Nvidia agreed this week to buy Hugging Face for $12.9 billion. For most people that is a transaction. For Michael Nuñez it is a closing scene: he was VentureBeat's editorial director from 2023 to 2026, and his first big story for the publication was a party that Hugging Face's chief executive convened almost by accident three years ago. He leaves the beat in the same week the company that hosted the open-model movement agreed to be bought by the company whose hardware sits underneath all of it.
The party is worth recalling in detail, because it dates the curve. Clem Delangue, Hugging Face's co-founder and CEO, was in San Francisco and suggested offhand that a few hundred open-source AI developers get together. Three weeks later about 5,000 people showed up at the Exploratorium. There were live llamas. The organisers made a joke of Meta's model, whose weights had leaked onto the internet a few weeks earlier. Someone called it the Woodstock of AI and the name stuck. Nuñez spoke to Delangue and Andrew Ng that day, and describes Delangue as calm, uninterested in the attention, and visibly embarrassed by the size of the thing he had accidentally organised.
Days before the event, Delangue had written to registered attendees that a meetup which started as a social media post might turn out to be the largest gathering of AI developers in history. Nuñez's verdict on that prediction, three years on, is that Delangue got it wrong in the other direction: it was the smallest gathering that ever mattered — the pond on the morning the water covered a quarter of the surface and the whole thing still looked like a party.
Set the acquisition price against Nvidia's own accounts and it shrinks. Nvidia booked $96 billion of revenue last quarter, more than double the year before. At that pace, the entire purchase price of Hugging Face is roughly twelve days of sales. This does not read like a company stretching for a strategic asset. It reads like a company buying the place where models are distributed at a price it will not feel, which is a different kind of statement about who sets terms in this market.
The organising idea of Nuñez's account of the last three years is a piece of 1970s psychology. In 1975 the Dutch psychologists Willem Wagenaar and Sabato Sagaria published a paper called "Misperception of exponential growth." Four years later Wagenaar and Hans Timmers turned it into the thought experiment that outlived the paper: imagine duckweed that doubles its area every day, give people a few starting values, and ask when it covers the pond. People answer confidently and wrongly, and the error grows as the process nears completion. They see the curve and mentally draw a straight line through it.
The bias is durable but not fixed. During the pandemic, Joris Lammers and colleagues showed in the Proceedings of the National Academy of Sciences that most Americans read COVID-19 case growth as linear, which partly explained resistance to social distancing — and that three sentences of instruction measurably corrected the perception. Political orientation mattered too: conservative participants made the error more often.
The best part of the story is the correction. In 2022 the legal scholar Hanjo Hamann went back to the 1975 paper and found a mistake in it. Wagenaar and Sagaria had carelessly rounded a starting value, so that by the tenth step their own exponential calculation was off by 168%. The researchers who demonstrated that people cannot extend exponential curves had failed to extend one themselves. The conclusion survived; the arithmetic did not.
That detail cuts against the tidy version of the lesson. The problem is not only that intuition is badly calibrated — it is that the correction requires doing the arithmetic carefully every time, including when you are the person who identified the bias.
Nuñez applies the frame to the numbers he spent three years writing about. ChatGPT went from research demo to more than 1 billion weekly users in under four years. Anthropic reached a $30 billion annual revenue run rate after growth its own executives described as insane. Nvidia's quarter more than doubled. Companies went from a Series A deck to a multi-billion-dollar valuation faster than most industries close a fiscal year. Individually, none of these facts surprises anyone now. Together they still do not feel real.
His own misses ran in one direction. In early 2024 he told a colleague that useful AI agents were one to two years out; they shipped that autumn. He expected the gap between open models and the frontier labs to widen, and it narrowed instead. He watched a model that could not count the letters in "strawberry" become a system that writes code for 30 hours unsupervised. Stories filed at 11 p.m. were stale by breakfast; benchmarks that set the quarter's agenda became footnotes the next day.
The reporting record is the part that holds up best. VentureBeat was first on DeepSeek V3.1 and the V3.2 models that matched GPT-5 despite export controls and shipped under an MIT license — the end, as Nuñez frames it, of the assumption that frontier models are necessarily American and closed. It broke the Salesforce and Anthropic deal that put an entire CRM inside Claude. It was among the first to report the joint warning from OpenAI, DeepMind and Anthropic that the companies might be losing the ability to understand their own systems. Nuñez interviewed Sam Altman, Dario Amodei, Andrew Ng, Marc Benioff, Mustafa Suleyman and Kai-Fu Lee, along with dozens of researchers whose names are not yet widely known.
Here is where I part company with the thesis. Exponential-growth bias is a genuine finding and an unusually convenient one for this industry to adopt about itself. It converts "we were wrong about how big this would get" into "human cognition is wired incorrectly," which is both flattering and unfalsifiable in the direction that matters: it supplies a standing explanation for why every skeptic is wrong, and none for when the extrapolation has run too far. Nuñez includes, among his errors, that he doubted the revenue figures Anthropic and OpenAI were citing almost until the companies reported them officially. That is not a cognitive failure. That is a reporter declining to print unaudited numbers, and filing it under bias is the point where the frame starts absorbing things that do not belong to it.
The account is also quiet about the obvious consequence of its own ending. Hugging Face mattered because it was the neutral ground where open weights lived — the counterweight to the closed labs, and the reason the DeepSeek story could be told as a shift in who controls frontier capability rather than as a curiosity. Nothing in Nuñez's telling addresses what that neutrality is worth once the repository belongs to the one vendor every lab in the ecosystem already depends on for compute. The deal is presented as a satisfying bookend to a personal story. It is at least as plausibly the moment the open-model layer stopped being independent infrastructure.
The public mood has already moved. Pew Research Center's latest survey, conducted in June, found 52% of Americans more concerned than excited about AI in daily life, against 37% in 2021. Only 9% feel the reverse. 71% expect job losses. For the first time, a majority of adults under 30 — the cohort that uses these tools most — sits in the worried camp. That is not technophobia; it is what a linear mind does when the ground moves exponentially, and it lands as vertigo rather than surprise.
The media rules moved too, and faster than the 2023 argument suggested they would. Last week the Wall Street Journal ran a column by the investor Stanley Druckenmiller that detection software flagged as machine-written; Druckenmiller did not deny it and said he had used AI and was not embarrassed about it. A week earlier the Financial Times corrected a column by a Harvard economist who had not disclosed using AI to shorten his text. In spring, the New York Times rewrote its rules for outside contributors after a run of its own awkward episodes. When Nuñez joined VentureBeat in early 2023 and published an editor's letter explaining how the newsroom would use these tools and where it would draw the line, he was criticised publicly on X.com for the disclosure. He still thinks disclosing was right; he is no longer sure he got the tone right. Three years later the argument is not whether a serious publication may use the tools but how it should say so — a better argument to be losing pieces of than the old one.
Nuñez is not leaving the industry, only the chair; he says he will describe the new position later. His pattern recognition was earned before this cycle — he spent a decade covering social platforms, and his 2016 Gizmodo investigation into Facebook's Trending Topics triggered a US Senate review and entered the Congressional Record. He saw the same dynamic here: a technology moving from niche to substrate faster than people could redraw their picture of the world.
Which leaves the asymmetry unresolved. The exponential frame is excellent at explaining why everyone underestimated the last three years, and it offers nothing at all for detecting the moment the line starts running above the curve instead of below it. The same instinct that made a $12.9 billion acquisition look inevitable in hindsight is the instinct that will make the next one look cheap.