Former Apple, Google
DeepMind, OpenAI, Meta Superintelligence Labs, and Scale AI researchers this week announced the launch of a startup called Trajectory that aims to help companies improve their AI
products by training on real-world user interactions.
Trajectory, a platform that will continually learn from its experiences, launched with $15 million seed round led by Conviction and
Bessemer, two venture capital (VC) firms that pool money from institutional investors to fund high-growth technology startups in exchange for equity.
“The research question motivating us
is simple: can AI systems improve in response to real-world experience?” Michael Elabd, co-founder of Trajectory and former DeepMind researcher, wrote on LinkedIn.
The company claims to
turn experiences into intelligence for more accurate and up to date information serving up in AI queries.
advertisement
advertisement
Elabd wrote that today’s agents are episodic. They complete a task, receive
feedback, and reset, which could prompt them to miss valuable learning signals, which he defined these learning signals as “retries, edits, user interventions” and more.
“By closing the loop between interaction data and model improvement, we believe that agents can continuously improve,” he wrote, explaining how AI research has long aspired toward
continual learning.
He is already seeing this becoming a “practical reality” in the company’s work.
Today, models are post-trained weekly, but the startup claims the models will
update hourly or at each query or interaction. That is the goal — to provide users with up-to-date information.
Early search engines operated in a similar way, using something called a batch
process to organize web information and then return it to users during queries. Trajectory is targeting hourly updates or an update at every interaction.
A model that keeps learning from
failures and fixes, combining its original output and feedback from users on an ongoing basis will become the answer to AI-search’s business.
Clay, Harvey, Decagon, and Rogo, with
Trajectory are early customers of the company.
Trajectory said funds will be used to build the platform that the company says will continually learn from each query
Rather than not
accept or understand the content and context, the AI will track organic behavioral signals to return correct information and use it to continually train the models to improve over time.
