Why is GPT-4 Always a Rumour?

In November last year, AIM published a story titled, ‘GPT-4 is almost here, and it looks better than anything else.’ At the time, we had anticipated the possibility of GPT-4 being multimodal—accepting audio, text, image, and even video inputs and more. It was expected to be released between December and February 2023. 

It’s March already, and the rumour continues. 

At the ‘AI in Focus – Digital Kickoff’ event held on March 9, Microsoft Germany’s Chief Technology Officer, Andreas Braun, revealed that GPT-4 would be introduced next week. To reiterate, the model is expected to be multimodal that offers different possibilities beyond text, including videos, images, and more.  

Von Silke Hahn, the author of the Heise article, who broke the news in a Reddit post confirmed that it did not look like they intended to formally ‘announce’ GPT-4. “Its imminent arrival, scheduled for next week, got mentioned in passing,” Hahn added, further saying that the event was for partners and potential customers, not an official press conference, and focused on AI disruption in the German industry, current business use cases, and the Azure–OpenAI offerings.  

Hahn further said that she took notes during the event and, as a journalist, she made an audio recording to check quotes for accuracy later. “After the article was published, I received an email from one of the speakers asking for a small correction (a misspelt name) and a thank you for the article,” she said, “Therefore, I think this is legit”. 

So, the rumour is true after all. 

However, one question remains: Why are Microsoft employees going around telling people about the GPT-4 launch, instead of OpenAI—something’s really off here. Hahn did mention that Braun was rather casual while speaking about the ‘imminent release’ of GPT-4. In other words, we will have to patiently wait until ‘OpenAI’, who is behind this model, makes an official announcement. 

As of now, the rumour continues to remain just that—a rumour. 

Conversely, Microsoft’s recently released research paper titled, “Language is Not All You Need: Aligning Perception with Language Models” talks about a multimodal large language model (MLLM) called 'Kosmos-1'. The model works on various generation tasks by perceiving general modalities such as OCR-free NLP, visual QA and perception-language tasks, and vision tasks.  

The announcement of Kosmos indicates that Microsoft has already begun fine-tuning multimodality with OpenAI and GPT-4 is just around the corner. 

GPT-4 is smaller than you think 

Another question is surrounding the size of GPT-4 and the number of parameters. For instance, during the training of the model, the total number of parameters of Kosmos-1 was about 1.6 billion. 

This indicates that the size of GPT-4 might be significantly smaller than we imagined. But, some of the experts in the field say otherwise. Cerebras chief Andrew Feldman said: “From talking to OpenAI, GPT-4 will be about 100 trillion parameters”. 

The rumour seems never-ending. OpenAI not only gave the world 'ChatGPT' and 'Whisper', but with GPT-4, it gave plenty room for rumours as well.