top of page
Writer's pictureBatiste Roger

Understanding Layer 2 AIs: The Critique.

Updated: Aug 22

A reflection by Batiste Roger, CTO of Odonatech.


In a world where artificial intelligence (AI) is revolutionizing our practices, it's crucial for leaders to understand emerging concepts. Today, let's dive into Layer 2 AI, focusing on a minimalist and straightforward implementation: the Critique.


The Critique
The Critique

This should provide you with concrete examples to understand

  • what it is,

  • how it works,

  • and why it's revolutionary in the financial sector,

all without needing any prior technical skills!


First and foremost, I would like to thank Daniel Guillemette from Diversico Finances Humaines, with whom I discussed this topic yesterday, and who inspired this article. Daniel is a visionary innovator and an inspiring optimist, a long-time partner of Odonatech. Perhaps because he is not a data scientist, Daniel has the talent to ask the right questions in universal terms.


Reminder of the definition of Layer 2 AI: Layer 2 AI represents a significant evolution from traditional generative AI models (Layer 1). It combines multiple Layer 1 models (and other types of functions, AI or not) to create a more intelligent, reliable system tailored to a specific use case.


In finance, this approach ensures the reliability, accuracy, and personalization that clients expect. AI-driven savings advice becomes feasible, along with many other seemingly utopian use cases. The key was to create intelligent AIs from simpler ones


Let's start with the basics: what is a Critique in AI?


A Critique is an AI that evaluates and critiques another AI to improve its performance.

As Daniel Guillemette would say: "Everyone knows that criticism is easy, but the art is difficult."


What is the purpose of a Critique in a Layer 2 AI (IAL2)?


In the example below, a Layer 2 system consists of two Layer 1 AIs. One AI is responsible for generating responses (like OpenAI’s ChatGPT), while the other AI critiques these responses (such as Anthropic’s Claude).


  1. User: "What is a PEL?"

  2. ChatGPT replies: "a Plan d'Epargne Long (Long-Term Savings Plan)" (which is incorrect).

  3. Instead of sending this mistake to the user, the system first passes it to the Critique (Claude) for verification. Claude responds: "No, PEL actually stands for Plan d'Épargne Logement (Home Savings Plan)."

  4. ChatGPT is then prompted again: "Answer the user’s question 'What is a PEL?', knowing that 'Projet d'épargne longue' is incorrect. Here is additional information to help you: 'No, PEL actually stands for Plan d'Épargne Logement.'"

  5. ChatGPT replies: "Plan d'Épargne Logement."

  6. The system then sends the validated response back to Claude for confirmation, and finally to the user. Happy ending.


 

This system is the prototype of a very simple Layer 2 AI.


  1. The user asks a question.

  2. The generative AI produces an answer (Layer 1 AI).

  3. The 'Critique' evaluates this answer (Layer 1 AI).

  4. If the answer is not satisfactory, the system asks the generative AI to try again.

  5. This process repeats until a quality answer is obtained.


Of course, this means there may be higher costs (both in terms of energy and API usage) and longer execution times. However, in financial consulting, reliability is always preferred over speed.


If you're wondering what prompt to give Claude to generate the critique above, that's a great question. I'll provide an example in the comments if this post gets 20 reactions.


Is LiLa just a Critique? (Hint: no)


Without revealing our secrets, I can share a few details about LiLa.


  • LiLa is much more complex than the example given. In this post and the following ones, I plan to show very simplified Layer 2 AIs to give you an insight.


  • That said, among all its modules, LiLa does indeed have a Critique. This module is responsible for verifying the validity of responses, applying security measures, and predicting whether users will be satisfied with the response (in short, it's not just "Claude").


  • However, relying solely on the Critique is costly. Therefore, we have developed techniques to generate accurate responses from the start whenever possible.


The Future of Generative AI in the Financial Sector


Key Takeaway: Layer 1 AIs are not inherently reliable; they need to be integrated into a system that supervises and utilizes them intelligently for specific use cases, balancing time, cost, and performance. This is what I call Layer 2 AI!


Layer 2 AI represents the future of artificial intelligence in business. By intelligently combining different AI models, it provides more reliable, adaptable, and effective solutions. Leaders who integrate this concept into their strategy will be better positioned to capitalize on the AI revolution.


Are you ready to explore the potential of Layer 2 AI for your business? Share your thoughts in the comments and feel free to contact me for further discussion.





Comments


bottom of page