AI 101 Live Session, 2024-03-01


Recorded live AI Salon / AI 101 session on AI topics.

Next: AI 101 Live Session, 2024-03-02

Previous: AI 101, Let's Learn About APIs (and a tiny bit of trying Claude API), 2024-02-24

AI Summary (hopefully useful, may be inaccurate)

Quick Recap

The team discussed image processing and workflow management. Pete shared his experiences and suggested using an AI app to enhance low-resolution images. R struggled with tracking tasks in different apps, and Leah recommended using Adobe Lightroom and actions in Photoshop to standardize the process. The team also discussed managing and selecting images in a workflow, with Pete suggesting the use of Phoenix Slides or xnView tools. The conversation concluded with a discussion on the pervasiveness and persistence of metadata in images.


Image Processing and Workflow Management

The meeting revolved around the topic of image processing and workflow management. Pete shared his experiences and potential topics for future discussions, while also providing a recommendation for an AI app to enhance low-resolution images. R shared his struggles with tracking tasks in different apps, which ended without a clear resolution. Leah suggested using Adobe Lightroom and actions in Photoshop to standardize the process and using a Google Doc to manually document steps. Pete agreed with these suggestions and discussed his workflow for his subscription service, which involved using tools like Typora, Obsidian, and a command line tool for auto-caption generation. He stressed the importance of keeping workflow documentation updated and demonstrated the use of Get and Github. The conversation concluded with a discussion on managing and selecting images in a workflow, with Pete suggesting the use of Phoenix Slides or xnView tools.

Image Workflow and Collaborative Tool Development

Pete discussed his method of working with a set of images, which involved marking images with different groups using a function key, clearing all marks, or selecting specific images. He also demonstrated how to organize images into different directories and tag them using different groups. Pete then explained his image workflow process, which included taking a series of images, applying automated processes, and selecting the best images for further editing in Photoshop. R brought up the use of similar methods when creating and tracking results from prompts using a tool called Hack MD. Pete and R also discussed the development of a collaborative tool for generating Gpt's, proposing the use of tables for deciding topics and tracking versions of prompts. They highlighted the importance of version control and the potential use of tools like Obsidian with the Gpt plugin. They also discussed the importance of real-time collaboration tools and the need to keep track of these tools. Finally, they agreed to work together to integrate these tools and discussions into the AI 101 website.

Image Post-Processing Workflows and Metadata Discussion

The team discussed workflows for image post-processing, with Kyle's list being appreciated. Pete emphasized the importance of order in these workflows and touched upon the use of AI enhancers. A doctor raised a question about metadata embedded in images, to which Pete confirmed there was no metadata. Pete highlighted the challenges of stripping metadata from an image to use in a tool like Mid Journey. They also discussed the pervasiveness and persistence of metadata in images, with Pete elaborating on how AI can upscale and invent detail in images. Towards the end, Paul inquired about the difference between the results obtained with blend and style reference, but no response was provided. The conversation concluded with a mention of a new system in development.

Image Generation and Spreadsheet Tools

Pete and the team had a discussion about generating images using prompts and various styles. They experimented with different aspect ratios and suffixes, and how these could affect the generated images. Pete also mentioned the possibility of expanding the prompt to multiple variations. Additionally, R shared about a Windows Power Toys app, Crop and Lock, which they found useful for focusing on specific items in a spreadsheet.

AI-Generated Image Naming and Pixie Tool Discussion

Pete and R discussed the categorization and naming of AI-generated images, considering style names and camera aspects. Pete demonstrated how to adjust settings for a cinematic effect and mentioned that he sometimes uses an AI to tag images. Later, Pete and Vikki had a discussion about Pete's tool, 'Pixie', which generates image descriptions based on user prompts. Vikki shared her experience of purchasing a book due to its descriptive writing and potential for AI prompts. They also explored the limitations of the tool, particularly its difficulty in replicating accents, and experimented with it, checking the number of tokens used and the cost.

Gender Representation in AI Images and Online Groups

Pete and Vikki discussed the representation of genders in AI images and online groups, expressing concern about the predominance of male groups and the underrepresentation of diverse individuals. Pete shared his approach of featuring more diverse individuals and his plans to compile insights on this issue, while Vikki suggested raising the issue in the AI ethics space. The conversation also touched upon cultural appropriation and sensitivity, with Pete seeking guidance on ethical considerations. Towards the end, Vikki and Dr. pointed out the need for more diversity, specifically requesting more representation of men in Pete's images.

Language Models' Functionality and Limitations

The participants discussed the functionality and limitations of language models. They clarified that these models are typically not in training mode during conversations but rather in inference mode, responding to questions and making inferences. They also noted that these models have a short-term memory and do not permanently learn from conversations. The conversation also touched upon the topic of AI training and the need for models to be exposed to diverse content. Towards the end, they discussed the challenges of maintaining workflows and the difficulty of going back to tasks once missed.

AI Training and Feedback Discussion

Pete and the team discussed their plans for the next day. There was also a conversation about AI and its training process, with Vikki emphasizing the importance of providing feedback to AI systems. She pointed out that just as people need to be corrected when they use words incorrectly, AI must also be trained to use language more appropriately. The team agreed on the need to adjust their approach to AI training.