AI Agents Still Struggling

Key Concepts

• Foundational AI vs. Applications of AI: Shift from building infrastructure to applying AI in specific use cases.
• AI Agents: Autonomous systems directed by LLMs to perform tasks across multiple steps.
• Enterprise ROI: Focus on achieving tangible returns on investment in AI within businesses.
• Physical AI: AI applied in the physical world, such as robotics and manufacturing.
• Open Source AI: Utilizing and contributing to publicly available AI models and tools.
• AI Infrastructure Stack: The underlying components (hardware, software, models) required to build and deploy AI solutions.
• Proprietary Data: Unique and exclusive datasets owned by specific organizations.
• Workflow Integration: Seamlessly incorporating AI into existing business processes.

AI Applications in Enterprise

• Coding, Customer Support, and Sales: These are the three main areas where the speaker's portfolio companies are seeing the most impact from AI agents.
• Maven AGI (Customer Support): This company is cited as a specific example of an AI agent successfully integrating with legacy systems to solve customer support tickets faster and at a lower cost. They are working with public companies.
• Autonomous Systems: AI agents are defined as autonomous systems where an LLM directs actions across a series of steps.
• Verifiable Loop: The speaker notes that verifiable loops are not yet widely seen outside of coding, customer support, and sales due to integration challenges.

Physical AI and Robotics

• Timeframe for Rewards: While text and image AI have shown repeatable results, physical AI (robotics, warehouses, manufacturing) is still early in terms of delivering enterprise ROI.
• Proprietary Data Troves: The challenge lies in accessing and integrating proprietary data and workflows in these physical domains.
• Workflow Integration: Integrating AI with the specific workflows of industries like biology, chemistry, and manufacturing is crucial for realizing value.

Consumer Applications and Challenges

• Complex Systems with Compounding Errors: AI agents can be complex systems, and even small errors can compound across multiple steps, leading to frustrating user experiences.
• Flight Booking Example: The speaker shares an example of trying to book a flight using an AI agent, highlighting the challenges of handling multiple criteria and preferences.

Opportunities for Smaller Companies

• New AI Infrastructure Stack: The emergence of a new AI infrastructure stack presents opportunities for smaller companies to provide solutions that complement hyperscalers and large chip companies.
• Together AI (Open Source AI Cloud): This company is an example of a smaller player that helps users run and productize open-source models, potentially reducing reliance on major cloud providers.
• Cheaper Chips and Smaller Models: The speaker notes that smaller companies are leveraging cheaper chips and smaller models to achieve similar outcomes as larger players.

Global Talent and Competition

• Expertise is Everywhere: The speaker emphasizes that AI expertise is not limited to specific regions.
• China's Deep Sea: China's Deep Sea project is highlighted as a "wake-up call" for the U.S., demonstrating hardware and engineering efficiency.
• International Examples: The speaker mentions companies in Japan (like the kind of AI) and France (Mistral) that are building strong AI capabilities.
• Leveraging and Doing More with Less: The focus is on using less compute, more efficient chips, and open-source technology to create a more efficient infrastructure stack.

US Government and AI

• AI Czar: The appointment of individuals like David Sacks to positions of influence in the AI industry is seen as potentially invigorating the field.
• Investment in Engineering Startups: There is excitement about investing more heavily in engineering startups, particularly those building on U.S. soil and leveraging open-source innovation.
• Public-Private Mix: The combination of public and private sector efforts is seen as promising for developing new applications in areas like robotics, space, defense, and manufacturing.
• Understanding Workflows: The speaker emphasizes the importance of understanding the specific workflows of industries like defense contracting to build effective AI applications.

Synthesis/Conclusion

The conversation highlights a shift from foundational AI development to practical applications, particularly within enterprises. While AI agents show promise in areas like coding, customer support, and sales, challenges remain in integrating them into complex workflows, especially in physical domains like robotics and manufacturing. The rise of open-source AI and a new AI infrastructure stack presents opportunities for smaller companies to compete with larger players. Globally, AI expertise is widespread, and countries like China are pushing the boundaries of hardware and engineering efficiency. The U.S. government's increased focus on AI and investment in startups are seen as positive developments for fostering innovation and developing new applications in both the digital and physical worlds.