The Hardware Lottery (hardwarelottery.github.io)

The essay coins the term "hardware lottery" to describe how research ideas win not solely because they’re better, but because they fit the prevailing hardware and software landscape. It traces historical examples — from Babbage’s unbuildable Analytical Engine to the decades-old backpropagation algorithm that couldn’t scale until GPUs (originally for graphics) unlocked massive parallelism — to show tooling determines empirical success. The piece also links incentive shifts (Moore’s Law and general-purpose CPUs) to a long period where hardware, software and algorithms evolved in isolation, and notes chip development now costs $30–80M and takes years, making hardware a durable gatekeeper of what research is feasible. For the AI/ML community this framing is significant: the recent swing to domain-specific accelerators (TPUs, specialized kernels) and the end of Dennard scaling mean hardware designers increasingly bias which approaches are practical. That boosts efficiency for mainstream deep neural nets but raises systemic risks — promising nonstandard ideas (capsule nets, unstructured pruning, alternative architectures) may be suppressed because current silicon doesn’t support them. The essay argues for closer cross-disciplinary collaboration so future tooling doesn’t just optimize today’s winners, but preserves flexibility to discover fundamentally different—and potentially superior—paths to intelligence.