“No traditional poll will exist by the time the next general election occurs” – that’s a bold prediction by AI forecasting startup Aaru’s founder Cam Fink. The startup wants to revolutionize forecasting in every application of business- and everyday life. For now they started with election outcome forecasts under their simulation tool Dynamo. Applying the software to several real presidential and state/congress elections, the deviation from the real results was on average 0.49%, while the polling average’s deviation came out to 1.31%, almost a threefold increase. The software also delivered very realistic results when applied to marketing performance assessments and even pop-culture problems (e.g. Gen Z’s opinion on skinny jeans or the trolley problem).

Aaru works by simulating real poll respondents (“agents”) using large language models (LLM). The agents are assigned demographically accurate data, hundreds of personality traits, and then fed the whole spectrum of existential knowledge. Also, the agents can access real-time information and update their beliefs constantly. For example when Donald Trump was subject to an assassination attempt, many of Aaru’s agents immediately switched voting preferences to Mr. Trump. The models are that realistic, that they can form their own opinion beyond what the real options are; one bot changed their voting preference to Mickey Mouse, explaining the move by stating it hates both Ms. Harris and Mr. Trump.

Many experts are certain that a model like Aaru’s is the future of polling. There are already now many Fortune 500 companies successfully leveraging Aaru’s solution. This is the case because LLMs are the most efficient forecasting device in existence – they are trained on every word that was ever written and update based on new information. All of that without being biased. While LLMs still have room for improvement, along human demographics like age, race, and gender, they definitely pose a novel avenue in forecasting that could change the field forever.

As of now, Aaru’s Dynamo predicts Kamala Harris to win the popular vote by 4.2% – a real-world case study that will be resolved in early November.

 

Articles:

AI startup Aaru uses chatbots instead of humans for political polls | Semafor

AaruWhitepaper.pdf (aaruaaru.com)