The Prediction Market of Tomorrow, Live Today
Yesterday, prediction markets were stuck. Burdened with opaque resolution, four-year seasonality, and insufficient incentives, prediction markets have failed to live up to their potential. They generate massive value for society as clairvoyant machines, but capture none of it for their participants, leaving them to dry out in a PvP landscape. There must be a better way forward.
We'd like to share our vision for the Prediction Market of Tomorrow. The holy grail experience seems clear: put money on ANY prediction and get rewarded if it comes true. The collective wisdom of the market generates valuable prophetic signals which earn fees for its participants, escaping the zero sum structure of today.
The Prediction Market of Tomorrow has the power to coax the world’s information from the minds where it hides today. To get there, we need a seismic shift in the way we think about prediction markets. To unlock liquidity for the long tail of interesting questions we need to move away from the paradigm of ‘one question one market’. Bootstrapping liquidity on new markets is expensive, especially for the zero sum prediction markets live today, thus making prediction markets an unlikely profit avenue for researchers & insiders with useful information.
However, moving from markets where participants can only meaningfully engage on a small set of questions with deep liquidity to ones that reward the long tail requires markets to be able to express information in higher dimensionality than the scalar markets of today. In short, we need markets with semantic awareness, able to judge the semantic similarity between predictions and trusted sources of truth. The challenge today is to infuse this semantic intelligence into prediction markets without sacrificing decentralization, censorship resistance, or security.
Pushing the Envelope with Intelligent Prediction Markets
Today we're unleashing the first iteration of an intelligent prediction market, securely supercharged with AI. TMR.NEWS - now live on base - is the first prediction market where all you have to do is type in any prediction. In typical Small Brain fashion, we’ve condensed this grand vision into a live, fun, and simple product we’ve built over the past month.
The TMR.NEWS mechanic is simple: submit a prediction - any prediction - for tomorrow’s New York Times front page article headline, and get paid out proportionally based on how semantically similar your prediction is to the true result. This relies on a useful property of LLMs that allows us to compute the similarity of text via computing the distance between their model-outputted embeddings.
Under the hood, the contracts are extremely simple and versatile. They allow anyone to permissionlessly create and resolve intelligent prediction markets for any piece of future content on the internet. Think markets on Elon’s next tweet, the next iPhone launch announcement, or changes to regulation. All you need is a url, regex, and expiry date. Markets can modularly declare their resolution verifiers - allowing anyone to customize their trust assumptions.
Building On The Technical Frontier
TMR.NEWS markets are resolved via web proofs, a new technology that enables cryptographically verifiable statements about the internet. We’re currently using Reclaim Protocol’s TLS Proxy implementation, exposing web proofs over a convenient zk fetch API. However, as the web proofs landscape heats up with more providers and stricter trust assumptions, modularity allows us to integrate Pluto, TLS Notary, or Opacity as their solutions become viable.
We also took an unorthodox approach to the implementation of on-chain AI, choosing to make web proofs over OpenAI API responses rather than implement ZK AI protocols. We experimented deeply with existing ZK AI providers but could not find anything that was remotely mature enough for our usage. Besides being the simplest implementation of trust-minimized AI, this choice also allows us to use the best available models, including those that are not open sourced.
Though perhaps the most interesting, web proofs are not nearly the most complex or compute-intensive component of the TMR.NEWS contracts. Since the payout function depends not only on the accuracy of your prediction, but its accuracy relative to everyone else’s, resolving the market requires computing a sum of participants’ scores. Only when the scoring function is linear the series collapses to a closed form expression computable onchain. However, these linear scoring functions result either in extremely small PnL or aren’t bounded below at 0. Furthermore, since the inputs to the scoring function are n-dimensional vectors (currently dim 768) gas usage quickly balloons to the block gas limit of 30M. This experience has opened our eyes to the need for more performant trust-minimized computing environments.
Small Brains, Big Plans
In the meantime, the Small Brain crew is full steam ahead building the frontier of digital experiences. We’re stoked to put this out there and see how far we can push prediction markets forward (we think pretty far). If you’re interested building intelligent prediction markets on some type of data we haven’t thought of or even building something cool in general, please reach out!