It is very difficult to determine the “best” model in artificial intelligence. Because it is difficult to determine the categories to be used as criteria here. The data on which the models are trained and the structure of this data directly determine the output obtained from the model. Therefore, the best way to rank these tools is to look at the accuracy of their outputs. In this context, Vectara has published an AI hallucination chart ranking various leading AI chatbots according to their ability to ‘not hallucinate’.The best artificial intelligence models
- GPT4
- GPT 3.5
- Llama 2 70B
- Llama 2 7B
- Llama 2 13B
- Cohere-Chat
- cohere
- Anthropic Calude 2
- Mistral 7B
- GooglePalm
- Google Plam-Chat
All artificial intelligence models tend to reflect a phenomenon referred to as “hallucination”. So these tools can occasionally make up facts to fill in the gaps. These fabricated facts are made so elegantly that an unwary mind can be deceived very easily.
According to the table published by Vectara, which tested the models on the market in this context, GPT-4 showed the best performance with the lowest hallucination rate and the highest accuracy. Vectara tested these models on fixed source material. The GPT-3.5 model managed to take the second place on the list. Both of these models were developed by OpenAI and are currently used on ChatGPT.At the bottom of the list are two major language models from Google, and they clearly perform poorly. Google Palm Chat was the most unreliable model, recording a hallucination rate of over 27 percent on the material given to it. According to Vectara’s measurements, Palm-Chat’s responses are full of traces of hallucination.
50890
