This week, I’d like to share some of my thoughts and observations from my time in San Francisco.
The Diverse World of AI Users
AI usage can generally be categorized into four distinct groups. Firstly, there are everyday users—like my mother and wife—who treat AI as an upgraded search engine and have no desire for more. Secondly, there are the curious individuals. These are people who work on computers and may dabble with platforms such as Lovable or Replit. A significant number, about 70%, identify themselves as non-technical, and while they may attempt to troubleshoot a minor glitch by typing “fix it,” they often become frustrated and abandon the process. Then, there are those of us who aren’t professional developers but have a basic understanding of concepts like localhost and how to guide a coding agent. Finally, we have seasoned developers.
My Journey from Builder to Enabler
Back in my Makerpad days, I initially thought my role was to transform people into builders. In reality, I was predominantly helping those who were already interested in learning the ropes of construction.
Many users remain unaware of the potential of AI agents. I believe that curiosity is often sparked by observing the creations of others. This inspiration drives me to create quirky projects, leading me to rename my presentations to focus on “how I build and play with AI.” Following one event, someone expressed their love for how frequently I emphasized the word “play.” This technology should be enjoyable and encourage experimentation, learning, and creativity.
The Future of Personal Agents
I envision personal agents as the user interface of tomorrow, but we’re not quite there yet. Current marketing positions them as tools that simply automate mundane computer tasks, like rebooking a flight. However, those of us who delve deeper recognize their potential to significantly enhance our computing capabilities. My primary frustrations with current agents include:
Unified Threads: Most systems tend to combine everything into one thread. I yearn for temporary side conversations, akin to Slack-style channels, along with lightweight widgets that can be created on the fly. Wabi is the only tool I’ve come across that offers a variant of this. If a personal agent could craft a small app tailored to specific needs, users would be more inclined to request modifications.
Lack of Memory Transparency: I’ve invested considerable effort into my “context garden,” yet personal agents fail to reveal what they ‘remember’ about me. Consequently, I’ve set up a system where every memory is synced to a Google Doc, allowing me to review it periodically.
Unfinished Loops: Agents often remind me about tasks I haven’t completed, such as signing a document, but these reminders occasionally arise from various platforms like Gmail or Slack, creating disconnects.
Building My Own Agent
I’m tempted to create my own agent (even though I know it will be a challenging endeavor). However, I’d choose to build it using foundational components like Pi rather than starting from scratch. My loyalty doesn’t lie with one specific tech provider, which allows me to curate the best elements. As Alex from OpenRouter highlighted, today’s landscape resembles a time when all SaaS applications featured just a login page and a dashboard.
Speed feels intelligent in this context. I’m not designing agents for lengthy overnight processes; I’m actively engaging with them throughout the day, and I prefer immediacy. Faster interactions can lead to smarter solutions, keeping me engaged and in the loop.
Growing Demand for Devices
There’s a noticeable demand for more devices. One example is the Microducks, which sold out within hours, and not to forget Ghost (details coming soon), also selling out rapidly. Despite having a Mac Mini that I hardly utilize, my desire for more gadgets to experiment with remains strong.
More computing capacity translates into more surfaces and more creativity.
Innovative Tools on the Horizon
Claude Code now allows custom modifications reminiscent of video game modding. These mods serve as small add-ons that adjust how Claude behaves or looks, including features like blocking sensitive commands or integrating additional buttons and panels. Claude can even generate a mod for you, enhancing the overall experience. Anthropic has introduced its first official mod, which flags elements in Claude’s responses that might otherwise go unnoticed.
Interestingly, many of these were points I raised during my talk last week, as I frequently request these types of outputs.
New Google Docs Features
Google Docs has introduced a feature that allows for direct opening of Markdown files without conversion. Users can now edit and comment on files like Agents.md collaboratively within Google Docs. This update will roll out over the next two weeks.
Enhancing Group Collaboration
Instinct can now participate in your group chats to assist with planning trips, splitting expenses, and managing fantasy leagues, even when your friends aren’t on the Instinct platform.
“I was traveling over the weekend with a group of strangers. Our guide was unable to organize our itinerary, so I consulted both ChatGPT and Instinct for an annotated interactive map. Instinct outperformed. While Codex/Claude Code might have provided a better solution, I found reaching out to them challenging during travel without a laptop.”— Keshav
Reflecting on this week, it’s clear the world of AI is evolving rapidly. As we navigate this ever-changing landscape, remaining curious and playful can lead to extraordinary innovations. Let’s keep experimenting, learning, and playing with technology!
