The question-and-answer website Stack Overflow was based 17 years in the past to permit programmers — human programmers — to publish questions on programming issues and get solutions from a neighborhood of like-minded people.
Since its founding, the world has turn into enamored with how ChatGPT and different generative AI packages can’t solely provide solutions however even carry out the work itself, handing you your individual customized code.
How ought to a neighborhood of individuals sharing information reply to the sudden enchantment of AI automation?
Stack Overflow’s CEO, Prashanth Chandrasekar, has been operating the corporate for nearly six years — and he has a plan. I sat down with him this month to find out how that plan is coming collectively.
“Issues have modified; we need to change with the instances,” Chandrasekar instructed me. “We needed to go together with the path of the circulation of the river.”
The rise of ChatGPT
He did not actually have a alternative. Beginning in 2022, the rise of ChatGPT had a right away influence on Stack Overflow’s public website visitors, which delivers the promoting that accounts for a big portion of its income.
A main means by which individuals got here to Stack Overflow was at all times a Google search. Paid search on Google was the “person interface to Stack Overflow,” as Chandrasekar places it.
With the ability to ask a chatbot as a substitute of looking out shortly reduce into these Google queries, and visitors started to say no.
Chandrasekar arrived at a broad philosophical conclusion about not simply AI and automation but in addition the web.
“Our view is that the character of the web has modified,” he stated. It is not principally about paid search from human queries driving website visitors. “The person interface has modified to be Gen AI instruments,” he noticed.
“And, so, we’re having to kind of be as conscious of that business change as doable. We have to additionally diversify” as a property.
Chandrasekar and staff realized there may be loads of worth in Stack Overflow’s 60 million solutions to handle the shortcomings of generative AI. One choice was to sue OpenAI and different makers of AI basis fashions. The “pre-trained” giant language fashions embody Stack Overflow conversations, that are legally the property of Stack Overflow.
“We stated, OK, we may do this, or we may do one thing a little bit bit extra, as an example, novel, relative to what everyone else was doing.”
As a substitute of submitting swimsuit, Stack Overflow put in code to dam scraping of the general public web site, “after which we stated, ‘when you’re constructing a chat bot or no matter, it’s a must to work with us. Let’s have a good enterprise settlement.'”
Data as a service
On prime of promoting and the paid enterprise model of Stack Overflow, referred to as Groups, Chandrasekar is constructing out a 3rd income often known as information options — or, knowledge-as-a-service.
Chandrasekar has been signing offers to license Stack Overflow’s content material, first with Google and OpenAI.
Previously 18 months, he is additionally been busy signing “all of the cloud hyperscalers, I am unable to identify them, however you understand them,” he stated. “We’re within the strategy of doing many extra.”
These offers embody integration with Microsoft’s Visible Studio Code instrument, GitHub CoPilot, and Google Gemini Code Help — to get solutions proper contained in the programming surroundings.
“All these firms are taking, are leveraging an official licensed model of our dataset to coach their fashions for pre-training to do issues like RAG and indexing in some instances.”
Solutions get surfaced on ChatGPT or different LLMs, with a credit score and a hyperlink again to Stack Overflow. “The aim is to really recapture the visitors that folks gave to our web site straight,” he stated. “We have gotten extra of a headless web site: if persons are spending all their time in gen AI bots, that is additionally nice.”
In fact, there’s the chance of not having direct relationships with the top person if they aren’t really on the Stack Overflow website. Chandrasekar stated the corporate has numerous agreements to get pertinent info from OpenAI and others about issues such because the immediate that the chatbot person is utilizing.
“There are loads of subtleties within the engagement between the accomplice and us,” he instructed me. “We’re working by way of every situation,” he stated, akin to how a lot of an LLM’s “context window” (the latest reminiscence of chats) is shared with Stack Overflow.
AI’s shortcomings
Chandrasekar stated that Stack Overflow is fixing three main shortcomings of the know-how for OpenAI and the opposite giants.
One shortcoming is what he calls the belief level. “You do not belief what’s popping out of it,” alluding to the notorious LLM hallucinations and confabulations.
The second shortcoming is LLM mind drain. “When you do not generate new info, these LLMs should not going to progress of their intelligence stage,” as is evident from the controversies over so-called artificial knowledge that may pollute LLMs.
Final, and maybe most vital, “the solutions popping out of the gen AI are literally not information,” stated Chandrasekar. “There could also be a solution, however they might be tapped out on complexity as a result of that is too difficult of a set of circumstances,” and subsequently, actually must have the wealthy context of Stack Overflow, he stated.
Groups integration
On the similar time that he has inked offers with the giants for the general public Stack Overflow, Chandrasekar has begun a second a part of the information options enterprise. Stack Overflow is integrating its public content material into the Groups product for firms that need to broaden their inside information sources for the aim of agentic AI.
The Groups product was launched as a result of firms stated they needed their very own model of Stack Overflow as a repository not for basic programming information however for his or her specific company processes. Now, stated Chandrasekar, the identical firms need to broaden that info pool as a result of they need to develop AI brokers that do much more than programming.
“We now have firms like Uber who’ve really performed this with us already, the place they’ve constructed an AI chatbot that is referred to as an assistant, or, in a really beneficiant sense, agent as a result of it is all about performing the motion. However that agent is leveraging the Groups knowledge for one thing like, ‘How can we really do that’ inside Uber? The bot serves up an computerized reply inside a company chat. “The bot has all of the information from the staff, so it is surfacing the proper info on the proper time.”
In different phrases, agentic AI automates what staff historically do with Groups. “We seen this as a result of our APIs are red-hot; our APIs are being hit always by the bot.”
On condition that development, Chandrasekar is including a brand new factor to Groups, licensing your complete Stack Overflow public website knowledge to the identical Groups clients.
“We stated, ‘Why do not we take our information options product, our public platform knowledge, and likewise current that to firms alongside Groups knowledge?’ An agent then can have the information from 60 million questions and solutions, and likewise all of the information particular to the corporate, after which it is much more armed with the proper reply on the proper time.”
The Groups product, like the massive licensing offers, is built-in into numerous merchandise, akin to Atlassian’s JIRA IT ticketing system.
Stack Overflow is negotiating licensing phrases with enterprise Groups customers. Chandrasekar declined to debate pricing particulars aside from to say, “It is a value-based pricing mannequin.”
“It is early levels,” he stated, in figuring out what the market will permit for such content material licensing.
Web site enhancements
Similtaneously information options are being developed, the general public Stack Overflow website is receiving some pretty important enhancements.
The corporate continues to be “working our means again” to the extent of visitors previous to ChatGPT, stated Chandrasekar, with out disclosing visitors numbers, including the positioning “has not but totally” made it again to the pre-ChatGPT stage.
Chandrasekar is constructing out the performance of the general public website to make it extra real-time.
The standard mode of Stack Overflow is one individual posting a query after which others posting their prompt solutions.
Two different types of trade are buried deep throughout the website — chats and discussions. Chandrasekar describes these as “swim lane” modes of interplay that do not present the perfection of the principle stack — overflow solutions — however can get a response to somebody far more shortly.
Chandrasekar’s philosophy is that “we need to present a number of kind elements and venture varieties for technologists of various varieties.”
The corporate can be serious about including tutorial video content material from customers. “There’s loads of nice content material,” he noticed. “Think about if someone is testing DeepSeek and we’re in a position to in a position to, as an example, reside stream that and seize it and someone else can be taught from that have. We actually need to go from being a information base into far more of a neighborhood website.”
If this sounds to you want Reddit, Chandrasekar stated he will get that lots. The distinction, he famous, is that “we’re clearly a really specialised viewers; we’re not attempting to be all issues to all individuals.” He admires the bigger social website. “I do know them very nicely, and so they’ve been nice. They’re very a lot an in depth cousin to us, or, possibly, an even bigger brother.”
He noticed that Reddit’s licensing offers with OpenAI helped pave the best way for the information options enterprise.
There may be at all times the hazard in increasing a profitable property that one can unfold one’s efforts too skinny. How does Stack Overflow place its bets?
“We need to choose those that resonate probably the most with our customers,” he stated. “I speak to the neighborhood lots, principally interact with them to know which of them to double down on. We are actually going and operating assessments on which of them really make a distinction.”
The connection internally with gen AI has additionally modified. Early on, when ChatGPT got here public, some customers of Stack Overflow had been grabbing ready-made solutions from the bot. The positioning responded by banning the cut-and-paste replies.
“However then, we stated, let’s speak to the neighborhood and see how they needed to do issues,” he stated. “One factor that grew to become apparent is that folks nonetheless discovered it to be pretty tough to only interact human to people on asking questions.”
Historically, human moderation on the positioning would possibly result in moderators scolding repetitive or beginner questions. “When you’re asking a query a few technical topic, if somebody had answered earlier than, someone would let you know, that is a incorrect query. Go search earlier than you ask.”
There got here “an enormous alternative to make use of AI.”
The positioning not too long ago went reside with “Gemini-powered” solutions. Now, “Gemini is supplying you with a prompting it is all been requested and answered ” and taking you to the related itemizing, “all in a non-public window of simply you and the AI,” so there is no disgrace in your beginner inquiry.
The modifications to Stack Overflow’s public website are the newest initiative, however the licensing offers and the additions to Groups appear to be serving to the enterprise already.
Going ahead
“We’re rising as an organization,” stated Chandrasekar, whereas declining to reveal financials. Stack Overflow is owned by European funding big Prosus NV of The Netherlands, which acquired it 4 years in the past for $1.8 billion.
Prosus is publicly listed, so the corporate could disclose precise monetary details about Stack Overflow when it declares its full fiscal yr report, which it normally does towards the top of June yearly.
The information options a part of the enterprise has turn into Stack Overflow’s fastest-growing enterprise, adopted by the Groups gross sales, and the promoting enterprise, which is a “very regular” enterprise due to the fixed demand to promote to programmers the place they spend time. Every of the three is a 3rd of income, roughly, stated Chandrasekar.
Maybe the stickiest a part of that is how the neighborhood handles it. There was pushback when the corporate first approached Google and OpenAI about knowledge licensing — not shocking, because the neighborhood considers the 60 million questions their neighborhood property in a way, even whether it is legally the property of the Stack Overflow company.
There was a lot pushback that some customers stated they had been banned from the positioning in the event that they induced a stir over the licensing offers.
With out entering into the small print of previous conflicts, stated Chandrasekar, at this level, Stack Overflow customers have come to comprehend that the corporate just isn’t a not-for-profit and that it must make an earnings to serve its goal.
“I did an AMA two weeks in the past,” an “Ask Me Something” with customers on Stack Overflow, he recalled. “I stated, look, we’re not like another website on the market, however we do not take donations. You will have a enterprise that helps the positioning, and one technique to drive a enterprise is to leverage what you’ve gotten that is helpful so as to add worth within the ecosystem.”
The end result, he stated, is that “they realized it is a good factor for Stack, and for them, as a result of by not doing this, we are literally not capturing the income that we want to have the ability to make investments again into the neighborhood, to construct the moderator instruments that they want. So, the neighborhood understands it now, slowly.”
Need extra tales about AI? Join Innovation, our weekly e-newsletter.
