Google Part III: The AI Company

The hosts frame Google's AI challenge as a classic innovator's dilemma. Imagine a company with a highly profitable business, huge market share, and customer lock-in, whose own researchers create a technology that could support a product better than the one generating those profits. If the company launches the new product, it may disrupt its existing cash engine; if it holds back, startups can commercialize the invention first. The episode applies that tension to Google: its researchers helped create the Transformer, while the AI products built on it began to challenge the search experience that funded the company.

This is not presented as an abstract threat alone. The hosts point to the Transformer paper's role in the wider AI ecosystem and Google's unusually strong position: Gemini, its own cloud infrastructure, custom TPU chips, and the search box that remains a front door to the internet for most people. These assets make Google capable of competing across models, chips, infrastructure, and distribution. Yet the opening question remains strategic rather than technical: should it risk its search economics to lead in AI, or protect the profits that could slow its response? The hosts treat that choice as unsettled, not as proof that either side has already won.

The episode argues that Google's AI position grew from a concentration of people as much as from one invention. It lists researchers who later founded or led organizations such as OpenAI and Anthropic, alongside figures in Tesla and DeepMind. The hosts compare this concentration to a hypothetical IBM that had hired everyone who knew how to program. The comparison is deliberately extreme, but it captures the scarcity of AI expertise at the time: the work depended on specialized PhD programs, a small number of advisers, and a field still arguing over which research directions were legitimate.

Larry Page's interest in AI also had roots before Google. His father researched machine learning and artificial intelligence at a time when the field was unpopular, and PageRank itself can be viewed as a statistical method related to AI. The hosts connect this background to Larry's 2000 description of an ultimate search engine that would understand the web and what a user wanted. Google therefore had a long-standing ambition to make search more intelligent; the dilemma was how to pursue that ambition once AI became capable of changing the product itself.

An early thread began over lunch in 2000 or 2001, when Google engineers George Herrick, Ben Gomes, and the newly hired Noam Shazeer discussed the idea that data compression might be technically equivalent to understanding. The analogy was learning material from a textbook and later reconstructing its concepts: if a system can reduce information and restore it, perhaps it has captured something about its meaning. The episode does not claim this debate settled what machine understanding is; it shows how a speculative idea led engineers toward probabilistic models of natural language and predicting what sequence of words would follow another.

Herrick and Shazeer pursued the work even though many colleagues thought it was a poor use of their talent. They built PHIL, a large language model for its time. In 2003, Jeff Dean used it to help implement AdSense, which let Google match ads to the content of third-party pages and expand the inventory for its existing advertising system. The same research also produced Google's 'Did you mean' spelling correction, reducing a frustrating extra search and avoiding work on results for mistyped queries. These examples give the episode a concrete bridge between exploratory research and product value: language models were useful to Google well before conversational AI became a consumer product.

Google's early language-model work began with a narrow, legible problem: recognize likely misspellings such as a query for a 'god groomer' that was probably meant to be 'dog groomer.' When confidence was high, the system could correct the query rather than ask the user to try again. The team called its relatively large model PHIL. In 2003, Susan Wojcicki and Jeff Dean used it to understand the content of third-party web pages so Google could match its AdWords inventory to publishers' pages. The hosts describe AdSense as a major new revenue source built by extending an existing ad system into a much larger inventory.

The same work also supported 'Did you mean' spelling correction. That feature improved the experience for people who mistyped a query and saved Google from serving results that would immediately be discarded after a corrected search. The hosts connect this practical value to Google's mission of organizing information: compressing and reconstructing language seemed relevant both to product quality and to understanding information. Yet the model had a substantial infrastructure cost; by the mid-2000s, PHIL reportedly used 15% of Google's data-center capacity, much of it for AdSense and related services. Early AI could create real economic value while also consuming scarce compute.

Google Translate became a useful test of whether larger language models could move from research into products. Franz Och's team trained an n-gram model on two trillion words from Google's search index. It performed exceptionally well in a DARPA machine-translation competition, but one sentence could take 12 hours to translate. That was acceptable for a benchmark with several days to submit results, not for a consumer service. Jeff Dean helped the team restructure the computation so words and sentences could be processed in parallel. On Google's distributed infrastructure, average translation time fell to 100 milliseconds, enabling deployment in Google Translate.

This example links algorithmic progress to Google's engineering advantage. The model itself mattered, but the ability to distribute work across CPUs and data centers made it fast enough for a product. Once the system was in production, Google could consider other uses, including predicting search queries and estimating AdWords click-through rates for ad-quality scores. The hosts frame that as a direct line to Google's bottom line: the model was valuable not only because it translated language, but because its predictions could improve the economics of advertising.

A second capability was Google's access to a dense academic network. In 2007 Larry Page recruited Sebastian Thrun from Stanford's AI Lab, SAIL, first part-time and then full-time. Thrun had led a lab with prominent researchers and a stream of Stanford undergraduates; the hosts name people who later became leaders in technology, including Chris Cox and Sam Altman. This network gave Google more than a conventional hiring pipeline. It let the company connect research, students, professors, and product teams, while allowing academics to retain their institutional positions and work on projects with the potential to reach millions of users.

Thrun's work on Ground Truth for Google Maps shows how that research network could solve a product bottleneck. Before Ground Truth, Google relied on map data from Tele Atlas and Navteq, which the hosts describe as expensive and not especially good. Thrun brought experience from the DARPA Grand Challenge and Street View to a project that rebuilt map data using Google's own photographs and dozens of other sources. It combined machine learning with substantial human labor: the hosts say Google hired about 1,000 people in India to resolve discrepancies and draw maps by hand. The result was not AI automation alone; it was an integrated system of data collection, algorithms, and human review.