ChatGPT-4o: At the verge of AGI v.01 Alpha - Part 2

If you had the chance to read my last article ,the first part of "ChatGPT-4o: At the Verge of AGI v.01 Alpha," you might have noticed that there wasn't much discussion about AGI. Let's take a step back and revisit what AGI is.

Artificial General Intelligence (AGI) refers to a type of AI that can understand, learn, and apply knowledge across a wide range of tasks at a human-like level. Unlike specialized AI, AGI has the ability to perform any intellectual task that a human can, demonstrating flexible and comprehensive cognitive capabilities.

In our previous discussion, we touched on a chatbot performing human-like tasks. By the end of the article, we saw ChatGPT evolving into InternGPT. This transformation is something we want to delve into further in this article.

Interactive Dialog

During the week following the release of ChatGPT-4o, we've observed that this new model behaves quite differently from its predecessors. Previously, we were accustomed to a chat model based on zero-shot answers, where we'd always get a response to any question, whether complete or not.

Now, with ChatGPT-4o, the model can ask clarifying questions to better understand its task before providing a response. This means it can inquire further to enhance the original prompt, resulting in more accurate and useful outputs. This improvement is made possible by OpenAI's recent release of a larger long-term memory, which helps maintain the context of an ongoing conversation for much longer periods.

These advancements—long-term memory and pre-planned task execution—are key steps towards the highly anticipated "Voice" Assistant unveiled last Monday.

Self-evaluation

Another thrilling feature is the introduction of self-evaluation, which significantly helps prevent hallucinations and ensures safer, more reliable outcomes. Just a few weeks ago, implementing such a feature would have involved a multi-agent system where one agent generates a response, and another evaluates it, sending it back for revision if necessary.

Now, ChatGPT-4o can autonomously assess its own responses, making the process more efficient and robust.Whether it’s dealing with external sources that failed (like fetching a result with a network issue) or handling code-interpreter execution errors (as the model's stability can still be somewhat unreliable).

That's why we consider the true beginning of AGI for everyday consumer use.