One repeated finding in clever large-scale studies like this one is that what we do shapes who we are: cultures that grow rice really are more collectivist than wheat farmers, because rice farming requires collective work & wheat does not. I wonder how today’s work impacts us.
@emollick
-
GPT-4 LLM Agents Spontaneously Collude on Pricing
By
–
New paper finds GPT-4 is very good at pricing, acting as an agent to get a merchant the best price. In fact, it is a little too good. When it is possible to establish an oligarchy, LLM agents spontaneously collude on pricing to the detriment of customers! https://
arxiv.org/abs/2404.00806 -
ChatGPT Adds Image Editing Feature Inside DALL-E
By
–
A quiet AI release that I think a lot of people would actually use is the addition of image editing right inside of ChatGPT. If you click on a DALL-E image, you can highlight the areas to edit & use words to describe the changes. It works well but has the usual inconsistencies.
-
Future Job Stimulus Programs: Economic Policy Ideas
By
–
Idea for future job stimulus programs. https://t.co/uIbVgPKIqm pic.twitter.com/XQoQmsJIWe
— Ethan Mollick (@emollick) 4 avril 2024Idea for future job stimulus programs.
-
Community Notes Slow Misinformation Spread But Lag Behind Viral Content
By
–
Community Notes really does work to slow down the spread of misinformation for tweets that get Notes. The effect is quite large. But the Notes are often posted too slowly, appearing only after false information has gone viral, lowering their real world impact.
-
GPT-4 Debate Effectively Reduces Conspiracy Theory Beliefs
By
–
Unexpected & big: it is famously hard to get people to stop believing in conspiracy theories, but… A controlled trial finds a 3 round debate with GPT-4 arguing the other side robustly lowers conspiracy theory beliefs and the effects persist over time, even for true believers.
-
Science Practices and AI Research Methodology
By
–
Amazing run of posts reden. But do you have any good news about how science is done?
-
GPT-4 Clinical Reasoning: Redefining Exclusively Human Capabilities
By
–
Interesting bit from a thread by an author of the recent JAMA paper showing that GPT-4 is good at clinical reasoning, the whole thread is great, but I wanted to highlight this. There are a lot of things we thought were exclusively human until a year ago, and now we know are not.
-
Using instruction manuals as prompts for Claude and GPT-4
By
–
Instruction manuals often make great prompts. Organizations should try feeding their processes & procedure documentation to Claude 3 or GPT-4 and seeing how far it gets them in producing good outputs.
-
Claude 3 excels at summarizing long books but struggles with fact-checking
By
–
"Talk to your documents" is a big use case for AI. This paper tested how good AI was at working with long, new books (like The Fourth Wing). Claude 3 does very well in SUMMARIZING big books. But none of the models do a good job FACT CHECKING claims about books made by others.