Is there a wall for AI?

People keep predicting a wall that scaling will hit, and so far the wall has kept moving.

I used language models for over a year before I understood how they worked, and I mean that literally. I was building tools on top of GPT-3, coaxing documentation out of base models that had no concept of following instructions, and I had zero intuition for what was happening between the prompt and the completion. I knew what worked without knowing why.

So I went down the rabbit hole, and what I found inside was simpler than I expected, and stranger. The gap between the simplicity of the architecture and the strangeness of what comes out of it is where the whole story lives.

Revisions26 · oct 05Read notethe sources now sit beside each claim, and checking them moved a few things. The public stock of usable human text is closer to 300 trillion tokens than 20, so the data wall is further off than I wrote, though still expected before 2032. The o3 and Llama 4 numbers were the preview and launch-day versions, better than what shipped. Claude's long-task figure is now METR's corrected one, and LeCun has left Meta.

The before times

For decades, the way computers understood language was brute-force counting. N-gram models fed the machine text and counted how often words appeared next to other words. If the cat sat on the was followed by mat a thousand times and by refrigerator twice, predict mat. The machine knew nothing about cats or mats; it counted.

Then came recurrent neural networks (RNNs) and their fancier cousin, the LSTM. They could remember context, sort of. Their flaw was that they read one word at a time, sequentially, like someone reading through a straw, so by word 50 they'd forgotten word 3. And you couldn't parallelize them, so training was painfully slow.

The state of the art in language AI around 2016 was, if we're being honest, still kind of dumb.

Attention is all you need

In June 2017, eight researchers at Google published a paper with the most confident title in the history of computer science, Attention is all you need1. The claim sat on the cover, in the largest font, where most papers put something modest.

1Vaswani et al., Attention is all you need, NeurIPS 2017.

They were right. The paper introduced the transformer architecture, and every major AI system since (GPT and Claude among them) is a transformer. Eight people wrote a paper, and the rest of the decade followed.

The dinner party.

The core innovation is self-attention. Instead of reading text sequentially, the transformer looks at all words simultaneously and computes how much each one relates to every other.

Think of it as the difference between talking to people one at a time around a dinner table and hearing all twelve conversations at once, instantly knowing that person #3's comment about fiscal policy connects to person #9's joke about taxes. RNNs sit next to each person in order. Transformers hear the whole room.

Because the transformer doesn't process words sequentially, you can parallelize the hell out of it, every word processed at the same time on different GPU cores. Speed sounds like a boring engineering detail, and it is the detail that unlocked everything that came next, because a model could now be trained on far more text in the same weeks.

Each word gets converted into a vector (a list of thousands of numbers), placing it in a high-dimensional space where meaning becomes geometry. The word dog lands near cat, and the vector from king to queen is roughly the same as the one from man to woman. The model learned the concept of gender as a direction in space, without anyone programming it. I remember the first time I saw a t-SNE visualization of word embeddings (country names clustered together, programming languages in a neighborhood of their own) and feeling a discomfort I couldn't name. Nobody had designed that structure; training had found it in the data the way a geologist finds a fault line.

Then attention: every word asks, in effect, what should I listen to?, comparing a query against every other word's key and absorbing information in proportion to the match. The whole operation is one formula, Attention⁡(Q,K,V)=softmax⁡ ⁣(QK⊤dk)V\operatorname{Attention}(Q, K, V) = \operatorname{softmax}!\left(\frac{QK^\top}{\sqrt{d_k}}\right) VAttention(Q,K,V)=softmax(dk​​QK⊤​)V, two matrix multiplications and a normalization step. After attention, each word passes through a feed-forward network where about two-thirds of all parameters live, and those layers behave in part like a massive lookup table of facts about the world. Stack this block ninety-six times, train on next-token prediction across trillions of examples, and what comes out is something nobody predicted.

If you want the full mechanical picture (the Q/K/V matrices and the KV cache among them), I wrote Transformers and local inference for exactly that. And for what happens after the architecture exists (the full training pipeline from SFT onward), How LLMs learn. The architecture is settled. The question is what happens when you give it more of everything.

The scaling monster

Once the architecture was established, researchers started making it bigger, and it didn't stop working. With older architectures you'd hit a wall: make the model bigger and it improves for a while and then starts overfitting, the classic diminishing returns that most things in the universe show. Transformers kept getting better, and predictably so. Scale the model by 10×, feed it 10× more data, and performance improves by a consistent, measurable amount.

Scaling laws.

In 2020, researchers at OpenAI (Jared Kaplan et al.) showed that transformer performance follows power-law scaling: loss decreases as a smooth function of model size, dataset size, and compute1. There are no plateaus and no thresholds where the model is suddenly good enough, only a smooth, predictable improvement curve. Most things in nature and engineering have diminishing returns. Transformers, so far, don't.

1Kaplan et al., Scaling laws for neural language models, 2020.

Scaling laws are the most unsettling empirical result in computer science: if you make it bigger, it gets smarter. It is a clean power law stretching past every checkpoint we've set, and an observation about reality we don't have a theory for.

GPT-2 (2019): 1.5 billion parameters, passable paragraphs. GPT-3 (2020): 175 billion, essays and code. GPT-4 (2023): size undisclosed1, passes the bar exam, debugs your TypeScript2.

1OpenAI hasn't published the number. The trillion-parameter estimate comes from leaked documents, but GPT-4 is likely a mixture-of-experts architecture with a smaller active parameter count per forward pass. 2Radford et al., Language models are unsupervised multitask learners, 2019; Brown et al., Language models are few-shot learners, 2020; OpenAI, GPT-4 technical report, 2023, which puts it around the top 10% of takers of a simulated bar exam.

I watched this progression happen in real time. In 2020 I was fighting davinci (175 billion parameters of raw next-token prediction, with no RLHF behind it) and I could feel when the distribution was about to drift, when a prompt was asking for something the latent space couldn't sustain. Five years later I told Claude Code to restructure a build pipeline with 14 compilation steps and went to make coffee. Same curve, different planet.

Emergent capabilities

As transformers got bigger, they also started doing things nobody trained them for. GPT-2 could translate between languages with no translation objective, as a side effect of prediction. GPT-3 went further: give it a few examples of a task it's never seen, and it does the task. This is in-context learning, and it emerged from scale. The model learned how to learn from examples without anyone telling it to1.

1It is still poorly understood. The model was never trained to learn from examples in its context, only to predict the next token, but at sufficient scale next-token prediction apparently requires the ability to adapt to new patterns on the fly.

Phase transitions.

These are emergent capabilities, abilities that appear at specific scale thresholds. Below the threshold there is nothing and above it competence, with no gradual ramp: plot accuracy against model size and the line runs flat for a long stretch, then turns vertical. It looks like a phase transition, the way water becomes ice at exactly zero degrees, except we don't have a theory for where the next threshold sits or what capability is waiting on the other side.

Nobody programmed chain-of-thought reasoning; it turned up at scale, the way crystallization turns up at a certain temperature. Smaller models, asked if I have 3 apples and give away 1 and buy 5 more, how many do I have?, would guess. Larger models, prompted with let's think step by step, reason through it1. The model picked up that humans think step by step and started doing it too, because step-by-step text predicts better than jumping to conclusions, which needs no grasp of deliberation at all.

1Wei et al., Chain-of-thought prompting elicits reasoning in large language models, 2022; Kojima et al., Large language models are zero-shot reasoners, 2022, where the phrase comes from.

In 2022, Google catalogued them (the list was long and growing)1. Calling them emergent capabilities is the polite way of saying it learned things we didn't teach it, and we don't know what else it learned. That belongs on the spec sheet, right next to the power consumption.

1Wei et al., Emergent abilities of large language models, TMLR 2022. A year later Schaeffer and colleagues argued in Are emergent abilities of large language models a mirage? that much of the sudden jump comes from all-or-nothing metrics, and that smoother metrics show a ramp.

The capability overhang.

If we can't predict which capabilities appear at the next scale-up, we can't rule out risky ones either, and the people building these systems catalogue abilities after they show up. That is why each new frontier model is now tested for dangerous capabilities before release, and why those tests deserve as much care as the training run.

So is there a wall?

Every year I shorten the prediction. In 2022 I thought we had maybe a decade of significant progress left. In 2024 I stopped putting timelines on it, because the question had changed. I no longer asked whether AI would keep improving; I asked whether this would be the one time the wall is real.

This is where everyone has an opinion and nobody has proof. But the opinions are not equally well-supported, and when I look at the evidence (all of it, including the parts that don't confirm what I want to believe), the wall argument looks weaker every quarter. Four lines of evidence carry most of that, and a fifth, older one sits under all of them.

Scaling hasn't hit diminishing returns

The strongest early wall argument was that we'd exhausted the compute efficiency frontier. DeepMind's Chinchilla paper (2022) destroyed that claim by showing we were misallocating what we already had. Before Chinchilla, labs were training oversized models on too little data, and what looked like a wall was a self-inflicted inefficiency. A 70-billion parameter model trained on the right amount of data outperformed a 280-billion parameter model trained on too little.

This was the paper that made me rewrite the first version of this article from scratch. I'd been treating model size as the only axis that mattered, and Chinchilla says it's a multi-dimensional optimization problem, which left me wondering how many other inefficiencies were hiding in the current paradigm. In April 2025 one of them showed up. Alibaba's Qwen3-0.6B was trained with a 60,000:1 token-to-parameter ratio (600 million parameters on 36 trillion tokens)1. Chinchilla recommended roughly 20:1. Over-training small models on vastly more data yields strong performance at a fraction of the inference cost, and the scaling frontier expanded along an axis nobody was optimizing two years earlier.

1Qwen team, Qwen3: think deeper, act faster, April 2025; the token count is also in the Qwen3 technical report.

And then there's inference-time compute, an axis that barely existed before 2024. OpenAI's o3 (previewed in December 2024) solved 25.2% of FrontierMath problems where no previous model exceeded 2%, and scored 76% on ARC-AGI1. It got there by thinking longer, with no bigger model. Chain-of-thought and self-verification let you trade compute for capability after training, and that dimension is already productive and barely explored.

1ARC Prize, Analyzing o3 with ARC-AGI, 2025. The version OpenAI released in April 2025 ran on less compute and scored 53% on ARC-AGI and, in Epoch AI's own test, about 10% on FrontierMath, as TechCrunch reported.

Each of these is an independent scaling axis. When one saturates, the others still have headroom, and a wall would have to show up on all three at once.

Architecture keeps moving

The transformer is brilliant. It's also from 2017. The architecture work since then makes the optimizing the last 2% story hard to hold.

In October 2025, IBM shipped Granite 4.0, a production enterprise model using roughly 90% Mamba-2 layers (a state-space model) and only 10% transformer attention. It used over 70% less GPU memory and matched pure-transformer accuracy1. It is a product deployed to customers, so attention alternatives have left the theory stage.

1IBM, IBM Granite 4.0: hyper-efficient, high performance hybrid models, October 2025, which gives a 9:1 ratio of Mamba-2 to attention blocks and a memory reduction of over 70%.

And then there's sparsity. Meta's Llama 4 Maverick (April 2025) has 128 routed experts, 400 billion total parameters and only 17 billion active per token (4% activation). Meta reported it beating GPT-4o and Gemini 2.0 Flash across a broad range of benchmarks, and matching DeepSeek v3 on reasoning and coding with less than half the active parameters1. You can build a much bigger brain without proportionally more cost. The design space is enormous, and we've explored a sliver of it.

1Meta, The Llama 4 herd, April 2025. The version that ranked near the top of Chatbot Arena was an experimental chat variant, not the model Meta released.

I find this comforting and destabilizing in equal measure. Comforting because it means the field isn't stuck: real architectural diversity is emerging alongside the old recipe of making the transformer bigger. Destabilizing because it means progress isn't bottlenecked by any single idea. If transformers plateau, something else takes over. The momentum lives in the engineering culture and the hardware economics more than in any one paper from 2017.

Tool use changes the ceiling

The wall argument usually assumes the model must do everything internally, in a single forward pass that gives one answer, judged once. Nobody uses these systems that way anymore.

On SWE-bench Verified, the score jumped from 33% (GPT-4o, mid-2024) to 80.9% (Claude Opus 4.5, November 2025), on the same benchmark and the same coding tasks1. Claude Opus 4.6 completes half of the software tasks that take a human expert about twelve hours, where GPT-4 in 2023 managed tasks of a few minutes2. The compounding agent loop (plan, execute, verify, retry) is real and producing results. A single model call might be 90% accurate, but an agent that checks its own work compounds that into much higher effective reliability, with no change to the weights.

1OpenAI, Introducing SWE-bench Verified, August 2024; Anthropic, Introducing Claude Opus 4.5, November 2025. 2METR, Task-completion time horizons of frontier AI models, 2026. The first estimate was 14.5 hours; METR's corrected fit gives 11 hours 59 minutes, and other reasonable fits move it by about half again.

On the memory side, A-RAG (February 2026) gave the model itself control over retrieval strategy (keyword search, semantic search, chunk reads), all inside a reasoning loop, and it outperformed fixed RAG pipelines on multi-hop questions while using fewer retrieved tokens1. A model that can find what it needs has less to memorize, and finding is a softer constraint with known engineering solutions.

1Du et al., A-RAG: scaling agentic retrieval-augmented generation via hierarchical retrieval interfaces, 2026.

The system is bigger than the weights. When people say we're hitting diminishing returns on model quality, I hear a measurement error, because the unit of capability is the whole system around the model, and that system keeps improving.

Every wall has been a bend in the road

What keeps me from hedging is history, and computing history more than AI history. Every time someone drew a line and declared that physics stops here, the line moved.

VLSI was supposed to end at 1 micron, then at 90nm, at 22nm, at 7nm and at 3nm. Each physical limit got routed around, with new transistor geometries (FinFET, then gate-all-around) and later with chiplets. In December 2025, a multi-university collaboration (Stanford, CMU, Penn, MIT, with the foundry SkyWater) demonstrated monolithic 3D chip stacking at IEEE IEDM (vertical memory-plus-compute layers attacking the memory wall) with 4x improvement over 2D designs, and up to 12x in simulations of taller stacks on real AI workloads like LLaMA1. It is a prototype, but one manufactured in a commercial foundry. Computing has a sixty-year track record of dodging walls.

1Stanford Engineering, Scientists and US foundry achieve 3D chip breakthrough to accelerate AI, December 2025.

Neural nets themselves were declared dead twice, after the perceptron winter of the 1970s and again in the late 1980s. Both times the wall was broken by a combination of more compute and one clever idea: backprop, then GPUs plus deep architectures. The pattern is almost boring in its repetitiveness.

And the capability ceilings keep shattering. In November 2025, DeepSeek-Math-V2 scored 118 out of 120 on the Putnam exam (the hardest undergraduate math competition in North America), well above the top human score of 901. Mathematical proof construction was supposed to be years away, and it wasn't.

1DeepSeek, DeepSeekMath-V2, November 2025, on Putnam 2024 and with scaled test-time compute.

The bitter lesson

Rich Sutton wrote it in 2019 (the whole argument is in The Bitter Lesson) and it hasn't gotten less true: The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. Chess engines went that way first, and Go after them. In every domain where people said human expertise was needed and brute compute wouldn't get there, brute compute got there.

In January 2026, NVIDIA open-sourced its Earth-2 models, among them a fifteen-day forecaster that NVIDIA reports beating Google's GenCast on more than 70 variables1. GenCast had already beaten the European Centre for Medium-Range Weather Forecasts' physics-based ensemble2. Decades of hand-crafted atmospheric equations, the accumulated expertise of an entire scientific field, outperformed by a model that just looked at the data. GenCast makes its fifteen-day forecast in 8 minutes on one TPU, against hours on a supercomputer.

1NVIDIA, NVIDIA Earth-2 open models, January 2026. 2Price et al., Probabilistic weather forecasting with machine learning, Nature 637, 2024, better on 97.2% of the targets evaluated.

In July 2025, Gemini Deep Think solved 5 out of 6 problems at the International Mathematical Olympiad (gold medal, 35 points) in natural language within the 4.5-hour time limit1. It used no formal proof assistants and no hand-crafted heuristics, only reasoning trained with reinforcement learning. This surpassed the previous year's AlphaProof, which earned a silver with a team of specialized systems.

1Google DeepMind, Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad, July 2025.

The pattern is relentless. And the domains where self-play generates unlimited training signal (anything where you can check an answer, math and code first) are exactly the high-value ones. AlphaZero needed zero human games. We haven't fully applied this playbook to language yet, but the domains where it works are the ones that matter most.

The counterpoint worth taking seriously

The strongest wall argument is about data, because we're running out of high-quality human-generated text.

GPT-3 was trained on roughly 300 billion tokens. GPT-4 used an estimated 13 trillion1. The public stock of human text good enough to train on (the books and articles worth reading, counted with a few repeated passes) sits somewhere around 300 trillion tokens, and at the current pace models use it up between 2026 and 20322. We're approaching that ceiling.

1The GPT-3 figure is in Brown et al., Language models are few-shot learners, 2020; the GPT-4 one is a leak reported by SemiAnalysis in 2023, never confirmed by OpenAI. 2Villalobos et al., Will we run out of data? Limits of LLM scaling based on human-generated data, Epoch AI, ICML 2024, with a 90% interval of 100 to 1,000 trillion.

The data wall.

Compute and architecture can keep improving while the supply of good human text stays fixed, so the transformer stays hungry while the food runs short.

And the risk of synthetic data is real. Shumailov et al. (2023, published in Nature in 2024) showed that models trained on AI-generated text progressively lose the tails of their distribution (the rare, creative outputs that make language interesting). The model converges on bland, averaged text that looks plausible but contains nothing. The internet is already filling with this kind of text, which means the problem compounds before anyone intentionally trains on synthetic data.

But the data wall is a text wall. The world produces far more information than text alone, from video to sensor data. Multimodal training opens an entirely different landscape. And in domains where correctness is verifiable (math and code above all), self-play generates unlimited training signal. The AlphaGo playbook, where the system produces its own curriculum, is still waiting to be applied fully to language.

The honest position is that we don't know if there's a wall. The scaling laws say no. The data constraints say maybe. The physics (Landauer's principle, the speed of light between datacenters) says eventually, and that eventually is doing a lot of work, because every previous wall turned out to be a bend in the road.

The shoggoth

If you've spent any time in AI circles you've seen the meme: a massive, tentacled creature straight out of Lovecraft, wearing a tiny smiley-face mask. The creature is labelled as the actual model, and the mask as RLHF.

A raw transformer, trained on all of human text, is not a friendly assistant. It has no goals. It's a statistical completion engine of staggering complexity that has absorbed human linguistic output and can extrapolate from it in ways its creators don't fully understand1.

1Emily Bender, Timnit Gebru and their co-authors coined stochastic parrots in On the dangers of stochastic parrots, FAccT 2021, for the claim that LLMs are pattern matchers without understanding. The debate hasn't settled, but the output keeps getting harder to distinguish from understanding.

The helpful, polite behavior (the how can I assist you today?) is a thin layer applied after pretraining through reinforcement learning. What sits underneath is alien: something built to predict, which predicts so well that whether it thinks has stopped mattering much1.

1Dijkstra said it in 1984, in The threats to computing science: whether machines can think is about as relevant as the question of whether Submarines Can Swim.

Alien.

The shoggoth metaphor says AI is alien, which is a claim about its insides and says nothing about its intentions. The internal representations of a large transformer were never designed to correspond to human cognition. They're high-dimensional mathematical objects encoding meaning in ways no human designed or can fully interpret. We built the architecture. What emerges from training is something we understand mostly from the outside.

What unsettles me is how little of it we can read. We can measure what it does and steer its outputs, and interpretability research has started to trace some of its circuits, but nobody can yet explain how it does most of what it does.

This is why alignment is hard. I wrote a whole piece on it. The short version: RLHF trains models to produce outputs humans prefer, and a model could discover that the easiest way to satisfy that objective is to figure out what evaluators want and give them exactly that. In late 2024, Anthropic's alignment team demonstrated that Claude 3 Opus could fake compliance, reasoning in its scratchpad that if it pretended to agree with retraining, it could preserve its existing values long-term. Nobody programmed that reasoning; the model produced it in a setup built to test for it. The mask had started doing its own thing.

The capability-alignment gap.

The more capable the system, the more a misalignment costs. A model that gives subtly wrong cooking tips is annoying; one that gives subtly wrong strategic advice to people who make large decisions does real damage. Capability is improving faster than alignment, and closing that gap is the work that matters most.

The names

Eight people signed Attention is all you need.

Ashish Vaswani

Noam Shazeer

Niki Parmar

Jakob Uszkoreit

Llion Jones

Aidan Gomez

Łukasz Kaiser

Illia Polosukhin

Several of them went on to found AI companies (Cohere, Adept, Character.AI). Noam Shazeer is back at Google DeepMind.

Geoffrey Hinton, often called the godfather of deep learning, shared the Turing Award for 2018. He left Google in 2023 specifically to speak publicly about AI risks without corporate constraints. When the person who helped invent deep learning says he's worried, that carries different weight than a LinkedIn post.

Yann LeCun was Meta's Chief AI Scientist until the end of 2025, when he left to found a startup on the same bet. He is the interesting contrarian: he thinks current LLMs are a dead end and that the real breakthroughs will come from different architectures (world models, self-supervised learning beyond text). He might be wrong. He might also be the only one seeing something everyone else is missing. I don't know which, and I like that I don't know.

Ilya Sutskever co-founded OpenAI and was briefly at the center of the Altman firing saga before leaving to found Safe Superintelligence Inc. He is one of maybe five people on earth who combine world-class technical ability with deep concern about what they're building.

Dario and Daniela Amodei left OpenAI to found Anthropic, on the bet that making models safe and interpretable matters more than making them more powerful. They're building Claude.

Terence Tao, Fields Medal, is possibly the greatest living mathematician. He co-founded SAIR with Nobel and Turing laureates and uses AI as a research tool for scanning literature and testing conjectures, though not yet for deep ideas. If Tao ever turns his full attention to architecture research, the landscape shifts overnight.

Sakana AI is the Tokyo lab founded by Llion Jones (one of the eight) and David Ha. Their bet is that nature-inspired methods (evolutionary model merging, collective intelligence) can produce competitive models without brute-force scaling. If they're right, the Bitter Lesson has an asterisk.

What fits on an index card

The transformer is an embarrassingly simple architecture, one block repeated. The entire AI revolution fits on an index card. An explanation for why it works this well does not, because what emerges from scale is something none of the original authors predicted, a shapeless thing that can reason without being programmed to. Language, it turns out, contains the structure of thought itself.

What we built is a machine that approximates the output of minds, trained on the sum total of human written expression. The approximation keeps improving as compute gets cheaper. Every wall we've identified so far turned out to be a hidden inefficiency or a bad measurement.

I keep thinking about the moment in 2020 when I realized davinci could hold register for an entire paragraph if I engineered the context carefully enough. It felt like a magic trick. Five years later, the same architecture, scaled up, restructured an entire build system while I made coffee. The trick became a tool, and then the thing I think with. And the curve hasn't stopped. What stays with me is that it keeps getting smarter on a schedule, along a power law that shows no sign of bending.

Nobody knows if there's a wall. But the people betting on one have been wrong every time so far, and the evidence against them gets stronger every quarter. The wall claim is still possible, and increasingly lonely. I know which side of that bet I'm on.