Why a Zapier founder thinks AGI research has been stalled for four years

Start with a number. In 2019 François Chollet, the creator of Keras and a Google deep-learning researcher, published ARC-AGI, a test built to measure general intelligence in machines. Four years later the best systems sit at 34%. When the eval first appeared, state of the art was 20%. In the same window GPT-4 and Gemini shipped, and Lab42 in Switzerland ran a small annual contest against the puzzles. The score barely moved.

Mike Knoop wants to make that number the point. He co-founded Zapier in 2011, has run the company's AI work for about two years, and has now joined Chollet to launch ARC Prize: a million-dollar-plus nonprofit public challenge to beat the eval, open-source the solution, and open-source the progress toward it. To his knowledge, ARC-AGI is the only true AGI eval in the world, because it measures the definition he thinks is correct.

He is not a keyboard critic. Days are spent building application-layer AI at Zapier; nights and weekends go to AGI, a curiosity that runs back to college and that he says is why Zapier was early to AI at all. The Chain of Thought paper in January 2022 shook him loose. He was running half the company, gave up the exec-team role, and went back to being an individual contributor to answer one question for himself: how close are we to AGI? The answer: not that close. Progress, he believes, has stalled for four or five years, mostly because the consensus definition is wrong, people are measuring the wrong things, and that makes the world think it is closer than it is.

The definition Knoop hears most in the industry is Vinod Khosla's: AGI is a system that can do the majority of economically useful work that humans can do. He does not call it useless. His day job lives on that line; language models in the current regime unlock real economic value, and Zapier is capturing some of it. He does not think it is a good AGI definition. It is a good definition of useful systems. He jokes that it says more about what many humans do for work than it does about actual general intelligence.

The wrong target has a research cost. If you think you are close, you over-invest in exploiting the large-language-model regime and stop exploring new ideas. Frontier labs have also, in his view, stopped publishing. The GPT-4 paper had zero technical details; the Gemini paper had zero technical details on the long-context work. He launched the prize because he wanted to see if he could help re-accelerate the search.

Chollet's definition is much shorter: general intelligence is a system that can efficiently acquire new skills. That is the whole sentence. Knoop's kitchen-table argument is the last decade of specialist wins. AI can now win at poker, fold proteins, drive cars, and win at chess, and yet you cannot take a system trained to beat poker and teach it to drive. A person is different. Pull them into a parking lot, teach them a car they have never driven, or show them a variant of poker, and they pick it up. Sample-efficient, energy-efficient acquisition of a new skill is what makes the ability general, and it is what pretty much every AI eval today does not measure.

ARC-AGI, built in 2019, is the formalization of that definition you can actually test against. It looks like an IQ test: a few input-output grids, a hidden rule, a new grid the system has not seen. Humans do not find the puzzles hard. Machines have been stuck for four years. Knoop stresses the timing. The benchmark was born before language models and before the scale wave, and it has successfully resisted both.

Sarah Guo puts the objection most listeners have. LLMs already look general: one system, trained broadly, that can do many specific subtasks. Why wouldn't more scale, or a multimodal version of the same idea, eventually get you there? Knoop's first answer is mechanistic. What large language models do today is high-dimensional memorization. They train on a lot of data, find and generalize patterns, and apply those patterns in new contexts. Memorization is a form of intelligence, he grants. It is not a form of general intelligence.

The AGI he wants is a system that can invent alongside people: new branches of physics, a better picture of the universe, new therapeutics. Those answers do not sit in high-dimensional patterns of existing training data, because the answers are unknown. Current agents work when the reasoning chain is simple enough that it shows up, in some abstract form, in the training set; the model plucks it and applies it. That, he says, is how Zapier's AI bots actually work. On open-ended problems whose reasoning chain exists nowhere in the data, language models fall flat. It is also why dazzling agent demos die on a task the model has never seen.

"At the end of the day I'm a purist," he says. In AI the only method that works is looking at what works and what does not. Objectively, language models do not beat ARC, and people have tried. Transformers might still be a component. The big gift of the last few years, in his view, is a robust perception stack: a deep network that can take multimodal data, turn it into numerical representations, and operate over them. That path might lead toward true AGI. The language-model version of it — next-token prediction trained on data — will not. Double the parameters, ten-x the parameters, ten-x the data: he claims you still never get AGI. Something additional is missing.

Asked what people should explore instead, Knoop names two old ideas. The first is program synthesis, a technique older than the current model stack and one that has already shown promise on ARC. A program searches the space of possible programs, assembling them so that a given input maps to a given output. There is no backprop gradient to follow. You sample the full search space, which, he argues, raises the odds of finding a general-form solution. That line of work got some of the mid-20% progress on ARC, and it is almost completely orthogonal to language-model chain-of-thought.

He offers an implicit contrast with gradient descent. Following a slope takes you to the nearest basin. Searching program space is more like sampling islands: the answer can live far from your starting weights. Because the search is relaxed and universal, it is less trapped by the local geometry of a neural net.

The second bet is neural architecture search: let computers discover the network architecture itself. The field is maybe ten years old and has never amounted to much. Knoop's diagnosis is resources. It has lived mostly in academia, and academic researchers lack large-scale compute, so they take shortcuts that produce publishable results. That constraint may have lifted. Four years of cheaper FLOPs, he suspects, is enough to revisit the old methods, relax the search, and take the bitter lesson seriously — stop biasing the search with human priors, and spend the cheap compute.

Why a prize instead of writing checks to startups? Because outsiders are needed. In 2023, 300 teams entered Lab42's smaller ARC contest. Look at the list, Knoop says, and you see one- or two-person teams, not employed at AI startups, often not even living in the Bay Area, Silicon Valley, or California. It is a globally distributed set of people with new ideas. He would bet that the solution to ARC comes from someone who has not been indoctrinated in language models and scale. He counts himself as half an outsider: Zapier is not an AI lab, and his AGI work is nights and weekends.

The puzzle itself is almost offensively small. He calls it a minimal reproduction of general intelligence: a grid that at most is about 15 by 15 squares, a dataset small enough to hold in your head. He thinks the solution can likely be written in 10,000 lines of code or less, and that it will not require a 200-billion-parameter model. That puts it in reach of tinkerers. The prize is meant to change the nights-and-weekends menu for the would-be AI researcher. The default is another LLM startup you might sell. The alternative he wants to make attractive is: try to beat ARC-AGI. Status used to be the only upside. Now there is money too.

Language models unlocked an economic incentive for startups. He is trying to use the prize to counterbalance some of that pull. If you care about new ideas more than another wrapper, the public challenge is where the upside is now attached to a problem the scale regime has not solved.

Half the inspiration for the prize is Knoop's day job. When companies adopt a new technology they often spend months just learning what it can do. Zapier's version of that education was brutal and short. In the summer of 2022, Knoop and co-founder and CTO Bryan Helmig gave up their exec-team roles, took no direct reports, and spent about six months as individual contributors. They implemented Chain of Thought and Tree of Thought, and built an internal ChatGPT-like system before ChatGPT shipped. The exercise gave them confidence they had, to the best of their ability, explored what a GPT-3-style model could do.

The gap they found was that the models were frozen in time. This was before tool use. Zapier had about 6,000 integrations sitting there. Hooking language models up to those tools is what made the company a launch partner for the ChatGPT plugin, and it is the first moment Zapier became widely associated with AI. The numbers are no longer a prototype. Over the year and a half they have been tracking, more than 50 million AI tasks have run on the platform. The dominant pattern is still a Zap with an OpenAI or ChatGPT step in the middle of a workflow: content generation, feature extraction, summarization. That, Knoop says, is how most people actually adopt AI today.

The newer product is Zapier Central: AI bots you program in natural language instead of clicking through the editor. There is not even a natural-language-to-structure mapping step. It is a pure inference engine that interprets the user's instruction and then uses the integrations and credentials the user has equipped it with. Ease of use, he says, jumped by an order of magnitude. As for agents, the future is not far away. It is happening: people are already paying for Zapier AI bots. He draws the product as concentric rings. Each time consistency and reliability mature, a new circle of use cases unlocks. Right now that circle is personal and team workflows where the cost of being wrong is relatively low.

The same bot template makes the risk gradient obvious. Watch the inbox for a payment or refund notification, and if it qualifies, post to Slack. A three-person startup with a handful of channels will let the bot pick the channel and write the message. Hand the identical template to a mid-market company with thousands of Slack channels, partner channels, and production traffic, and they will want the possibility space clamped down. Zapier's product thesis is to let individual users install those clamps — to decide how tight the bot's behavior space is. Better clamps enlarge the circles you can unlock. Product work still has reliability left to squeeze; the underlying models will also get more consistent. The two curves, together, pull higher-risk use cases inward.

Guo also asks a more philosophical question, citing Peter Watts's novel *Blindsight*: intelligence and self-awareness are not the same thing, so does AGI include sentience? Knoop is not a philosopher and has not thought deeply about whether the system would be conscious. What he cares about is living in the future, and he thinks one of the best ways to pull that future forward is to invent systems that can invent and discover alongside us. That requires the general form of intelligence: efficiently acquiring new skills and helping with open-ended problems. On risk, he wants an empirical lens. Predicting future capabilities and then writing laws that prohibit research directions through a theoretical frame, he says, has not worked. Nobody sitting today would have called where AI would be five years ago.

Open source sits in the same argument. AGI progress has stalled; the field still needs fundamental new ideas. The internet and open source are, in his view, among the world's best inventions for generating them. If you care about discovering AGI in our lifetime, it is incumbent to promote anything that raises the odds of new ideas and puts problems in front of many researcher brains, not behind a big lab's hiring process. The proof is OpenAI itself: the company's genesis came out of a published research result from Google. Commercialization has now made frontier labs play close to the chest. GPT-4 and Gemini papers withheld the details that used to seed the next lab. The trend upsets him because it stalls progress. ARC Prize is a small counterweight.

Elad Gil adds the historical rhyme. The internet won with open protocols against closed ones, and with Linux as the workhorse of early servers. In the 1990s the U.S. government tried to regulate cryptography for overlapping reasons — malicious actors, threats to security — and cryptography's spread was net positive. Knoop's policy split follows. Narrow AI has not created harms that existing agencies lack the power to handle; regulate it through the frameworks already on the books. On AGI, prescriptive legislation ahead of any empirical evidence of what the systems can do is, he thinks, really dangerous. His personal line: he would not trade personal independent freedom for what it would take to prevent AI from ever being developed. Discover what the systems look like, then decide how to release and control them. Predicting that in theory is the move he is strongly against.