Slack, Instagram, ChatGPT: Who Really Decides How We Use Tech Products?

One November morning in 2012, Stewart Butterfield asked Johnny Rodgers to join him for a walk along the waterfront in Vancouver’s Yaletown district. The designer, who had joined Tiny Speck eight months earlier, was surprised by the invitation, as he was unaccustomed to private strolls with his CEO.

Butterfield appeared both exhausted and determined. He explained that the game their team of forty was developing, Glitch, lacked a sustainable business model. It was too expensive to maintain, had too few players, and relied on Flash technology, which was incompatible with smartphones where casual gaming was now prevalent. Continuing would mean burning through the remaining cash reserves in a matter of months without anything to show for it.

He then suggested that one of the tools developed internally might be useful to others. As he later recounted on the blog Building Slack, Johnny Rodgers simply responded: “Our IRC server?” Enhanced with some custom scripts, this makeshift IRC server, originally used only by the Tiny Speck team, became Slack. At the time of Glitch, it was not intended to become a business messaging service, but observed usage and financial necessity dictated otherwise. And this case was not unique.

Lack of funds, not intuition, triggers a pivot

Ben Horowitz, whose fund had invested in Tiny Speck, shared the investor’s perspective during a talk at Berkeley in December 2020. The company had raised over 15 million dollars to develop its game, Glitch. One day, Stewart Butterfield called him to report that only 6 million remained, with no possibility of raising more due to lack of results to show. Horowitz identified two main reasons for the failure: the development had begun using Flash before Steve Jobs effectively killed the market on the iPhone, and players completed the game in just two days, which slightly harmed player retention.

Other cases followed a similar pattern, always stemming from a situation of failure or impasse: the growth of Justin.tv, a lifecasting platform launched in 2007, eventually stalled; Fates Forever, a game by studio Hammer & Chisel, received positive reviews but failed to monetize. As for the founders of YouTube, who initially wanted to create a video dating service, they posted ads on Craigslist in Las Vegas and Los Angeles offering 20 dollars to women to post a video of themselves. No one responded.

None of these companies pivoted due to foresight. They did so when there was nothing else left to try. Ben Horowitz offers a more flattering interpretation, summarized by the question he posed to the founder when the failure of Glitch became apparent: “Do you have a secret?” According to him, it was through the struggle to launch an impossible game that Stewart Butterfield learned where software development was lacking.

Every pivot starts with what is already working

At Tiny Speck, an internal tool became a product

The story recounted by designer Johnny Rodgers details the direction taken by the company. At Tiny Speck, no one used email. Everything was managed through an IRC server, a messaging protocol from the 80s, organized into thematic channels. There was a channel for general discussions, one for code deployments, one for incoming support requests, and so on.

Around this core, the team added what was missing. A FTP server that published files sent to it in a dedicated channel, a system that organized conversation histories into a searchable database, and integrations that reported every game sign-up, every credit purchase, every code deployment, and every App Store review directly into the channels.

The designer described a system that only a developer could appreciate, held together by custom scripts and scheduled tasks, with no graphical interface. He also noted that after just two weeks at the company, he couldn’t imagine working without it. A similar situation existed at Hammer & Chisel, where the voice chat, which would become Discord, functioned well, whereas the game itself did not.

Instagram and YouTube scaled down their product

Jawed Karim, cofounder of YouTube, summarized the mechanism that led to the platform we know today in a speech to graduates of the University of Illinois in 2007: “Our users were one step ahead of us.” People were posting videos of their dogs and vacations on what was supposed to be a dating service, and the team eventually wondered why not just let them define what the site was for. Steve Chen, another cofounder, described the same thing at SXSW in 2016, simply by dropping the dating aspect to open the platform to any kind of video.

Kevin Systrom, cofounder of Instagram, quantified the sacrifice on Quora. About a year of work had produced a complete version of Burbn as a native iPhone app. “It was cluttered and overloaded with features,” he wrote. The team hesitated to start over, but eventually kept only the photo, comment, and like features. What remained became Instagram, launched eight weeks later. None of these products were born by adding a feature. All are survivors of a series of eliminations.

Who really believed in it within the teams?

What the pivot stories often omit is that no one really believed in them at the time. Johnny Rodgers writes candidly about the future of Slack. He couldn’t see why a gaming company would start producing a communication software, nor why anyone would pay for the makeshift assembly he knew from the inside. It was mainly for lack of better options that he agreed to stay at Tiny Speck.

A detail in his account goes further. He notes that his boss’s speech that morning seemed already rehearsed. Therefore, the decision had not been made during the walk; it had been simmering for some time, silently, while forty people continued to deliver features for a doomed game.

Jason Citron took the same risk when switching to Discord, except he talked about it as he was doing it. Interviewed by GamesBeat in 2015, he presented the voice chat as a side project to the game, which he described as “the most promising product” his team had built. He also described stopping the game as an emotionally very difficult moment, a confession that no retrospective account reports.

By 2026, publishers continuously measure usage

Observing the practices of users, whether internal or external to the company, has become common, even clearly formalized, in current structures. AI labs, in particular, keep a close eye on what people do with their tools with numerous possibilities.

OpenAI and Anthropic publish quarterly reports

The two labs have made observing usage a permanent function of the company. OpenAI publishes under the name “Signals” quarterly reports accompanied by downloadable time series. Anthropic has released three editions of its Economic Index between January and June 2026.

In its first quarter report, OpenAI states that its users are moving from the experimentation phase to establishing recurring uses. The company specifies that the analysis excludes Codex as well as its business and education offers, thus underestimating professional usage. Anthropic, on the other hand, replaced a weekly sampling with continuous hourly recording, revealing an unanticipated data point: personal use accounts for about 35% during the week and nearly 50% on weekends. Research teams are now in place to continuously monitor user behaviors.

Vibe coding, a practice named by a user

The example of vibe coding is also revealing: the term emerged from a message published by Andrej Karpathy on X in February 2025, where he described his own weekend practice. No publisher had launched a product bearing or using this name, but the industry followed the hype started by a user (albeit a cofounder of OpenAI). Cursor, one of the coding publishers that became emblematic of this wave, claimed 100 million dollars in annualized revenue in January 2025. It announced it had surpassed one billion in November of the same year.

The gap between discourse and usage persists

While these companies measure everything and publish their results, they also sometimes miss what their product is becoming. But agreement on how to measure remains elusive. The ranking AI in the Wild places therapy and companionship at the top of the uses of generative AI for the second consecutive year, with a share that has doubled in a year. A study conducted with OpenAI, on the other hand, attributes only 1.9% of messages to relationships and personal reflection. Despite measurements and publications, it remains difficult to agree on a seemingly simple question: what is the product for?

However, some findings are hard to contradict, especially the use of AI assistants for coding. Yet, the entire industry has sold the redesign of software development with the advent of artificial intelligence. But in a study by OpenAI co-authored with researchers from Duke and Harvard involving about 1.1 million conversations, computer programming accounted for only 4.2% of all messages sent to ChatGPT. Writing, on the other hand, represented 40% of work-related messages, while personal uses surged. Thus, publishers like OpenAI or Anthropic quickly shifted to dedicated solutions, such as Claude Code and Codex, separating them from their more “general public” products.

Uses continue to lead to new features or products. The health tab or the option to add a trusted person in ChatGPT are examples. But between 2012 and 2026, what has changed is not the foresight of companies, but the speed at which they measure usage. Nonetheless, some aspects still elude those who design the products.