GPT/LLM links, 4/30

Ethan Mollick on how to use chatbots; The Zvi has another big roundup; Ethan Mollick on Bing as a creative teacher; Tim B. Lee on economic implications; Chatbot fails Steven Landsburg's economics test

Mark your calendars: This Monday, May 1, Tim B. Lee of Full Stack Economics will join our Zoom for paid subscribers at 8 PM New York Time to talk about his new Substack, Understanding AI. Then on May 8 Bryan Caplan will join us to talk about his latest collection of essays, Voters as Mad Scientists.

Ethan Mollick writes,

The common misconception about utilizing ChatGPT is that users should aim for the perfect prompt. However, this approach is not that useful, because there is no one-size-fits-all prompt. Instead, engaging in a conversation with the AI and asking questions can lead to better results.

It’s hard to excerpt from this essay, which is Mollick’s most useful post yet.

Zvi Mowshowitz writes,

From IWF’s Patricia Patnode: I Rode in a Driverless Car, This is the Future. I mean, yes, of course it is, the question is how far away is that future. The two most hilarious things in this write-up are the argument for the safety of driverless cars - gesturing at the fact that one might do the math without actually doing that math - and the author’s relatives being terrified of the driverless car, with two of them following behind her in another car, and her mother texting ‘get out now.’ Wow, everyone, if you feel that way about driverless cars, do I have news for you about non-driverless cars.

But most of the long post is about GPT stuff.

Ethan Mollick writes,

Once an exclusive privilege of million-dollar budgets and expert teams, EdTech now rests in the hands of educators. While it's important to remain vigilant for potential hallucinations, errors, and biases, GPT-4 enables teachers to craft personalized prompts tailored to their local contexts, significantly bolstering their resources in the pursuit of quality education.

It’s too bad that the media is filled with noise about large language models. Ethan’s post is filled with signal, and I hope that you will read it and share it.

Timothy B. Lee writes,

Computers and smartphones have become ubiquitous across the economy. But this has led to only modest changes for established industries like health care, education, housing, and transportation.

…most of the American economy is not information-focused: It’s focused on delivering physical goods and services like homes, cars, restaurant meals, and haircuts. It will be hard for AI to have a big impact on these industries for the same reasons that it’s been hard for Internet startups to do so.

Good point. Large language models are surprisingly good at simulated conversation, but they have little or nothing to do with robots, which is what you need to affect the delivery of physical goods.

I think that the industry that is most up in the air is education. Simulated conversation could change the game there. If it does, the cascading effects would be large.

Steven Landsburg writes,

our commenter John Faben, who has an existing subscription, offered to submit the exam for me.

The result: Whereas the older ChatGPT scored a flat zero (out of a possible 90), GPT-4 scored four points (out of the same possible 90). [I scored the 9 questions at 10 points each.] I think my students can stop worrying that their hard-won skills and knowledge will be outstripped by an AI program anytime soon.

On what should we blame the failure? Is the body of economics content on the Internet just really bad? Or is the LLM approach to AI simply not rigorous enough to deal with a sophisticated economics exam? Would a different prompt (telling the Chatbot that it is supposed to be a Ph.D economist, for example) have gotten a better result?

Substacks referenced above:

@

Why I'm not worried about AI causing mass unemployment

@

Democratizing the future of education

@

AI #8: People Can Do Reasonable Things

@

One sentence.