No. AI models are pre-trained, they do not learn on the fly. They are hoping to discover General Artificial Intelligence, which is what you are describing. The problem is that they don’t even understand exactly how training even works. While engineers understand the overall architecture, the specific “reasoning” or decision-making pathways within the model are too complex to fully interpret, leading to a gap between how it works and why it makes a particular decision.
My assumption wasn’t that they learned on the fly, it was that they were trained on previous interactions. Eg the team developing them would use data collected from interaction with model v3 to train model v4. Seems like juicy relevant data they wouldn’t even have to go steal and sort
That’s true to an extend, but the interactions are only useful for training if you can mark it as good / bad etc (which is why sometimes apps will ask you if they were useful). But the ‘best’ training data like professional programming etc is usually sold at a premium tier with a promise not to use your data for training (since corporations don’t want their secrets getting out).
First: that’s wrong, every big LLM uses some data cleaned/synthesized by previous LLMs. You can’t solely train on such data without degradation, but that’s not the claim.
Second: AI providers very explicitly use user data for training, both prompts and response feedback. There’s a reason businesses pay extra to NOT have their data used for training.
I mean - yeah, it is? This is a well-researched part of the data pipelines for any big model. Some companies even got into trouble because their models identified as other models, whose outputs they were trained on.
It seems you have a specific bone to pick that you attribute to such training, but it’s just such a weird approach to deny pretty broadly understood results…
No, it doesn’t. Unless you can show me a paper detailing that literally any amount of synthetic data increases hallucinations, I’ll assume you simply don’t understand what you’re talking about.
No. AI models are pre-trained, they do not learn on the fly. They are hoping to discover General Artificial Intelligence, which is what you are describing. The problem is that they don’t even understand exactly how training even works. While engineers understand the overall architecture, the specific “reasoning” or decision-making pathways within the model are too complex to fully interpret, leading to a gap between how it works and why it makes a particular decision.
My assumption wasn’t that they learned on the fly, it was that they were trained on previous interactions. Eg the team developing them would use data collected from interaction with model v3 to train model v4. Seems like juicy relevant data they wouldn’t even have to go steal and sort
That’s true to an extend, but the interactions are only useful for training if you can mark it as good / bad etc (which is why sometimes apps will ask you if they were useful). But the ‘best’ training data like professional programming etc is usually sold at a premium tier with a promise not to use your data for training (since corporations don’t want their secrets getting out).
You can’t train ai on ai output. It causes degradation on the newly trained model.
First: that’s wrong, every big LLM uses some data cleaned/synthesized by previous LLMs. You can’t solely train on such data without degradation, but that’s not the claim.
Second: AI providers very explicitly use user data for training, both prompts and response feedback. There’s a reason businesses pay extra to NOT have their data used for training.
yep ai training on ai is totally making things better…
I mean - yeah, it is? This is a well-researched part of the data pipelines for any big model. Some companies even got into trouble because their models identified as other models, whose outputs they were trained on.
It seems you have a specific bone to pick that you attribute to such training, but it’s just such a weird approach to deny pretty broadly understood results…
i’m aware i also know it increases hallucinations
No, it doesn’t. Unless you can show me a paper detailing that literally any amount of synthetic data increases hallucinations, I’ll assume you simply don’t understand what you’re talking about.
what paper? no one in industry is gonna give you this shit, it’s literal gold
academics are still arguing about it but save this and we can revisit in 6 months for a fat i told you so if you still care
ai is dead as shit for anything that matters until this issue is fixed
but at least we can enjoy soulless art while we wait for the acceleration