Using AI In Audio Debates & GR Research [Video]

[PristineSound]

He responded to me.

View attachment 532916

I suspect there may be another YT response from Danny soon.

I think at this point, it's just a merry-go-round. I don't think Amir should respond again, but if he does, he should be a very short video, like 10 mins, restating all the facts.

 

[Brian Hall]

AI is too polite to tell you that your question is stupid, let alone throw a direct insult at you, while Danny is certainly arrogant enough to do so. He could try and ask AI weather this kind of behavior automatically makes him win the debate with grown ups.

View attachment 532831

I asked a locally installed model that question. It is one of the free open models released by Alibaba:

 

[Somafunk]

My best case for using AI
View attachment 532860

I now want a levitating table, imagine if it followed you around as you go about your daily chores?, that’d be so handy

 

[Brian Hall]

There are different types of AI models. The larger "thinking/reasoning" models are more likely to get the correct answer. The non reasoning (smaller / faster) models go straight to the obvious answer.

The Gemma4 open and free models released by Google come in different versions. The small gemma4:e4b model failed the trick question and immediately recommended walking. (The e4b indicates effective parameters used by the model)

The qwen3.6 model is labeled 35b-a3b. 35 billion parameters. It is a mixture of experts model so only 3 billion parameters are used by the "expert" it decides to use to answer the question.

The larger gemma4:e31b model is designed for "thinking/reasoning". This is a 31 billion parameter model that doesn't fit in my 16gb of vram so it runs a lot slower (46 seconds to respond). I clicked on the option to "show the thinking":

 

[amirm]

One thing I wonder about. What AI engine (Google AI?) is generating a summary when you do Google searches these days?

It is called Gemini from Google but it is running in fast mode, designed to answer quickly. You can change it to thinking mode which slows it down and provide better answers. I find Chatgpt paid version to be superior however.

 

[Rayman30]

I asked ChatGPT if I would notice a difference upgrading my Topping D30 Pro/A30Pro stack to the DX9 Discrete.

It said...

• slightly deeper stage
• more separation
• more effortless dynamics
• cleaner imaging

 

[voodooless]

I asked ChatGPT if I would notice a difference upgrading my Topping D30 Pro/A30Pro stack to the DX9 Discrete.

It said...

• slightly deeper stage
• more separation
• more effortless dynamics
• cleaner imaging

Please see:

Using AI In Audio Debates & GR Research [Video]

I have access to gpt 5.5 pro model, let me know if you want me to run something through it. I also work for LLM company as a scientist.

www.audiosciencereview.com

I asked it the exact same question and it said (summarized):

Probably no meaningful audible upgrade in normal listening

… it’s like Jackal & Hide inside of those neural nets…

 

[Brian Hall]

I asked ChatGPT if I would notice a difference upgrading my Topping D30 Pro/A30Pro stack to the DX9 Discrete.

It said...

• slightly deeper stage
• more separation
• more effortless dynamics
• cleaner imaging

I asked the new Grok 4.3 beta that question. It doesn't think there would be any audible difference and referenced information from here (ASR). It talked about diminishing returns with more expensive gear.

 

[xanalog]

You Go, @amirm!
Where have you been for the last few months? While I have been trying to drill some of your video's AI discussions to my mate?

I have come to the realization that you can lead a horse [no dis] to the water, but you can't make them drink from it... until they are ready.

 

[xanalog]

Please can you also comply with the ASR policy - and at least provide the prompt you used.

Easy to request, much harder to accomplish.
Especially when you have to rinse/flush/repeat your real question (prompt) umpteen times, for due diligence.
By which time, you've become more of an 'expert' than it.

 

[Gremlins]

Man, that Danny guy is really a joke ...

 

[testp]

good video.

probl. mentioned, but also a good thing in prompting is to give AI an actor/role, so from start it knows who/what it is. if im asking about weather, you are an weather expert etc.

 

[Gary_G]

Please see:

Using AI In Audio Debates & GR Research [Video]

I have access to gpt 5.5 pro model, let me know if you want me to run something through it. I also work for LLM company as a scientist.

www.audiosciencereview.com

I asked it the exact same question and it said (summarized):

… it’s like Jackal & Hide inside of those neural nets…

I asked my wife, "Does not compute." was the return.

ASR. Where NI - Natural Intelligence rules.

 

[BobbyTimmons]

Please can you also comply with the ASR policy - and at least provide the prompt you used.


EDIT - Also bear in mind it will have taken the whole of Amir's video as part of its prompt. For example - in your post the AI states categorically (bolding it as a key statement):


It doesn't know that. Nor do we. None of us has seen Richie’s prompt. It took the implication of that being possible from Amir's video (or from your own prompt, because we've not seen that either) and then stated it as a fact.

Because of this, the rest of the response becomes equally suspect. In fact, a lot of it is just restating what Amir stated in the video, and flattering Amir, thus illustrating its tendency towards sycophancy.

The prompt I used was downloading the video and asking what it thinks honestly.

And you can click to expand its internal monologue as it was crafting the reply to the video.

Then it posted the answer I posted an excerpt from.

My next prompt was.

And it responded with what I already posted an excerpt from.

Then I just uploaded a screenshot of Amir's response to the excerpts.

And it responded.

You could argue the video itself as acting like a leading question and I should have included both Amir and Danny's videos in the original prompt. Although the state of the art models are not just a reflection of the prompts. They do have preferences and different weightings for different kinds of sources.

 

[haris525]

There are different types of AI models. The larger "thinking/reasoning" models are more likely to get the correct answer. The non reasoning (smaller / faster) models go straight to the obvious answer.

The Gemma4 open and free models released by Google come in different versions. The small gemma4:e4b model failed the trick question and immediately recommended walking. (The e4b indicates effective parameters used by the model)

The qwen3.6 model is labeled 35b-a3b. 35 billion parameters. It is a mixture of experts model so only 3 billion parameters are used by the "expert" it decides to use to answer the question.

The larger gemma4:e31b model is designed for "thinking/reasoning". This is a 31 billion parameter model that doesn't fit in my 16gb of vram so it runs a lot slower (46 seconds to respond). I clicked on the option to "show the thinking":

View attachment 532939

The question itself is so ill-posed, but it is what it is. This is with 4b, I see people do these type of things all the time at work. Be clear.

 

[xanalog]

The question itself is so ill-posed, but it is what it is. This is with 4b, I see people do these type of things all the time at work. Be clear.

4. Identifying the Intent:

• Intent D (The Stinky Car): I just wanted to go to the carwash to pick-up a

 

[AaronJ]

He responded to me.

View attachment 532916

This response was predictable. But addressing him directly serves little purpose because he is never going to be convinced. He needs to be ignored, not addressed.

 

[CD2]

I just want to underline the sources point. LLMs make sources in exactly the same way they make text: what would be statistically likely to see in that place. That does not mean they are "getting" stuff from those places. They also should not be relied on for explanations of their own workings. Everything is what kind of text would a reader expect to see. People are being played.

When I first saw people using LLMs for audio I had to laugh because yeah, it would be like asking an LLM for tne truth about UFOs or Shakespeare or the Federal Reserve. There is voluminous crankery out there because cranks have a lot of spare time. All that goes into the training set.

 

[amirm]

On the disconnected AC cable, after I did my video about that, folks complained about current there just as well. So I repeated my test with an amp using the AC cord, and then measuring the impact on the AC cable:

The spectrum of the interference changed some but not the overall picture (blue). The amp itself (red) was pumping out far more than that than the speaker cable picked up.

 

[dualazmak]

My best case for using AI
View attachment 532860

Sorry for my being out of the scope of this thread, but...
Really nice simple setup and wonderful listening room as well as beautiful garden you have.

Let me ask which model is your Accuphase integrated amp?
I have my treasure Accuphase E-460 in my multichannel multi-amplifier setup; if you would be interested, please visit #931 and #1,009 on my project thread.