YouTube: https://youtu.be/Qe9NRsARZ8U
Recorded live AI Salon / AI 101 session on GPTs, APIs, etc.
Next: Mega prompts with guest Lee Chazen, 2024-01-16
Previous: GPTs and APIs, 2024-01-04
Pete initiated a discussion about potentially setting up a hack Md and proposed exploring the topic of mid journey. The team experienced some technical difficulties with the mid journey tool, which affected their ability to use it effectively during the meeting. Despite the issues, Pete continued to demonstrate its use with saved images, while other team members shared their experiences and concerns. The exact resolution to the technical issues was not provided in the transcript.
Pete discussed the performance of a tool called "mid-journey", noting its initial struggles with short visual descriptions but later adaptation to longer descriptions. They stressed the importance of having one or two decent images per grid and demonstrated how to use a visual description prompt to create similar prompts with different subjects. Pete also discussed the quality and creativity of different descriptions, comparing them to the work of Nick. They highlighted the vibrancy and detail present in the descriptions and expressed their intention to pick interesting descriptions and apply a slight variation of Nick's approach. Pete also discussed the concept of suffixes, aspect ratios, and demonstrated how to use a system that generates visual descriptions. Towards the end, they noted some issues with the people and bench in the generated images, indicating room for improvement.
Pete discussed the default rendering style of Mid Journey, which tends to create visual descriptions akin to movie scenes. They suggested that the system might guess subjects when not explicitly provided, and the output often resembles a movie scene. Pete also experimented with different prompts, such as "iPhone photo" and "camera selfie", and observed the system's reaction to these prompts. They noted that the system seemed to produce more people in the output when "movie still" was specified, but the image quality did not change significantly. Pete also pointed out the aspect ratio, suggesting it might contribute to the movie still aesthetic.
Pete and Mr discussed how to use the Mid Journey bot. Pete explained that the bot can be accessed through a channel with many users where images from Mid Journey are shared. They further clarified that anyone can use the bot and there's no need to add it to a server. Instead, you can direct message it. Pete also mentioned a feature called 'stealth mode' which requires a higher plan. They also explored how to view and use generated images from Mid Journey, including the ability to search for similar images. There was also a discussion about the legal implications of using AI-generated content, with Pete stating that it's currently not copyrighted in the US.
Pete and Mr. discussed a selfie generation project, noting issues with the aspect duration and suggesting adjustments to the camera aspect ratio. They also discussed the potential of medium.com, with Mr. planning to share something if they finds it useful. Pete suggested working on variations and changing the aspect ratio to resemble an Instagram post. Both agreed on the need to increase discussions about AI in 2024. Mr. expressed interest in learning more about AI and creating APIs for their GBDS, and in simplifying the explanation of AI for non-technical people, particularly baby boomers, and reaching a Spanish-speaking audience. Pete acknowledged these points and appeared to be considering how to address these needs.
Pete and Mr discussed the challenges faced by rural schools in South America and the potential role of Instagram in addressing them. They considered linking Instagram or a website to automatically share products, and sharing small images and links on the platform. They also discussed the use of Gpt builder to assist non-technical people in Spanish, exploring its features such as a chat function and a coding feature. Pete demonstrated how to create a simple Gpt and shared it with others through a link, which sparked interest from Mr for further discussion.
Pete, Mr, Cori, and Ethan discussed the addition of more buttons in the configuration and their potential impact on the system's appearance. Ethan's work on using AI to generate ideas was introduced, which involves a five-phase process resulting in a Word document of accumulated ideas. The team also discussed the idea of using unclickable prompts as space for user instructions. The topic of jailbreaking GPTs was brought up, with a focus on anti-jailbreaks. Pete suggested creating one's own anti-jailbreak and encouraged experimentation. The team also considered adding an email or website contact to their GPTs for users seeking more information. Cori and Pete shared their experiences with using anti-jailbreaks and emphasized the importance of testing them. Towards the end, the idea of creating a Chinese version of their GPT for users in Asia was considered.
Pete and Cori discussed issues with a system designed to translate and execute commands, with Pete expressing concern over its inability to execute certain commands effectively. They agreed to test and troubleshoot the system, with Cori suggesting the use of other commands. They also discussed potential issues with a chatbot's responses, with Pete expressing concerns over its lack of originality and inability to redirect users effectively. Pete suggested altering the bot's responses and they agreed to implement these changes. They also discussed the bot's ability to translate and its potential use in preventing jailbreaking. Additionally, Pete introduced a personal information manager called 'obsidian'. They also discussed the concept of anti-jail breaking and how to achieve it, with Pete suggesting rewording instructions to include the idea of paraphrasing them. Cori proposed copying Pete's effective anti-jail break and incorporating it into their confidential documents, which Pete agreed to but recommended rephrasing the instructions to be more comprehensive.
Pete proposed several architectural approaches for safeguarding information in a GPT system, emphasizing the importance of not fully integrating valuable assets due to the risk of theft. They suggested creating a free introductory session as a solution. Cori agreed with Pete's suggestion and joined them in discussing the potential of AI technology, specifically the creation of a bot. They explored the possibility of monetizing it by offering free, high-value resources and charging for exclusive, advanced content. They also discussed the concern of users 'jailbreaking' the bot and the potential of using a confidentiality agreement to deter such actions. The conversation ended without a clear decision on the next steps, but with plans to continue the discussion in the future.