Google announces LaMDA and TPU v4 at I/O

At its I/O developer conference, Google previewed LaMDA — Language Model for Dialogue Applications — and announced the fourth generation of its in-house TPU accelerator. Chief executive Sundar Pichai presented the two together as evidence that Google’s research and its hardware were advancing in step.

LaMDA was built on the Transformer architecture Google had open-sourced in 2017, but trained specifically on dialogue so that it could “engage in a free-flowing way about a seemingly endless number of topics,” in the company’s phrasing. Google judged responses on two axes it said mattered more for conversation than raw accuracy: sensibleness, whether an answer made logical sense in context, and specificity, whether it engaged with the particulars of what was asked rather than answering generically. It acknowledged unresolved problems with factuality, bias and misuse, and did not release the model publicly. LaMDA would not become a public product until the ChatGPT-driven Bard launch nearly two years later.

TPU v4 was the infrastructure Google set alongside it. Pichai said a single pod combining 4,096 of the new chips delivered over one exaflop of computing power — “the fastest system we’ve ever deployed,” and, by Google’s account, more than double the throughput of TPU v3. He said Google already ran dozens of v4 pods internally, mostly on carbon-free energy, with cloud availability to follow later in the year.

Neither announcement carried an independent benchmark. The pairing was nonetheless a preview of the shape training would take for the rest of the decade: purpose-built accelerator generations feeding successively larger dialogue and reasoning models, with Google positioning custom silicon as a counterweight to its rivals’ reliance on NVIDIA GPUs.