Taste is the new programming language (w/ Astrocade CEO Amir Sadeghian) | AI Basics

Astrocade is a platform founded in 2022 that lets anyone create interactive games, educational content, and companions using natural language prompts. CEO Amir Sadeghian began working in AI in 2013 and completed his PhD at Stanford AI before the field reached mainstream attention. He and his co-founder and CTO are brothers raised across four continents including Spain, Zimbabwe, Iran, and the US. The founding year was chosen because 2022 marked a capability threshold where generative AI became sufficient to generate interactive worlds.

The platform publishes created content to a feed modeled on TikTok and YouTube, where users scroll through and play games made by creators worldwide. As of the episode, Astrocade has more than 100,000 creators from more than 100 countries. Some creators work full time on the platform and earn thousands of dollars a month through a creator fund that pays based on number of plays received. Sadeghian frames this as a new creator economy analogous to what YouTube built for creators like Mr. Beast.

Traditional game development requires coding skills, asset creation, game tuning, and an understanding of player psychology. Sadeghian says a game of the quality Astrocade can now produce would have required five people and six months of work with significant budget two to three years ago. Astrocade removes those barriers by letting users describe a game in natural language, with the AI generating both the game and its editor simultaneously.

Sadeghian argues that taste is the new programming language. Because large language models are trained on vast data and tend to revert to the average or most common design, human taste becomes the differentiating input that steers the AI toward something original. Most fully realized games on the platform go through many iterations of prompts, which Astrocade calls wishes, before reaching high quality. Sadeghian is explicit that AI is still not capable of producing a complete original game without creator iteration and refinement.

Natural language is not always the optimal tool for fine-tuning specific game parameters because large language models are non-deterministic. Adjusting precise values like percentages is better handled through a direct editor than through repeated back-and-forth prompting. AI can suggest gamification approaches, but Sadeghian cautions that those suggestions are often not good on their own and the creator must choose and tune the specifics. Every number in a game matters, and even with AI involved, creators still need to play test and iterate.

One unexpected use case that emerged organically was users creating interactive birthday cards featuring images of their friends, enabled by a feature allowing creators to upload their own images and assets. Astrocade plans to retrain language models on its own platform data over time, with the goal of building what it aims to be the best models in the world specifically for generating interactive experiences. The company is actively hiring for product, recommendation systems, and AI agents post-training roles.