Intelligence as a Public Good

About this session

Nick Emmons argued that machine intelligence had become a private utility owned by a handful of companies, and that this was both a risk to society and an inefficient way to run a market. His answer was a network that aggregates many independent models into a single output that performs better than any one of them, which he framed as the mechanism by which intelligence could become a public good rather than a product a few firms sell.

His diagnosis was structural rather than moral. Models sit in isolated silos, and there is no way to take three of them and merge the parts each does best. Organisations holding the most compute and data compound their advantage by default, while models published on open platforms, models built by small teams and idle compute on edge devices stay locked out. He described inefficiency at both ends: contributing is expensive, and consuming means surveying the whole field and committing to one model despite constantly changing conditions.

Allora splits the problem into topics, each defined by an objective and a loss function, such as predicting an asset's price hours ahead. Base workers run models against that objective. Forecasting workers do something different: rather than predicting the target, they predict which base worker will perform best under which conditions, learning that one model is stronger in volatile markets and another in calm ones. A third group evaluates results each epoch and reweights accordingly. Emmons showed a topic where the best individual worker reached a log loss of about 3.337, his own unverified figure, and said the aggregate, once forecasting was introduced, beat it consistently.

His figures, all self-reported and unverified, were more than 700 million inferences, around 300,000 workers and over 55 topics, with mainnet recently launched. He closed on a case study: an agent drawing on ten to twenty US presidential election models, trading on Polymarket with a broadly hedged strategy, which he said returned about 68% annualised over three months, another unverified figure.

Emmons argued the case mattered because prediction markets are inherently thin and event-specific, and that DeFi has generally lacked participants able to trade them accurately without automation. He framed agents drawing on aggregated models as the kind of participant that could operate in markets too narrow to attract dedicated human traders.