The Newest New Thing: Generative AI inside the Enterprise
Generative AI has the potential to finally unlock the disconnected reams of data inside information-heavy enterprises such as consulting firms, investment firms and banks.
Generative AI is The Newest New Thing. Sonya Huang and Pat Grady of Sequoia have generously shared an outstanding thesis that lays out the application or interaction layer (and therefore, the data layer on which the models train on) as well as their respective models.
With models continuing to get better/faster/cheaper, and model access trending to free and open source, the application layer is ripe for an explosion of creativity.
– Sonya Huang & Pat Grady, Sequoia Capital
Now, the application layer (“make a picture of a chimp in a spacesuit”) is exciting enough as it turns every single user into the woodcarver Geppetto who can generate infinite Pinocchios at will.
Sequoia’s report focuses on the the white spy application layer for consumers to generate text, images, speech, video and code, but the black spy layer to identify the use of Generative AI for cheating on student submissions, in images (think photo alibis in court proceedings), and in code (say, job screenings) is likely to be an industry in itself.
Black spy monitoring is not happening somewhere in the future. Google, who pioneered some of the early research into AI techniques, is already able to weed out AI-generated content for its SEO rankings.
But apart from the application layer, there’s the enterprise layer, which is just as interesting — and where much of the monetization could sit.
Let me explain.
While Google has already indexed the internet where you can find what you’re looking for in from the public domain, it remains incredibly hard to query relevant information inside the enterprise. Enterprise databases and records are often silo-ed, mostly structured, and incredibly painful to query and search (despite $BOX being one of the best tools for searching for knowledge in documents within a firm).
Fortunately, Ankur Goyal has thought through how GPT-3 could work as a “better database” …
GPT-3 inside an application could evolve into a new front-end to query the application:
You could open up Salesforce and ask, “What is my division’s revenue per salesperson trend?”
Inside Workday: “How much would I get if I cash out my leave for the year?”
Inside Bloomberg: "Draw a chart of AAPL P/E vs XOM since 2010"
Inside SAP: “How many inventory days do we hold now vs the last four holiday seasons?”
Inside Power BI: “Show me a table of cost of raw material by product” (h/t @guruchahal)
But the opportunity for institutionalizing knowledge inside information-heavy enterprises such as consulting, investing and banking feels like an even more powerful use case.
These are enterprises that have endless reams of data; indeed, you could say that information arbitrage is the reason they exist. Yet every large bank, investment firm and consulting firm has been hobbled by their inability to institutionalize this knowledge base that sits inside proprietary databases to their enduring benefit.
Could Generative AI fix this, abstract away the complexity of SQL, and answer questions in natural language?
Inside Fidelity: “What factors helped us outperform in EM last time the Fed hiked rates but the ECB did not?”
Inside Bain & Co: “What are the possible county regulatory and state employment law red flags for gig economy companies in North Carolina and Tennessee?”
Inside JP Morgan Chase: “What country did most of our customers travel to this summer, and what is the decline rate on cards used there?”
In the last 30 years, the primary human interface moved from Rigid Syntax Text (DOS) → Click (Apple, Windows) → Touch (iOS), and made computing accessible to billions. Generative AI has the potential to move the primary way we interact with applications (not just search!) towards natural language text-first questions.
Generative AI is an exciting development that makes querying easy to master for those with the ability to be specific with language — but could perhaps be a step backward for the masses who are used to interacting a computer only by touch.
* The Pinnochhio image was generated at dreamstudio.ai
I’m surprised software doesn’t commonly have natural language interfaces even today for simple actions typed into a text box.
Google calendar used to have a great feature where you could type in “schedule a meeting with Kunal for next Wednesday at 3pm” but for some reason they took it out.