Multiple AI agents working together on a single task can outperform a single powerful model. Andrew Ames' framework shows that assistive agents like code checkers and security engines built around GPT-3.5 beat one-shot GPT-4, with even bigger gains when stacked on GPT-4.
Every application that is being built today and in the future will have not just one foundation model but multiple foundation models. So you might use Visual for this, you might use Cohere for that, you might use OpenAI for something else within the same application. But agents is where the big step function is. Go and listen to Andrew Ames multi shot agentic framework approach. He looks at code development and he says let me develop code using GPT 3.5 and GPT 4. Dramatic difference in in terms of GPT scale and size and accuracy and so on. But then let me add assistive agents to each code development process. So with GPT 3.5 they built a code checker, they built a code security engine, they built a bunch of these things that do assistive help on the code development life cycle. They did that for 3.5 and they realized that it's getting you way better results than a one shot GPT4. Of course when you apply it to GPT4 it'll do even better. Multiple agents working together to solve a problem, it is where the industry is headed towards.
