-
Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
Using AI In Audio Debates & GR Research [Video]
- Thread starter amirm
- Start date
OP
- Thread Starter
- #22
You can ask it if it can truly be helpful in settling debates about audio. 
Whatever you do with AI, make sure your base prompt is setup correctly. This will root out 99% of the subjective audio nonsense right away. You can just ask it to help you setup a proper base prompt.
I asked GPT 5.5 & Claude Sonnet 4.6 (not even the most advanced model) the exact same question and both gave a very nuanced and above all correct answer from the get go.
I asked GPT 5.5 & Claude Sonnet 4.6 (not even the most advanced model) the exact same question and both gave a very nuanced and above all correct answer from the get go.
Fidji
Addicted to Fun and Learning
- Joined
- Feb 17, 2019
- Messages
- 599
- Likes
- 1,168
I think we can all meanwhile agree that Danny is just lost in modern world. You can almost feel sorry for him as he keeps getting into emotional spiral that just forces him to go and shout more and more nonsense,
Artificial intelligence and natural ignorance is deadly combination.
Already looking forward to Amir’s repsonse video to Danny’s response to this Amir’s video ;-).
Artificial intelligence and natural ignorance is deadly combination.
Already looking forward to Amir’s repsonse video to Danny’s response to this Amir’s video ;-).
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.
Last edited:
DatSpeakerTho
Active Member
- Joined
- Apr 29, 2026
- Messages
- 180
- Likes
- 346
As a former user of the PE tech talk forum, I kinda have to laugh at the "no one answers the question in this thread" comment. Yeah, that was most threads there lol.
lets do it! I will give it as vague questions as possibleYou can ask it if it can truly be helpful in settling debates about audio.
Last edited:
milanjeromos
Member
- Joined
- Jan 4, 2026
- Messages
- 13
- Likes
- 8
While the general SOTA thinking models will continue to just be statistically models to generate text responses that the extend the 'context' of the conversation - so can easily be made to mostly say whatever opinion they are coerced to generate. I have been studying how for example the coding tools work so well and I feel we should see allot more of that capability in other domains like audio. There are really 3 things they do very well that allow these models to generate 'consistent' results ( Note: the early versions coding agents were like everything with LLM chat and just would make up code that might work but also might be terrible ). But, the recent 'tricks' are really outside the LLM itself - it about harness engineering, context engineering and having the thinking / planning phase pull in deterministic data.
Harness engineering - is really about putting specific 'guardrails' around the conversation to stay on topic and focused on a specific result - typically this is in the system prompt ( that you can't really change with the general models - as people would abuse these ).
Context Engineering - is really about keeping only relevant text in the chat history ( which is really the context of what is generated next ) and to guard against the model loosing the purpose of the conversation - there is recency bias to how the context is used generate the next response ( which often means the initial question and purpose of a chat is completely lost ) - so there are now clever techniques in the tools ( like Claude Code ) to spin off sub-agents to do the 'dirty' work while leaving the main context fairly clean and concise.
Skills and Tools - with decent harness engineering you can force the thinking phase of a chat to pull in focused relevant deterministic answers - especially if you build a glossary or in coding terms a 'graph' of the relationship of this part of the code. The general chat models will (as @amirm showed in the video will pull in sources from everywhere ) - but if you control the system prompt - you can force specific focused tools calls ( and skills ) to be used to only pull in the text you believe to be objectively correct.
The nice thing with these techniques is they can all be applied to much smaller models and give those models focused specialised knowledge.
Harness engineering - is really about putting specific 'guardrails' around the conversation to stay on topic and focused on a specific result - typically this is in the system prompt ( that you can't really change with the general models - as people would abuse these ).
Context Engineering - is really about keeping only relevant text in the chat history ( which is really the context of what is generated next ) and to guard against the model loosing the purpose of the conversation - there is recency bias to how the context is used generate the next response ( which often means the initial question and purpose of a chat is completely lost ) - so there are now clever techniques in the tools ( like Claude Code ) to spin off sub-agents to do the 'dirty' work while leaving the main context fairly clean and concise.
Skills and Tools - with decent harness engineering you can force the thinking phase of a chat to pull in focused relevant deterministic answers - especially if you build a glossary or in coding terms a 'graph' of the relationship of this part of the code. The general chat models will (as @amirm showed in the video will pull in sources from everywhere ) - but if you control the system prompt - you can force specific focused tools calls ( and skills ) to be used to only pull in the text you believe to be objectively correct.
The nice thing with these techniques is they can all be applied to much smaller models and give those models focused specialised knowledge.
audiofooled
Addicted to Fun and Learning
- Joined
- Apr 1, 2021
- Messages
- 742
- Likes
- 790
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.
Wow, is interesting that Danny chose AI to discuss speaker electrical phase as he never shows any in his measurements! Does he imagine what the phase is? For that matter, he has a double standard on other company speakers and his own. If you go look at his own design measurements, they are all over the place in terms of impedance. His most popular is the X-LS Encore and Amir measured it…
If you trust Danny’s assertion about imaging, this one must really suck around 1000 Hz. Look at his $5000 NX-Otica. Here is Danny’s own impedance measure…
He allows the impedance to rise significantly in the midrange. Note the lack of a phase trace compared to more Amir’s standard measure above.
If you trust Danny’s assertion about imaging, this one must really suck around 1000 Hz. Look at his $5000 NX-Otica. Here is Danny’s own impedance measure…
He allows the impedance to rise significantly in the midrange. Note the lack of a phase trace compared to more Amir’s standard measure above.
Watch this from 6:25 on:
Reminiscing about the late great speaker designer Peter Snell
Peter proved that using very expensive crossover components in his Type A speaker did not make it sound any different.
Reminiscing about the late great speaker designer Peter Snell
Peter proved that using very expensive crossover components in his Type A speaker did not make it sound any different.
I think we need to start by laying the groundwork and clearly defining AI.
AI does not reason with intentions, it lacks consciousness and does not understand.
AI is merely a statistical learning system, or at best, a cognitive amplifier. Prompting, if poorly executed, can induce an "influenced" response, and this is how we often fall into confirmation bias.
Danny often falls into confirmation bias because he probably doesn't use AI wisely...I therefore presume that the way in which he prompts is not well thought out...
AI does not reason with intentions, it lacks consciousness and does not understand.
AI is merely a statistical learning system, or at best, a cognitive amplifier. Prompting, if poorly executed, can induce an "influenced" response, and this is how we often fall into confirmation bias.
Danny often falls into confirmation bias because he probably doesn't use AI wisely...I therefore presume that the way in which he prompts is not well thought out...
A good clarifying video Amir.
One thing I wonder about. What AI engine (Google AI?) is generating a summary when you do Google searches these days? And if we quote from it what should we state?
So it follows the policy regarding use of AI at ASR, that is.
This AI type of summary, which I assume pops up in most Google searches these days:
One thing I wonder about. What AI engine (Google AI?) is generating a summary when you do Google searches these days? And if we quote from it what should we state?
So it follows the policy regarding use of AI at ASR, that is.
This AI type of summary, which I assume pops up in most Google searches these days:
Last edited:
PristineSound
Major Contributor
Great video and much needed. So many people don't understand how to use AI. And so many simply just go with the results without checking.
You did let him off the hook on all other lies and misinformation. I suspect Danny will claim victory on the topics you didn't bother running back. But no surprises with Danny.
You did let him off the hook on all other lies and misinformation. I suspect Danny will claim victory on the topics you didn't bother running back. But no surprises with Danny.
Buckchester
Active Member
- Joined
- Sep 30, 2021
- Messages
- 278
- Likes
- 520
He responded to me.
I guess I’m in the minority. I thought this response was a bit muddled and tries to have its cake and eat it too.
It’s fine if a little bit laborious to use AI to show that AI cannot be trusted for this sort of thing, but much of the video essay was about how you need to have neutral prompts to get quality responses.
I’m no expert on leading edge AI but current mainstream models simply shouldn’t be used this way. It’s necessary but not sufficient to use neutral prompts and disclose what the prompts are. AI has a memory, it knows what you’ve asked before and what your preferences and predilections are. We don’t know Danny Richie’s prompt but it would not surprise me if it was worded neutrally. It could easily have been colored by previous prompts unrelated to the Ascend speaker.
Even if we could establish best practices that maximizes objectivity, argument-by-AI can and should be dismissed out of hand. We shouldn’t have to process a wall of text for every little thing. AI will sound authoritative even when there’s no authority to be had.
It also pulls the same BS that people do, such as citing references in support of contentions that the reference does not support (as Amir pointed out with the ASR reference). But at least when people do that, it reflects on their lack of honesty or understanding. When AI does it, it’s just the mysterious workings of a marvelous statistical correlation device. People need to be accountable for their own arguments and reasoning.
It’s fine if a little bit laborious to use AI to show that AI cannot be trusted for this sort of thing, but much of the video essay was about how you need to have neutral prompts to get quality responses.
I’m no expert on leading edge AI but current mainstream models simply shouldn’t be used this way. It’s necessary but not sufficient to use neutral prompts and disclose what the prompts are. AI has a memory, it knows what you’ve asked before and what your preferences and predilections are. We don’t know Danny Richie’s prompt but it would not surprise me if it was worded neutrally. It could easily have been colored by previous prompts unrelated to the Ascend speaker.
Even if we could establish best practices that maximizes objectivity, argument-by-AI can and should be dismissed out of hand. We shouldn’t have to process a wall of text for every little thing. AI will sound authoritative even when there’s no authority to be had.
It also pulls the same BS that people do, such as citing references in support of contentions that the reference does not support (as Amir pointed out with the ASR reference). But at least when people do that, it reflects on their lack of honesty or understanding. When AI does it, it’s just the mysterious workings of a marvelous statistical correlation device. People need to be accountable for their own arguments and reasoning.
Similar threads
- Poll
