GPT4: AI Technology Aims to Improve on Imperfect Models, Nuance Still a Challenge
- Chat GPT struggled to understand the concept of generating a text of the same length in response to a question
- Chat GPT’s 3.5 version had bias which has been improved in GPT4
- Building AI in public allows people outside to provide feedback on both the good and bad parts of technology, however it leads to imperfect models being released
- Small details like counting characters are hard for these models to do well
- Jordan Peterson asked GPT to say positive things about two different presidents and asked it to rewrite with an equal number of characters, but it failed
- The nuance that these models can bring back is exciting.
OpenAI ushers in GPT4 with User Control and Safety Considerations
- GPT4 was released after extensive safety considerations
- Open AI developed a system message allowing users to have more control over their interactions with GPT4
- Alignment between human values and the model is still an ongoing issue
- System message has authority over GPT4, giving users more steerability
- There will always be jailbreaks of the system, but OpenAI is developing it to learn from the system message.