If I follow through on even half of these projects and regularly publish/publicize them then I think that I can gain quite a following as a document ai thinker. The missing bits from my performance has always been writing/communication and follow-through. It's time to change that with a regular writing habit, no need to focus on getting papers published. # Home Page - [ ] add publications to the home page - [ ] add sub-headings to the tinyhnsw posts - [ ] point jbarrow.ai to obsidian publish instead - [ ] add subheadings to AllenNLP posts - [ ] fill in more detail in the About Me section # Gemini Structured Mode - [x] decide on a dataset and task - [x] write the code for the dataset/task - [x] copy the code over - [x] write a tutorial for how to get an API key - [x] post on Twitter/Bluesky/LinkedIn # Slylock Fox - [ ] create a new series: Sunday Paper LLM Tasks - [ ] collect slylock fox comics - [ ] get bounding boxes for each panel - [ ] get answers for each panel - [ ] code to run on gemini flash - [ ] code to run with 4o - [ ] run through o1 - [ ] write intro about why it's a challenging but important task # Cryptoquips - [ ] collect a bunch of cryptoquips - [ ] come up with an evaluation metric # LMM Landscape - [ ] come up with a final set of models to track - [ ] get the multimodal metrics from each model - [ ] add a graph for each metric in its own subset - [ ] write a small blurb about each metric in its own subset - [ ] download all the papers - [ ] link all the papers - [ ] outline - [ ] writing: introduction - [ ] writing: measured vs reported # DocVQA - [ ] read paper on layout-infused llms - [ ] take notes and publish those notes - [ ] publish the notes on twitter/bluesky/linkedin and @ the authors # Chart QA - [ ] take notes on each of the chart papers and publish those notes - [ ] synthesize those notes into writing - [ ] publish the notes on twitter/bluesky/linkedin and @ the authors # ANLS vs ANLS* # Localization in LMMs # PenPusher # Meaningful Unit of Work # The Diversity Package # TinyHNSW