Too many things
Favorite art, yoga retreat, AI bot updates
Dear Friends,
More than a month has lapsed since my last update, the longest time between posts since I started writing in December 2022.
It’s Sunday morning. Five inches of fresh white snow blanket the yard, brighten the windows, and muffle what little noise there is outside, making the house feel even quieter than a typical Sunday.
The more often I write, the more I enjoy it.
Writing keeps the thought-to-keyboard pathway clear, like a well-trodden trail. But when too much time passes without practice, weeds grow in the cracks, and brush creeps in from the sides. The path narrows, making it harder to walk. I can push through at first, but if I wait too long, I’ll need a machete to hack my way back.
When I let too much time pass, I end up with too many topics I want to write about. Since my last update, I’ve traveled to Utah for a quick ski trip, Colombia for a yoga retreat, and I’ve made good progress on my coding projects.
I named this Substack some of the things for a reason—I don’t have to write about all the things. So for this update, I’m going stream-of-consciousness and sharing some of what’s been on my mind. I’m reminding myself why I started this in the first place: not to be comprehensive or perfect, but simply to write. To build a habit.
On Art
I am a big fan of Kirill Polevoy’s photography so when his 2024 Annual Reflections magazine came out, it was an insta-buy for me. Right after I put my order in, I got a message from him asking:
“Do you have a favorite artist? Or maybe type of art you really enjoy? Need it for a surprise for you!”
It took me more than a few minutes to answer his question. Asking what type of art someone likes isn’t like asking what their favorite color is, or season, or food - things that have some sort of boundary or known list to choose from. Art is infinite. Art is everywhere. Art serves many needs. Is my favorite art the art that is uniquely or universally beautiful? The kind that makes me feel happy? Or is it the art that hits my heart with a heaviness no physical force could match? Is it the art that is so intricate I could never dream of creating myself, or the art that is accessible to all?
I did eventually answer his question.
The reason I’m writing this is not to share what my favorite art is, but to encourage others to spend time with this prompt. How fast can you answer this question, and if you sit with it longer, does your answer change?
What type of art do you enjoy?
Yoga Retreat & Writing Prompts
I returned to Cartagena for a yoga retreat for the second year in a row. Travel creates neuroplasticity; I find myself more open to ideas when I’m out of my daily routine. The further you get from routine, the easier it is to be open. The retreat’s theme this year was about reinvention and entering a new stage in life. We were given journals and prompts to reflect on during each yoga session. Some of the prompts included:
How do you find the difference between happiness and joy?
Who is someone who transitioned into their second stage in life beautifully and how did suffering help them?
What is something you went through, that while you were going through it, you thought you would never get through it. What is something beautiful that came out from one of those moments of suffering?
What is a little bit of service you can offer your community?
What is one thing you need from your community?
What was a time you had a meaningful coincidence (synchronicity) and why do you think a higher being made it so?
What are you bringing to the table? What is your superpower?
Can I also share something I learned that I think is so cool?
Look at this picture: Can you guess what this small hole filled with water, in the center of a fort built in the 1600s on an island just off the coast of Colombia was used for?
Hint #1: it was used to get information. Hint #2: there are several small narrow tunnels that shoot out from the hole outwards in different directions
Answer: Small narrow tunnels shot outward and connected to the sea. As boats approached they could tell which direction the boats were coming from by observing the water’s movement in the hole. The local guides called it a sefon, and we made a joke it was a sea-phone, because it was a communication device for the sea. Pretty neat engineering design.
AI Bot Update (part 3)
I’ve written part 1 and part 2 about my adventures building an AI bot. I really need to take all my activities and learnings and compile into its own dedicated blog post, but until I make time for that I’ll keep sharing progress.
Since my last update two months ago, here’s what I’ve shipped:
Contextual memory: The bot can now engage in back-and-forth conversations with users.
Timestamps: When the bot recommends a YouTube video, it updates the URL to start at the relevant timestamp.
Transcript expansion: The chunk size I used for embeddings was the right size for semantic search and the bot often pulled the correct episodes, but the chunk was too small to use as enough context on its own. Now, the code expands the chunk, adding 10 sentences before and after the snippet for better context, resulting in better bot responses.
Automated 6-step pipeline: Built a pipeline to generate transcripts from new videos and give the bot instant access. The process can run in batch or single-video mode and includes:
Generate transcript from a YouTube video
Clean transcript for known spelling errors (e.g., dGen → degen)
Chunk the transcript to prepare for embeddings
Generate embeddings for semantic search
Store embeddings in a vector database
Store full transcript and metadata in a database
Safe reruns: Added functionality to rerun the pipeline for specific episodes, useful for manual transcript improvements and updating embeddings.
Metadata overrides: Added the ability to override YouTube metadata when needed.
Expanded content: Previously, the bot only had access to Farcaster 101 content. Now it includes all 200+ GM Farcaster episodes, as well as Vibe Check, The Hub, and Here for the Art.
Prompt iteration: Continuously refined the prompt for better responses.
Gaps & What’s Next
The biggest challenge with the bot right now is its inconsistency in answering different types of questions.
It handles questions like “What are some best practices for new users?”, “What is Clanker?”, and “What are warps?” very well because semantic search can easily find when those topics were discussed across all of our content.
However, it struggles with questions that are about specific episodes, or ones that are more meta-data driven, such as:
“Who were the guests on episode 100?”
“How many times has DWR been on the show?”
“What was discussed recently?”
“Can you summarize the clanker episode?”
To improve, the bot needs the ability to take different paths depending on the question type. Currently, it follows only one path with a single call to GPT. Moving forward, I’ll likely need additional GPT calls to determine which path to follow.
Path 1 (current path): Semantic Search with Transcript Context
Use semantic search across all transcripts to find when a topic was discussed. Pass the excerpt of the transcript, along with episode metadata (e.g., date, YouTube URL, episode title) into the prompt.
Example questions:
“What are some best practices for new users?”
“What is Clanker?”
“What are warps?”Path 2: Metadata Search + Transcript Enrichment
Start with a metadata search and enrich the result with relevant details.
Example questions:
“Who were the guests on episode 100?”
“Can you summarize the clanker episode?”
“How many times has DWR been on the show?”
“What was discussed recently?”
The key distinction is whether the bot should look across all content in our library, or whether it should narrow down to specific content first.
Someone in an AI agent builders group I’m in recommended this Anthropic blog post and it’s GOLDEN! It outlines several patterns for working with LLMs. I’ll be starting with the Workflow: Routing pattern. (did you know software patterns are my love language?)
The second type of question the bot struggles with is longitudinal questions. Consider two questions:
“Tell me about toe spacers (or mole, wowow, tongue scrapers, etc.)”
“Tell me the most iconic cultural moments on Farcaster in 2024.”
Toe spacers was an iconic cultural moment on Farcaster. We covered it heavily on our show. Right now, the bot is great at answering the first question, where the focus is on specific topics that were mentioned on our show. But the second type of question - summarizing broad trends or capturing key moments over time - is more challenging.
I’m excited to figure out how to improve this because /gmfarcaster has chronicled so much of Farcaster’s history and lore since the early days. I know it’s technically possible to get the bot to handle both kinds of questions, and I’m going to make it happen. Eventually.
Who Were the Cypherpunks?
I really enjoyed this post about the history of the cypherpunks written by Jason Chaskin from the Ethereum Foundation. Some excerpts:
The cypherpunks were a group of nerds driven by one simple motivation: they wanted the world to be freer. They weren’t building tools to get rich or for recognition—they were building because they believed freedom and privacy were fundamental rights. And they were willing to go to jail to defend those rights.
The cypherpunks embraced a technology-driven view of historical change. They believed that meaningful progress doesn’t come from lobbying or electing the right people—it comes from technological innovation and adoption. If you want the future to unfold a certain way, you have to build it yourself.
Do you believe freedom and privacy are fundamental rights? But would you go to jail for those beliefs? I’m grateful there are people in the world who would.
The internet is still whimsy
Lumon Industries source code leaked
Freecompliment.com - I own this domain and quickly rebuilt this website to make it compatible as a Farcaster Frame. I also added a backhanded compliment option.
Some of the pics
Row 1: Walking on water, colorful food, mug cakes on a snow day
Row 2: Cartagena streets, lunch at Celele
Row 3: Art at blue apple, an arch looking out at the sea, the blue apple beach at sunset
A Note to My Readers
I started this substack in December 2022 as an experiment to see if developing a writing habit would help clarify my thinking and/or provide other benefits. You can read about my original intentions in my first post or my more recent reflections after sticking with it for a year.
I write about twice a month and share musings, meditations, and links to things I’m finding interesting as I build out a farcaster-native media company, a modern technology consulting company, raise my kids, and have fun creating and learning in the worlds of crypto, tech, finance, science and wellness.
Thank you for supporting my writing and journey. If you’d like to get in touch you can reply to me here or find me on X and farcaster.
Until next time, keep putting good into the world. —adrienne🌏❤️