GPT-5 Outlook Predictions! What New Capabilities Will GPT-5 Possess?
-
Sam Altman is regarded as an extremely influential figure in the entire AI field and even the broader technology sector. The dramatic power struggles at OpenAI in 2023 further amplified Sam Altman's presence, earning him the title of "CEO of the Year 2023" by Time magazine.
Therefore, a single tweet from Sam Altman can instantly become a signal that shakes the entire AI industry, especially when it relates to the highly anticipated "GPT-5".
On Christmas Eve 2023, Sam Altman boldly announced his ambitious plans for 2024 on social media. The keywords he shared not only outline OpenAI's overall strategy for 2024 but also align with users' pressing needs. These include: AGI (Please wait patiently)
Better voice mode
Higher rate limits
Better GPT
Better reasoning Control of work/behavior level
Better browsing
"Login with OpenAI"
Open source Sam Altman revealed that OpenAI plans to achieve several impressive milestones in the coming year. These developments are not just simple technical updates but represent a crucial transformation in AI. Of course, the most anticipated aspect is GPT-5.
The Debate Between Open Source and Closed Source in AI: Similar to the Android vs. iOS Rivalry.
For the AI community, beyond whether GPT-5 can break through technical barriers, a key concern is whether GPT-5 will be open-sourced. The debate between open-source and closed-source large models has long been a contentious topic in the industry, reminiscent of the Android vs. iOS rivalry during the mobile internet era. Interestingly, major AI companies have taken different stances on whether to open-source their models.
Currently, leading models like OpenAI's GPT-4 and Baidu's ERNIE Bot remain closed-source, while Meta has opted for open-source, releasing the LLaMA and LLaMA-2 models for "academic research purposes." Baichuan Intelligence, on the other hand, adopts a hybrid approach—open-sourcing 7B and 13B models for academia while keeping the 53B model closed-source to safeguard commercial interests and technological advantages.
GPT's closed-source strategy has proven lucrative for OpenAI. According to The Information, OpenAI CEO Sam Altman revealed to employees that the company is generating revenue at an annual rate of $1.3 billion (approximately ¥9.493 billion), averaging over $100 million per month—a staggering 4,542% increase from last year's $28 million. This figure also surpasses the revenue projection from three months ago by 30%, making 2023 the fastest-growing year in OpenAI's eight-year history. These gains are directly attributable to GPT-4's closed-source model. Keywords: Can GPT-5 achieve open source?
So, can GPT-5 achieve open source? Probably not.
Regarding the business model, OpenAI has explicitly stated on its official website that it "intends to continue providing ChatGPT for free," but it also plans to generate revenue from paid premium services for users and enterprises. Moreover, although OpenAI claims it "does not expect to profit in the near term," given the high costs of developing and providing large models, survival remains a challenge it must confront. Moreover, despite OpenAI's rapid growth, the industry costs behind it cannot be ignored. Public information shows that in 2022, OpenAI spent approximately $540 million just on training GPT-4. By April 2023, OpenAI was paying about $6.944 million per day in operational costs for ChatGPT (primarily electricity), with annualized operational costs around $250 million. The total annualized costs could exceed $1.3 billion. Undoubtedly, OpenAI remains in a loss-making phase.
Therefore, without the support of commercialization, OpenAI could soon face bankruptcy. More importantly, having already benefited from keeping GPT-4 proprietary and generating substantial revenue, OpenAI has little incentive to open-source GPT-5 entirely. Doing so would undermine its competitive edge in the large model race. From this perspective, the probability of GPT-5 being open-sourced is low.
Even though Sam Altman highlighted "open source" as a key term in his tweet, this seems more like a response to industry demands rather than a definitive corporate goal for 2024. However, the possibility of "partial open-source" is not out of the question. While the likelihood of GPT-5 being fully open-sourced is relatively low, there is a high probability that OpenAI may open-source toolsets related to GPT to facilitate ecosystem development. Perhaps OpenAI will provide open-source access to a smaller portion to aid developers in development, debugging, and sharing.
What new capabilities will the future GPT-5 possess?
Recently, the Allen Institute for Artificial Intelligence released Unified-IO2, which is highly significant as it helps us better anticipate the capabilities of GPT-5. Why is this said? What is the relationship between Unified and ChatGPT?
In fact, as early as June 2022, the Allen Institute for Artificial Intelligence introduced the first generation of Unified-IO, one of the first multimodal models capable of processing both images and language. Around the same time, OpenAI was internally testing GPT-4, which was officially released in March 2023. Therefore, Unified-IO can be seen as a forward-looking model for future large-scale AI models. In other words, with the emergence of Unified-IO2, we can roughly predict one thing: OpenAI may be internally testing GPT-5 and is likely to release it in a few months. Unified-IO2, developed by the Allen Institute for Artificial Intelligence, is the first model capable of processing and generating text, images, audio, video, and motion sequences. This new advanced AI model was trained using billions of data points. Despite its compact size of 7B, it demonstrates the most extensive multimodal capabilities to date. Its training data includes: 1 billion image-text pairs, 1 trillion text tokens, 180 million video clips, 130 million images with text, 3 million 3D assets, and 1 million robotic agent motion sequences. The research team combined over 120 datasets into a 600TB package, covering 220 visual, language, auditory, and motion tasks. Unified-IO2 employs an encoder-decoder architecture with modifications to stabilize training and efficiently utilize multimodal signals. The model can answer questions, compose text based on instructions, and analyze textual content; it can recognize image content, provide image descriptions, perform image processing tasks, and create new images from text descriptions; it can generate music or sound based on descriptions or instructions, as well as analyze videos and answer questions about them. Additionally, by training with robotic data, Unified-IO2 can generate actions for robotic systems, such as converting instructions into sequences of robotic movements. Due to its multimodal training, it can also handle different modalities, such as labeling the instruments used in a musical track on an image.
Overall, Unified-IO2 performs well across over 35 benchmarks, including image generation and understanding, natural language comprehension, video and audio understanding, and robotic operations. In most tasks, it matches or even surpasses specialized models. In the GRIT benchmark for image tasks, Unified-IO2 achieved the current highest score. Through these achievements, we can also get a better glimpse of what GPT-5 might look like in the future. For AI development, both the technology ecosystem and commercialization are indispensable core elements. The advancement of technology and applications requires commercialization to provide necessary support and safeguards, while the success of commercialization depends on the construction of an ecological environment. The two must complement each other and be organically integrated. It is hoped that in the upcoming GPT-5, OpenAI can take a leading role in achieving a balance between ecosystem and commercialization.