About 900 million people use ChatGPT every week. The number who use AI the way its builders do, handing an agent a goal and letting it run, is a few million. That gap is most of the story. The world has met a chatbot; far fewer have met what the technology has actually become, and from the chatbot alone it is reasonable to suspect the whole thing is overhyped. It isn’t, and what follows is the case for why, without the hype and without the hand-waving.
The chatbot isn’t the frontier
How most of the world meets AI
For most people, AI is a chatbot: a question typed into a box, an answer read back. That is genuinely useful, and it is also the smallest, most familiar corner of what these systems do. Judging the technology by it is a little like judging the internet by a single web page.
What it can actually do
The shift that matters is from systems that answer to systems that act. Given a goal, a current model will work through it in steps, write and run its own code, use other software, check its results, and return something finished. The clearest measure of this comes from METR, an independent lab that asks how long a task, measured by the time it takes a skilled human, an AI can complete on its own. In March 2025 the answer was about an hour. By January 2026 it was more than five, and that horizon has been doubling every few months.
What it still gets wrong
Now the honest part, dated on purpose because it moves fast. As of May 2026, these systems still get plenty wrong. They are overconfident, giving a wrong answer in the same even tone as a right one. They hallucinate, stating invented facts as if they were certain; OpenAI’s own researchers have shown this is structural, the product of training that rewards a confident guess over an admission of uncertainty. They are uneven in a way that catches people off guard, cracking a genuinely hard problem and then fumbling a simple one. The people who use them most say the same: in a 2025 survey of nearly fifty thousand developers, the most common frustration was AI that is “almost right, but not quite.”
Getting real value out of them works like getting it out of any powerful tool. It means knowing where they are weak, checking their work, and building that checking into how they are used. Done that way, what is left is genuinely powerful. For a reminder of how confidently wrong they can still be, the creator Husk has made a small art form of it:
The weaknesses are real. So is the power. The mistake is to let the first one hide the second, and that is what the rest of this is about.
Pace: how fast this is moving
The pace of change in AI is hard to compare to prior technology cycles. Companies are laying off staff and posting higher margins in the same quarter, a pattern that hasn’t coexisted at this scale in any prior tech wave. Frontier model capability is roughly doubling every twelve months on the relevant benchmarks.1 Compare that to the PC era’s 10-20 year doubling cycles, the web’s 5-10 year cycles, or mobile’s 3-5 year cycles. This isn’t a faster version of those curves; it’s a different curve.
Why it’s not a bubble
The financial question and the technology question are not the same. Whether today’s valuations hold, and which companies survive, is anyone’s guess, and we won’t pretend to know. But they are different questions, and the market, real as it is, isn’t the one we’re pointing at. What matters here is the technology itself, and even more, what applying it does: the impact on how we work, how we live, and how society changes. That is what we’re asking you to focus on.
Different from every prior wave
The reason isn’t pace alone. Every earlier wave, the PC, the internet, smartphones, the cloud, handed people better tools to do human work. This one is different in kind. It is unprecedented in both capability and pace, and what it reaches for is the reproduction of human intelligence itself. Intelligence is the thing that produces new technology, and for the whole history of the world we have been the only ones able to do that at any scale. A technology now closing that gap is not a fad waiting to deflate.
The architectural turning point came in June 2017, in a single research paper from Google.2 From it, in less than a decade, came GPT-scale models, multimodal understanding, and the first agentic systems now showing up across consumer products. Eight authors, twelve pages of math, a deliberately provocative title, and that is the seed of everything since.
To understand why a single 2017 research paper from Google is the architectural turning point of modern AI, it helps to know what the field looked like before. From the late 1990s through the mid-2010s, the dominant approach to sequence-based tasks (translation, text completion, voice transcription) used recurrent neural networks (RNNs) and their more sophisticated descendants, long short-term memory networks (LSTMs). These architectures processed sequences one element at a time, maintaining a hidden state that was updated as each new input arrived. Conceptually elegant; computationally a nightmare. They were inherently sequential, which meant they couldn’t take advantage of parallel compute hardware. Worse, they struggled with long-range dependencies: the further apart two related words were in a sentence, the harder it was for the network to connect them.
Researchers had been trying to fix this for years. Bidirectional LSTMs, attention mechanisms layered on top of RNNs, convolutional approaches to sequences: incremental progress, but no fundamental shift. Then in June 2017, eight researchers at Google Research and Google Brain (Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Lukasz Kaiser, and Illia Polosukhin) published a paper with a deliberately provocative title: Attention Is All You Need. The argument was simple. We don’t need recurrence. We don’t need convolution. The attention mechanism, previously a useful addition to recurrent architectures, could be the entire architecture.
Mechanically, attention is about asking, for every word in a sequence, which other words it should pay attention to. The transformer architecture introduced in the paper used something called self-attention: every token in an input sequence attends to every other token, with the strength of those connections learned from data. Three matrices (query, key, value) compute these relationships in parallel across the whole sequence at once. That parallelism was the second breakthrough. Where RNNs had to process word one, then word two, then word three, transformers process the whole sequence simultaneously. Modern GPUs and TPUs are built for parallelism. Training got much faster, which meant you could train on much more data, which meant you could build much larger models.
The downstream effects were almost immediate and remain ongoing. In 2018, Google published BERT, a transformer trained to understand language bidirectionally, and it broke most of the natural language understanding benchmarks of the time. The same year, OpenAI published the first GPT paper (generative pre-trained transformer), taking the same architecture in a different direction: autoregressive generation. GPT-2 followed in 2019, GPT-3 in 2020, and the qualitative leap between generations surprised even the people building them. Multimodal models followed: vision and language together, then audio, then video. Agentic systems built on transformer cores started showing up around 2023, and the curve hasn’t bent down since.
Why does this matter for the bubble argument? Because the transformer wasn’t a marketing breakthrough or a hype cycle. It was an architectural unlock, the kind of fundamental discovery that happens maybe two or three times a decade in any given field, and rarely with such direct and rapid applications. Computer scientists who lived through the deep learning revolution of the early 2010s often compare the transformer moment to the introduction of the relational database, or to the development of the public-key cryptography that enabled the modern internet. These are infrastructure-level changes that quietly reshape what’s possible for decades afterward.
It’s worth being honest about what the transformer didn’t do by itself. Reinforcement learning from human feedback (RLHF), refined by Anthropic and OpenAI in different ways, was a separate breakthrough. The aggressive scaling of compute and data was a parallel story; researchers have been arguing for years that scale alone explains a remarkable fraction of capability gains. The infrastructure to train models with tens or hundreds of billions of parameters required separate innovations in distributed computing, custom silicon, and data pipelines. And the techniques that made these systems usable in conversation (instruction tuning, constitutional AI, the various forms of alignment work) were their own contributions.
But none of those other things work without the transformer. They are extensions and refinements of an architecture that itself was a clean break from the previous paradigm. When you hear someone claim that “AI is just statistics,” or “we’ve had neural networks for decades, this is incremental,” that’s a position worth taking seriously, but it doesn’t quite hold up against the evidence of the past eight years. The architecture that you can trace in a direct line from a June 2017 paper has produced systems that pass professional licensing exams, generate functional code, hold sustained conversations about complex topics, and increasingly take actions in software environments with limited human guidance. That’s not a continuation of the prior paradigm. That’s a different kind of thing.
If you read one paper to understand where the current moment came from, read Attention Is All You Need. It’s available free on arXiv (2017). For a visual companion that’s accessible without a CS background, 3Blue1Brown’s transformer chapter is the best explainer I’ve found.3 It’s technical but not impossibly so; anyone with a CS background and some patience can follow the core argument. Once you’ve read it, the framing of “hype cycle” becomes harder to sustain. What’s happening is the consequence of a specific architectural insight that solved a real problem and unlocked a generation of capability we’re still working out how to integrate.
People tend to fixate on the line between our intelligence and the intelligence we are building, often in its sharpest form: is the thing actually sentient? It is a genuinely interesting question, philosophically, and we are not alone in finding it so. But the more useful test is an older one. Alan Turing’s question was never whether a machine is “really” thinking, but whether its responses can be told apart from a human’s. Once they cannot, whether it is “really” sentient stops mattering much. Several of the field’s leaders now speak of the Turing test as effectively behind us, with little disagreement. The striking part is how little was made of it: we crossed one of the founding milestones of computer science and simply kept going, which says something about the pace.
Many of the leading minds of this era of AI, all of them standing on the shoulders of giants, have grown blunt about how little separates it from us. Geoffrey Hinton, often called the godfather of AI:
“This question that people often raise of, do these LLMs really understand what they’re saying? The answer is yes, they understand what they’re saying, and they understand what they’re generating, and they understand it pretty much the same way we do.”
Geoffrey Hinton, public lecture, “AI and Our Future”
More plainly, with Jon Stewart:
“This idea there’s a line between us and machines. We have this special thing called subjective experience and they don’t. It’s rubbish.”
Geoffrey Hinton, The Weekly Show with Jon Stewart
And it is far from only him. Plenty of the field’s most respected researchers say versions of the same thing. Demis Hassabis, who shared a 2024 Nobel Prize for his AI work at DeepMind, reaches it from the other direction:
“In the end, it’s all information, and we’re information-processing systems. And I think that’s what biology is.”
Demis Hassabis, interviewed by Hannah Fry, “The future of intelligence”
The line we like to draw between human and machine intelligence keeps getting blurrier, and it is getting harder to name something we can do that AI cannot.
What we’ve lived
We are not writing this from the outside. Our own use of AI has moved in step with the models, and about nine months ago, in the fall of 2025, the pace jumped. A run of real gains from Gemini, OpenAI, and Anthropic arrived close together, enough to let our team take on meaningfully harder work, and we rebuilt our internal tooling around the new capabilities. The effect has been to multiply what our people can do.
What it mostly took off our plates was the rote. With less of our time going to syntax, minutiae, and the uncreative parts of building, more of it goes to the parts that are unmistakably human: judgment, choosing the right problem to pursue, questions of taste. (We also talk to our computers a great deal more now. For the Wispr Flow users among you, can we get an amen.) Those human strengths show through more clearly than they used to, and they matter more, not less, as the raw building gets handed to AI.
We will go deeper on the skills that keep their value in a companion piece, Skills That Matter in the Age of AI.
Where there were doubters, a consensus is forming
The doubters are converting
It was reasonable to doubt this a couple of years ago. In 2023 and 2024 the people with a real view of where it was headed were mostly the ones building it, and from the outside the models were genuinely uneven. Skepticism was fair. What has happened since is that the models got dramatically better, fast, and the doubters have begun to change their minds in public.
Andrej Karpathy is a useful case, because he is no hype man; in October 2025 he dismissed the autonomous agents of the day as “slop.” Three months later:
“Given the latest lift in LLM coding capability, like many others I rapidly went from about 80% manual and autocomplete coding to 80% agent coding.”
Andrej Karpathy, on X, January 2026
Thorsten Ball, an engineer who wrote two well-regarded books on building programming languages in Go and spent two years unconvinced by AI, put it more bluntly:
“For two years I fought with AI. These LLMs are slot machines, I thought… Then, throughout the last six months, I changed my view on LLMs completely.”
Thorsten Ball, “There’s Beauty in AI”, 2025
It’s every camp, not one
The most telling part is who agrees now. Not only the people building it, though they are worth hearing despite the obvious interest: Sam Altman at OpenAI, Dario Amodei at Anthropic, Demis Hassabis at Google DeepMind. The people whose job is to scrutinize it arrive at the same place. Tristan Harris, co-founder of the Center for Humane Technology and the team behind The Social Dilemma, has no stake in the hype:
“AI dwarfs the power of all other technologies combined.”
Tristan Harris, TED, 2025
So do the people writing the rules. Alex Bores, a New York State Assembly member, co-wrote the RAISE Act, one of the first U.S. state laws regulating frontier AI models:
“This technology is moving very quickly, comes with massive amounts of risk, and we need to have a government functioning and regulating it.”
Alex Bores, NY Editorial Board, 2026
(We have more to say about what to do in a world of AI abundance, in a companion post coming soon.)
Builders, critics, and regulators share neither incentives nor opinions about what to do. They no longer disagree about whether it is real. When people that different arrive at the same place, that is about as close to the truth as we get.
If you take away one thing
Read, listen, watch, discuss, debate, engage, and keep current with where AI is going. Stay open-minded and stay skeptical, the way good science is and the way we try to be.
We don’t believe there is a job, an industry, or a corner of society that AI leaves untouched, and we think the reshaping comes fast, within our lifetimes and probably within single-digit years. That is a strong claim, and the right response to a strong claim is not to take our word for it. It is to stay close to it and judge for yourself.
The dotcom era had a bubble. It also had the largest economic force of the next 25 years building underneath it. The bubble popped; the internet did not. AI has the same shape, on a steeper curve, and it works on cognition itself rather than distribution.
Here is where we go to keep up:
- Dwarkesh Podcast: the deepest long-form interviews with the people building and studying AI.
- Andrej Karpathy: a founding OpenAI researcher who teaches how the technology actually works, from the ground up.
- 3Blue1Brown: the clearest visual explanations of the math under neural networks and transformers.
- Hard Fork: a weekly read on AI’s impact on society and policy, skeptical and accessible.
