Open AI's GPT models: Microsoft's Swiss Army Knife? Google vs Microsoft & Sparrow VS GPT-4?

Previous articles (1 & 2), briefly covered the intentions of Microsoft to integrate Open AI’s GPT?models into Bing, as well as the office productivity suite. That was far from unexpected especially after ChatGPT caught the world by storm and Open AI witnessed an incredible growth of awareness for it. This was also confirmed when Google asked via its official Twitter account?what people are going to Google search, first thing in 2023 and many responded ChatGPT.?

It seems that for Microsoft, Open AI’s GPT?models have currently become its Swiss army knife of AI.?A multitool that brings endless functions/uses and can be applied in many scenarios to build the next-gen applications, just in perfect sync as we are entering the age of mass adoption of AI solutions.

In the next months, the rapid advancements expected in AI due to the competition between Microsoft, Google and others will increase the availability of such AI tools but also potential impacts and risks. Microsoft’s recent promotion of ChatGPT in public, introduced extra pressure on Google that had no other option but to respond with its recent announcement of its own ChatGPT like chatbot, named “Sparrow”. The latter which might go into “private beta” this year, is trained with human feedback, and will access the Internet through Google, allowing it to tap into up-to-date information for its responses (i.e. a plus compared to ChatGPT). It is based on Deepmind’s Chinchilla language model and although it comes with fewer parameters than Open AI’s largest models, it is still expected to perform similarly or even better than ChatGPT.

On the other hand, there is also lots of speculation on Open AI's GPT-4 (i.e. the next gen supposed to outperform the latest GPT-3.5) which is expected to be released soon (i.e. perhaps 1st quarter of 2023). GPT-3 has 175 billion parameters and although there are rumours that GPT-4 will have trillions, it’s more likely that Open AI will follow a different approach and try to get more out of similar parameter numbers instead. As a result, the image below is probably not what is going to happen.

The reason is that the number of parameters is not the only factor and does not guarantee that the AI model performance will bring huge benefits especially since there is a trade-off to be made against costs. Larger models often come with increased costs to fine-tune and GPT3 was such a case where it was trained only once for that reason. Indeed, Open AI’s co-founder, Sam Altman mentioned that GPT-4 might focus on higher-quality data and better algorithms to bring more progress.

Regarding the timeline, if we take into account Microsoft’s plans to integrate into Bing and/or the Office suite, it is possible that GPT-4 release plans might be slightly shifted at a later time. Add to this also the fact that after all the hype and the numerous articles and opinions on GPT risks, shared on the social media, it goes without saying that Open AI will first emphasize on accuracy, safety and responsibility. The company wishes to continue capitalizing on the positive fame of their technology and not risk a rushed release of GPT-4 that might not produce equally or even better results (i.e. than GPT3.5 or Sparrow) or worse fail in sensitive areas. For this reason, Open AI researchers collaborated with Georgetown University’s Centre for Security and Emerging Technology and the Stanford Internet Observatory to investigate how large language models might be misused for disinformation purposes (here).?

Sam Altman also talked about a generative AI system for video (i.e. multimodal) as Google’s text-to-video like Imagen Video and Phenaki, and Meta’s, Make-a-Video. Such multimodal models accept text, audio, image, and even video inputs to generate audiovisual content. Once again competition pushes innovation & improvements in AI solutions.

In any case, although such solutions open up new possibilities in many diverse fields (i.e. workplace, health, law, education, science etc), they still need to earn their way into people’s trust. If nothing else, history has taught that new technologies can be easily misused and in this case the more AI penetrates in our lives, the more the risks and the impacts.

Technological advancements are usually rushed with the intention of profit and at the expense of being much more ahead of any ethical, social, psychological or other important adaptations and considerations first. ?

Following the analogy of the Swiss army knife, it is such a useful tool but for anyone who owned one, they know you need to be extra careful when handling, as the blades are extremely sharp.?