GPT-5 Preview! What new capabilities will GPT-5 have?

Sam Altman is regarded as a very influential figure in the entire field of AI and even the entire technology field.OpenAIThe endless reversals of the palace fighting made Sam Altman's presence felt, and he was even named "CEO of the Year 2023" by Time magazine.

For this reason, a tweet from Sam Altman can instantly become a signal that will shock the entire AI industry, especially when this tweet is also related to the much-anticipated “GPT-5”When it’s relevant.

On Christmas Eve 2023, Sam Altman boldly announced his ambitions for 2024 on social media. The keywords he published not only covered OpenAI's overall plan for 2024, but also met the urgent needs of users. These include:

AGI (Please be patient)

GPT-5

Better voice mode

Higher rate limits

Better GPT

Better reasoning

Level of control over work/behavior

video

Personalization

Better browsing

"Log in with OpenAi"

Open Source

Sam Altman revealed that OpenAI plans to achieve several impressive milestones in the next year. What is involved is not just a simple technical update, but a crucial AI revolution. Of course, the most popular one is GPT-5.

The open source and closed source debate in the AI field is similar to the battle between Android and IOS.

For the AI community, in addition to being concerned about whether GPT-5 can make a breakthrough in technical barriers, they are more concerned about one key point: Can GPT-5 be open source?

The debate over whether to open source or close source large models has always been a focus of debate in the industry. This debate is similar to the debate between Android and IOS in the mobile Internet era. Interestingly, the choices of various AI giants on open source or closed source are also different.

Currently, OpenAI's GPT-4 and Baidu's Wenxin Yiyan, which are both in the lead, insist on closed source. Meta has chosen the open source path and has successively opened the LLaMA and LLaMA-2 models for "academic research purposes". Baichuan Intelligence has both open source and closed source. In the academic field, it has chosen open source and uses two large models of 7B and 13B. In commercial exploration, it has closed the 53B model to protect commercial interests and the competitive advantage of technology.

GPT's closed source has brought OpenAI considerable income. According to The Information, OpenAI CEO Sam Altman told employees,The company is generating revenue at a rate of $1.3 billion (about RMB 9.493 billion) per year, with an average monthly revenue of more than $100 million, which is more than 450 times the $28 million for the whole of last year, reaching 45,42%. This figure is also 30% higher than the annual revenue expected three months ago. This also makes 2023 the year with the fastest revenue growth in the eight years since the establishment of OpenAI.And these are exactly what the closed-source nature of GPT-4 brings.

Keywords: Can GPT-5 be open source?

So, can GPT-5 be open source? Not necessarily.

Regarding the business model, OpenAI has clearly stated on its official website that it "intends to continue to provide ChatGPT for free," but will also choose to sell it to paid users.advancedIncome is earned from the users and businesses it serves. Moreover, although OpenAI says that it "does not expect to make a profit in the near future", considering the high cost of developing and providing large models, survival is still a challenge it has to face.

In addition, despite OpenAI's rapid growth, the industry costs behind it cannot be ignored. According to public information, in 2022, OpenAI developed GPT-4, with training costs of about $540 million alone. In April 2023, OpenAI paid about $6.944 million in operating costs for ChatGPT every day (mainly electricity costs), with an annualized operating cost of about $250 million, and the comprehensive annualized cost may exceed $1.3 billion. There is no doubt that OpenAI is still in a loss-making stage.

Therefore, without commercial support, OpenAI may soon go bankrupt. More importantly, OpenAI, which has already tasted the sweetness of closed-source GPT-4 and earned a lot of money, obviously has no sufficient reason to fully open source GPT-5, which is tantamount to self-destruction and losing its leading edge in the competition of large models. From this perspective, the probability of GPT-5 being open source is not high.

Even though Sam Altman marked "open source" as a keyword in his tweet, it was more of a response to the industry's call. We cannot interpret it as "the company's development goal for 2024."

However, the possibility of "partial open source" is not ruled out. Although the possibility of GPT-5 being fully open source is relatively small, for the sake of GPT-related ecological construction, the possibility of providing open source for GPT-related tool sets is very high.Perhaps, OpenAI will facilitate developers' development, debugging, and sharing by providing open source for a smaller number of parts.

What new capabilities will GPT-5 have in the future?

Recently, the Allen Institute for Artificial Intelligence released Unified-IO2, which is of great significance because it can help us better predict the capabilities of GPT-5.

Why do you say that? What is the relationship between Unified and ChatGPT?

In fact, as early as June 2022, the Allen Institute for Artificial Intelligence launchedFirstUnified-IO, the first generation of Unified-IO, is one of the first multimodal models capable of processing images and language. Around the same time, OpenAI is testing GPT-4 internally and will officially release it in March 2023. Therefore, Unified-IO can be seen as a preview of future large-scale AI models. In other words, because of the emergence of Unified-IO2, we can roughly predict that OpenAI may be testing GPT-5 internally and is likely to release it in a few months.

GPT-5 Preview! What new capabilities will GPT-5 have?

Unified-IO2, launched by the Allen Institute for Artificial Intelligence, isFirstA model that can process and generate text, images, audio, video, and action sequences.This newadvancedThe artificial intelligence model is trained using billions of data points, and although the model size is only 7B, it demonstrates the most extensive multimodal capabilities to date.Its training data includes: 1 billion image-text pairs, 1 trillion text tags, 180 million video clips, 130 million images with text, 3 million 3D assets, and 1 million robot agent motion sequences.The research team combined a total of more than 120 datasets into a 600TB package covering 220 visual, language, auditory, and action tasks. Unified-IO2 adopts an encoder-decoder architecture and makes some changes to stabilize training and effectively utilize multimodal signals.

The model can answer questions, write text according to instructions, and analyze text content; it can recognize image content, provide image descriptions, perform image processing tasks, and create new images based on text descriptions; it can generate music or sounds based on descriptions or instructions, as well as analyze videos and answer questions about them. In addition, by training with robot data, Unified-IO2 can also generate actions for robotic systems, such as converting instructions into robot action sequences. Due to multimodal training, it can also handle different modalities, for example, marking the instruments used in a certain track on an image.

Overall, Unified-IO2 performs well on more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. In most tasks, it is able to match or even outperform specialized models. On the GRIT benchmark for image tasks, Unified-IO2 achieved the currentHighestThrough these, we can also get a better glimpse of what GPT-5 will look like in the future.

For the development of AI, the science and technology ecology and commercialization are indispensable core elements. The development of technology and applications requires the necessary support and guarantee of commercialization; and the success of commercialization is also inseparable from the construction of the ecological environment. The two must complement each other and be organically combined. It is hoped that in the future release of GPT-5, OpenAI can play a leading role and take the lead in achieving a balance between ecology and commercialization.

Text reference:

https://baijiahao.baidu.com/s?id=1787599025284931811&wfr=spider&for=pc&searchword=GPT-5

https://k.sina.com.cn/article_1667925927_636a87a70190118py.html

https://baijiahao.baidu.com/s?id=1786220479790922625&wfr=spider&for=pc&searchword=GPT-5

statement:The content is collected from various media platforms such as public websites. If the included content infringes on your rights, please contact us by email and we will deal with it as soon as possible.
Information

Apple closes San Diego AI team, data operations annotation team moves to Austin and merges

2024-1-15 11:47:32

Information

IMF warns: 40% jobs worldwide will be impacted by AI

2024-1-16 9:54:22

Search