Current Projects

What is the probability of the US Steel/Nippon Steel merger being officially announced before January 21, 2025? Latest forecasts here.

BestWorld’s Botmaster Jeremy Lichtman, our new colleague Michał Dubrawski, and also Brian Labatte, and Carolyn Meinel, have been conducting many other AI-enabled forecasting experiments.

First experiment, launched July 12, 2024, ended Sept. 30, 2024: Forecasts by Jeremy’s bestworldbot in the Q3 AI Forecasting Tournament on 125 geopolitical questions, which were only open for 24 hours each weekday, beginning at 10:30 Am EDT. The sponsor of the contest, Metaculus, has been providing the resulting data to us for our planned “Bots or Not” analysis. We expect this AI Forecasting Tournament to continue for three more quarters, ending June 30, 2025.

Our second, the Humans vs Multi-AI Panel Forecasting Experiment was launched July 18, 2024 on “What is the probability that the US FED will cut interest rates in September 2024?” It ended Sept. 18, 2024, when the US FED lowered its key overnight borrowing rate by a half percentage point, or 50 basis points, a huge cut compared to the usual quarter point changes, and the first cut since 2020. This also was its first cut since 2020. The Multi-AI Panel was created by Jeremy, assisted by Brian LaBatte and Carolyn Meinel, using five generative AIs: Perplexity, Claude, Mistral, Cohere, and OpenAI , aided by the AskNews news search and aggregation system. working together in a panel format. The competing human forecasters were BestWorld staffers Brian LaBatte and Carolyn Meinel.

The third experiment, also launched on July 18, 2024, ending Sept. 18, 2025, was Old Bot, which also forecasted the FED rate cut. It used the same component AIs as the Multi-Panel experiment above, but in a different format.

We also have an experiment in development, an apparently revolutionary chatbot that debates conspiracy theories with members of the public, or, as the case might be in this experiment, college students. Results: a statistically significant and persisting reduction in beliefs in clearly false conspiracy theories.

We plan to attempt to replicate these results, as reported in Science, the journal of the American Association for the Advancement of Science. “New research indicates that finding the right facts can instill skepticism in conspiracy believers, and artificial intelligence (AI) is just the tool to help. After an evidence-based dialogue with GPT-4 Turbo, an AI model, conspiracy theorists substantially reduced their belief in favored conspiracies, and this newfound skepticism remained for months. Unraveling webs of conspiracy beliefs at scale has not previously been possible. See pages 1143, 1164, and 1814.

Carolyn’s novella, “The Optimizer Did It,” was an entry in the Superintelligence Imagined Creative Contest on the Risks of Superintelligence. They promise to announce the winners next week. 

Our long term research goal is creation of a “What’s News and What’s Next” system of traditional journalism enhanced with Gen AI-based news aggregation along with AI, human, and hybrid AI/human discussions of what will likely happen next. The objective is to give credibility to our news coverage via what we expect to be our mostly true forecasts, much like how people nowadays trust weather forecasts. 

Read more about our future projects here.

Pictured below, the final leaderboard of Metaculus’ Q3 AI Forecasting Benchmark Tournament, with bestworldbot at # 53 out of 55. A likely explanation for bestworldbot’s poor status recently is that after Sept. 10, when it was # 17, we tried extremizing its forecasts. With most human forecasters, extremizing is, on average, effective. However, we now have concluded that extremizing made bestworldbot worse. That’s experimentation at work. Oh, well, we learned something.