Exploring AI's Pace of Evolution, AGI's Future, and Data Dominance with Adept CEO David Luan | E1855

Episode Summary

Executive Summary: David Luan, CEO of Adept AI, discusses the path to AGI through AI agents that can automate complex workflows across multiple software tools. He argues that training models to understand pixels and actions, not just text, is key to building reliable agents. Luan predicts that within 1-2 years, AI could handle tasks like prioritizing emails and scheduling, and within 5 years, act as a capable chief of staff. Adept focuses on enterprise deployments, targeting operations workflows that span many tools, with early tests showing tasks reduced from 90 to 30 minutes.

Main Topics: Definition and Timeline of AGI (Priority: 5/5): Luan defines AGI as an AI that can proactively manage daily work tasks, like prioritizing emails and scheduling. He predicts such capabilities within 1-2 years, with chief-of-staff-level performance in under 5 years. AI Agents and Multimodality (Priority: 5/5): The next battlefield in AI is building agents that understand images and can take sequences of actions to achieve goals. Adept trains its own models to understand pixels and generate actions, achieving high reliability. Data Challenges for AI Training (Priority: 4/5): Access to clean, high-quality training data is becoming the number one problem. Public internet data is closing off, and the smartest knowledge worker data is never public. Adept focuses on learning from enterprise workflows. Adept's Product and Enterprise Focus (Priority: 4/5): Adept builds AI agents that automate tedious tasks across multiple software tools. They target operations workflows like invoice processing and data entry, with early tests showing time savings from 90 to 30 minutes per task. Business Model and Pricing (Priority: 3/5): Adept focuses on high-ACV enterprise contracts rather than per-seat pricing. They charge for custom deployments that eliminate or augment jobs, with customers willing to pay for significant productivity gains. Comparison with Verticalized Solutions (Priority: 3/5): Adept differentiates by handling workflows that span multiple tools (e.g., Redfin to Google Sheets), which no single vertical solution can do. Custom workflows per company are also a key advantage. Future Vision: AI Teammates (Priority: 3/5): Long-term, Adept aims to build AI teammates that can brainstorm, plan, and collaborate with humans. This requires reliable agents as a foundation, with abstraction levels rising each year.

Key Arguments: AGI is achievable within 1-2 years for task management and under 5 years for chief-of-staff-level performance, based on current progress in AI agents. Training models on pixels and actions, not just text, is essential for building reliable AI agents that can interact with any software. Access to clean, high-quality training data is the number one problem for AI progress, as public internet data is closing off and enterprise data is private. Adept's focus on enterprise workflows that span multiple tools gives it an advantage over verticalized solutions like HubSpot or Salesforce. Human-in-the-loop systems are necessary for reliability, with Adept building models that are 90%+ accurate for specific use cases after fine-tuning. AI will not destroy jobs but will make teams 30-50% more efficient, allowing companies to grow without adding headcount.

Data Points: Time savings per task: 90 minutes to 30 minutes - A workflow that took an hour and a half can be reduced to 30 minutes using Adept's agents. Number of software tools per knowledge worker: 17 - The average knowledge worker uses 17 different software tools daily, creating opportunities for cross-tool automation. Accuracy of GPT-based agents: 60% - Most agents built on GPT are 60% accurate, making them unreliable for enterprise use. Adept's model reliability after fine-tuning: Very reliable - After custom fine-tuning per use case, Adept's models achieve high reliability, enabling enterprise deployments. AGI timeline for task management: 1-2 years - Luan predicts that within 1-2 years, AI can handle tasks like prioritizing emails and scheduling based on learned context. AGI timeline for chief-of-staff level: Less than 5 years - Luan believes AI will match a $150K/year chief of staff in under 5 years, with conservative estimates at 5 years.

Pivotal Quotes: "I think that we'll be at a spot where you would be able to get that within the next one to two years." — David Luan: Responding to a question about when AI could proactively manage daily work tasks like prioritizing emails and scheduling. "I think it's at the pace of progress. I think right now the field is still split between people that are like, wow, I see how this stuff is going to keep compounding. And then people who are like, well, just because the last three years has been crazy doesn't mean the next three years will be crazy. Capabilities are slowing down. Like models aren't going to get too much smarter anytime soon. I think the first group is correct." — David Luan: Discussing the debate on whether AI progress will continue at the same rapid pace, with Luan arguing for continued compounding progress. "I think the key with the agents thing is just getting them to actually be reliable. And that's like, I think that's the key advantage that we're really trying to run at: you got to control the whole model stack to do that." — David Luan: Explaining why Adept trains its own foundation models in-house rather than relying on third-party APIs, to achieve reliability.

Implications: AI agents that automate cross-tool workflows will transform knowledge work, making teams 30-50% more efficient. Enterprises will invest heavily in custom AI deployments, while verticalized software companies must adapt or partner. The race for clean training data will intensify, and AGI may arrive sooner than expected, reshaping job roles and productivity.