How Google Plans to Crush OpenAI Before 2023 Ends
The five forces that could take down OpenAI—part 3
Google has been sleeping through all the milestones OpenAI has achieved in the last two years, critically lagging behind. OpenAI released and open-sourced GPT-2 and then Google made Meena, but only as a research project, inaccessible to users. OpenAI released GPT-3 (with an API and playground) and then Google made LaMDA and PaLM, but only as research projects, inaccessible to users. OpenAI experimented with DALL-E in the visual front and then cemented its position with DALL-E 2—all before Google made Parti and Imagen, but only as research projects, inaccessible to users. OpenAI released ChatGPT and GPT-4 and then Google made Bard and PaLM 2, inaccessible to users (at first).
I just summarized the last few years of the relationship between the unlikely AI leader that is OpenAI and the search giant just by changing the names there. Annoying to read, right? Imagine how annoyed Google devs and researchers felt when they realized their work and ingenuity were transformed into real stuff by the competition that they started to leave the company in troops—some of them, perhaps still feeling bitter, to go ship things at OpenAI.
Googlers in the audience could rightfully criticize my partly-satire, partly-not depiction of what’s happened since the GPT boom started in 2018. Google has done incredible things—like experimenting with the Mixture of Experts paradigm six years before OpenAI applied it on GPT-4—and it was them, not OpenAI, who invented the transformer. That may not be a strong enough defense anymore, though; all eight co-authors of the transformer paper have left Google, adding to the attrition.
In August, Llion Jones, the last remaining author of the paper still at Google left the company to start Sakana AI, a startup focused on bio-inspired AI. Łukasz Kaiser joined OpenAI (alongside other researchers not involved in the transformer paper, like Liam Fedus, Barret Zoph, and Luke Metz), Noam Shazeer co-founded Character.ai, Aidan Gomez co-founded Cohere (now a Google partner), Ashish Vaswani and Niki Parmar first co-founded Adept AI Labs and recently left to co-found Essential.ai (still in stealth)—pretty much all left to create or join the competition.
Perhaps Google is just the perfect place to hone AI-related skills and then jump to more risky but surely fulfilling endeavors. Perhaps the reality that outsiders don’t often have access to is that Google is a slow bureaucratic machine that slept in its leadership; twisted incentives to get promoted, a non-existent risk budget, and endless barriers to shipping anything in case it could generate bad press are the perfect recipe to lose against a young, fast, and bold startup with an ambition larger than the sun. Perhaps Google didn’t realize OpenAI had changed the rules transforming the field from a global academic research lab into a full-blown production machine. Perhaps Google knew all along and simply wanted to avoid entering into a risky game that could force it to redefine its entire suite of products with a not-yet-mature technology while threatening its primary source of revenue.
Whatever the case, people promptly jumped to claim that Google was dead. Google management freaked out and declared ChatGPT was a “code red” for their search business. Presumably not because of the chatbot’s outstanding abilities—given that Google owned similar tech—but because of the apparent impotence of the company to respond with the required dynamism and decisiveness.