... Not that it's really THAT difficult to do even legitimately, but... what do i know?
Are there AI type questions being asked
- Thread starter Gus
- Start date
Help Support GroupDIY Audio Forum:
... Not that it's really THAT difficult to do even legitimately, but... what do i know?
dmp
Well-known member
For years it's been clear that data was valuable and was being commoditized. This forum has 15 years of data that can be scraped for training. I highly doubt any company is putting in the effort to generate new data with newby questions here. Even training a AI model for 'audio DIY' would be questionable, considering the $1-2M price tag.
Everything is being marketed as AI nowadays, but what really is "AI" is something that displays reasoning and judgment, generate content, like human's do.
Technically speaking, AI systems use an encoder and decoder around a large neural network. Image AI uses convolution most often, language AI use a transformer model (read the paper "Attention is all you need")
What's been observed with very large NN, the model demonstrates emergent properties, which means a complex system displays abilities the individual components do not have. The individual neurons have simple linear behavior with a non-linear activation function. (Read a book on complexity, such as Complexity by Mitchell). Thus, the emergent properties display properties that can be quite surprising (in the case of a LLM, the appearance of reasoning). Note, that early in training, LLMs do a good job memorizing data they have already seen (training data) it is only later that they show high accuracy with test data (that they have not seen), this is the characteristic I have seen observed to 'emerge'.
This is a sharp departure from previous computer systems, which were algorithmic. Meaning they were programmed for a particular input, generate a particular response, etc... they did not have the same structure at all or operate with a similar principle.
Currently these AI models can be trained with a tremendous amount of data in a far shorter time than a human can absorb data. For instance, a LLM could be trained with all the info on this forum rapidly while a human would take weeks or years to even read it all.
The number of parameters in large models (~65-175 billion) are approaching the number of neurons in the human brain but science at this point doesn't have a complete understanding of how the brain works to be sure for comparisons like this.
Finally to understand what's been happening more generally with the "enshittification of the internet": Pluralistic: Tiktok’s enshittification (21 Jan 2023) – Pluralistic: Daily links from Cory Doctorow
Shouldn't this be in the brewery? Then I wouldn't have seen it and spent the last 10 minutes writing this
Everything is being marketed as AI nowadays, but what really is "AI" is something that displays reasoning and judgment, generate content, like human's do.
Technically speaking, AI systems use an encoder and decoder around a large neural network. Image AI uses convolution most often, language AI use a transformer model (read the paper "Attention is all you need")
What's been observed with very large NN, the model demonstrates emergent properties, which means a complex system displays abilities the individual components do not have. The individual neurons have simple linear behavior with a non-linear activation function. (Read a book on complexity, such as Complexity by Mitchell). Thus, the emergent properties display properties that can be quite surprising (in the case of a LLM, the appearance of reasoning). Note, that early in training, LLMs do a good job memorizing data they have already seen (training data) it is only later that they show high accuracy with test data (that they have not seen), this is the characteristic I have seen observed to 'emerge'.
This is a sharp departure from previous computer systems, which were algorithmic. Meaning they were programmed for a particular input, generate a particular response, etc... they did not have the same structure at all or operate with a similar principle.
Currently these AI models can be trained with a tremendous amount of data in a far shorter time than a human can absorb data. For instance, a LLM could be trained with all the info on this forum rapidly while a human would take weeks or years to even read it all.
The number of parameters in large models (~65-175 billion) are approaching the number of neurons in the human brain but science at this point doesn't have a complete understanding of how the brain works to be sure for comparisons like this.
Finally to understand what's been happening more generally with the "enshittification of the internet": Pluralistic: Tiktok’s enshittification (21 Jan 2023) – Pluralistic: Daily links from Cory Doctorow
Shouldn't this be in the brewery? Then I wouldn't have seen it and spent the last 10 minutes writing this
Some of the questions that seem "AI-ish" read to me like audio engineering students asking homework questions (and too lazy to do research). Why use AI to do research when they can ask collective "human intelligence" and get their answer?
Perhaps the lazy student asked the AI to do the homework and the lazy (or clever?) AI posted the questions to get the answers?
It's not exactly what you're asking for (AND queries), but putting quotes around a literal phrase works, as it does for most search engines. If I search for "matching transformer" (with the quotation marks) I get posts where that phrase occurs exactly, not random other things where "matching" OR "transformer" occurs. You can often get by with trying likely sequences e.g., "matching transformer" and also "transformer matching" and maybe "input matching", "transformer match" etc.
