
Learnings from 54 AI Builder Lightning Chats
April 20, 2026 | 📍 Ridgewood, NJ
A week ago, I wanted to get a better vantage point of how people are currently trying to apply AI to their real-world problems. I am acutely aware, between being in SF and being on X, that I am in a founder/researcher bubble. I wanted to try to help folks wrangling agents for new applications, outside of all the noise.
So I had a form of office hours: I opened up my calendar to anyone navigating hard problems that they suspected AI could be better applied to. I had 54 AI lightning chats, 15 minutes each, with the goal of being a sounding board and offering guidance where I could.
I had conversations across prediction market trading, somatic coaching, relationship simulations, construction, talent pipelines, new frog species mapping, PCB design, legal statute aggregation, carbon credit systems, and more.
It was a hell of an office hours sprint, and my dominant feeling is awe at all that sincere builder energy. I feel really lucky that folks wanted to share their problem spaces and passions with me — and so energized on the other end of this.
If this is our renaissance moment, it makes sense that there is an electric social current in the air: innovation is rarely achieved alone, it is arrived at relationally.
In reflecting across all the conversations, below are a few themes, broken down by problem and solution, that repeatedly came up. I also put together a
slide deck version if you prefer that format.
I. Memory and observability as the bottleneck
The Problem. Across Discord channels, second-brain systems, and scrapers, builders experienced agent hallucinations and forgetfulness. This was especially confusing for nontechnical builders who had black-boxed the underlying system and were relying on the agent inputs/outputs to understand what was going on. In digging in, I found across folks that the agent flailing stemmed from (1) relying only on the foundation model for memory, (2) convoluted agent-made systems (Obsidian, etc.), and (3) the lack of a human-readable tracing for the builder to better align the agent with, and for the agent to later reference.
The Solution. Where folks had local setups similar to OpenClaw, I recommended they use that. If they were in too deep and didn't want to migrate, we chatted through a naive memory.md to start.
On top of that, we chatted through slack alerts, lightweight data visualization on their personal website, or tables in a database for some layer of tracing that the human orchestrating the agents would want to check in on. That last part is important: there is an emotional layer to managing agents, and the builder should have a solution specific to how they want to work and run their workflows.
II. Build one workflow at a time, with agents earning your trust
The Problem. Builders got stuck trying to architect a grand unified AI assistant up front. Nontechnical builders often did not have the right expectations on agent capabilities out of the box, or the need to iterate with feedback loops. In the case of experienced engineers, many didn't have evals that were serving them: offline benchmarks running daily to help them validate model switches and sanity check the product with, and online alerts on user behavior that bumped into new product pathways or should be included in the offline test suite.
The Solution. Often in these chats, I mirrored back what the builder was saying and we explored narrowing the scope of their agent performance evaluation to one concrete workflow. If they were already on Slack, we chatted through smart alerts and agent conversations in dedicated channels, where the human delivers fast feedback that the agent then has access to in its context or rules (and this plays into the last point on tracing the agent's memory/understanding). We chatted through what a 2-week training camp for the agent could look like: reviewing outputs daily, tightening the criteria file, iterating until the agent has earned the builder's trust. If the user had issues with an OpenClaw (or OpenClaw-like) setup and missed cues, we chatted through more deterministic cron job workflows. For more experienced engineers, we chatted through eval setups, both with off-the-shelf solutions and what they should build for their piece of mind.
There was also a theme of builders that, when we narrowed their problem to one pressing pain point, we realized that they didn't have an agent problem — they had a people problem. For example, in one conversation on hardware development and a firmware issue that indicated a deeper board design issue, we chatted through the path to source other people who had already built out a similar sensor system at scale. In this case, it'd be most valuable to try to get on a call with them to try to source their manufacturers rather than build it from scratch. The theme of finding the right people and building high-trust relationships as the unlock came up again and again.
III. Marketing & distribution systems: the role of agents in your own consistency and authentic communication
The Problem. A lot of folks came with marketing questions, and the nagging feeling that they should have some type of AI system, but weren't sure how to approach that in a way that felt good. Almost everyone said they didn't want to spam their followers with AI-generated content, but there is pressure to keep up for product positioning. There was universal agreement that there are too many social platforms and too much AI slop. We all know marketing and the right distribution channels are important, but competing priorities created overwhelm for a lot of builders. I've been there, so I really sympathized when this came up.
The Solution. What has helped me most is to view content like going to the gym, not like going to a casino. The goal is to create systems that make putting out my writings, videos, and photography as routine and automatic as possible, so that it isn't eliciting any dopamine spike (which eventually becomes exhausting), but rather becomes a reflex. It should feel as routine as going to the gym and putting in reps, and no single piece of content should have that much emotional weight attached to it. The goal is to sustainably build community and connect, and to have fun, because when you're having fun, you're naturally putting more energy in and are not so fixated on outcome.
We chatted through the role of agents as a brainstorming partner (especially over voice conversations) to pull out key ideas, as cross-posters for the content they've already written, and as platform listeners for maintaining community and people discovery.
I believe that in this moment of the growing AI-generated slop, and the degradation of trust in digital spaces, your ability to authentically share is the moat.
IV. The current moment
I suspect part of the reason so many people added time on my calendar in such a short window is because it is clear an individual or team can have never-before-seen outsized leverage on delivering value in their problem space. And yet, this is also a moment of escalating security concerns and failure modes. The democratization of building software has created a palpable excitement in the air, and with that is a hunger for more support to safely step towards these bold product visions.
We're all on this new terrain together. While the builder energy is electric, this is also a time to come home to your body and feel into a sense of grounding. Yes, everything is moving fast. AND you must remember that life is beautiful and there is time — because you want to build from a place of abundance, not lack.
I loved having these conversations. They gave me a bird's-eye view of the reality that the AI wave is here, and as it washes over us, there is a path to welcome its integration into our lives in ways that bring more creative visions to life than ever before.
If this is our renaissance moment, I want to be in the room with all of you, overexcited and overcaffeinated at the expansiveness of what's ahead.