🧠 Compound AI: Complex and Coordinated

Subscribe | AI Toolkit | Meet The Team

Welcome back AI prodigies!

In today’s Sunday Special:

• ⚙️How Does It Work?

• 🤖What Does Compound AI Look Like?

• 🦾How To Overcome Its Unique Risks

• 🔑Key Takeaway

Read Time: 7 minutes

🎓Key Terms

• Generative AI (GenAI): Uses AI models trained on text, image, audio, video, or code data to generate new content.

• Compound AI: Combines multiple AI models, high-quality databases, and external tools to solve more complex problems.

• Artificial General Intelligence (AGI): AI systems that perform tasks as well as humans and exhibit human traits such as intuition, sentience, and consciousness.

• Computed Tomography (CT): Creates detailed images of a patient’s bones and soft tissues to detect injuries and diseases.

• Electronic Health Records (EHRs): Digital versions of a patient’s medical history, including symptoms, diagnoses, and medications.

🩺 PULSE CHECK

⚙️HOW DOES IT WORK?

Should we consider how to apply advanced AI applications to make high-impact decisions or complete high-value tasks? Some would say: “No, AI will never be strong enough.” Others would say: “No, an AI agent would do whatever it wants.” Let’s consider another scenario: “We will be able to apply advanced AI applications to large, consequential tasks.”

But how could we apply advanced AI applications to large, consequential tasks without losing control? Today, we’ll describe one approach pioneered by Professor K. Eric Drexler of the University of Oxford (“Oxf”) at the Future of Humanity Institute (FHI): organizing tasks into Compound AI systems.

When explaining this concept, we’ll assume that developers continue to produce flawed AI models that usually do what we want and can perform an ever-wider range of tasks in aggregate. For instance, OpenAI’s ChatGPT generates text, Meta’s AudioCraft generates audio, and OpenAI’s Sora generates video. Together, they can make a movie. Though this simplifies the capabilities of a Compound AI system, it underscores how imperfect, differentiated AI models can work together to accomplish large, consequential tasks.

🤖WHAT DOES COMPOUND AI LOOK LIKE?

People can follow simple routines, like brushing their teeth, with minimal cognitive effort. However, building a space telescope requires the coordination of countless specialists and technologies, operating beyond the comprehension of any individual. Similarly, AI agents might excel at straightforward tasks, such as setting a calendar reminder, but deploying advanced AI applications to manage the complexity of an entire organization raises the challenge: how can we scale AI’s utility without resorting to risky superhuman AI agents equipped with AGI? First, let’s recall how we use GenAI today. Some GenAI models, such as OpenAI’s GPT-4o (“o” for “omni”), are multimodal, meaning they can process virtually any input type (e.g., text, image, audio, video, or code) and convert these inputs into virtually any output type. By carefully crafting your inputs (i.e., prompts), you can “train” the GenAI model to perform desired outputs, such as writing a poem, designing a flow chart, or summarizing an article.

In theory, advanced AI applications would have access to multiple GenAI models that generate engineering designs, system architectures, and implementation plans with schedules, budgets, and timelines. They’d also employ domain-specific, industry-expert AI agents to refine and evaluate sub-plans (i.e., smaller plans contributing to larger, overarching plans). Researchers have found no need for a monolithic, omnicompetent AI model with this division of labor. According to the Berkeley Artificial Intelligence Research (BAIR) Lab, State-of-the-Art (SotA) advanced AI applications are increasingly achieved by Compound AI systems with multiple components, not just a monolithic, omnicompetent AI model. In other words, BAIR Lab’s findings suggests the future of advanced AI applications lies in breaking down complex tasks into smaller, specialized subtasks and assigning them to domain-specific, industry-expert AI agents. Let’s explore a real-world example and hypothetical scenario:

Lab of AI Agents?

Researchers at Stanford University (“Stanford”) developed “The Virtual Lab,” a framework that enables four specialized AI agents to collaborate on scientific challenges.

“The Virtual Lab” relies on the Principal Investigator (PI), which coordinates tasks between the four specialized AI agents. For example, the PI holds structured meetings that allow the four specialized AI agents to discuss and refine their work.

Here’s the team of four specialized AI agents:

1. Immunologist: Proposes a scientific hypothesis related to the immune system, a complex network of cells, tissues, and organs that protect the human body from harmful infections.

2. Scientific Critic: Acts as a critical thinker, evaluating the strengths and weaknesses of scientific arguments and experimental designs.

3. Data-Driven Specialist: Extracts meaningful insights from complex datasets by identifying reoccurring patterns.

4. Computational Biologist: Applies computational techniques to biological problems.

Over 90% of the AI-designed molecules in “The Virtual Lab” were stable and worked as intended when produced in labs by human scientists. It even successfully designed 92 Nanobody candidates to fight the SARS-CoV-2 Virus, a strain of the Coronavirus (COVID-19). A Nanobody is a small, artificially designed antibody that can fight against various diseases.

A Hospital Run On AI?

Imagine if a hospital employed a Compound AI system for patient care (i.e., the core elements of the Compound AI system are highlighted).

The hospital uses Natural Language Processing (NLP) to scan patient records and extract relevant medical history, transforming unstructured text into actionable data. Computer Vision (CV) analyzes medical imaging, such as CT Scans, to detect abnormalities like tumors. Machine Learning (ML) frameworks use this combined data to predict patient outcomes and recommend personalized treatment plans, such as specific medications or interventions based on a patient’s relevant medical history and current condition. Optimization Algorithms then schedule doctor appointments. Finally, the Integration and Orchestration Layer (OL) seamlessly connects this entire process, ensuring that outputs from one component, such as medical imaging results from CV, flow directly into the ML frameworks to create a coordinated patient care system.

🦾HOW TO OVERCOME ITS UNIQUE RISKS

Compound AI systems have ongoing technical, social, and legal challenges (e.g., accuracy, trust, and clinician liability). However, these challenges also exist in independent GenAI models. What challenges are unique to Compound AI? Compound AI accomplishes large, consequential tasks with elements rarely mentioned in the AI safety literature: documentation of tasks, auditing, reporting, budgets, schedules, contingency plans, regulatory compliance, and mechanisms for ongoing reviews and revisions.

Bounded Tasks?

These elements require a new framework for organizing GenAI models to align with the organization’s objective: bounded tasks. Organizing work into bounded tasks creates a new framework where each GenAI model is connected to the next GenAI model but can only proceed if the previous GenAI model’s outputs are correct. Tasks have bounded goals to be accomplished in bounded time and with bounded resources because people demand results without too much cost and delay.

Data Access?

Compound AI systems should only have access to data as needed. This opinion is a fundamental principle of computer security called the Principle of Least Authority (POLA): users should be granted the minimum level of access necessary to perform their required tasks.

By ensuring that each GenAI model or each AI agent only has the permissions necessary for their specific task, POLA limits the potential for damage when failures occur. In large-scale projects, if a GenAI model encounters an issue, POLA ensures that the problem doesn’t cascade through the entire system, keeping other elements unaffected and maintaining overall system integrity.

🔑KEY TAKEAWAY

In this approach, superintelligent AI isn’t about creating a single, all-knowing AI model. Instead, it’s about building a system where specialized GenAI models and domain-specific AI agents work together, each focusing on their task. We can tackle complex challenges by organizing these AI applications into a Compound AI system with clear boundaries while keeping humans in the driver’s seat through crafting prompts and discarding outputs. Even as tasks grow in scope, Compound AI systems can remain efficient, secure, and aligned with our goals.

📒FINAL NOTE

FEEDBACK

❤️TAIP Review of The Week

REFER & EARN

🎉Your Friends Learn, You Earn!

You currently have 0 referrals, only 1 away from receiving ⚙️Ultimate Prompt Engineering Guide.

Refer 3 friends to learn how to 👷‍♀️Build Custom Versions of OpenAI’s ChatGPT.

Copy and paste this link to friends: https://theaipulse.beehiiv.com/subscribe?ref=PLACEHOLDER