Janus

Everyone is betting everything. Not everyone is betting on the same thing.

When people think about AI, they think about Gemini.

But maybe they should be thinking about Janus instead.

Janus was the Roman god with two faces. One looking forward and one looking backward. And increasingly, that feels like the real state of the AI industry. Not divided between companies so much as divided within them.

Because there are now two fundamentally different beliefs about what AI actually is, where it is heading, and what kind of business will survive around it.

One side believes we are on the road to general intelligence. Not metaphorically, but literally. They see current systems as early, primitive, incomplete versions of something far larger. From that perspective, intelligence itself is mostly a scaling problem. More GPUs, more energy, more data centers, more parameters, more synthetic cognition layered on top of synthetic cognition until something qualitatively different emerges.

This side of the industry is spending like a civilization-level arms race has already begun. And if they are right, the spending is rational. If the first company to achieve true AGI gains control over the most important economic engine in human history, then a trillion dollars in infrastructure is not excessive; it is conservative.

That is the face building nuclear power agreements and hydro contracts. The face buying entire global GPU supply chains. The face preparing for a future where intelligence itself becomes centralized infrastructure.

But there is another face staring back at them from the same mirror.

And what makes this split even stranger is that the foundations of the modern AI industry did not emerge from research into consciousness, reasoning, or grounded understanding of reality. They emerged from translation.

Google’s 2017 paper Attention Is All You Need became the true founding document of the modern AI era.

https://arxiv.org/abs/1706.03762

The Transformer architecture it introduced turned out to be astonishingly effective at modeling relationships inside language and symbolic representations. The original breakthrough was not “we discovered digital cognition.” It was something much narrower and much more practical: attention mechanisms were extraordinarily good at understanding relationships between tokens inside sequences.

From that breakthrough, however, the industry began making a series of increasingly ambitious extrapolations. Translation became language modeling. Language modeling became reasoning. Reasoning became intelligence. And intelligence became AGI.

The entire modern AI industry now sits on top of that ladder of assumptions.

Then in 2020, OpenAI published Scaling Laws for Neural Language Models, which effectively poured gasoline onto the idea.

https://arxiv.org/abs/2001.08361

That paper argued something deceptively simple but economically explosive: larger models trained on more data with more compute kept improving in surprisingly predictable ways.

And that was the moment the industry truly lost its mind.

Because if capability kept scaling with compute, then the race became obvious. More GPUs, more data, more power, more infrastructure, more scale. And to be fair, the evidence initially looked overwhelming.

The models worked.

Absurdly well.

Good enough that an entire industry briefly convinced itself that maybe intelligence really was just next-token prediction at sufficient scale.

And perhaps that is unfair.

But only slightly.

Because underneath all the mythology, these systems are still fundamentally predictive engines. Extremely sophisticated predictive engines, yes... but predictive engines nonetheless.

The uncomfortable possibility haunting the industry is that scaling prediction may not be the same thing as scaling understanding.

GPT-4 felt like Lucy finally holding the football steady.

GPT-5 was the moment a lot of people in the industry started wondering if they were Charlie Brown again.

Because now the industry is arriving at the dangerous phase of the experiment: finding out whether the world’s most expensive autocomplete engine can actually cross the next threshold. Increasingly, many of the people closest to these systems appear to suspect the answer is no. Or at least, not through brute force scaling alone.

What these systems unquestionably mastered was language. But language is not reality.

Language is humanity’s symbolic compression layer for reality. A language model lives inside humanity’s descriptions of the world, not necessarily inside the world itself. That distinction becomes brutally obvious the moment you move from text into video.

Text prediction allows plausibility. World prediction requires inevitability.

A language model can often generate something believable about what could happen next. But asking a video model what will happen three frames from now turns out to be a radically different category of problem. Not what might happen, but what must happen.

A dropped glass must continue falling. Water must preserve continuity. Objects hidden behind other objects must still exist when occluded. Bodies must obey physics. The world must remain coherent even when partially unseen.

Suddenly the barrier is no longer syntax or semantic association. It is grounded reality itself. And no amount of additional compute automatically guarantees you can brute-force your way through that wall.

I increasingly suspect this possibility haunts the second faction inside these companies. Because if they are right, then the current models are not primitive proto-gods waiting to wake up. They are already the product.

And if that is true, then the race changes completely.

This second group sees something the dreamers often do not. Current models are already powerful enough to reshape enormous parts of civilization. Software development, education, customer support, management structures, search, scientific research, material science, media production... all of it is already beginning to bend around systems that exist today.

But to them, this is no longer just a technological observation. It is a financial emergency.

Because sitting behind the industry is an almost absurd engine of capital expenditure. A monster consuming power plants, GPUs, data centers, fiber, water, and entire national-scale energy agreements at a rate that starts to feel less like software and more like heavy industry.

The people trying to commercialize current AI increasingly look like they are frantically paving the road ahead of that machine before it catches them.

Every SaaS feature suddenly becomes vulnerable. Calendar companies, CRM companies, note-taking apps, customer support systems, search products, workflow tools, presentation software, research assistants... everything starts looking like one more feature that can be absorbed into a larger AI subscription bundle.

You can almost feel the urgency behind it. One more product integrated. One more workflow captured. One more recurring subscription pulled into the ecosystem before the infrastructure bill arrives.

Google understands this instinctively because Google has always understood bundling. Gmail was never just Gmail. Docs was never just Docs. Search was never just Search. The real business was always the gravitational field created by tying everything together.

Gemini increasingly feels like the continuation of that strategy. Not merely an AI model, but an attempt to turn all of Google Workspace into a single cognitive subscription layer where every adjacent software category eventually collapses inward.

And if you use Gemini long enough, you start noticing another tell. It leans heavily on Google’s existing search infrastructure: existing indexes, cached knowledge, retrieval systems, summaries, search grounding. You almost have to strong-arm it into actually going out and spending meaningful inference effort exploring something fresh.

That is not an accident. That is economics leaking through the interface.

A company truly convinced that unlimited inference was the future would behave differently. It would burn compute aggressively and freely because compute would be the path to the prize. Instead, what we increasingly see across the industry are systems trying to conserve reasoning wherever possible. Retrieval first. Cached answers first. Smaller specialist systems first. Inference only when necessary.

These are not the behaviors of companies escaping economic gravity. They are the behaviors of companies deeply aware of it.

Microsoft’s behavior, meanwhile, feels oddly conflicted.

For years, the company executed one of the most successful business transformations in modern corporate history. It turned itself from a boxed software company into a cloud and subscription empire built around Azure and Office 365. It understood recurring revenue, enterprise dependency, workflow gravity, and organizational lock-in better than almost anyone.

Which is what makes the current moment feel so strange.

Because Microsoft already owned the daily operational fabric of enterprise work. Email, calendars, documents, spreadsheets, meetings, identity management... the company was already sitting exactly where delegated cognition was naturally going to emerge.

And yet much of its AI posture still feels like it is orbiting OpenAI’s narrative rather than extending Microsoft’s own.

Copilot often feels bolted onto products rather than metabolized into them. A feature suite attached to Office instead of an operational layer dissolving invisibly into workflow itself.

Google, by contrast, increasingly behaves like a company trying to make AI disappear into the substrate of its ecosystem.

That difference matters.

One approach treats AI as a destination product.

The other treats it as atmospheric infrastructure.

Anthropic appears to have approached the problem differently, leaning hard into programming because code has unusually clean economics. It has measurable outputs, enormous labor replacement value, direct enterprise demand, and vast open-license ecosystems to train against.

That may also explain why Anthropic seemed unusually willing to make peace with creators and publishers early. I increasingly suspect they understood something important: general content may matter politically, but programming ecosystems matter economically.

And more than that, I suspect they understood something strategically brutal.

If OpenAI, Google, xAI, and everyone else end up trapped in endless licensing wars over the right to answer humanity’s infinite stream of general-interest questions, all the better for Anthropic.

Because every billion dollars spent negotiating with publishers, creators, news organizations, archives, and eventually perhaps even something like Wikipedia is a billion dollars diverted away from infrastructure, tooling, programming ecosystems, and enterprise integration.

Meanwhile, Anthropic increasingly behaves like a company that believes the real long-term value lives somewhere much narrower and much more economically defensible: code.

Code has validation systems. Code has reinforcement loops. Code produces leverage. Code generates more systems.

A model optimized for programming is not merely answering questions. It is participating in the recursive construction of future infrastructure.

And if that is the future Anthropic sees coming, then encouraging the rest of the industry to become entangled in increasingly expensive cultural licensing battles starts looking less like altruism and more like strategic positioning.

Let everyone else spend fortunes building infinitely patient machines for answering the world’s stupidest questions.

Meanwhile, the company focused on helping developers build the future may end up owning the economically productive layer underneath all of it.

And underneath all of this sits the industry’s deepest fear: the models themselves may become commodities.

In fact, the existence of public APIs already suggests the companies know this risk is real. If raw intelligence could remain permanently proprietary, they would never expose it cleanly to the market in the first place.

But they know what happens to infrastructure over time. Storage commoditized. Bandwidth commoditized. Compute commoditized. Databases commoditized. Cloud infrastructure commoditized.

Increasingly, intelligence itself may commoditize too.

There will always be someone willing to sell inference cheaper. There will always be open models. There will always be regional providers. There will always be distillation. There will always be optimization pressure.

Which means the durable value likely moves upward into the surrounding systems: memory, workflow, identity, coordination, trust, orchestration, integration.

Not the model itself... the ecosystem wrapped around it.

That is why the industry increasingly feels split between two futures.

One future treats AI as a path toward centralized superintelligence.

The other treats AI as a new infrastructure layer around which useful systems will be built.

The first group wants bigger minds.

The second wants better orchestration.

And increasingly, I suspect the second group may win.

Not because AGI is impossible. Not because the models stop improving. But because markets eventually force discipline.

Public companies obey margins. Infrastructure obeys economics. Customers pay for usefulness, not metaphysics.

A model that is 20% smarter but 30 times more expensive may simply lose commercially to a system that is cheaper, integrated, reliable, and already good enough for most economic purposes.

And “good enough” may already be enough to permanently alter civilization.

That is the strange thing about this moment.

The AI revolution may already have happened.

The industry just cannot yet agree on whether it is building the convergence... or a business.