Huang Renxun talks about OpenAI joint creation: GPT-4’s reasoning ability has not met expectations – WSJ

“As early as 2015 and 2016, we couldn’t even imagine that ChatGPT could reach the current level.” OpenAI co-founder and chief scientist Ilya Sutskever said in an “interview” with Nvidia founder and CEO Huang Renxun.

This conversation took place on the second day after the launch of the multimodal model GPT-4, on March 22 local time, Nvidia officially released a complete video of the conversation. During the nearly one-hour conversation, Huang Renxun and Sutskever talked about the development process of the chat robot ChatGPT, the progress of GPT-4, and the prospect of the AI ​​(artificial intelligence) industry.

ChatGPT is a chat robot released by OpenAI in November last year. It has become one of the most popular computer applications less than half a year after it came out, and it has also driven major technology companies to compete in the field of generative AI.

Huang Renxun mentioned many times that “the iPhone moment in the AI ​​industry has arrived”, but as one of the founders of OpenAI, Sutskever said that around 2016, they did not dare to imagine that ChatGPT could reach the current level.

“One of our ideas at the time was to allow machines to have unsupervised learning capabilities (unsupervised learning). Now it seems that this is a matter of course, people can use natural language models to train AI. But in 2016, this is an unknown field that no scientist has had the experience and insight to do so,” Sutskever said.

“Earlier, I also thought that ‘learning’ can only be done by humans, and computers cannot.” But between 2002 and 2003, his thinking changed, “If computers can be allowed to learn continuously, perhaps It will bring changes to the AI ​​industry.”

He believes that the introduction of neural networks into the field of AI is a very important advancement. It works like a human brain, capable of “deep learning” so that its logic can be applied to a computer.

There is a very solid theoretical foundation behind “deep learning”. If the computer’s neural network is deep enough and large enough, it can solve deep-level hard-core problems, but this requires sufficient database and computing power to support it. Therefore, OpenAI has put a lot of effort into optimizing the data model, such as making “neural network feedback”, and training the neural network to make it larger and obtain more data.

The working principle of ChatGPT is actually to train the neural network system and let the neural network predict the next word. Take mystery novels as an example. A novel contains various characters and plots. Readers can speculate on the murderer based on clues during the reading process. This is very similar to the working method of the GPT series.

“We want it to predict words with a certain logic, which is consistent with past training texts.” Sutskever said, “We are not simply completing AI learning based on human experience, but reinforcement learning based on human feedback. Feedback is very important. Importantly, more feedback makes the AI ​​more reliable.”

The latest large-scale multimodal model GPT-4 was also one of the topics of this talk. OpenAI started training GPT-4 6-8 months ago, and it has made considerable improvements over ChatGPT in many ways, most notably that this new model can handle image content.

Sutskever said that the world is made up of pictures, and humans are visual animals, and one-third of the gray matter of the human brain is used to process images, so multimodal models can provide more help to users. For example, in math competitions, many questions require diagrams to be answered, and GPT-4 can interpret the diagrams, greatly improving the accuracy of responses.

He boldly predicted that some future version of GPT might not only be able to read images, but also generate graphs in responses.

Another feature of GPT-4 is based on a better neural network that can predict the next word more accurately. However, when asked whether GPT-4 had reasoning capabilities, Sutskever said it was difficult to define the term, but he thinks such capabilities may be realized in the near future.

“We believe that the reasoning ability of GPT has not reached the level expected before. If the database is further expanded and the business operation model is maintained, its reasoning ability will be further improved. I am confident in this.” He said.

At the end of the talk, Sutskever also predicted the future of generative AI, “Currently we just feed AI text and let it draw conclusions, but AI cannot verify the authenticity of these texts and their origin. The next step is to let neural The network verifies the authenticity of the data source and makes it aware of the user’s needs.”

Author of this article: Yang Yang, source: The Paper, original title: “Huang Renxun talks about OpenAI joint innovation: GPT-4 reasoning ability has not met expectations”