Welcome to our latest engineering update. At Cykel, we're working towards an AI agent that can assist anyone with any online task, and our latest research is another step towards building the most useful co-pilot on the web.

In this update, we'll share some recent advancements in leveraging transfer learning, new ways to apply language models and promising results from implementing AI self-reflection.

‍

Transfer learning for better performance

In our recent research, we've found that focusing on transfer learning allows our language models to decipher raw HTML and execute web tasks with 50% higher success while using 192x less data compared to other methods.

This approach enables our co-pilot to quickly comprehend most web pages and show broader generalisation abilities with significantly reduced efforts.

‍

‍

New ways to apply language models

We're also exploring new ways for language models to understand the web beyond normal language skills. Our experiments prove that pre-trained models can classify forms, describe content, and navigate sites better than custom HTML networks.

We're using models like T5 and PaLM with minimal adjustments for better performance. These updates help our co-pilots handle various tasks with high success rates.

‍

Self-Reflection in AI

We've integrated "self-reflection" into our AI. This means it critiques its own suggestions before moving on, improving accuracy. Through tests, we've initiated the AI to reanalyse its suggestions and action history before executing subsequent task steps.

This iterative self-checking process has significantly enhanced accuracy and reduced confusion, especially in handling more complex tasks.

By doing this, our AI becomes better at understanding instructions and doing a wider range of tasks. It's also fascinating as an observer to see how it learns and corrects itself while narrating its "thought" process while we debug.

‍

‍

What's next

Moving forward, our aim is to expand the capabilities of the co-pilot improving its ability to navigate ever more complex interfaces and tasks.