Who's liable when your AI agent burns down production?
Summary
The article analyzes forward propagation of errors through time (FPTT) as a forward-time alternative to backpropagation through time (BPTT) for training recurrent networks. It presents two main insights, demonstrates finite, but unstable, gradient propagation in practice, and concludes that while the idea is academically interesting, it is not practically usable with current tools. It also discusses multi-layer extensions and future research directions relevant to AI tooling and hardware.