OpenAI was the pioneering platform to publicly release its ChatGPT Artificial Intelligence platform for free, utilizing the GPT-4 model. On the other hand, Google took more than six months to launch Bard, its conversational AI platform based on PaLM, distinguishing it from OpenAI’s GPT model and Microsoft‘s OpenAI and Bing Chat. While both platforms offer similar capabilities, the underlying learning models employed differ. Understanding these distinctions is crucial to maximizing their potential according to individual needs.
Regarding the amount of training data used to develop their respective AI systems, both Google and OpenAI have not disclosed specific information. This omission is likely intended to circumvent potential concerns and controversies regarding the usage rights of the employed data. As Bard from Google is now available, albeit not in the European Union, and ChatGPT has encountered usage restrictions in certain countries, it is highly probable that copyrighted materials have been incorporated into their training processes.
ChatGPT excels at providing information on a wide range of topics without specific subject expertise. However, Google has dedicated a significant portion of Bard’s training to logic, mathematics, reasoning, and science, all interconnected domains. Consequently, if we seek reliable and accurate information in these subjects, Google’s solution would be preferable as it is less prone to inventing answers when faced with unfamiliar topics, unlike ChatGPT.
Both ChatGPT and Bard utilize natural language processing to facilitate information retrieval, with similar capabilities in the Cervantes language. However, when it comes to other languages, ChatGPT’s functionality is more limited, as it has not been extensively trained on multiple languages. In contrast, Google’s solution boasts training on over 100 languages, allowing for better comprehension of questions and translation capabilities, as claimed by the search giant.
To enhance the responses of OpenAI’s GPT-4, a significant portion of its training has involved human supervision to foster a deeper understanding of the text and enable more comprehensive and coherent responses. While Google hasn’t explicitly mentioned this aspect, it is highly likely that they also incorporate human feedback to ensure not only correct but sensible answers.
In terms of performance, GPT-4 outperforms Google’s PaLM2 in the MMLU test (language compression) with a score of 86.4 compared to PaLM2’s 81.2. Similarly, in the HellaSwag test, which assesses AI’s common sense, GPT-4 achieves a score of 95.3, surpassing PaLM2’s 86.8. These results suggest that currently GPT-4 holds a slight advantage over Google’s solution. However, it is probable that Google’s Bard will improve over time and approach the performance levels of GPT-4.
A significant distinction between PaLM2 and GPT-4 lies in their potential future business models. While Google’s Bard is not available in Europe, it is entirely free and not restricted by any paywalls. The same applies to Microsoft’s Bing Chat, which is also based on GPT-4. On the other hand, accessing OpenAI’s ChatGPT based on GPT-4 requires a mandatory checkout, as the free version of the platform currently relies on the previous GPT-3.5 version.