"CoreWeave's Serverless RL platform, launched on Wednesday, builds on two of CoreWeave's most recent acquisitions: OpenPipe, which specialized in using RL to build custom AI agents, and Weights & Biases, which offers, among other things, a serverless platform for GPU-accelerated workloads. Serverless in this context makes a lot of sense because workloads can be distributed across available free or underutilized GPUs, eliminating resource stranding."
"According to CoreWeave, this eliminates the need for customers to manually provision virtual machines or bare metal servers in order to build custom AI agents using RL. Instead, they only pay for the tokens generated as part of the fine-tuning process. If CoreWeave is to be believed, the approach is also nearly 1.4x faster and about 40 percent cheaper than using locally hosted Nvidia H100s."
Reinforcement learning trains models through trial and error by rewarding positive outcomes and penalizing negative ones, and it has recently been used to fine-tune language models. CoreWeave launched a Serverless RL platform built on acquisitions OpenPipe and Weights & Biases to make RL more accessible to enterprises. The serverless approach distributes workloads across free or underutilized GPUs to eliminate resource stranding and leverages the stateless nature of many AI tasks. Customers do not need to provision virtual machines or bare metal servers and instead pay only for tokens generated during fine-tuning. CoreWeave claims the platform is about 1.4x faster and 40% cheaper than locally hosted Nvidia H100s and currently offers the service through the Weights & Biases platform while expanding AI services with acquisitions such as Monolith AI.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]