Will LLMs Be the Betamax of AI?

Every technology cycle has a first miracle.

The thing that makes the future visible. The thing that attracts the capital. The thing everyone starts building around. The thing that suddenly makes the old world look obsolete.

For AI, that miracle has been the large language model, or LLM. Most of today’s leading LLMs are built on the Transformer architecture introduced by Google researchers in 2017. OpenAI’s GPT models, short for Generative Pre-trained Transformers, turned that architecture into the first mass-market interface for AI.

OpenAI made it mainstream. Anthropic positioned it for enterprise. Google, Meta, xAI, DeepSeek and others joined the race. Capital flooded in. Data centres became strategic assets. Chips became geopolitical infrastructure. Every boardroom started asking what its AI strategy should be.

But there is a dangerous assumption hiding in plain sight: that the technology which makes a new market visible will also be the technology that owns it.

History says otherwise.

The Better Technology Does Not Always Win

When I was a kid, my parents bought a Sony Betamax video recorder. Many of my friends’ families had VHS. At the time, Betamax was widely understood to be the better technology.

But that did not help me much when more and more of my friends had VHS tapes I could not play!

That was the real lesson. The best technology in isolation is not enough. If the ecosystem moves elsewhere, you are stuck with the better product that fewer people use.

Betamax lost because VHS became the market standard. VHS machines were cheaper, easier to use and, crucially, could record longer programmes and films on one tape. The format war was not settled by technical elegance alone. It was settled by compatibility, convenience, availability and network effects.

Technology markets do not reward purity. They reward timing, distribution, ecosystem formation, standards, interoperability and customer behaviour.

Kodak offers the same warning from another angle. Kodak did not miss photography. Kodak was photography. It popularised consumer photography, built one of the most trusted brands in the world and even had deep digital imaging capabilities. Yet Kodak was brought down by competition, by the fear of cannibalising its own analogue film business, and by its inability to keep pace with the move from film to digital. It eventually filed for bankruptcy protection in 2012.

The demand did not disappear. Photography exploded. The value simply moved elsewhere and different winners emerged.

That is the question now for GPTs and LLMs.

Are they the final architecture of artificial intelligence, or simply the first mass-market interface that made AI legible to the world before being superseded?

The Limits of Next-Token Prediction

Yann LeCun is not another commentator looking for a contrarian angle. He is one of the people who made the current AI era possible. He shared the 2018 Turing Award, computing’s closest equivalent to a Nobel Prize, with Geoffrey Hinton and Yoshua Bengio for breakthroughs that made deep neural networks central to modern computing. He founded Facebook AI Research and helped create one of the most influential AI research organisations in the world.

His view is not that LLMs are useless. It is more precise than that.

LLMs are extremely useful for what they do. They manipulate language, code, mathematics, documents and structured text with extraordinary power. They are already changing work. They are already valuable.

But LeCun’s argument is that they are not the route to human-level intelligence. Once you understand how LLMs actually work, his view starts to make sense. A LLM is trained to predict the next token. It looks at the sequence so far and estimates what should come next. That works brilliantly when the world you are operating in is made of language, symbols or code.

But the physical world is not a sentence, and an artificial intelligence needs to understand it to decide how to act and respond.

A car approaching a junction does not need the next word. It needs to know what might happen if a cyclist turns, if a pedestrian steps out, if the road is wet, if another driver hesitates, if a van blocks the view, or if braking now creates danger two seconds later.

A chatbot predicts tokens.

A world model predicts consequences.

From Words to Worlds

A world model is a system that can anticipate the consequences of actions before taking them. It can imagine possible futures, search across them, and choose a sequence of actions that moves towards a goal.

Which is exactly what humans and animals do.

A 17-year-old does not learn to drive by consuming millions of hours of labelled driving data. They already understand objects, movement, danger, space, people, cause and effect. They know a ball rolling into the road may be followed by a child. They know a wet road changes braking distance. They know a lorry can block sight lines. They still need training, but they are not learning the physics of the world from scratch.

Autonomous vehicles, by contrast, have had more than a decade of investment, huge data collection and some impressive deployments in constrained environments. Yet full, everywhere, all-conditions autonomy remains a harder problem than many expected.

Data is not the same as understanding. Imitation is not the same as planning. Scaling is not the same as intelligence.

This is the technical hinge in the AI debate. LLMs can sound intelligent because language contains vast amounts of compressed human knowledge. But sounding intelligent and acting intelligently in the world are not the same thing.

If you are asking AI to summarise a document, draft a contract, generate code, translate a language, or produce a marketing plan, an LLM may be enough. If you want AI to control a robot, optimise a factory, help design a course of treatment, manage an industrial process or operate safely in the physical world, it needs more than fluent output. It needs memory, perception, planning and a model of what happens next.

LeCun left Meta to start Advanced Machine Intelligence, a venture explicitly aimed at building systems that learn world models from video and physical interaction rather than scaling text prediction. AMI is not a bet against AI. It is a bet against LLM-centrism. It is a bet that the next step is not a larger prompt box, but systems that can reason, plan and operate in the real world.

What If LeCun Is Wrong?

The opposite case deserves a hearing.

Language is not just a description of the world. It is a compressed encoding of human reasoning, causality, social dynamics, scientific knowledge and practical know-how. Every chess game, surgical procedure, legal argument and engineering decision has been described somewhere in text. Add tool use, memory, real-time retrieval, multimodal perception and agentic planning, and the case can be made that LLMs do not need to be the whole intelligence stack. They simply need to be the reasoning engine that orchestrates everything else.

On this view, world models are not a replacement for LLMs. They are something LLMs increasingly incorporate. The autoregressive Transformer becomes the substrate, and capability accrues through scale, tools and scaffolding rather than architectural revolution.

Throwing more compute, capital and data at LLMs is clearly making them more capable. The open question is whether they are becoming fundamentally more intelligent, or simply more convincing.

LeCun may be right. He may also be wrong. The honest answer is that nobody knows yet. Which is precisely why building a strategy on the assumption that today’s architecture is permanent is dangerous.

The Internet Did Not Stop at Cisco

This is where my Betamax analogy breaks down as our Sony Betamax lost outright. LLMs are unlikely to vanish because language is too important. They are good at what they do. The more likely outcome is demotion, not extinction.

LLMs may move from being the centre of the AI story to being one layer inside a broader intelligence stack: memory, planning, perception, simulation, search, tool use, robotics, agents and real-time interaction.

The better analogy is not Betamax. It is Cisco.

Cisco sold the infrastructure that allowed the internet to scale. At the height of the dotcom boom, it briefly became the most valuable company in the world, the equivalent of NVIDIA today. Sun Microsystems, whose slogan “The Network is the Computer” was almost prophetic, was another foundational player. Neither was a weak technology or a weak company. They were essential to the first phase.

But the internet did not settle there. Linux, open source, cloud, SaaS, search, marketplaces, mobile and application platforms moved the economic value elsewhere. Cisco remained important, but it did not own the internet economy. Sun was eventually acquired by Oracle.

Adobe is the counterexample because it moved. It shifted from packaged software to Creative Cloud, took the pain, absorbed the backlash and repositioned itself as a workflow and subscription platform. It survived because it adapted to the next layer of value creation rather than defending the old one. Adobe is doing it again as LLMs are extremely capable of generating and editing content.

The real question for OpenAI and Anthropic is whether they are Adobe, able to evolve as the stack changes, or Sun and Cisco: essential to the first wave, but not necessarily the long-term control point.

When the World Model Arrives

The capital pouring into the LLM ecosystem makes today’s leaders feel inevitable. In the internet boom, money poured into access, networking, servers and portals. Once those layers became cheaper, standardised and widely available, the value moved upwards.

The same dynamic is already visible in AI. Smaller models are improving. Costs are falling. Open models are getting stronger. Many tasks no longer require frontier-model capabilities. Enterprises are moving from experimentation to deployment.

A research lab asks whether the dominant architecture is wrong. A product company asks how to scale and defend what already works. OpenAI and Anthropic are now both - and that tension matters. Defending today’s revenue engine and re-architecting around a fundamentally different approach are very hard to do at the same time. Kodak could not. Few do.

If a materially better architecture does emerge, the re-rating event would be brutal. Companies built around specific LLMs, prompt chains, model-specific integrations and brittle agent workflows would have to rebuild.

Some would adapt. Many would not.

Enterprise buyers would have to revisit procurement, governance and integration assumptions. Investors would discover that what looked like a moat can evaporate overnight.

This is why technology agnosticism matters when the winners have not yet been crowned. If your strategy depends entirely on one vendor, one model class or one interface, you are not building long-term capability. You are betting all of your chips on Cisco in 2000.

Build Capability, Not Dependency

Many businesses are making the wrong assumption. They think adopting AI means choosing a model. It does not.

If your AI strategy is a chatbot licence or access to a coding platform, you do not have an AI strategy. You have access to a tool. If your AI strategy is tied entirely to one model provider, you are not becoming AI-native. You are becoming dependent on a single version of the future.

Three failure modes are being signed off this quarter that will look like obvious mistakes in three years.

The first is single-vendor agent stacks dressed up as strategic partnerships. The integration cost is real. The lock-in is real. The architectural assumption that today’s leading model will still be the right answer in 2028 is a bet, not a plan.

The second is prompt-chain businesses that confuse a clever workflow inside one model with a defensible product. Capabilities improve. Margins compress. What looks like an AI company is often a thin wrapper around someone else’s roadmap.

The third is governance and evaluation frameworks bolted on after deployment rather than designed in. When the model changes, and it will, there is no way to know what broke, what regressed, or what now needs to be re-approved.

The real question is not, “Which model should we use?” The real question is, “Where will our advantage sit when every competitor has access to capable AI?”

It will not sit in access alone. Access commoditises. It will sit in proprietary data, workflows, customer relationships, governance, distribution, domain expertise and the ability to redesign work around intelligence.

That is the move from tool adoption to operating leverage. A tool helps one person move faster. Operating leverage changes how the organisation works. A chatbot can draft an email. A redesigned workflow can change how customer service, legal review, sales qualification, procurement, product development or compliance operates end to end.

Leaders should use GPTs aggressively. Waiting for the perfect architecture is a mistake. By the time it is obvious, the advantage will be gone. But they should not worship GPTs. They should build AI systems that can swap models as the market changes. They should own their data layer. They should build evaluation and governance early. They should understand which decisions can be automated, which need human judgement, and which will become machine-led over time.

The next AI wave may not be conversational. It will be operational.

The first wave made intelligence accessible through language. The next wave will embed intelligence into systems that remember, perceive, reason, plan and act. Some of that will be powered by LLMs. Some of it will not.

The mistake is to assume that because GPTs started the AI revolution, they must also finish it.

Here is the signal to watch out for: If a model emerges that learns to drive from a hundred hours of footage rather than a hundred million miles, or learns a manufacturing process from a week of observation rather than a year of labelled data, the stack is moving. That is what consequence prediction looks like in practice. It will not arrive with a chat interface.

Language may be how we communicate with intelligence. It does not follow that language is intelligence itself.

The companies dominating AI today may not be building the final form of intelligence.

They may simply be building the bridge to it.

Thank you for reading.