Large language models are brilliant at any text-based task and mediocre at reasoning. They draft an essay in seconds and stumble computing a tip. They write a perfect email and lose track of a five-step logical chain. This weakness, which used to look like a size problem —”more parameters will fix it”— is increasingly clearly an architecture problem. And in recent months several news items point in the same direction: the current path has a ceiling, and half the industry is starting to hunt for “the next AI beast that will change the world”. Many are reopening old folders and dusting off ideas that had been left behind.
David Silver, the creator of AlphaGo, has left Google DeepMind to found Ineffable Intelligence at a $5.1 billion valuation. Subquadratic has stepped out of its ‘incognito mode’ with SubQ, a model that processes 12 million tokens of context at a fraction of the cost of frontier models. Yann LeCun, Meta’s chief AI scientist for a decade, walked away in November 2025 to raise $1.03 billion and build AMI Labs, another open bet against LLMs. Three moves, three different angles: Silver questions how models learn, SubQ questions how they process, LeCun questions what they predict. But they share one diagnosis: the autoregressive LLM as we know it today is not the right tool for the next big leap.
Key takeaways
- Three frontier bets against LLMs in 2026: Silver / Ineffable (RL in simulation, $5.1B), LeCun / AMI Labs (world models and JEPA, $1.03B), and SubQ / Subquadratic (sparse attention, 12M token context).
- The shared diagnosis: human data is a finite fossil fuel, transformers waste compute, and the autoregressive model accumulates errors exponentially.
- Three beasts to watch: reinforcement learning in simulation, world models with joint embeddings, and ARC-AGI-style program synthesis. None of them assume language is the natural substrate of thought.
- General reasoning outside the linguistic paradigm —multi-step planning, formal deduction, causal modeling— remains unsolved, as Judea Pearl has been arguing for years.
- 498A has accumulated over 9 million LLM-produced brand mentions and 2.2 billion tokens in its database: vertical fuel that already trains internal agents and will power fine-tuning of local models.
Silver: the fossil fuel of LLMs
Silver has a metaphor you grasp immediately: human data is fossil fuel. It has given today’s AI an astonishing shortcut —the entire web, every book, every line of code— but it is finite, and worse, capped by the ceiling of human intelligence it imitates.
His favorite thought experiment makes the point crystal clear. Imagine training an LLM in a time when everyone believed the Earth was flat. No matter how much the model refines its own code, it would remain a flat-earther. It has no mechanism to discover the Earth is round, because its only source of truth is what humans already thought.
Silver’s alternative is reinforcement learning inside simulations (‘The Matrix is coming’ — there seems to be no other way out). An AI agent that learns by trial and error, the way AlphaGo mastered Go in 2016 without studying human matches, discovering moves no human champion had seen. Silver calls these “superlearners”: systems that don’t imitate, they discover. He compares the current moment to biology before Darwin: many people trying to understand life, but no unified framework, no explanation that sketched a viable future.
Ineffable Intelligence has raised $1.1B in a single seed round —a record for European AI— and Silver has pledged to donate every dollar he personally earns to charity (we’ll see). It will probably take years before we see a product. But the diagnosis is sharp: if you want an AI that discovers new things, it cannot learn only from what humans already know.
And what humans already know seemed like a lot until LLMs arrived and devoured it all. In less than five years they’re already asking us “what other data do you have for us?” — and the reality is that we have almost none left.
SubQ: the transformer’s wasted compute
While Silver attacks the source of learning, Subquadratic attacks the architecture. The problem they identify is more technical, but just as fundamental: the transformer’s attention mechanism is O(n²).
Translation: every token in the context is compared with every other. With 1,000 tokens, the model computes a million relationships. With one million tokens, it computes a trillion. Doubling the context quadruples the cost. That’s why long context windows are slow, expensive and impractical. And it’s why an LLM “gets dumber” when you push it on deep reasoning: it spends its compute budget comparing relationships that mostly don’t matter.
SubQ’s thesis is elegant: the vast majority of token-token pairs are irrelevant. Only a fraction matters. Processing the rest is pure waste. Their architecture, Subquadratic Sparse Attention (SSA), dynamically identifies which relationships deserve compute and discards the rest. It’s not a Longformer-style patch or a Mamba-style hybrid: it’s a redesign from scratch. In any case, our deeper problem isn’t only cost or wasted compute. Even a much faster system doesn’t fix the underlying issue: the challenge is structural — it’s about how the model works internally.
There are caveats: no peer-reviewed paper yet, benchmarks are self-published. But the direction is right: today’s transformers burn fuel on relationships that lead nowhere.
LeCun: what LLMs cannot predict
If Silver and SubQ might sound like outsiders, Yann LeCun adds institutional weight to the diagnosis. Turing Award winner, Meta’s chief AI scientist for a decade, he has spent years repeating publicly what the industry preferred not to hear: autoregressive models are doomed. His argument is structural. Every token an LLM generates is a probabilistic prediction about the next. The probability of error compounds exponentially with output length. It’s not a bug, it’s the geometry of the model.
LeCun left Meta in November 2025 and in March 2026 announced AMI Labs with $1.03 billion in seed funding to build what he has been proposing for years: JEPA —Joint Embedding Predictive Architecture. The difference from an LLM is radical. JEPA doesn’t generate, it predicts in an abstract representation space. Instead of guessing the next pixel or word, it learns which world representations follow which. That gives it three things LLMs don’t have: causal understanding, planning by simulating alternative futures, and robustness to noise because irrelevant details are abstracted before prediction.
It is, fundamentally, a world model. And here the loop closes with Silver. Where Silver proposes agents that learn by reinforcement in simulations, LeCun proposes architectures that learn predictive models of the world from observation. Two forms of the same intuition: intelligence requires modeling reality, not just modeling the text that describes it. Pretending to understand things only holds for a while: when you actually have to solve real problems, it shows that you can talk fluently but haven’t grasped the essence — and so you can’t actually solve the challenges.
An eighty-year arc shifting phase
This didn’t come out of nowhere. Simulation has been part of computing since Monte Carlo (Ulam and von Neumann, 1947), through SPICE (1973), industrial CFD (1989), PhysX and Bullet (2008), all the way to Digital Twins (2022). For seventy years the pattern was the same: simulate by writing equations that describe the physics of the world —circuits, fluids, collisions— and let the computer solve them.
In 2024 something changed. Sora (OpenAI) and Genie (DeepMind) showed that a model trained on enough video learns world dynamics without anyone programming physics into it. Simulation stops being programmed with differential equations and starts being learned from observations of the world. In 2025, NVIDIA Cosmos brought the idea to robotics: if you need millions of trajectories to train a robot and the real world only gives you thousands, learned simulation generates the rest.
What Silver proposes with his superlearners and what LeCun proposes with JEPA are the next turn of that same arc: learned simulation as the substrate of reasoning, not just perception. It’s not a 2026 fad. It’s an eighty-year transition coming to a head.
Why I changed my mind
A personal aside, because I’ve lived this pattern from inside before, and that shapes how I read it.
My modest research training was in the computer architecture lab, specialized in parallelism and simulation of parallel systems. Our comfort zone was solvable problems with clear paths and easy to represent with everyday maths: load balancing, scheduling, architectures, inter-process communication. When the question was how to get many components to reason as a single entity —how unified cognition emerges from a distributed substrate— we didn’t know how to answer (yet). Emergence shone on paper and worked in simple simulations, but we couldn’t see how to translate it scalably into useful reasoning through multilayer neural networks. Reaching anything serious without some symbolic foundation seemed unthinkable.
That’s why expert systems, symbolic AI and rule-based approaches looked more promising. Not because they were more powerful —they weren’t— but because they didn’t depend on pieces that didn’t yet exist. They were pragmatism over the available. Deep neural networks needed compute, data and gradient methods that would arrive two decades later. Meanwhile, the community pronounced them dead — and even if anyone had believed in them, there was no compute power to bring them to the real world.
I suspect we’re doing something similar today. We’re uncomfortable thinking about the future of AI without transformers because the alternatives require unsolved pieces. But that same discomfort justified abandoning deep neural networks in the 80s and 90s, and we learned the hard way that dismissing promising lines of work just because they depend on future pieces is a costly mistake. That’s why I find it so exciting to see decades-old papers being revisited. Siamese networks (Bromley, LeCun et al., 1993), for example, are the direct ancestors of the joint embedding that JEPA is built on. LeCun isn’t inventing a new architecture: he’s returning to an idea he himself helped define thirty years ago, now with the compute and data to push it further.
When I read SubQ’s argument —”most token-token pairs are irrelevant, processing all of them is waste”— I think: true, not very relevant, but true. We know models will keep optimizing, specializing, running on-device, working in syndication. That’s the comfortable part of the prediction. The question that matters to me is different: which line will start delivering serious reasoning results without using a language model architecture?
What we see from the trenches
This isn’t theory. We see it every time a serious project lands at 498A.
In a recent engagement with IMAGIN —which had spent years using IBM Watson, the rule-based system that for two decades was the star of corporate tech, until they migrated to Watson X— we collected thousands of question-answer pairs from that legacy installation and pitted them against a local LLM deployed under their compliance requirements. The result was decisive: LLMs win on quality, reliability and nuance by a wide margin. The reason is structural: that task is made of language. When the task is genuinely linguistic, and despite known issues like hallucinations and the rest, there’s no debate — current models are a qualitative leap with no going back.
The trap is assuming any problem is. In another project, this time with AstraZeneca, we worked on a multimodal study of clinical data. What surprised us most was how easily the models handled heterogeneous formats —structured and unstructured data, images, free text, medical records— without flinching. But we couldn’t ignore how weak they became when the task moved from digesting to inferring. They generalize well over surfaces; they reason poorly about depths. This is exactly the critique Judea Pearl has been making for years: modern statistical systems are brilliant at detecting correlations and terrible at modeling causality, and without causality there is no medical, scientific or clinical reasoning that holds up. A system that suggests you walk to wash your car because the carwash is 200 meters from your house is not the system you want recommending whether an oncology treatment approved for one indication can be applied to another disease. And that leap —from the first case to the second— isn’t covered by adding more parameters.
Three beasts to watch
If I had to bet on which beasts enter the arena in the coming years, I would watch three lines, all of them anti-LLM by nature:
- Reinforcement learning in simulation —the Sutton-Silver line, now capitalized by Ineffable Intelligence. The intuition: useful knowledge isn’t in the data, it’s in interaction loops. AlphaGo didn’t read Go textbooks; it played.
- World models and joint embeddings —the LeCun line, capitalized by AMI Labs and JEPA. The intuition: predicting abstract representations of the world is more informative than predicting pixels or tokens. Intelligence requires modeling the territory, not the map that describes it.
- Program synthesis —the François Chollet line with ARC-AGI. The intuition: general-purpose reasoning looks more like writing a program that solves a novel problem than retrieving an answer from a corpus. ARC-AGI has worked as a filter for years: no pure LLM passes it well; approaches that combine search and symbolic abstraction make progress.
The three share something. None of them assumes language is the natural substrate of thought. All three accept, as neural networks did in the 80s, that they depend on still-immature components. And all three are, precisely for that reason, the ones worth watching.
I’ve changed my mind about something in recent years, and it’s this: every time the world thinks it has found THE solution, it’s usually wrong. Final answers are rarely monolithic or simple. Gary Marcus and Ernest Davis have been making the same case from a more combative angle: deep-learning monoculture is fragile because it ignores decades of work in symbolic reasoning, world models and hybrid architectures.
The future of AI cannot be what we have today, because we don’t have a diversity of solutions while we still have a huge diversity of problems.
Brand visibility in generative AI
We audit how brands appear inside LLMs and design their competitive positioning. This is what we do at GEOradar.
This is what we do →The missing layer: simulating users to understand markets
While these bets play out at the research frontier, there’s a much less visible layer already building something any of those paradigms will need once it matures: vertical data on real behavior inside LLMs.
At 498A, Zoopa’s R&D division, we’ve spent nearly two years building tools that, on the surface, look like something else: GEOradar measures how brands appear in LLMs like ChatGPT, Gemini and Claude. Prompt Atlas structures strategic prompt datasets. S.A.M. —Semantic Alignment Machine— maps how AIs understand a client’s competitive category. LEO generates transmedia content at scale with narrative connections that were previously impossible to manage.
What we do when running a GEO audit is, in essence, simulate real users inside the LLMs to obtain 100% real responses. Not abstract simulations like Silver’s, nor compute filters like SubQ’s: simulations of market behavior. Different personas, at different funnel stages, asking different questions of the models about a specific category —banking, FMCG, automotive, retail. Each simulation produces answers, cited sources, mentioned brands, detected intentions, associated sentiments, semantic ground gained or lost.
Brands using GEOradar today think they’re measuring their AI visibility. They are. But they’re also doing something else: co-building a proprietary dataset of their own competitive category at a tiny cost, impossible to replicate from outside, and which in the near future will be the key to feeding their specialized AI models.
Source: 498A, May 2026
This isn’t a pilot. It’s an operational dataset already working as an advance party behind the scenes: training the agents that plan audits, run them and evaluate results. In the coming months that same data will fuel the fine-tuning of local models specialized in brand visibility analysis. Data is today the fuel and the essence of models. Accumulating your own fuel, in your own terrain, framed by each brand’s or institution’s own view of the world, is the only way not to depend on someone else’s tank when the next engine arrives.
The dataset’s two lives
What makes this more interesting than “we have nice data” is that the dataset has two distinct lives, separated by years:
Life 1, the one we’re living now. The data helps brands position themselves better in current LLMs. You know how Gemini and Claude see you, what sources they cite, what competitive gaps you have, what information gaps to fill, and you design content and information architecture to improve your visibility, positioning and conversion. Applied work, immediate ROI.
Life 2, the one coming. When the paradigms Silver, LeCun, SubQ and others are building mature, brands won’t settle for measuring how they appear in an AI. They’ll want to simulate how consumers will behave in environments saturated with intelligent agents. For that you need world models applied to the market. And world models, as LeCun says, are trained with data from the world. In this case, from the specific world where a brand competes.
Brands that have spent two years accumulating this kind of data show up to that conversation with a full tank. Ready for the agentic web and the new AIs coming. The ones who start then, won’t.
The right question
The public conversation about AI is dominated by two wrong questions. “Which architecture will win?” and “Which company will dominate?”. Both assume the game is picking sides.
The right question is different: which assets will be impossible to replicate when the paradigm shifts?
When the next leap arrives —Silver, LeCun, SubQ, or something nobody is predicting yet— the models will change, the costs will change, the architectures will change. What won’t change is that to understand a specific market you’ll still need data on that specific market. And data on brand behavior inside LLMs is generated only one way: by running studies and user simulations on real brands across real LLMs, month after month, vertical by vertical.
You can’t buy that. You accumulate it.
Silver says human data is fossil fuel. LeCun says LLMs predict tokens when they should predict worlds. SubQ says the transformer engine wastes fuel. At 498A we’ve spent nearly two years distilling high-octane fuel in a very specific terrain: how brands exist inside LLMs. When the new engines arrive —Silver’s, LeCun’s, SubQ’s, or others not yet on the map— that fuel will already be in the tank.
That, more than any architecture debate, is the bet that matters.
Let’s talk
If your brand is starting to wonder how it shows up inside ChatGPT, Gemini or Claude —or how it will defend its positioning when the next AI paradigms arrive— this is a conversation worth having before your competitors have it. Here’s how we do it at 498A.
Let’s talk →Glossary
Large-scale language model trained on vast text corpora to predict the next token. The dominant architecture in generative AI since 2022 (GPT-3.5 onward). Examples: GPT, Claude, Gemini, Llama.
Neural network architecture introduced in 2017 (“Attention is All You Need”, Vaswani et al.) that underpins almost all current LLMs. Its core mechanism, self-attention, has quadratic complexity O(n²) with respect to context length.
Paradigm where an agent learns by trial and error in an environment, optimizing a reward function. Theoretical basis of AlphaGo (Silver, 2016) and canonically developed by Sutton and Barto, awarded the Turing in 2025.
Architecture proposed by Yann LeCun from 2022 onward that predicts abstract representations in an embedding space rather than generating pixels or tokens. It aims to capture causality and enable planning. Foundation of LeCun’s new venture, AMI Labs (2026).
AI model that learns a predictive representation of the world —how states evolve over time and under what actions— rather than learning only surface correlations. Examples: Sora (OpenAI, 2024), Genie (DeepMind, 2024), NVIDIA Cosmos (2025).
Variant of the attention mechanism that computes only a subset of token-token relationships, discarding the irrelevant ones. SubQ (Subquadratic, 2026) elevates it into a complete architecture: SSA (Subquadratic Sparse Attention).
Benchmark created by François Chollet (2019) to measure general abstract reasoning. It tests a system’s ability to solve novel problems unseen during training. Pure LLMs perform poorly on it; program synthesis approaches advance.
Brand visibility audit platform inside LLMs developed by 498A. Measures how brands appear, what sources models cite, and what semantic share brands hold within specific competitive categories.
AI R&D division of Zoopa, specialized in brand visibility within generative AI systems. Develops GEOradar, S.A.M. and Prompt Atlas.
Company founded by David Silver in 2026 after leaving Google DeepMind. Raised $1.1B in seed funding at a $5.1B valuation, betting on building “superlearners” based on reinforcement learning in simulations.
Company founded by Yann LeCun in 2026 after leaving Meta in November 2025. Raised $1.03B in seed funding to build world models based on JEPA.
Frequently asked questions
Why does LeCun say autoregressive LLMs are doomed?
Because every generated token is a probabilistic prediction, and error probability compounds exponentially with output length. It’s a structural problem of the model’s geometry, not a bug fixable with more data or parameters. That’s why LeCun proposes JEPA, an architecture that predicts in an abstract space instead of generating token by token.
What is the “next beast” in AI?
It’s the metaphor we use to ask which AI architecture will start delivering serious reasoning results without using a language model. The three lines with the most traction are: reinforcement learning in simulation (Silver, Sutton), world models with joint embeddings (LeCun, JEPA), and program synthesis (Chollet, ARC-AGI). All three share the assumption that language is not the natural substrate of thought.
How do world models differ from LLMs?
LLMs learn to predict the next token from human text. World models learn to predict how the state of the world evolves from observations —video, sensors, actions. The difference matters because it enables planning, causal reasoning, and robustness to noise. Sora (OpenAI), Genie (DeepMind) and NVIDIA Cosmos exemplify the state of the art in world models trained on video.
How does SubQ’s sparse attention affect LLM performance?
SubQ dynamically identifies which token-token relationships deserve compute and discards the rest. The result: 12-million-token context window (10× current frontier models), 52× faster than FlashAttention at 1M tokens, and less than 5% of Claude Opus’s cost for equivalent inference. No peer-reviewed paper yet, but the data points to today’s transformers wasting compute on irrelevant token-token relationships.
Why do LLMs fail at causal reasoning?
Because they are statistical systems that detect correlations, not causality. This is the critique Judea Pearl has been making for years: without a formal model of cause and effect, the system can explain “what” but not “why”. This is especially serious in domains like medicine, where deciding whether a treatment works for another condition demands real causal inference, not statistical extrapolation over surfaces.
What does 498A do with the data it accumulates in GEOradar?
Today, that dataset —more than 9 million brand mentions and 2.2 billion tokens— serves two purposes: positioning client brands better in current LLMs, and training internal agents that plan, execute and evaluate audits. In the coming months it will be used to fine-tune local models specialized in visibility analysis. It’s an accumulative asset specific to each market vertical, impossible to replicate from outside.
Published by Carlos Ortet | 498AS
Carlos Ortet is CEO of Zoopa and director of 498A, its AI R&D division. Before founding Zoopa he did some research in computer architecture, specializing in parallelism and parallel systems simulation. 498A develops GEOradar, S.A.M., LEO and Prompt Atlas, advanced brand visibility analysis tools for generative AI systems.
Sources and references:
- Will Knight — AlphaGo’s creator bets $5 billion that the AI industry is going the wrong way, WIRED, April 2026.
- Subquadratic — Introducing SubQ and The New Stack, May 2026.
- LeCun raised $1B to replace LLMs — Menon Lab Blog; JEPA Deep Dive — Rohit Bandaru.
- Sutton, R. S. & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press. Available at incompleteideas.net.
- Pearl, J. & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.
- Marcus, G. & Davis, E. (2019). Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon.
- Chollet, F. (2019). On the Measure of Intelligence. arXiv:1911.01547.
- Bromley, J., Bentz, J. W., Bottou, L., Guyon, I., LeCun, Y., et al. (1993). Signature Verification using a Siamese Time Delay Neural Network. NIPS.