> I have two awesome guests on my podcast: Oliver Kim (who has a brilliant new paper on Taiwan’s land reform)
I'm really looking forward to this. We should really compare and contrast with African countries who still suffer from low agricultural productivity which lead to many more downstream problems. It's also important to compare modern Ethiopia with the Asian tigers when they were still focused on agriculture.
>Claude often makes up references, so this is a good signal of plagiarism.
>especially as it does not necessarily have the latest evidence
tools like perplexity.ai can guarantee citations to real articles. they can also take in web pages as input to "read" for answers. so i fear that the citation component of these assessments will likely not hold up against LLM improvements for long. and note that OpenAI is already prototyping similar functionality (https://openai.com/index/searchgpt-prototype/).
perhaps the best counter-measure would be to assign sufficiently new / obscure books that they are not already widely available online in a way the LLM can ingest. of course that channels scholarly skills in a very particular direction...
makes me wonder what the relevant work context will evolve into. i already feel like many academics are happy to cite sources without reading, and these tools will expedite summarizing a text and grabbing supportive quotes for citation.
RE: individuality
>let students showcase their unique brilliance!
i wonder how well prompting the LLM with particular personalities will substitute for this.
stellar! i'd love to hear about your experiences over the next 6-12 months as these tools evolve and you get to see how students adapt.
it came to mind in part from reading papers like this (https://github.com/tencent-ailab/persona-hub). the core idea being to direct LLMs to adopt a persona, then use that to generate new ("synthetic") data that ends up getting fed into training other LLMs. this kind of generation is becoming increasingly common, so my guess is the LLMs will become increasingly adept at adopting a persona over time (and for far more niche personas).
A note on the margins: the only only intervention that can drastically increase fertility in South Korea is a DPRK intervention. It would tackle binding constraints of peace and prosperity that are suppressing birth rates in all First World countries.
These are all great suggestions—I’ve been doing something similar in my classes, where I ask students to generate a summary of readings with GPT/Claude and then ask them to evaluate the summary. Project-based work is also good, like proposing or even implementing a novel research study.
Where it gets harder (at least in my experience) is assessments that scale to class sizes of 400-500 students (that also don’t over burden the limited teaching staff we might have on hand). Though some ways this is not a new problem, it just exposes the paucity of those assessment methods.
Either way I agree that it’s not something that can really be ignored!
"As educators, our role is to help students build AI-complementary skills." This may be the most depressing thing I've read on the internet in quite a long time.
This is an excellent approach to evolving education to respond to new tools and technology. LLMs are a resource and I want my kids to know how to use them to their advantage, without outsourcing all of the thinking and all of the work. I applaud!
These are always a joy to read, but this one in particular has helped prompt me to think more about how we could use LLMs in my workplace, especially for non-native speakers. I have one team member whom I am fairly sure is using it to put his contributions into better English, which has resulted in quite an idiosyncratic writing style.
> I have two awesome guests on my podcast: Oliver Kim (who has a brilliant new paper on Taiwan’s land reform)
I'm really looking forward to this. We should really compare and contrast with African countries who still suffer from low agricultural productivity which lead to many more downstream problems. It's also important to compare modern Ethiopia with the Asian tigers when they were still focused on agriculture.
great read! some thoughts:
RE: providing references
>Claude often makes up references, so this is a good signal of plagiarism.
>especially as it does not necessarily have the latest evidence
tools like perplexity.ai can guarantee citations to real articles. they can also take in web pages as input to "read" for answers. so i fear that the citation component of these assessments will likely not hold up against LLM improvements for long. and note that OpenAI is already prototyping similar functionality (https://openai.com/index/searchgpt-prototype/).
perhaps the best counter-measure would be to assign sufficiently new / obscure books that they are not already widely available online in a way the LLM can ingest. of course that channels scholarly skills in a very particular direction...
makes me wonder what the relevant work context will evolve into. i already feel like many academics are happy to cite sources without reading, and these tools will expedite summarizing a text and grabbing supportive quotes for citation.
RE: individuality
>let students showcase their unique brilliance!
i wonder how well prompting the LLM with particular personalities will substitute for this.
"you are Daron Acemoglu (https://economics.mit.edu/people/faculty/daron-acemoglu). research and answer the following question. be sure to frame it in your own unique voice and approach"
Ha! I already trialed impersonation! We think alike
stellar! i'd love to hear about your experiences over the next 6-12 months as these tools evolve and you get to see how students adapt.
it came to mind in part from reading papers like this (https://github.com/tencent-ailab/persona-hub). the core idea being to direct LLMs to adopt a persona, then use that to generate new ("synthetic") data that ends up getting fed into training other LLMs. this kind of generation is becoming increasingly common, so my guess is the LLMs will become increasingly adept at adopting a persona over time (and for far more niche personas).
Great article, thank you.
A note on the margins: the only only intervention that can drastically increase fertility in South Korea is a DPRK intervention. It would tackle binding constraints of peace and prosperity that are suppressing birth rates in all First World countries.
These are all great suggestions—I’ve been doing something similar in my classes, where I ask students to generate a summary of readings with GPT/Claude and then ask them to evaluate the summary. Project-based work is also good, like proposing or even implementing a novel research study.
Where it gets harder (at least in my experience) is assessments that scale to class sizes of 400-500 students (that also don’t over burden the limited teaching staff we might have on hand). Though some ways this is not a new problem, it just exposes the paucity of those assessment methods.
Either way I agree that it’s not something that can really be ignored!
"As educators, our role is to help students build AI-complementary skills." This may be the most depressing thing I've read on the internet in quite a long time.
This is an excellent approach to evolving education to respond to new tools and technology. LLMs are a resource and I want my kids to know how to use them to their advantage, without outsourcing all of the thinking and all of the work. I applaud!
These are always a joy to read, but this one in particular has helped prompt me to think more about how we could use LLMs in my workplace, especially for non-native speakers. I have one team member whom I am fairly sure is using it to put his contributions into better English, which has resulted in quite an idiosyncratic writing style.