Is GPT-5 still far away?

By: Kyle

On August 1st, it was reported that OpenAI has officially filed a trademark application for "GPT-5," which covers the following:

• Software for artificially generating human speech and text
• Converting audio data files into text
• Sound and speech recognition
• Language and speech processing based on machine learning

According to the trademark application for GPT-5, it covers the functionality of AI-generated speech and text, the ability to convert audio files into text, achieve sound and speech recognition, and utilize machine learning technology for language and speech processing.

This may indicate that GPT-5 will support speech capabilities, bringing users a more advanced and efficient experience in speech and text processing, further enhancing multimodal capabilities.

When will GPT-5 arrive?

When GPT-4 was released in March 2023, it was expected that OpenAI would release the next generation model in December 2023. Siqi Chen, co-founder of Runway, previously stated that he was informed that GPT-5 is scheduled to complete training in December of this year, and OpenAI expects it to achieve Artificial General Intelligence (AGI). This means that there will be intense debate about whether it has truly achieved AGI.

However, during an MIT event in April, when asked if OpenAI was training GPT-5, OpenAI CEO Sam Altman stated, "We are not, and we won't for a while." In an interview in June of this year, Sam Altman, founder and CEO of OpenAI, when asked about the release of GPT-5, said, "I'm also curious, we don't have an answer, we won't have GPT-5 soon, we have to make safety a very important part."

Nevertheless, some believe that OpenAI may release GPT-4.5 before October 2023, which would be an intermediate version between GPT-4 and GPT-5, similar to GPT-3.5. It is said that GPT-4.5 will eventually bring multimodal capabilities, i.e., the ability to analyze images and text. OpenAI had already announced and demonstrated the multimodal capabilities of GPT-4 during the GPT-4 developer livestream in March 2023. Microsoft has now released the multimodal capabilities of GPT-4 in Bing Chat. It seems that the next major update for GPT-4 is on the horizon.

In addition, before starting research on GPT-5, OpenAI still has a lot of work to do on the GPT-4 model. Currently, the inference time for GPT-4 is very long, and the operating costs are quite expensive. Access to the GPT-4 API is still difficult. Furthermore, OpenAI has recently just opened access to the ChatGPT plugin and code interpreter, which are still in the testing phase. The internet browsing feature has been removed from GPT-4 because it displayed content from paid websites.

Although GPT-4 is very powerful, I believe OpenAI realizes that computational efficiency is one of the key elements for sustaining a model. By adding new features and functions, you can handle larger infrastructures while ensuring that all checkpoints start up and run reliably. Therefore, speculating boldly, if we assume that government agencies do not set regulatory barriers, GPT-5 is likely to be released in 2024.

Prediction: GPT-5 Features and Functions

• Reducing Illusions

The industry is hotly debating whether GPT-5 will achieve AGI (Artificial General Intelligence). In addition to this, GPT-5 should be able to reduce inference time, improve efficiency, and reduce illusions, among other things. Let's start with illusions, which are a key reason why most users are not very trusting of AI models.

According to OpenAI's data, in a factual assessment of internal adversarial designs in all nine categories, GPT-4 scored 40% higher than GPT-3.5. Now, the likelihood of GPT-4 responding to inaccurate and disallowed content has decreased by 82%. In accuracy tests across various categories, it scored very close to 80%. This is a huge leap in combating illusions.

It is expected that OpenAI will reduce illusions to below 10% in GPT-5, which would be significant in making the LLM model trustworthy.

• Computational Efficiency Model

We already know that the operating cost of GPT-4 is high (0.03 USD per 1K token), and the inference time is longer. The older GPT-3.5-turbo model is 15 times cheaper than GPT-4 (0.002 USD per 1K token). According to a recent report by SemiAnalysis, GPT-4 is not a dense model, but is based on an "expert mix" architecture. This means that GPT-4 uses 16 different models for different tasks, with 1.8 trillion parameters.

With such a massive infrastructure, the cost of running and maintaining the GPT-4 model becomes very expensive.

In fact, many new large models have begun to pursue being "small and precise," allowing large models to have as few parameters as possible, rather than more.

In a recent explanation of the Google PaLM 2 model, the PaLM 2 parameters are quite small, but the performance is fast.

• Multisensory AI Model

Although GPT-4 has been announced as a multimodal AI model, it only handles two types of data, namely, images and text. With GPT-5, OpenAI may take a big step towards achieving true multimodality. It will also be able to handle text, audio, images, video, depth data, and temperature. It will be able to interconnect data streams from different modes to create an embedding space.

• Long-term Memory

With the release of GPT-4, OpenAI introduced a maximum context length of 32K tokens, with a cost of 0.06 USD per 1K token. We have rapidly seen a transition from the standard 4K tokens to 32K in a few months. Recently, Anthropic increased the context window of its Claude AI chatbot from 9K tokens to 100K tokens. It is expected that GPT-5 may bring support for long-term memory through a larger context length.

This will help AI characters and friends remember your character and memories, and can last for many years. In addition, you can load books and text document libraries in a single context window. With support for long-term memory, various new AI applications may emerge, and GPT-5 can make this possible.

When do you think GPT-5 will be released, and what disruptive innovations do you think it will bring?

Reference: Beebom

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。