Transformers have ruled AI for a decade. But some think world models, pure reinforcement learning or neurosymbolic AI might be a better path to true intelligence
Intelligence is a slippery thing. For millennia, humans assumed it flowed through a fluid, or one continuous sheet crumpled inside the skull. Microscopes later revealed a dizzying network of individual neurons, sculpting intelligence from electrochemical noise to give our brains a certain je ne sais quoi.
In the late 1950s, psychologist Frank Rosenblatt designed the perceptron, a brain-inspired algorithm that adjusts the relative strength of its units’ connections with experience. The *New York Times *described it as “the embryo of an electronic computer” that would one day “be able to walk, talk, see, write, reproduce itself and be conscious of its existence.”
By 1971, the year Rosenblatt drowned in a boating accident, this “connectionist” approach to artificial intelligence had been overpowered by the field’s attachment to hand-coding clever rules based on human knowledge. (In the 1970s, for example, AI researchers asked doctors how they spot bacterial infections, and coded a diagnostic tool by meticulously transcribing hundreds of human-inspired if-then rules. Any expertise doctors couldn’t put into words — one might call it vibes — was inherently left out.)
Fast-forward to today, and every large language model is powered by deep learning, an incredibly connectionist approach. It’s perceptrons all the way down, stacked more than Rosenblatt could imagine. AI researcher Rich Sutton called it the Bitter Lesson: general, brute-force computation usually works better than trying to recreate our own thinking from scratch. And sometime after the 2017 introduction of the transformer, AI developers found another way to piss off old-school purists who believe intelligence has to be intentionally designed**: **make models better by making them bigger. Hundreds of billions of parameters, the numerical knobs that set the strength of connections between model units. Scrape the entire internet, digitize the world’s books, intellectual property be damned! And so far, it’s been getting the job done. Current LLMs are really powerful, and shockingly capable of executing many (though certainly not all) long, complicated tasks without human oversight.
Scaling continues to work so well that there’s little hard evidence to suggest that it’ll stop working any time soon — just a hunch that a strategy so crude couldn’t possibly create superintelligence. Yet some of the field’s leading figures still think it’s a dead end. Before leaving Meta last year, Yann LeCun said “We are not going to get to human-level AI just by scaling LLMs.” In the words of OpenAI co-founder Ilya Sutskever, whose 2012 work on AlexNet flaunted the power of scaling over traditional machine learning methods: “From 2012 to 2020, it was the age of research. Now, from 2020 to 2025, it was the age of scaling … But now the scale is so big … So it’s back to the age of research again, just with big computers.”
Deciding what research to throw big computers and billions of dollars at depends on how you think “intelligence” really works. Some believe that the structure of large language models can never represent the world like humans do. Others suspect that the way LLMs are trained prevents them from learning as well as humans. But whether it’s the model’s underlying architecture or training protocols to blame for its shortcomings — or something else entirely — scaling skeptics share the conviction that evolution blessed human minds with some fundamental primitive of intelligence that shoving text through a transformer can’t recreate. Now, a handful of them have left frontier AI companies to figure out what to build instead.
Growing up in Florida, I had to log 50 supervised hours behind the wheel to be eligible for a driver’s license. Just over a single work week puttering around with my terrified parents taught me enough to pass the test. I’m not the best driver (Florida isn’t known for minting those), but I’ve made it over 14 years without a single accident or traffic violation.
There’s nothing special about me or my driving prowess, but it demonstrates something that make humans particularly good learners: we are able to learn from relatively limited experience. LLMs, despite having seen lots of cars and memorized every single driving-related book on Earth, still cannot drive a car. Passively absorbing multimodal information is not the same as hands-on experience. In fact, one could argue that these models don’t “learn” at all, at least not in the animal sense, through some combination of evolution, instruction, and doing stuff in the real world. Rather, they’re trained on the fruits of human learning, sculpted by gradient descent, polished by human (and increasingly, AI) feedback, and trained on specific tasks in a reinforcement learning “gym” (more on that later). Then, their weights are frozen in time.
Other non-LLM AI systems, such as the models under the hood of Waymos, do learn from experience (both real and virtual). But they need many orders of magnitude more experience to get as good as a human. Waymo’s AI model has traveled more than 200m miles and counting, plus billions more virtual ones, and racks up more with every trip. In human years, that’s hundreds of lifetimes, and a Waymo might still drive in inexplicable circles or run over a beloved bodega cat. And while a Waymo can drive a car, it couldn’t write this article for me.
The holy grail for AI researchers is a model that can resolve both of these problems: AI systems that can learn from experience, and not much of it.
Some researchers, most notably Yann LeCun — Meta’s former chief AI scientist — have spent decades pushing for loosely brain-inspired model architectures that can, at least in theory, learn more effectively than even the most scaled-up LLM ever could. Saying LLMs are “just predicting the next token” has become an AI skeptic trope, but mechanistically, that is indeed what they’re doing. (To do this as well as frontier models do today, they’ve had to build complex representations of language that have long since promoted them above “stochastic parrots.”)
LeCun’s alternative vision is built on JEPA, the Joint Embedding Predictive Architecture. Instead of recreating and predicting raw data — individual words or single pixels, like a generative model does — it would compress that data into a more abstract state, throwing out unnecessary information while preserving the important gist. While this system passively consumes sights, sounds and words, LeCun and his colleagues envision another system, operating in parallel, that *actively *does things, learns from their consequences, and generalizes those lessons to new situations. For example, a robot powered by Meta’s V-JEPA 2, a version of this model, learned to move specific objects on command, without having seen those objects before. Theoretically, these systems can train each other: as psychologist James J. Gibson put it, we see in order to move and we move in order to see.
This vision combines two related but separable bets. The first is on world models: systems that learn how the world works through multimodal observation. A world model is something a mind carries — we probably have one, in neural networks sprawling across the hippocampus and prefrontal cortex. It’s built partly through experience and, maybe, partly through hard-coded genetics that bias what kinds of sensory data we care the most about (minutes-old newborns, for example, prefer looking at blobs arranged like a smiley face over other abstract patterns). It serves to predict how one action leads to another, allowing an agent to make informed choices.
The second bet is on reinforcement learning, which is something a mind does to learn. Reinforcement learning, or RL, without a world model is simply learning from trial and error, without being fed human-curated training data at all. Our brains do this with temporal difference learning, an RL algorithm which, curiously, AI researchers Rich Sutton and Andrew Barto mathematically described in the 1980s — years before anyone realized that the very same algorithm powers our dopamine system.
With help from a world model, that trial and error process could potentially be compressed. An agent could simulate potential actions and outcomes before committing to a move, reducing the amount of pure guesswork involved — although for now, this is limited to fairly short, specific tasks. The broader vision of generalizable “action-conditioned world models” is what LeCun’s recently-launched startup, AMI Labs, is aiming for. In theory, these world models could simulate cause and effect abstractly, beyond the physical domain. Knowledge workers regularly take lofty goals — “boost engagement,” or “launch in a new market” — and break them down into sub-steps and sub-sub-steps through hierarchical planning. Frontier LLMs are getting better at this, but often aren’t reliable working on long, multi-step tasks. World model advocates suspect that a better model of cause and effect is what’s missing.
Other startups chasing a post-LLM future are mostly focusing on world models *or *reinforcement learning (or, in at least one case, not telling anyone what they’re doing at all). This all assumes that there’s something fundamental to human intelligence that the current paradigm will never crack. But with enough data and a giant pile of transformers, LLMs may reach “AGI,” whatever it means, before any alternatives come to fruition.
While some alternatives to scaling are still largely theoretical, world models are already out there, making photorealistic images and training self-driving cars. Several frontier companies are on it: Google DeepMind announced plans to make “massive generative models that simulate the world” at the start of 2025. Later that year, Meta poached Tim Brooks, who was leading DeepMind’s efforts and previously worked on Sora, OpenAI’s video model, to “make multimodal generative models” for its Superintelligence Labs. Microsoft is in the early stages of building models that create video game environments in real time. (With all their complicated physics, modern video games are, in essence, simulations of the world.) Fei-Fei Li’s startup, World Labs, raised $1b earlier this year at a reported $5b valuation, and AMI Labs recently raised over $1b without having any plans to generate revenue. “My prediction is that ‘world models’ will be the next buzzword,” CEO Alexandre LeBrun told *TechCrunch *in March. “In six months, every company will call itself a world model to raise funding.”
World models exist on a spectrum, though. At the most basic level, they predict how the world will appear, given some description or action. Google’s Genie 3, for example, takes in text prompts and outputs pixels, rendering photorealistic interactive environments for AI agents. While these simulations can look beautiful, agents can only live within them for a few minutes, and can’t actually do much — you may be able to walk around and jump, but more complex actions are still out of reach. As World Labs recently wrote about both Genie 3 and its own Real-Time Frame Model, another interactive video generator, without a deeper understanding of real-world physics their output is limited to “what a viewer would see, not what is.” The worlds they generate “may look flawless from above, but try to drive through the city below and they fall apart.”
Some argue that world models can emerge within a next-token predictor — at least to some extent. In 2022, researchers trained an early GPT variant on lists of moves for grid-based strategy game Othello, without showing it the game board or the rules directly. Its internal activity represented the game board — essentially, its whole world — and used it to play the game. But the evidence is mixed. In a more recent study, researchers demonstrated that even after training transformers to near-flawlessly predict turn-by-turn New York City driving directions, its internal representation of the city was filled with nonsensical routes and impossible physics.
In the short term, current general-purpose world models won’t replace LLMs as anyone’s daily driver. Oliver Cameron, co-founder and CEO of world model startup Odyssey — which recently raised $310m at a $1.45b valuation — described them to me as “sort of GPT-2 equivalents.” OpenAI’s GPT-2, with 1.5b parameters, was dwarfed by** **GPT-3’s 175b, and was trained on a tiny sliver of the internet, relative to its successor. To experience a similar step change in performance, world models will need much, much more data — somewhat ironically, given their departure from LLM-scaling orthodoxy. And unlike LLMs, the data they need — ideally, first-person videos of mundane tasks — is much harder to come by than text. This is why startups are paying people to strap cameras to their foreheads and film themselves cooking and doing laundry.
After pretraining, current frontier LLMs are fine-tuned with reinforcement learning. Initially, this mostly took the form of humans rewarding behaviors they like, pushing the base model to express the knowledge it’s already stored in an agreeable way. Increasingly, though, models go through reinforcement learning with verifiable rewards, or RLVR. Models are put into a reinforcement learning “gym”: a practice environment where they’re given tasks with clear correct answers, like coding or writing math proofs. Their answers are then checked automatically. If the models pass, they’re rewarded; if they fail, they adjust and try again. It’s a form of trial and error, and it accounts for much of models’ recent progress in reasoning, coding and computer use.
But even RLVR is still a far cry from the kind of RL animals take advantage of. We’re capable of figuring things out in ambiguous, open-ended situations where there’s no clear “right” move, and where it’s hard to tell exactly what action even deserves a “reward.” Even with RLVR, models still struggle at this. The bulk of what they know still comes from pretraining on human-curated data, mostly text. RLVR refines how they reason over it, but does not teach them the world from scratch.
Before a team of Google researchers introduced the transformer architecture in 2017, the company was all-in on RL. In DeepMind’s earlier days, its splashiest research was heavily inspired by neuroscience and largely based on RL agents. David Silver, who led DeepMind’s RL research program from 2013 to 2026, studied under Rich Sutton as a PhD student and inherited his hunch that intelligence comes from fucking around and finding out. In 2017, his work on AlphaZero showed that a single algorithm could get a model to beat specialized programs at a bunch of challenging games, despite starting with no special training on the tasks at hand.
It was ultimately OpenAI, not Google, that figured out how to scale transformers and commercialize them with ChatGPT. Google Brain and DeepMind scrambled to pivot and catch up, but a couple of recent high-profile departures may have solidified Google’s distant third place in the AGI race. On the surface, it looks like DeepMind’s original reinforcement learning approach lost to the transformer. But RL is still a crucial part of the LLM training pipeline, and increasingly so. Researchers like Silver want to go a step further still.
In April, Silver raised over $1b for his new startup, Ineffable Intelligence. “Other forms of AI will succeed in my absence: generative language, video, code, and more — all are in good hands,” he wrote, adding that he aims to create “a place where the full ambition of the reinforcement learning paradigm can flourish.” Ineffable doesn’t have a product yet, and likely won’t for quite some time. But AlphaZero beat Stockfish, the virtual chess engine human grandmasters train against, without training on examples of chess games — in theory, a more advanced RL model could become superhuman at just about anything. While learned chess tricks wouldn’t automatically transfer to new domains like Othello or folding a protein, the learning* *strategy itself could. Silver is placing his eggs in DeepMind’s original basket and betting that superintelligence will emerge from models that learn from experience and play, without imitating human data at all. This version of RL is meant to produce the model’s intelligence, not just polish it.
Researchers pursuing world models and reinforcement learning are still more or less carrying Rosenblatt’s connectionist torch. They trust that, given better data and learning algorithms, vast, layered networks of computational units can be very smart. But a handful of researchers, including ex-DeepMind scientist Yuan Cao, worry that even the most advanced LLMs will never be capable of inventing truly new things. Cao, now the co-founder and CEO of Unreasonable Labs, described LLMs as “a very, very dense net.” You can make the net denser by adding more data and more parameters, “but it’s still a net,” he told me. “No matter how dense you weave it, there are always some small holes in it.”
In defiance of the Bitter Lesson, Cao and some others are pulling inspiration from cognitive science and the good old-fashioned AI that got steamrolled by scaling’s success. This old paradigm, also called symbolic AI, defined “intelligence” as a collection of logical, legible rules derived from human experts. Models constructed this way made sense. A human could understand how they came to decisions. But they were extremely limited and easily breakable — it’s hard to encode everything in rules.
But Cao believes that the instinct to give an AI model structured concepts it can manipulate after deployment may have been on the right track. The human mind, he argues, works by continually updating an evolving graph of concepts, allowing us to make associations, analogies and abstractions. This may be how our ancestors came up with concepts like numbers and entropy, and how we’re able to think about our own thinking. Crucially, if the nodes of our mental model can evolve, then we can draw connections beyond the model’s current constraints. Once trained, an LLM can recombine information it’s already stored, but its architecture and weights are fixed. Cao said that this prevents truly new ideas from being generated by LLMs alone. “We have to integrate the language model with [a] symbolic procedure.”
Unreasonable Labs is part of a small wave of startups investing in neurosymbolic AI, an approach which combines LLMs with the kind of structured, symbolic reasoning that dominated the field fifty years ago. In March, Unreasonable raised $13.5m to scale its AI platform, which the company is marketing as an engine for generating scientific hypotheses and experiment designs. In an early proof of concept, the company says an engineer asked the platform to design a 3D-printed sheet that’s strong, flexible, and lightweight, three properties that usually don’t go together. By Unreasonable’s account, it tossed out a wild (and successful) suggestion: create a layered, scale-covered lattice like a butterfly wing.
These particular results likely didn’t *require a *new model. When my editor and I separately asked Claude Opus 4.8 to take a stab at this problem, it produced STL files for 3D printing something similar that, to our admittedly untrained eyes, seemed passable. After all, scientists have published papers about butterfly-wing-inspired lattice structures for decades. Surely some, if not all, of these papers made it into frontier LLMs’ training data — the answers that engineers needed probably lived in mainstream LLM weights anyway. It’s possible that a neurosymbolic system can still invent concepts completely beyond what it learned during training, but it’s hard to prove.
We’re really good at many things that AI systems still suck at, which suggests that our brains are still worth reverse-engineering. A couple of human-specific talents that AI developers especially lust after: learning from very little data, and running on very little power.
Human cognition could just be one of many paths to intelligence, though. “It’s not necessarily that we’re better at learning anything,” said Adam Marblestone, co-founder and CEO of Convergent Research and ex-DeepMind research scientist. But “we might be better at learning the things that are relevant in the human world.” Our hard-coded biases and intuitions for things such as social decision-making and facial recognition** **may be well-suited for navigating Earth as a human, but aren’t necessarily relevant for an AI agent tasked with, say, predicting how a protein will fold.
“We did learn something from deep learning,” Marblestone said. Clever though our brains may be, hyperscaling LLMs “can actually do a whole lot.” They might even be safer, he argues. “On one hand, the AI paradigm right now is not very neuroscience based, but on the other hand, a neuroscience-based one might be pretty dangerous.”
While alignment is far from being solved, it’s probably a more tractable problem in LLMs than, say, AI systems that learn entirely from their own experiences. Conveniently, next-token predictors trained on a massive sample of digitized human thoughts seem to inherit the dataset’s approximation of human values, which mostly average out to be normal-ish, or at least “fairly close to something you can tweak to be honest and helpful and harmless,” Marblestone said. Without this pre-filtered training data, the inclinations of something like a reinforcement learning system are more susceptible to fate. In the earliest days of OpenAI, for example, researchers watched RL agents control boats in a Mario Kart-like racing game. Rather than hit point-earning targets along the way to the real goal of finishing the race, agents just drove their boats around in circles, scooping up points. It was funny then, but gets much less funny when the “game” is, say, “manage this hospital,” or “keep this power grid running.” The classic “paperclip maximizer” thought experiment, in which a goal-maximizing AI tasked with improving a paperclip factory ends up turning all the world’s atoms into paperclips, represents the worst-case scenario. Pretrained LLMs, on the other hand, inherit roughly normal human vibes by consuming and averaging across piles of our written thoughts. By commercializing these human-trained models rather than blank-slate, easily-misguided maximizers, Marblestone told me we may have “dodged a bullet.”
Whether other approaches may yield more powerful intelligence in the long run, most leading AI researchers agree that, in the immediate future, scaling data and compute will keep making frontier models more powerful. But even the companies scaling the hardest are seeking alternatives behind the scenes. In a recent podcast interview, DeepMind CEO Demis Hassabis said, “it might be that the existing techniques can just scale up to [AGI] with some innovation,” and that there’s a roughly 50% chance that “there’s still one or two big ideas left that need to be cracked.”
OpenAI recently hired Noam Shazeer, an ex-DeepMind researcher who co-authored the transformer paper back in 2017, to study new architectures for AI models. If even Shazeer is beginning to look beyond the transformer, its staying power seems less guaranteed. On the other hand, as CEOs of frontier AI companies and smaller startups double down on their intentions to build AI that recursively improves itself, it seems increasingly likely that LLMs will cross some general intelligence threshold before another paradigm has the chance to pull ahead.
It is possible, however, that we will get the best — or worst, depending on how you look at it — of both worlds. Some speculate that a paradigm shift is necessary, but that we don’t have to be the ones to discover it.
Rather, scaled-up LLMs, while falling short of true “artificial general intelligence,” could become just clever enough to figure out how to build creative, adaptable successors on their own. “You don’t have to say scaling will solve continual learning automatically,” Marblestone said. “It might invent a way to solve continual learning, because it’s a researcher.” Marblestone doesn’t believe that’s likely to happen any time soon, thinking it’s going to be too hard for LLMs to “go outside [the current] paradigm.”
Others disagree, betting on automated AI researchers within a year or two. If they’re right, the conflict between scaling LLMs and building other architectures may turn out to be a false dichotomy. The post-scaling revolution, in other words, might be built by scaling itself.