Scientists from the DeepMind lab trained artificial intelligence to efficiently distribute funds in an online game. The results of the research were published in the journal Nature.

To create the algorithm, the researchers conducted an experiment using an online economic game, in which participants choose options to allocate their resources for mutual benefit.

The simulation involves four people, each receiving a different amount of money. Players must decide to keep them or contribute to a public fund that generates a return on investment. At the same time, the ROI can be adjusted so that some players benefit others.

Possible ROI mechanisms include three models:

• Strict egalitarian, in which income from public funds is evenly distributed regardless of contribution;
• libertarian where payments are commensurate with contributions;
• liberal-egalitarian, in which each player pays in proportion to the share of their personal funds contributed.

The researchers asked groups of people to play multiple rounds with different levels of inequality and different profit-sharing mechanisms. They then chose their preferred method.

Thanks to this data, the scientists taught the AI ​​to mimic human behavior in the game, including the voting phase. While pulling agents in thousands of rounds, another algorithm set up a vote-based redistribution mechanism.

As a result, AI settled on a model that was as similar as possible to the liberal-egalitarian one. However, the algorithm returned almost nothing to the players unless they invested about half of their capital.

The scientists then used the AI ​​method against the other three in a new phase of the human game. According to the report, the algorithmic method tended to win the most votes as a result of the blind poll.

According to the researchers, the AI-enhanced mechanism is likely to work well because payments based on relative contributions help correct the initial wealth imbalance. However, imposing a minimum deposit prevents less affluent players from using the wealthier ones’ deposits.

The DeepMind researchers noted that their work is not “instructions for creating an AI government.” Nor are they considering developing AI tools for politics.

Recall that DeepMind introduced a visual language model with 80 billion parameters in May.

In February, AI Lab developed AlphaCode, a tool that can code itself.

In August 2021, DeepMind introduced a universal architecture for building artificial intelligence.