Orchestrating Smarter AI Systems with Yoav Shoham | AI Basics with Google Cloud

Key Concepts:

• Large Language Models (LLMs)
• Reasoning in AI
• AI21 Labs (Jurassic-2, Wordtune, Jamba)
• Orchestration and Planning of AI Systems (Maestro)
• Hallucinations/Reliability of LLMs
• Agents and Agent Washing
• Small Language Models (SLMs) vs. Large Language Models (LLMs)
• Context Length
• Transformer Architecture
• Hybrid State-Based Models
• Verticalization of Knowledge
• Agent-to-Agent (A2A) Protocols
• Semantic Web
• Adaptive Learning

1. State of Reasoning in AI and AI21 Labs

• The Problem of Reasoning: While there's significant experimentation with LLMs in corporate America, concerns remain about their reasoning abilities, reliability, and "hallucinations."
• AI21 Labs' Approach: AI21 Labs, co-founded by Yoav Shoham, focuses on building LLMs and tools for collaborative reasoning. They've taken a different approach with their Jamba family, using a hybrid architecture instead of a pure transformer for efficiency.
• Orchestration is Key: AI21 Labs is focusing on orchestration and planning of complex AI systems with their product Maestro.
• Reliability Gap: While LLMs can produce brilliant output, their unreliability (e.g., being wrong 5% of the time) hinders deployment in mission-critical enterprise applications. The ratio of experiments to deployments is high (10:1 or 20:1).
• Maestro's Solution: Maestro aims to improve reliability by having multiple LLMs work in concert, checking each other's work and justifying their answers.

2. Orchestration and Validation

• Beyond Guardrails: Simply adding guardrails to LLMs won't eliminate variance. External logic is needed for orchestration.
• Orchestration Components: Orchestration involves routing between different LLMs, running code, using tools, accessing databases, and calling APIs.
• Testing: Unit testing at every step of the process is crucial, along with system testing. Validation can be done with judge language models or by simple methods like counting words to meet output length requirements.

3. Agents and Automation

• Agent Promise: Agents are seen as a way to automate repetitive tasks and chores, freeing up humans for more valuable work.
• Agent Washing: The term "agent" is often overused for any kind of automation.
• Agent Definition: True agents are more than just transactional calls to an LLM. They are ongoing, proactive, execute complex flows, and use multiple tools.
• Complexity and Uncertainty: Composing multiple LLM calls increases uncertainty, so caution is needed. Simple, repetitive tasks are the most suitable for automation.

4. Small Language Models (SLMs) vs. Large Language Models (LLMs)

• Consumer vs. Enterprise: For general-purpose chat experiences (like ChatGPT), large language models are necessary. However, for narrower enterprise needs, smaller models can be more effective.
• Advantages of SLMs: SLMs offer cost savings (both in dollars and latency).
• Jamba Model: AI21 Labs' Jamba model is a hybrid state-based model and transformer that offers competitive quality with lower latency and memory footprint, especially with increasing context length.
• Transformer Limitations: The transformer architecture, while revolutionary, has quadratic complexity in the input (context length), making it expensive for long inputs.

5. Verticalization of Knowledge

• Domain-Specific Models: Verticalizing knowledge by creating models focused on specific domains (e.g., accounting, legal) can improve accuracy.
• Benefits of Constraint: Constraining models to specific domains reduces the need to worry about irrelevant information, leading to higher fidelity content.
• Retaining Common Sense: It's important to retain general common sense and language skills even when specializing in a domain.

6. Agent-to-Agent (A2A) Protocols

• A2A Vision: A2A protocols aim to enable interoperability between agents, allowing them to interact between companies (e.g., for RFPs).
• Early Stage: A2A is still in its early stages, with no significant deployments yet.
• Google's Role: Google is playing a key role in shepherding A2A development.
• Challenges:
  • "Something from Nothing" Fallacy: The idea that simply building multiple agents will automatically lead to magic is often disappointing.
  • Semantic Understanding: The A2A protocol specifies syntax but not semantics (meaning), which can lead to misunderstandings between agents.
  • Shared Incentives: Agents from different companies (or even different units within the same company) may not have the same incentives, requiring controls and game theory to ensure cooperation.

7. Overhyped vs. Underhyped Opportunities

• Underhyped: Reliable Workflows: Focusing on making AI-powered workflows reliable and customizable for each deployment is a major blocker in the enterprise. This is not a "sexy" area, but it offers real value.
• Underserved: Education: Rethinking education with AI offers a significant opportunity. The focus should be on using AI to proactively teach students, not just preventing cheating.
• Adaptive Learning: Adaptive learning, where AI tailors the learning experience to each student's needs, has immense potential.

8. The Future of Education with AI

• Beyond Televised Radio: New technologies are often initially used in the same way as old technologies. It's time to rethink education with AI.
• Personalized Tutors: AI can provide personalized tutoring to anyone with a phone, offering customized Q&A and guidance.

9. Conclusion

The discussion highlights the current state of AI, focusing on the importance of reliability, orchestration, and domain-specific knowledge. While agents and A2A protocols hold promise, they are still in early stages and face significant challenges. The most immediate opportunities lie in making AI-powered workflows more reliable and in rethinking education with adaptive learning technologies.