CognitionAI Unveils AI Model for Prompt-Based Website and Game Creation

Startup CognitionAI has announced its AI assistant Devin, capable of generating complex code for websites or games based on prompts.

The AI model is a next-level chatbot. Instead of merely offering coding advice, it can independently complete an entire project.

Bloomberg journalists conducted several experiments with Devin. The chatbot was prompted to “Create a website listing all Italian restaurants in Sydney.” The AI model searched for restaurants, compiled their addresses and contact information, and created and published the website.

During its operation, Devin displayed all tasks being performed, autonomously identified and corrected errors, and tested the written code. The chatbot created the website from the prompt in 10 minutes and also managed to recreate a web version of the game Pong in the same timeframe.

“Teaching AI to program is a very deep algorithmic task. It requires the system to make complex decisions and look several steps ahead to choose a path,” said developer Scott Wu.

According to CognitionAI’s founders, one of the company’s main advantages is its approach to teaching the chatbot to reason.

“In AI language, to ‘reason’ means that the system can not only predict the next word in a sentence or the next piece of code but also solve problems rationally,” added Wu.

Importantly, during its operation, Devin displays its plan in text, followed by commands and code. It also shows how the created website will appear in a browser. If something looks incorrect, a corrective prompt can be sent to the chatbot.

Most modern AI systems struggle to maintain a sequence of actions, but Devin manages this effectively.

Silas Alberti, co-founder of AI startup PIGEON, stated that CognitionAI’s technology is a real leap forward. According to him, it resembles less a chatbot assisting with code and more a genuine worker tackling assigned tasks.

It is not specified how CognitionAI achieved this breakthrough in such a short time. Wu declined to disclose the technology’s foundations, stating that the team found a unique way to combine large language models with reinforcement learning methods.

In May 2023, Microsoft integrated the GitHub Copilot X coding assistant into the Windows terminal.