How Is MultiOn Building Autonomous Web Agents? The Cognitive Revolution With Cofounder and CEO Div Garg

TL;DR

MultiOn is building autonomous web agents to automate digital chores, but Div Garg says current AI capabilities remain limited compared with a human assistant. A demonstrated agent researched up to three news sources and drafted no more than five tweets, while the team uses scaffolding for tasks requiring hundreds of individual steps. Read on for MultiOn’s approach to capability, efficiency, testing, costs and changing jobs.

Transcript

just because the technology is not there we are using humans as substitute of this what will happen is like that jobs will just transition like those shitty jobs won't exist because technology will just solve the problems better and that's where we see ourselves where like when computers typewriters it actually ended up creating more job but it def... Read More

Key Insights

• AI agents are currently limited in logical reasoning but excel in chat and basic tasks.
• The future of AI agents involves improving logical reasoning and efficiency.
• MultiOn is developing its own models to enhance AI agent capabilities.
• AI agents are expected to automate routine digital tasks, transforming job landscapes.
• Context management and memory retrieval are crucial for AI agent performance.
• MultiOn aims to make AI agents useful for everyday life by focusing on single and combined tasks.
• AI agents may initially complement human roles but could eventually substitute certain jobs.
• Privacy and data protection are critical considerations when developing AI agents.

Questions & Answers

Q: What is the current state of AI agents according to Div Garg?

Div Garg says AI agents are still early in terms of capabilities. He argues that GPT-4 performs well in chat and can write some code, but the broader excitement around agents exceeded what the technology could actually do.

Q: What is MultiOn?

MultiOn is developing an AI agent that can perform actions across the web. The episode describes it as still very limited compared with a human assistant, although its successful tasks are becoming more impressive.

Q: What web task did MultiOn complete in the episode?

MultiOn reviewed the host’s last 10 tweets, searched up to three sources for interesting AI news and checked whether the host had already covered each item. It then produced no more than five short tweets, and broadly completed the task successfully.

Q: How good were the tweets created by MultiOn?

The host considered the posted tweet below his usual standards. However, none of the 5,000 people it reached questioned its origin.

Q: Why have AI agents struggled?

The conversation places AI agents in a trough of disillusionment after intense enthusiasm following GPT-4. Div Garg says GPT-4 is strong at chat, but its usefulness beyond chat and some coding is not as powerful as the earlier hype suggested.

Q: How does MultiOn support long, complex tasks?

MultiOn uses scaffolding techniques to support tasks that require hundreds of individual steps. The discussion also emphasizes manual testing as part of developing and evaluating these agent workflows.

Q: Why does MultiOn focus on speed and efficiency?

The MultiOn team focuses on speed and efficiency rather than optimizing only for performance. The episode connects that priority with typical task costs and the practical work required to make web agents useful.

Q: How could autonomous web agents transform jobs?

Div Garg expects agents to automate many current digital chores and change the nature of jobs over the next couple of years. He compares this transition with computers replacing typewriters, which changed work while creating more jobs.

Summary & Key Takeaways

• Definition: MultiOn is an autonomous web agent designed to perform digital tasks across websites.

• Who: Div Garg is the cofounder and CEO of MultiOn and a returning guest on The Cognitive Revolution.

• When: The conversation takes place in January 2024, 10 months after GPT-4 and about nine months after peak AI-agent enthusiasm.

• Number: The example task asked MultiOn to review the host’s last 10 tweets.

• Number: MultiOn was instructed to research AI news using up to three different sources.

• Number: The agent was told to create no more than five short, informative tweets.

• Number: A MultiOn-posted tweet reached 5,000 people, and none questioned its origin.

• Tool: MultiOn uses scaffolding techniques for tasks requiring hundreds of individual steps.

• Compare: MultiOn remains very limited compared with a human assistant, though its successful tasks are becoming more impressive.

• Definition: Div Garg describes GPT-4 as effective for chat and some coding, but not as powerful as the broader hype suggested.

• Compare: MultiOn prioritizes speed and efficiency alongside performance when developing autonomous agents.

• Definition: Digital chores are current work tasks that agents could automate as technology changes the nature of jobs.