New preprint and blog post! We explore the social disincentives of warning about unlikely risks. Many people are reluctant to warn about large but unlikely risks because they could look bad if the risk doesn’t occur 1/5

Philipp Schoenegger Spencer Greenberg 🔍 Alexander Grishin Lucius Caviola Thanks for making the code/data so clear! Replicated your main analysis here (github.com/seantrott/pers…), and I also tried implementing my "need to beat framework" I introduced here (direct.mit.edu/opmi/article/d…).

Sean Trott Philipp Schoenegger Alexander Grishin Lucius Caviola I find this graph you made - very interesting:

Cool research by @schoeneggerphil et al compares human lay people/experts with LLMs/specialized AI Systems predicting personality correlations. LLMs outperform lay folk but not experts; specialized AI outperforms individual experts, but not groups: buff.ly/3RymA3s

This AI Forecasting Benchmark Series from @Metaculus looks exciting! It will start with a $30k contest on July 8 and then go from there, providing a broad comparison of AI prediction performance (over time) to that of human forecasters. metaculus.com/notebooks/2552…

Almost comical chart

‼️New preprint: Scaling laws for political persuasion with LLMs‼️ In a large pre-registered experiment (n=25,982), we find evidence that scaling the size of language models yields sharply diminishing persuasive returns: arxiv.org/abs/2406.14508 1/n

Check out our preprint 👇 Ft. a nice foreword by GPT 🤖 "This groundbreaking study delves into the persuasiveness of AI-generated political messages, uncovering a log scaling law that suggests limited gains from increasing model size. A crucial read for AI ethics and policy."

The large-scale collaborative AI Persuasion Project has a website now! Head over there for information about what we're up to, our team, and (eventually) our outputs as they come out! sites.google.com/view/ai-persua…

Excited to be collaborating on a very interesting global #responsibleAI project with 50+ leading researchers across behavioural sciences, CS, AI, ML, and more... Led by Philipp Schoenegger LSE et al. sites.google.com/view/ai-persua… ai@cam Responsible Ai UK Gates Cambridge

In September 2022 Swift Centre forecasters looked at Biden's re-election chances under 14 different conditions and estimated he was likely to lose in all of them. Some of the most potentially damaging issues identified were Hunter Biden's legal issues and Biden Snr's health

This was indeed an amazing workshop with an incredible group of people that I've learnt a huge amount from!

Concur ;)

The conference of alums of LSE Psychological & Behavioural Science Executive MSc in Behavioural Science continues with Philipp Schoenegger talking about ‘judgmental forecasting tournaments: applications & limitations’

New paper in TACL with Cameron Jones and Ben Bergen comparing humans and LLMs on a range of tasks designed to assess theory of mind. Check out Cameron Jones’s thread for more details!

Large Language Models can produce responses to debate questions posed in a UK political TV-debate show that were judged as more authentic and relevant than the original responses from the people who were impersonated—research by Steffen Herbold et al: buff.ly/3LBlKzk

Today, we're releasing a report on our study using "conditional trees" to generate high-value forecasting questions about AI risk. Our method produced more informative questions than those on existing forecasting platforms. Here are the key findings: 🧵

Can we figure out how to incentivize not just accuracy of forecasts but also quality of questions? Here's a new FRI report that describes a conditional tree method of generating high-quality questions that beats business-as-usual on a major forecasting platform. I hope this will

The results in "LLMs Are Superhuman Forecasters" don't hold when given another set of forecasting questions. I used their codebase (models, prompts, retrieval, etc.) to evaluate a new set of 324 questions—all opened after November 2023. Findings: Their Brier score: .195 Crowd