Back to The history and future of AI at Google, with Sundar Pichai, with Sundar Pichai.
Why Didn't Google Ship ChatGPT First?
Sundar Pichai says the story that Google invented transformers and then left them unused is misunderstood. Transformers and the TPUs were built for product problems: better translation, and speech recognition that had to serve two billion people when Google did not have enough chips for inference. The research teams were guided by those product needs, and transformers went into Search at once. BERT and MUM produced some of the biggest search-quality jumps of that period, because Google measured search quality so strictly.
Google was also building the chat product internally, called LaMDA. An engineer thought the model was sentient. That was an early ChatGPT he could already talk to inside the company. At Google I/O in 2022, AI Test Kitchen was LaMDA, held back because there was no end-to-end RLHF version yet. The version he saw was too toxic to ship, and a search-quality bias set a higher bar for what counted as good enough. OpenAI's deal with Microsoft came only a couple of months earlier. A clearer signal was on coding, through GitHub and through the jumps from GPT-2 to GPT-3 to GPT-4. The miss was not research that nobody applied. People who worked on ChatGPT launched it the week of Thanksgiving, as a buried test, not as a declared bet on the future.
The history and future of AI at Google, with Sundar Pichai ·
From the transcript
A bit of history that people talk about a lot in the context of Google and AI is the fact that transformers were invented at Google but then productized outside of Google with mostly chat GPT and kind of that style of product. How do you reflect on that now? I think it's actually worth talking about. It's a bit misunderstood. You know, transformers was done in the context of AI.
a lot of TPUs, transformers were all done to solve a specific product need to some extent, like the teams thinking about how to make translation better. In the case of TPUs, how do you, hey, speech rec works, but you suddenly have to serve it to two billion people. We don't have enough chips for it. It's like, how do you solve inference for it? I hadn't known that.