In 2024, I committed to the crazy goal of being the most informed person in any room when it comes to artificial intelligence. Not the smartest, but the most informed. I pledged to average two hours of daily learning (not counting podcasts). Beyond that, I resolved to use AI to publish what I learned.
I share this happy update to point out how foolish I felt almost the entire year. If you are in the middle of a crazy goal, I hope you double down and turn off the voice in your head that is saying you can’t make it. You can.
Running 1000 miles in one year a few years ago showed me that consistent daily goals add up to big results. 1000 miles is 2.7 miles per day. Every single day. If you miss a day, you have to make it up, and still keep the 2.7 mile average.
Our daughters dance every night until 9:30 PM. That gives me a large window to work, and if I get behind, I can catch up on weekends. I decided to use that time alone to my advantage.
Over the past year, I organized 15,600 articles about artificial intelligence into 34 categories! Please take a moment to check out the weekly archive:
https://ethanbholland.com/category/this-week-in-ai/
If you’re curious, my takeaway after all of this reading about AI is: robots are coming quickly, and AI agents will do everything. Almost everything in AI is leading to robots.
If you want to read about robots and agents, here are the two category archives:
Robots: https://ethanbholland.com/category/ai/robotics-embodiment/
Agents: https://ethanbholland.com/category/ai/agents-and-copilots/
Two keys to robotic embodiment are: training in simulation (AR/VR) and the ability to visually recognize the world around them (multimodality).
If you are interested in those two topics, here are the category archives:
AR/VR: https://ethanbholland.com/category/ai/augmented-reality-ar-vr/
Multimodality: https://ethanbholland.com/category/ai/multimodal/
Over the year, I cut the average time to organize a week’s worth of headlines from 13 hours per week to 5 hours. At one point, I was ten weeks behind. Now, I am two only weeks behind.
I plan to use the time to write a feature on the evolution of two key elements of multimodality: segmentation and depthing. Segmentation is the ability to track and identify objects, and depthing is the ability to track how close or far away an object is from the camera. I have collected over 200 links over 18 months that demonstrate how these two fields contribute to embodied robots. Both are currently hiding in plain sight, being trained, as Snapchat and TikTok filters.
Here are a few posts that helped me realize the connection to robot training:
Viggle: https://ethanbholland.com/2024/04/08/the-viggle-ai-memes-impact-on-image-to-video-awareness/
SAM2: https://ethanbholland.com/2024/09/14/trying-metas-segment-anything-2-sam2-demo/
AR Glasses: https://ethanbholland.com/2024/04/14/metas-ar-glasses-are-actually-robot-training-tools/
Jim Fan: https://eureka-research.github.io/dr-eureka/
Here’s my current newsletter workflow, which took a year to evolve:
• Each week I go through 10 Twitter accounts and seven daily newsletters. I open and read the links and news, which ends up being about 700 items. Anything I feel is important, I leave open in a new tab.
• I built a custom Chrome extension that saves all of the open tabs into a text file with headlines and URLs.
• A Python script sorts them and runs a regular expression to clean and standardize their structure, then removes any duplicates. After this script, there are between 300-400 links per week.
• Next another Python script reads all of the links and sorts them into 34 categories as well as an unknown pile.
• The categories are: Agents, Amazon, Anthropic, Apple, AGI, Audio, ARVR, Autonomous Vehicles, Business, Chips, Consumer, Education, Ethics, Google, Images, Inflection, International, Local, Meta, Microsoft, Mobile, Multimodality, OpenAI, OpenSource, Perplexity, OpEds, Publishing, RAG, Robots, SSI, Science, Video, Twitter, Tech
• Another script works with Claude 3.5 and the Ideogram API to create cover images for all 34 categories and store them locally.
• I then manually put all of the links into a Google Doc that structures the files into the final categories. I highlight anything I feel should be called out in an executive summary or a top link or visual for the week.
• A Google Apps script then exports the data into a JSON file which I save on my local machine.
• Another Python script uses the WordPress API to create drafts for each category, upload the category image, set the timestamp for the post, and the category.
• A second Google Apps script searches the Google doc for any of the links I highlighted and extracts them into a separate list which I organize by hand.
• A GPT then helps me write executive summaries for the main themes, and I edit them by hand.
• I give Claude my main takeaway, and it gives me six image ideas for the main cover. I pick my favorite and run it through MidJourney, Ideogram, and Flux to test them.
• I edit the best choice in Photoshop and manually create a top six category mosaic.
• Then I upload the executive summary and top links to WordPress, and it’s done.
• Five hours total per week!
I also am my 135th week of working out five days per week. That’s an important balance to so much time on the computer.
Leave a Reply