Ronak Malde, co-founder of Trajectory.ai, wasn't supposed to end up at Google. Back in 2021, when his previous company, Windsurf (then Kodium), was being acquired, he thought it was OpenAI calling. Instead, Demis Hassabis and the DeepMind team were on the line. The unexpected Google acquisition was a twist, but it only sharpened Malde's conviction about a deeper problem facing AI: most models are dead on arrival.

He watched as Sonnet 3.5 emerged, and the world of AI began its explosion. Yet, even as models grew more powerful, Malde saw a critical flaw: they were static. “It's still that AI like it's it's obviously very powerful, but it kind of acts like any other software uh in that it's very static,” Malde explains. “The model that you used yesterday, it's going to be the same model making the same mistakes tomorrow.” This means every user correction, every edit, every nuanced interaction is wasted, never feeding back into the system to make it smarter.

Malde’s big bet is that the future belongs to “living systems”—AI products that constantly adapt and learn from real-world user interactions. He argues that waiting for the next foundational model update is a losing strategy when your product could be improving itself every second.

Key Takeaways

  • Static AI is a Dead End: Current AI models, despite their power, act like traditional software—unchanging. They repeat mistakes because they don't learn from real-world user corrections or interactions, wasting valuable feedback.
  • The Google Plot Twist: Ronak Malde's journey took an unexpected turn when his previous company, Windsurf, was acquired by Google's DeepMind, not OpenAI as he initially expected in 2021. This experience further solidified his vision for dynamic AI.
  • Giving Up a Payout for a Bigger Vision: Malde chose to forgo his acquisition money from the Google deal to start Trajectory.ai, driven by a deep conviction that continual learning from user interactions is the "next major unlock" for AI across all domains.
  • The "Living System" Thesis: Trajectory.ai's mission is to build a platform that transforms every AI product into a self-learning "living system"—one that continuously adapts and grows intelligence based on human-in-the-loop interactions.

The AI Models Are Dead. Long Live the Living Systems.

The genesis of Trajectory.ai wasn't some abstract market analysis; it was born from Malde's frustration with AI's inherent shortfalls, even at the highest levels. He saw firsthand the immense power of models like Sonnet 3.5, but also their limitations. Imagine building an incredible product, only to have its core intelligence make the same obvious mistakes day after day because it's unable to learn from its users.

Malde recognized this as AI's biggest missing piece. “All of those corrections you gave it, the edits like in any product is all just being put to waste,” he laments. This isn't just inefficient; it's an existential threat to AI products that need to operate in messy, unpredictable real-world environments. The goal isn't just better models; it's a dynamic feedback loop that treats user interaction as the most valuable training data.

Giving Up Acquisition Money for a Bigger Bet

Malde's conviction was so strong that he made a jarring personal decision. After the unexpected DeepMind acquisition, he chose to give up all his acquisition money to fund Trajectory.ai. “I decided to give up all the acquisition money to start trajectory,” Malde says, explaining that the explosion of AI capabilities in coding, legal, healthcare, and finance showed a clear path forward. He believed that the time was ripe for a platform that could supercharge these domains with continuous learning.

This wasn't about building a slightly better model; it was about reimagining the very nature of AI products. His vision is for every product to become a 'living system,' constantly observing, adapting, and growing from every user interaction. Trajectory.ai's work on technologies like Self-Distillation Policy Optimization and Continuous LoRA are directly aimed at making this vision a reality, turning static software into perpetually improving intelligence. “We realized continual learning is kind of the ultimate like paradigm to do that,” Malde concludes, seeing it as the next major unlock for AI.

What to Do With This

Stop treating your AI product as a finished piece of software. This week, identify one core user interaction that always requires a human override or correction, then design a simple feedback loop to capture that human input. Redirect that specific data stream back into your model's learning pipeline, even if it's just a small, focused adjustment. You're not just fixing a bug; you're building a muscle for continuous adaptation.