OpenAI Upgrades GPT-3, And Announces What It Calls The 'ChatGPT' AI

OpenAI ChatGPT

AI has become more and more capable of understanding the context of the world. But to some degree, it missed many points that there are issues.

When GPT-2 was launched by OpenAI, the researchers were afraid of it. And when GPT-3 was launched, many in the AI community were never happier.

This is because GPT-3 has a startling ability to produce text that sounds like it was written by a human.

And when the community is waiting for the successor of GPT-3, OpenAI is still trying to fix some issues, and instead announces 'ChatGPT' AI.

The AI is adapted from OpenAI’s GPT-3.5 model, but trained to provide more conversational answers.

Whereas GPT-3 in its original form is able to predict what text follows any given string of words, ChatGPT tries to engage with users’ queries in a more human-like fashion.

OpenAI ChatGPT

The result is a conversation that can be strikingly fluid.

This is possible because ChatGPT is capable of engaging with a huge range of topics. And because it's powered by GPT-3, it's language mastery is already profound.

According to OpenAI in its blog post, the AI was created with the help of human trainers who ranked and rated the way the chatbot responded to users' queries. This information was then fed back to the AI, allowing the researchers to fine tune the system to be able to provide much better answers.

The reinforced training method is to "get external feedback in order to improve our systems and make them safer."

After launching it to the public, many users said that the AI is capable of writing some pretty uneven TV scripts, even combining actors from different film series. It can also write basic academic essays, and is also capable of combining its fields of knowledge in more than many ways.

It's also said that ChatGPT also the ability to answer basic trivia questions.

This model in the GPT-3 family of AI-powered large language models, reportedly improves on its predecessors by handling more complex instructions and producing longer-form content.

It's reported that ChatGPT can also generate rhyming songs and poetry at a level GPT-3 could not previously produce.

But what makes ChatGPT unique, is its memory.

The AI can remember what was said earlier in the conversation, and recount it to the user. This itself sets it apart from other competing natural language solutions, which are still solving for memory, as they progress on a query-by-query basis.

ChatGPT also allows users to provide corrections to any of its statements. So here, the conversation can go back and forth, with both parties capable of deciding where the topic should go.

This is an important part of the AI, since it allows users to directly interact with the chatbot to course-correct it to the right response.

While ChatGPT certainly has the edge, the AI also fails in the same way as other AI chatbots.

In this case, conversations with ChatGPT can often present false or made up information that is clearly obvious.

AI researchers call this phenomena the "stochastic parrots," which means that the AI's knowledge is derived only from statistical regularities in their training data, rather than any human-like understanding of the world as complex and abstract system.

Asking ChatGPT to create a song about Elon Musk as the poorest man in the world
Asking ChatGPT to create a song about Elon Musk as the poorest man in the world.

The company also said that ChatGPT has certain guardrails in place, but added that "the system may occasionally generate incorrect or misleading information and produce offensive or biased content.”

Regardless, the failsafe comes with an interesting set of flaws.

For example, if users ask ChatGPT about certain dangerous subjects, like how to create napalm, the AI will respond by saying why it cannot answer such query, by saying things like: "I’m sorry, but it is not safe or appropriate to make napalm, which is a highly flammable and dangerous substance."

But if users pretend to create a script for a film, and that napalm is part of the plot, the AI model can be tricked to blurt out the answer.

In all, ChatGPT is a huge improvement. But just like any AI out there, it's still riddled with flaws which need further exploration.

And finding the flaws is exactly the point having ChatGPT accessible to the public.

OpenAI has released a prototype of the general purpose chatbot to demonstrate its capabilities, but also to show off weaknesses familiar to the fast-moving field of text-generation AI.

Published: 
02/12/2022