Today is my last day as Editorial Director at VentureBeat.

When I try to make sense of the last three years, my mind keeps returning to my first weeks on the job. It's a writer's instinct to look for narrative symmetry, but sometimes the universe just hands you the perfect bookend.

Three years ago, my first major story at VentureBeat started with a tweet. Clem Delangue, the co-founder and CEO of Hugging Face, was in San Francisco and casually floated the idea of getting a few hundred open-source AI developers together for a meetup. Three weeks later, roughly 5,000 people descended on the Exploratorium. There were live llamas on site, a joke about Meta's model, whose weights had leaked onto the internet weeks earlier. They called it the "Woodstock of AI."

It became part of San Francisco legend. More importantly, it was a breakout moment for artificial intelligence and a resurgence for the city. For those of us who had weathered the darkest, quietest days of San Francisco in 2021, watching the city jolt back to life in a single afternoon was electric. I sat down for an interview with Delangue and Andrew Ng that day. Delangue was quiet, unassuming and genuinely bashful about what he had accidentally orchestrated. He was just happy they had pulled it off.

It all comes back to that moment today because, this week, Nvidia agreed to acquire Hugging Face for $12.9 billion. I couldn't have scripted a better bookend if I tried.

A tweet became 5,000 people. Five thousand people became $12.9 billion.

That is the story of the last three years, and it's the one I want to tell on my way out, because I think it explains why this period has been so hard for smart people to reason about.

ChatGPT went from a "research preview" to more than one billion weekly users in less than four years, roughly the time it took many media companies to convene a task force. I reported on Anthropic reaching a $30 billion revenue run rate after growth its own executives described as "crazy." Nvidia booked $96 billion in revenue last quarter, more than double the year before. I watched companies go from a Series A pitch deck to a multibillion-dollar valuation in less time than it takes most industries to close a fiscal year.

None of these facts is individually surprising anymore. Cumulatively, they still don't feel real. There's a reason for that, and it isn't a character flaw.

The pond, the duckweed and the line we all draw

In 1975, the Dutch psychologists Willem Wagenaar and Sabato Sagaria published a paper with a title that reads like a diagnosis: "Misperception of Exponential Growth." Four years later, Wagenaar and Han Timmers turned it into a thought experiment that has outlived the paper itself. A patch of duckweed doubles every day. Given a few data points, when does it cover the pond? People guess wrong, confidently, and the further along the growth is, the worse they do. They see a curve and draw a line.

Exponential growth bias, as it's now called, is the tendency to underestimate compounding processes by extrapolating them linearly. It is robust, but it is not destiny. During the pandemic, Joris Lammers and colleagues showed in the Proceedings of the National Academy of Sciences that most Americans saw COVID-19 case growth as linear, that the misperception partly explained resistance to social distancing, and that three sentences of instruction measurably corrected it. Politics mattered too; conservatives were more prone to the error. The brain ships with a linear default. The default can be overwritten, but only on purpose, every single time.

There is a humbling coda. In 2022, the legal scholar Hanjo Hamann went back to the 1975 study and found it was flawed. Wagenaar and Sagaria had rounded their starting value so carelessly that their own exponential was off by 168% by the tenth step. The researchers who proved people can't extrapolate exponentials had failed to extrapolate their own. Their conclusion survived; their arithmetic didn't. Nobody is immune. That, more than anything, is what the last three years taught me.

I've come to believe that exponential growth bias is the master key to this period. It explains the industry, which kept underestimating its own products. It explains the media, which kept treating those products as a fad. And it explains the public, which is now, understandably, uneasy about a technology that seemed to arrive all at once. Nobody was stupid. Everyone was running the same default. The only question was who caught themselves, and when.

What it cost to bet on AI in 2023

When I joined VentureBeat in early 2023, saying AI would define the next era of media was not a neutral statement. Serious technology journalism treated large language models as a novelty at best and a plagiarism engine at worst.

I said it anyway, on the record, to Bloomberg, the San Francisco Standard and the Palo Alto Daily Post in my first weeks on the job: this is bigger than social media. I had spent a decade covering social platforms, including a 2016 Facebook investigation for Gizmodo that triggered a U.S. Senate inquiry, and I recognized the shape. A technology was moving from fringe to substrate faster than anyone's mental model could update.

I also published a letter from the editor explaining how our newsroom would use AI tools and where we'd draw the line. The transparency earned me a public dragging on X.com (formerly Twitter), which stung, and which I've thought about more than I'd like to admit. I still believe disclosure was the right call. I'm less sure than I was that I made it in the right tone.

The press was running the same linear default as everyone else. Last week, the Wall Street Journal ran an op-ed by the investor Stanley Druckenmiller that detection software flagged as machine-drafted. Druckenmiller shrugged: "Of course I used AI. I'm not embarrassed by it." The week before, the Financial Times corrected a Harvard economist's column for undisclosed AI condensing. The New York Times had already rewritten its rules for outside contributors in the spring, after a run of its own embarrassments.

An opinion page is not a newsroom, and a billionaire contributor is not a beat reporter. But the debate has moved. In 2023 the question was whether a serious publication could touch these tools at all. In 2026 the question is how to disclose them. That's a better argument, and I'd rather be part of it than be proven right about the old one.

Three years that felt like one long week

Here is what the curve looked like from inside it.

We were first on DeepSeek V3.1 and on the V3.2 models that matched GPT-5 under an MIT license despite export controls, the end of the assumption that frontier meant American and closed. We broke the Salesforce-Anthropic deal that put an entire CRM inside Claude. We were among the first to cover when OpenAI, DeepMind and Anthropic jointly warned they might be losing the ability to understand their own systems. I sat across from Sam Altman, Dario Amodei, Andrew Ng, Marc Benioff, Mustafa Suleyman and Kai-Fu Lee, and from dozens of researchers whose names you don't know yet and will.

I also got things wrong, always in the same direction. I told a colleague in early 2024 that useful agents were one to two years out; they shipped that fall. I thought the open-source gap with the frontier labs would widen. It closed. I doubted the revenue numbers that Anthropic and OpenAI were floating right up until they reported them. Every miss was a line drawn through a curve.

The strange part was never the volume. It was the clock. I watched a model that couldn't count the letters in "strawberry" become one that codes unsupervised for 30 hours, and it did not feel like three years. It felt like one long week that kept resetting. Stories filed at 11 p.m. were stale by breakfast. Benchmarks that defined a quarter were footnotes by the next. Time doesn't move faster at the front of an exponential. It compresses, because the distance between "impossible" and "shipped" keeps shrinking, and a brain built for lines keeps waiting for a pause that never arrives.

The public isn't wrong to feel dizzy

The same bias that made technologists underestimate their own technology now makes everyone else feel ambushed by it.

Pew Research Center's latest survey, conducted in June, finds 52% of Americans are more concerned than excited about AI in daily life, up from 37% in 2021. Just 9% say the reverse. Seventy-one percent expect fewer jobs. And for the first time, a majority of adults under 30, the generation most fluent in these tools, has crossed into the concerned column.

That isn't Luddism. It's what a linear brain feels when the ground moves exponentially: not wonder, vertigo. People were told for years this was a toy, and then the toy started writing their performance reviews. The whiplash isn't a failure of public understanding. It's a rational response to a bad map, and the people who drew the map, including the press, owe them a better one.

It changed my job. For most of my career, the task was to be skeptical of hype. For the last three years, the harder task was to be skeptical of my own instinct that the hype couldn't be true, while staying honest about who was getting hurt on the way up. Both at once. Most days. The reporters who managed it weren't the ones with the best sources. They were the ones who accepted that their gut would be wrong about the slope, and checked the arithmetic anyway. I got to work alongside a lot of them.

Thank you, and what comes next

None of this happens alone. To the reporters and editors who made VentureBeat the fastest room on the fastest beat in the world: Sharon Goldman, Carl Franzen, Sean Michael Kerner, Ben Dickson, Dean Takahashi, Gina Joseph, Victor Dey, Shubham Sharma, and to Matt Marshall, who made the bet. To every source who trusted me with a scoop, every reader who argued with me in the comments and every communications lead who took my 11 p.m. call: thank you.

I'm not leaving the industry. I'll be telling these stories from a different vantage point, and I'll have more to say about that soon. Until then, find me at nunez-sf.xyz, through my newsletter Nuneybits, on LinkedIn and on X at @MichaelFNunez.

A few days before the Exploratorium, Delangue told his RSVP list that what started with a tweet might become the biggest AI meetup in history. He had it backwards. It was the smallest one that would ever matter: the pond on the morning it was a quarter covered, when it still looked like a party.

Every room since has been filled with smart people, me included, waiting for the curve to slow. It never did for long.

The curve sets the speed. People choose the direction. The journalist's job is to write down both while it still looks like a party.

Michael Nuñez is an editor and investigative journalist covering artificial intelligence and enterprise technology. He served as Editorial Director of VentureBeat from 2023 to 2026, leading its coverage of OpenAI, Anthropic, Google DeepMind, Microsoft, Nvidia and the enterprise AI ecosystem. His 2016 investigation into Facebook's Trending Topics for Gizmodo prompted a U.S. Senate inquiry and was entered into the Congressional Record. He has previously worked at Forbes, Mashable, Gizmodo, Popular Science and the deep-tech venture firm Playground Global. He is based in San Francisco and serves on the board of the San Francisco Press Club. Portfolio: nunez-sf.xyz. Newsletter: Nuneybits.