In a recent conversation with Vance Spencer of Framework Ventures, AI Eye got an inside look at the potential of an upcoming game called AI Arena that the fund invested in. This unique game allows players to train AI models how to battle each other in an arena, and is compared to a mix of Super Smash Brothers and Axie Infinity. What sets this game apart is the tokenization of AI models as NFTs, allowing players to train and sell them for profit or rent them out.
Spencer, known for early investments in Chainlink, Synthetix, and an NFL platform similar to NBA Top Shots, believes that tokenized AI models will be some of the most valuable assets on-chain. Chief Operating Officer of AI Arena, Wei Xi, elaborated on the origin of the game concept, revealing that the team wanted to incorporate NFTs and AI into a game. The game aims to reveal the process of AI research through its core loop, which involves iterating through three steps of training, calibrating, and diagnosing the performance of the AI.
The game utilizes a custom-built feed forward neural network, making the AIs lightweight and constrained, meaning the winner won’t just be determined by throwing the most computing resources at the model. AI Arena is currently in closed beta testing and plans to launch on Ethereum scaling solution Arbitrum in the first quarter of next year. It will feature a browser-based version for casual players as well as a blockchain-based version for competitive eSports players. As with many crypto projects, there is a token involved that will be distributed to players and used to pay entry fees for competitions.
Spencer envisions a big future for the technology behind AI Arena, seeing potential for its use in other game genres as well as in a marketplace for AI models trained for business tasks. Although still in development, the game looks promising for introducing a new dimension of competition and creativity to the AI and NFT space.
In other developments in the world of AI, a recent analysis from AI startup Vectara revealed that output from large language models like ChatGPT and Claude may not be entirely trustworthy in terms of accuracy. The analysis quantified the amount of “bs” each model generates, with GPT-4 being the most accurate and Google’s PaLM being the least accurate.
AI images have also made headlines in recent events, particularly during the Israel-Gaza war. While they haven’t played a huge role in the war, AI-generated images have been shared on social media, often to stir emotions. These images, while created by AI, have been used to discredit real images from the conflict, raising concerns about the spread of disinformation.
On a more positive note, Google has announced tools to help users spot fake images, which comes as a welcome development given the prevalence of AI-generated fakes.
OpenAI also recently unveiled GPT-4 Turbo, a faster and more powerful model that can accept longer text inputs and generate captions or descriptions of visual input. The model has been trained on data up to April this year and will be more cost-effective for developers to access. Additionally, OpenAI has revealed their own version of the App Store, called the GPT Store, allowing for developers to access and create various applications using AI. These recent advancements indicate the continued growth and expansion of AI technology in various industries.