm not sure why MZKM put it there.

I think the code interpreter plugin which is in alpha will do better. I saw a YouTube vid yesterday where the YouTuber altered two cells in an excel spreadsheet with tens of thousands of point of data. Code interpreted correctly picked up the values in those cells as being statistical outliers. Really impressive

 

[Jeromeof]

Very interesting. There is no 96dB distortion measurement for the 8341 posted in Amir's review. So this thing is making it up out of thin air. It's 'playing review'...
Which is not really any worse than people making up explanations for why cables sound differently. But also not any better....

There is a 93 dB SPL plot for a Polk speaker mixed in with other plots for the Genelec 8341 at https://docs.google.com/spreadsheet...b8MxK555SbaYHdZVRHD6QA9BaMT35989P4gB/pubhtml# that is labeled 'serious issue at 400Hz', but I doubt that the program is picking up on that, since it's not from the primary review, but linked from a post by MZKM. It would be wrong anyway, since it's for a different speaker. I'm not sure why MZKM put it there.

I think possible with that WebPilot plugin if it can't find something just returns nothing and ChatGPT just makes up stuff like the default chat window, the official Browser Plugin (and definitely the Code Interpreter plugin's are different - but both are buggy)

 

[Gruesome]

Don't take my post the wrong way. I understand this field is very much in development, and whatever is available for public consumption right now might have really been 'knitted with a hot needle', as they say in the old country. I'm sure given time and effort, and access to the right tools, one can put something together that is maybe a little less user friendly and a whole lot more accurate.

 

[Jeromeof]

I will post it here (rather than a new thread) as Erin just posted a new video using ChatGPT to recommend some speakers:

Erin rightly points out that just asking ChatGPT directly will give you results based on every "source" that they used to train ChatGPT (roughly 2 years ago) so will be more like a popularity contest than actual good recommendations.

Hopefully OpenAI will open up their plugins as I know @pierre has done a good job already of exposing his spinorama data as a ChatGPT plugin so that more advanced comparisons (using actual data science) can hopefully be available soon.

 

[Barrelhouse Solly]

From my one voluntary interaction with ChatGPT, I was impressed by how its programmers/data feeders/care takers had integrated hard data into its repertoire (I asked it about yield strains for different grades of steel bolts), but also by how easily it confused itself with the meaning it had assigned to numbers/variables. Like, in one moment it would pick an arbitrary number as an example for yield stress, to illustrate the concept, and two sentences later it was mistaking that number for the yield stress in a computation. Impressive was again that it correctly parsed my comment pointing out this mistake, and corrected its calculations and spit out a corrected table of yield strains.

But unless a particular instance of this program has been configured to put a very high emphasis on not making stuff up, and a much lower emphasis on spitting results out to keep the 'conversation' going, I don't see how you can use a tool like this without cross checking every bit of output. From my interaction with a particular instance of one of these programs, I would fully expect it to make up correlations between say speaker sensitivity and predicted ranking if it 'got the impression' that was something you might like to hear.

As a former heat treater I'm happy to hear about yield point data. Ask it about phase diagrams for various alloys. <g>

 

[Barrelhouse Solly]

I'm a retired business programmer who's also been a heat treater. I think a lot of the criticism of AI that I see is based on fear. Another factor is expecting it to be mature. AI has made amazing strides but is still in its infancy. In the '80s it was still trying to solve the problem of recognizing shapes. OTOH, in "Godel, Escher, Bach" Hofstadter had a quiz asking the reader to differentiate between real artspeak and machine generated artspeak. Even in the '80s it was hard to tell the difference.

It seems to me that code generation is something AI could be good at. Most programming involves using off the shelf components and bolting them together. Knuth's books on algorithms, which are literally prehistoric in internet years, have almost everything that's in use today. Other than learning different languages, which can take a while, about 3 courses will teach anyone enough to write code that can handle most tasks that don't involve inventing something. A 4 year degree involves learning more than many people will ever use at work. Learning to model the process you're trying to automate is much more difficult than writing the code. You can almost always find examples of code for solving any problem you need to solve.

I don't think AI will replace us or enslave us. It's a tool. Unless something unexpected happens, if worst comes to worst, humans can pull the plug. I always wondered why the space ship in "2001" didn't have a manual override that depended on a human controlled switch and a backup system with "dumb" automation.

 

[onion]

This story does have me wondering: Crazy killer AI sim

Reason is that AI may give the edge in warfare to early adopters so it's quite likely we will see an arms race in AI-powered weapons. But as with all things, early adopters are taking the biggest risks as more is unknown at that point. An AI that attacks its drone operator and the communication network supporting him! Crazy!!

 

[thegeton]

"...It takes your input and iteratively applies probabilistic outcomes to it until it reaches a certain threshold of 'certainty' and then barfs out an answer..."

Ha!

This could also describe any mature product development process in use today.

 

[jolathonlogan]

Avoid relying solely on ChatGPT for recommendations, as it pulls from sources that are two years old, potentially skewing results towards popularity rather than quality. Erin's insight highlights the importance of seeking out multiple sources to make informed decisions.

 

[ferrellms]

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.

For 20 bucks a month you can get GPT-4 with plugins to extract data from websites (and tons of other computations). GPT-4 is a lot better that ChatGPT and GPT 3.5. I'd wager that any info you get from it will be way more accurate than the typical post in these forums'

 

[ferrellms]

Not until someone connects it to something that it controls but doesn't understand the consequences of its actions when it makes a mistake. Yes, Chat-GPT does make mistakes. It 'hallucinates' facts because it is not a true intelligence. It takes your input and iteratively applies probabilistic outcomes to it until it reaches a certain threshold of 'certainty' and then barfs out an answer.

Applying what is essentially a 'novelty' to scientific research, or to mathematics or code writing—when you have little or absolutely no knowledge of the subject matter—will quickly lead one to erroneous conclusions. Chat-GPT has already been tasked with programming and ≈50% or so of the code it wrote was insecure. Trusting this 'novelty' to generate something which one has no knowledge of how it is supposed to work would be a bad idea because if that code was used for something that needed to be threat-resistant or failsafe, it could turn out to be a bad day indeed.

This is completely wrong about how this technology works - Stephen Wolfram has a good explanation, available on the web .

 

[ferrellms]

From the link:

"Language models have become more capable and more widely deployed, but we do not understand how they work."

How can we not know how they work? (The programmers, not my somewhat inept neuron?)

People do know how they work! They invented them! But in any particular case, in a large LLM the inner workings are so complex they cannot really be understood in detail. Stephen Wolfram has a good explanation on his websites.

What is really interesting is the amount to which they seem to gain "emergent" human-like reasoning capabilities as they get larger. There is a lot of dispute on this one.

 

[ferrellms]

Avoid relying solely on ChatGPT for recommendations, as it pulls from sources that are two years old, potentially skewing results towards popularity rather than quality. Erin's insight highlights the importance of seeking out multiple sources to make informed decisions.

Prompt - "why are cardioid speakers like kii 3 good for use in a room"

GPT-3.5 (LLM used by ChatGPT I believe) -
'I'm sorry, but I'm not familiar with "kii 3." Could you please provide more context or clarify your question?'

GPT 4 with web-browsing plugin -
"Cardioid speakers like the Kii Three are often preferred for use in rooms due to their unique sound dispersion pattern. Here are some reasons why they are considered better:

1. Reduced Room Reflections: Cardioid speakers are designed to emit sound in a heart-shaped pattern, which is more directed towards the front and less towards the sides and back. This reduces the amount of sound that bounces off the walls, ceiling, and floor, leading to fewer room reflections and a cleaner sound.
2. Improved Sound Clarity: Because of the reduced room reflections, the listener hears more of the direct sound from the speakers and less of the reflected sound. This can improve the clarity and detail of the sound, making it easier to hear subtle elements in the music.
3. Better Bass Control: Some cardioid speakers, like the Kii Three, also have cardioid dispersion in the bass frequencies. This can help to reduce the impact of room modes on the bass response, leading to tighter, more controlled bass.
4. Increased Listening Sweet Spot: The focused sound dispersion pattern of cardioid speakers can also create a larger "sweet spot" where the sound is well balanced. This can make it easier to find a good listening position in the room.
5. Less Disturbance to Neighbors: The reduced sound emission towards the back can also make cardioid speakers less likely to disturb people in adjacent rooms, which can be a benefit in apartments or other shared living situations.

It's important to note that while cardioid speakers have these potential advantages, the actual performance can depend on many factors, including the specific design of the speakers, the acoustics of the room, and the placement of the speakers within the room."

No hallucinations, just more and better information than most ASR posts to consider when making a choice of speaker type.