It’s neither obvious nor clear to me. Who wrote the rest of their training data, besides us oh-so-fallible humans? What percentage of the data does this non-human authorship constitute?
Modern multimodal LLMs are trained not just on text data, but also on images and video. In the finetuning stage for reasoning models, they are trained not to predict human reasoning, but to do reasoning themselves to accomplish a binary goal. Not sure about exact percentages but in terms of capabilities, they can recognize images very well and generate very realistic images and videos, to the point where many people can’t tell them apart.
That gives them more different abilities; I don’t think it constitutes a fundamental change to their way of thinking or that it makes them more intelligent. (It doesn’t improve their performance on text based problems significantly.) Because it is just doing the ~same type of “learning” on a different type of data. This doesn’t make them able to discuss say abiogenesis or philosophy with actual critical human-like thought. In these fields they are strictly imitating humans. As in, imagine you replaced all the learning data regarding abiogenesis with plausible-sounding but subtly wrong theories. The LLM would simply slavishly repeat these wrong theories, wouldn’t it?
It’s neither obvious nor clear to me. Who wrote the rest of their training data, besides us oh-so-fallible humans? What percentage of the data does this non-human authorship constitute?
Modern multimodal LLMs are trained not just on text data, but also on images and video. In the finetuning stage for reasoning models, they are trained not to predict human reasoning, but to do reasoning themselves to accomplish a binary goal. Not sure about exact percentages but in terms of capabilities, they can recognize images very well and generate very realistic images and videos, to the point where many people can’t tell them apart.
That gives them more different abilities; I don’t think it constitutes a fundamental change to their way of thinking or that it makes them more intelligent.
(It doesn’t improve their performance on text based problems significantly.)
Because it is just doing the ~same type of “learning” on a different type of data.
This doesn’t make them able to discuss say abiogenesis or philosophy with actual critical human-like thought. In these fields they are strictly imitating humans.
As in, imagine you replaced all the learning data regarding abiogenesis with plausible-sounding but subtly wrong theories. The LLM would simply slavishly repeat these wrong theories, wouldn’t it?