The static era of artificial intelligence is over. Or at least, Trajectory wants it to be.
Former researchers from the big leagues — Google, Apple, OpenAI — are betting the farm on one idea. Real-world feedback. They launched Trajectory this week, aiming to close the loop between user error and model improvement. No more training once and hoping for the best.
Trajectory just raised $15 million. Seed round. $115 million valuation. Conviction led the charge. Bessemer and Radical VC joined. Big names in individual investing? Jeff Dean, Google DeepMind chief scientist. Fei-Fei Li, the “godmother of AI.” The credibility stack is thick.
Ronak Malde is CEO. He came from Windsurf, then jumped to DeepMind during that messy $2.4 billion hiring sweep last year. His cofounders include Arjun Karanam from Apple’s Vision Pro team and Michael Elabd, formerly in DeepMind’s robotics arm. Small team, only eleven people right now. Ambition? Huge.
They argue the industry is broken. Why?
Because your AI stops learning when the training phase ends. OpenAI, Anthropic, Google — they build beasts. Coding models that write better than you. Math models that solve proofs. But after launch? Frozen. A statue. The model that failed you on Tuesday will fail you the exact same way on Friday.
“Even the most powerful AI today is static.”
Malde points to coding assistants like Cursor as proof of concept. They already use user data to tweak models constantly. That’s why coding AI exploded so fast. It’s not just base model strength. It’s the continual learning layer on top. Trajectory wants to bring that same mechanic to every other vertical. Not just code. Customer support. Sales. Law.
It’s harder, though.
Code either runs or it crashes. Binary. Clean. A return request handled by a support bot? That’s messy. Ambiguous. Success is fuzzy. Karanam admits the challenge. Trajectory helps companies define their own success metrics. You don’t start with GPT-4. You start with an open-source model, post-trained for your specific job.
Take Decagon. They build AI agents for customer support. When a bot fails and hands off to a human, Trajectory logs that failure. Next week? They train on it. The new model doesn’t make that mistake. They claim these post-trained, narrow models beat frontier giants at specific tasks.
Executives love the idea. Right now, deploying AI requires hiring “forward-deployed engineers.” Fancy consultants embedded in your company to babysit the tech stack. OpenAI and Anthropic are busy building teams to do this manual labor for you. Elabd hates this. He wants the software to fix itself. No engineers needed for the upkeep. Just plug it in. It learns. It gets better.
Is it true continual learning? Critics might roll their eyes. The updates happen weekly. Between those updates, the model is still static. It’s intermittent. A patch. Not a stream.
Elabd brushes this off. Says weekly is just the starting gate. The goal is daily updates. Hourly updates. Maybe even per-interaction learning. He envisions a world where companies don’t train one giant AI for everyone. They train specific AIs for each individual employee. Personalized. Adaptive. Live.
Sounds risky. Sounds incredible. Which one will it be?
We’ll have to wait for the next weekly patch to find out.
