As someone who has regularly trained language models, fine tuned and used LLM's including GPT3 almost daily, I think I can comment.
1. When you prompt the algorithm to "write an essay on" you won't get much success, without showing it some examples first and with a bit of breakdown at the paragraph level. Finding an efficient way of breaking down what "analysis" is and feeding it into the model in a way it understands (few - shot learning) will help it learn what you want.
2. In fact, not just analysis, anything else that fits the marking criteria - quotes form the book, clarity of language. There is an emerging field of AI engineering called "prompt engineering" geared towards how to get these generative models to produce what you want.
3. It is possible to fine tune davinci to generate analytical paragraphs. With enough labelled data, writing an analytical paragraph is possible.
The rumor is that GPT-4 is coming out soon, maybe early next year. It'll be better, but my guess is that it will have some of the same pitfalls of GPT-3.
As a secondary math teacher homework is practically useless for assessing student learning. Between messaging solutions to your friends, solution apps like Photomath and Brainly, and even teacher editions of curricula Googleable you can’t trust what students submit as their work.
Having students do exercises in class has painfully but lucidly revealed what they know and can do; a big hill to climb, but bring your water and boots cause we’re hiking it.
The problem isn’t that AI has surpassed the analytical capability of a focused, articulate, engaged young reader -- it’s that it is very nearly surpassing what many instructors are forced by long experience to accept from the freshman students admitted to large state universities (at least in the US). That generated sample was boring and repetitive, but more robotically competent than many of the essays I have to grade every semester.
As someone who has regularly trained language models, fine tuned and used LLM's including GPT3 almost daily, I think I can comment.
1. When you prompt the algorithm to "write an essay on" you won't get much success, without showing it some examples first and with a bit of breakdown at the paragraph level. Finding an efficient way of breaking down what "analysis" is and feeding it into the model in a way it understands (few - shot learning) will help it learn what you want.
2. In fact, not just analysis, anything else that fits the marking criteria - quotes form the book, clarity of language. There is an emerging field of AI engineering called "prompt engineering" geared towards how to get these generative models to produce what you want.
3. It is possible to fine tune davinci to generate analytical paragraphs. With enough labelled data, writing an analytical paragraph is possible.
The rumor is that GPT-4 is coming out soon, maybe early next year. It'll be better, but my guess is that it will have some of the same pitfalls of GPT-3.
As a secondary math teacher homework is practically useless for assessing student learning. Between messaging solutions to your friends, solution apps like Photomath and Brainly, and even teacher editions of curricula Googleable you can’t trust what students submit as their work.
Having students do exercises in class has painfully but lucidly revealed what they know and can do; a big hill to climb, but bring your water and boots cause we’re hiking it.
The problem isn’t that AI has surpassed the analytical capability of a focused, articulate, engaged young reader -- it’s that it is very nearly surpassing what many instructors are forced by long experience to accept from the freshman students admitted to large state universities (at least in the US). That generated sample was boring and repetitive, but more robotically competent than many of the essays I have to grade every semester.