This startup is helping companies train AI with an old but buzzy technique. Read the pitch deck it used to raise $7.5 million.

"A London-based startup that provides software to help companies speed up AI training has raised $7.5 million in funding. AgileRL has created a platform called Arena that engineers and data scientists can drop their AI models into, run simulations, fine-tune them before deployment, and monitor them while they are running. The startup, cofounded by Param Kumar and Nicholas Ustaran-Anderegg in 2023, focuses on reinforcement learning, or RL, an AI training technique in which systems learn by trying actions and improving based on the feedback they receive."

"Kumar, AgileRL's CEO, told Business Insider that after ChatGPT launched in late 2022, companies "moved their budgets from working on RL" to focus on transformers, the technology underpinning large language models. Where transformers learn patterns from large datasets all at once, RL learns one step at a time. Now, he says more companies are realizing that transformers can only get them so far."

AgileRL raised $7.5 million to expand its Arena platform for accelerating AI model training. Arena enables engineers and data scientists to drop models into simulated environments, run simulations, fine-tune parameters before deployment, and monitor models in production. The company was co-founded in 2023 by Param Kumar and Nicholas Ustaran-Anderegg and focuses on reinforcement learning (RL), a training approach in which systems learn via actions and feedback. Reinforcement learning traces back to the 1950s and is seeing renewed interest as teams recognize limitations of transformer-only approaches. Arena targets task decomposition in robotics and other domains to speed development and improve task-specific performance.

Read at Business Insider