For decades, AI (artificial intelligence) has offered the promise of new opportunities, but now with LLMs (large language models) we have an opportunity to take AI to the next level. Stanford University suggests chatbots are now turning into action bots—but the challenge persists: if systems are trained with unknown or bad data, will it still be helpful?
Researchers at Stanford aim to address this challenge, having developed NNetNav, which is an AI agent that learns through its interactions with websites, which is open source and uses fewer parameters. Professor Chris Manning says NNetNav could become a lighter weight, faster, privacy-preserving alternative to OpenAI’s recently released Operator.
Here’s the difference: It is not trained by how humans behave online. Rather, it gathers synthetic training data by exploring websites much the way a young child might. It clicks all the buttons and types into all the fields to see what will happen. It then prunes out the pathways that don’t help achieve the user’s goals.
The team collected 10,000 positive demonstrations of NNetNav on 20 websites. Those successful trajectories were then used to fine-tune the model. When the team looked at NNetNav’s performance before and after fine-tuning, the model compared favorably with GPT-4 and did better than other open-source, unsupervised methods. It also used about one-third fewer parameters than the next-best performing model.
Here is how this can help:
- Ensure more privacy while still leveraging AI.
- Become more accurate and efficient through learning.
- Learn through interaction as you go.
Looking to the future, we can expect exponential growth for AI in the future—but there will be opportunities to use different agents. AI agents are here. How will you proceed in a new era?