2024 passed quickly, anchored by work that was equal parts challenging and rewarding. Early in the year, after Cogno gained some traction, momentum slowed. We continued to code and talk with customers, but progress came in uneven cycles. Our focus on multi-agent systems felt directionally right, yet we never found a defensible niche in sales conversion.
One lesson stood out: it isn’t enough to oversee product development—you need to be hands-on with the code. Direct involvement not only accelerates iteration but also sharpens your sense of what’s truly risky about shipping. I realized this later than I’d have liked, but it reshaped how I think about building.
That realization also clarified another decision. I had once thought about taking time off school to pursue “the next big thing,” but this year I chose to stay. I still wanted the full college experience—structure, humanities, friendships, and Chicago’s unique energy. Dropping out wasn’t the path I needed right now.
In March, after more than a year on Cogno, I joined TikTok to build agent systems for Creative Copilots and Insights. I’m grateful to Zhengjin and Caoye for the opportunity—even though it cut short a long-anticipated Puerto Rico trip. The experience was worth it: building at scale, learning from colleagues across countries, and attending Nvidia GTC and Google Next in Vegas. The work on multimodal insight extraction opened new ways of thinking about how LLMs interact with content. Whether probabilistic models can be “creative” remains debated, but their ability to surprise is undeniable. I’ve come to believe the future lies in combining reasoning with synthesis—likely with greater weight on synthesis—and in multimodal output.
Returning to Beijing and then back to school later in the year felt like a blur, but I enjoyed academics more than expected. Great teammates and LLM-focused project-based classes brought a level of intensity often missing in industry. This period also sparked Crowdlistening, an exploration of how to extract meaning from the flood of unstructured, multimodal social content. While I’m now focused on AI features for a stealth startup, I still see that problem as deeply important.
Looking back on Cogno and my earlier projects, I realize most didn’t fail for lack of innovation. Since I began with LLMs in late 2022, the pace of GenAI experimentation has been relentless—domain-specific prompting, data flywheels, agents, multi-agent orchestration. Everyone was building; few created real value. Being first to market rarely matters. What matters is building moats—durable advantages that outlast hype. Technology only matters when it solves hard problems or drives efficiency at scale. In every other case, survival comes from patient, sustainable advantage, even if it means being last.
I’m still figuring things out, but 2024 was a year of exploration and growth—one I’m deeply grateful for. I’m excited to see how these lessons evolve in 2025.