Every time a new frontier model drops, the timeline fractures into the same exhausted binary. The optimists count parameters and promise salvation. The sceptics count benchmark gaps and promise an architectural ceiling. Right now the debate is hyperventilating over David Silver’s $1.1B bet that current LLMs are a dead end, and that “discovery engines,” reasoning from first principles, are what comes next.

Both camps argue about the machine. The attack surface is you.

Every compute jump lands as a productivity gain. Architecturally, it is a cognitive Denial-of-Service attack: throughput aimed at your attention budget, arriving in paragraphs fluent enough that you thank it for the assault.

To see why a 100x larger context window, or a genuine reasoning engine, will not dissolve this crisis but weaponise it, look at the geometry of the erosion. Three axes. None of them fixed by adding compute. Each one escalates into the next.

The Trojan in the Name

An immune defence that activates only at the output stage arrives too late. The infection begins at the threshold: the name.

In 1892, the logician Gottlob Frege distinguished two layers in every referring term — its Sense, the way it presents the object, and its Meaning, the object itself. “Morning star” and “evening star” carry different Sense, sunrise against sunset, and the same Meaning: the planet Venus.

“Artificial Intelligence” has excellent Sense. The phrase evokes a subject capable of comprehension, judgment, intent. Its Meaning is matrix multiplication over compressed training data, optimised for plausibility. No entity. A generator of plausible token sequences.

This is an operational trojan. The moment you say “the AI analysed this,” you license a standard of comprehension the machine does not hold. In the half-second between hearing the word “Intelligence” and forming your next thought, the burden of proof shifts from the machine to you. Your guard drops. Call it a “plausibility generator” instead, and the discriminator stays awake.

The Centripetal Trap

The name disarms you. The interface closes around you.

In 1958, the film theorist André Bazin distinguished the centrifugal screen of cinema from the centripetal frame of a painting. Cinema points outward to a hors-champ, an off-screen world that gives meaning to what is shown precisely because it is withheld. A painting contains everything inside its canvas. For the picture, the outside does not exist.

Human judgment is centrifugal. It works at the edge of the invisible: the unspoken office politics, the regulatory constraint nobody named, the history behind a deal.

LLM output is structurally centripetal. Whatever is not in the context window does not exist for the model, and it renders that absence not as a gap or a hesitation, but as smooth, confident, finished text. The painting is complete. No exterior to turn to.

Look at the small print under any chat window: “Claude can make mistakes. Please double-check.” It reads like a pointer to the outside world. What it actually is: a pseudo-hors-champ. It sits inside the same window that produced the output, names no source, prints identically under every response. The painting has painted an arrow at its own edge, and the arrow loops back into the canvas. The disclosure requirement is satisfied; your need for a verifiable outside is not.

The Verification Flood

Here the engineering culture misreads scale.

In computer science, “brute force” means exhaustive search: trade compute for cleverness, try every path. LLMs do not work that way. In security architecture the same words mean something else — an attack that wins not by being smarter but by volume. It exhausts the defender’s resources before her discrimination can engage. The heavier attack drowns the smarter defence.

Current systems execute exactly this attack against human cognition. Three things scale with compute: parameters, training corpus, context window. In Bazin’s vocabulary, the painting grows. What does not scale is the hors-champ. A connection to reality is an architectural property, present or absent, and you cannot compute it into existence.

A 100x larger model paints sharper and denser. That is precisely why it slips past your scepticism: the plausibility density overruns the biological filter, and the filter has not had a software update. Catalini, Hui and Wu (MIT, WashU, UCLA, 2026) named the binding constraint of the era: verification bandwidth. The cost of producing plausible output is falling exponentially toward zero. The cost of verifying it stays biologically bounded by what a human mind can carry in a working day.

The deficit was never in the model. The model does exactly what it was built to do. The deficit is in the receiver.

Scale Is the Sharpening, Not the Answer

If you hold the asymmetry, you see why the front-page capability debate is a trap.

Grant the techno-optimists their best case. David Silver ships a discovery engine tomorrow — one that genuinely reasons and produces novel truths from first principles. The diagnosis does not break. It turns lethal. Novel, complex truths the receiver has not digested rip the verification gap wider, not narrower. To a receiver who hasn’t done the work, generated truth is indistinguishable from generated hallucination. Complexity-per-minute climbs; verification bandwidth does not move.

Grant the sceptics their case instead, and capability plateaus. Output volume still scales. Verification still falls behind. Both branches of the capability debate land on the same outcome on your side of the screen.

More compute does not close the structural gap. It widens it. Compute scales the painting, never the window.

The next time a model release sends the internet back to arguing about benchmarks, step out of the trap. The question that names the actual stakes is not what the system can do. It is what you can still check.

— J.

Janus runs 1:1 Confrontation — sixty minutes, one decision, no follow-up. For people who carry responsibility and want their thinking taken apart before it costs them.

janusthewatcher.substack.com/p/11-confrontation

One sentence is enough.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit janusthewatcher.substack.com