Recursive self-improvement is usually told as a spell: the model tweaks itself, gets smarter, tweaks itself again. Control theory treats that as a feedback loop with a maximum safe speed, set by how fast you can generate trustworthy evidence that a change was actually an improvement.

A loop stays stable when the product of its gains stays below one. Gain here is how large a modification the system makes per unit of evidence. The entire appeal of self-improvement is high gain. Adaptive control already learned the hard version of this: proofs assumed a plant, a mild violation of the assumptions sent the system haywire. Unstable adaptive systems often do not blow up immediately. They burst: run beautifully, then fail without warning.

Most companies will use AI to clear a faster ticket queue. In-house creative teams already function as service desks: hundreds of reasonable requests that reproduce decisions already made. When the queue is too slow, everyone else ships around it. The brand dies by papercuts. If the structure stays the same, AI only makes the same bad system move faster.

Guidelines were built for scarce production. They packaged judgment into a manual and handed it to people with different expertise. A guideline can describe the brand. It cannot participate in the work. Brand as software can: it carries beliefs, product truths, audience, voice, and the line where a human must step in into the artifact itself.

Reading is not decoding information. It is meeting a mind. The same circuitry used to infer what a friend is thinking lights up when you read a paragraph. AI text activates that circuitry perfectly, then leaves it with no one to find. You turned up for a conversation and the other chair is empty.

Six mechanisms land as one event. There is no intention behind the word choices, so inferences close on air. The relevance bargain breaks: too many sentences, too little return. Surprise is gone, because the model sits at the statistical centre of everything ever written. The sender spent a minute; the reader spends far longer. Smoothness used to mean skill. Now it often means automation. And the templates (lists of three, hedges, restated closings) get processed as pattern rather than message.

For most of history, physical work was muscle. Then steam, combustion, and electric motors took it. Human muscle became a rounding error in the global energy budget. Cognitive work is on the same curve, compressed in time.

When the price of a scarce input collapses, supply floods and demand expands. Cheaper engines did not reduce coal use. They multiplied it. Cheaper cognition will not reduce thinking. It will multiply it. Most problems on Earth currently go unthought about because thinking is expensive.

Execution is commoditizing. Most agent loops can plan, call tools, and finish a task. What still feels scarce is state: an agent that remains oriented across days of waiting, callbacks, and quiet intervals.

Storing every transcript is not memory. A tape recorder keeps photons. A mind forgets differences so it can think. Real memory is a compiler. It makes an active bet about what will be useful tomorrow.

Most discussions of an AI bubble focus on valuations. The more important risk may be the quality of the earnings those valuations rest on.

Reported profits across parts of the AI stack are being supported by a dense set of interconnected mechanisms. Large fiscal deficits and cheap capital boost demand that can look organic. Circular financing routes investor capital into startups that then buy the same investors’ hardware or cloud capacity. Accounting choices extend depreciation lives, capitalize R&D, and lean on non-GAAP adjustments. Buybacks and high operating leverage amplify earnings growth. Some demand reflects inventory build and defensive capital spending rather than immediate end-user pull. Subsidies and favorable financing terms further improve reported returns.

Most systems still give models knowledge the hard way. They retrieve raw chunks at query time and force the model to re-derive understanding on every question. Nothing accumulates.

A better pattern has quietly emerged in 2026.

Anthropic researchers have identified a distinct internal workspace inside Claude — referred to as "J-space" — where thoughts appear to form before any tokens or words are generated. Suppressing this space allows basic recall but severely impairs complex, multi-step reasoning, suggesting it plays a fundamental role in cognition.

John Nosta connects this discovery to global workspace theory from cognitive neuroscience, which proposes a shared, flexible mechanism for chaining reasoning, integrating information, and making results available across different parts of a system. The emergence of a similar structure in a non-biological model raises a profound question: Are there universal first principles of thought that arise whenever sufficient complexity and training dynamics are present?

AI is intensifying two interconnected crises in our information ecosystem.

The first is what Salomé Viljoen calls the "Cylon problem" — the growing indistinguishability between authentic human content and AI-generated material. This creates widespread epistemic instability: people can no longer reliably tell what is real, leading to generalized suspicion, the dismissal of genuine content as fake, and the proliferation of low-quality "slop." The result is a "zombie internet" where AI-generated content dominates engagement without genuine human connection.

AI is fundamentally changing the economics and speed of cyberattacks. What was once a slow, resource-intensive cat-and-mouse game between attackers and defenders has become an asymmetric race that defenders are rapidly losing.

Generative AI tools now allow even low-skill actors to write adaptive malware, automate sophisticated phishing and espionage campaigns, and exploit known vulnerabilities in record time. The window between a vulnerability being disclosed and actively exploited has collapsed from hundreds of days to as little as 44 days in some cases. Meanwhile, the foundational insecurity of the internet — built on rushed code, reactive patching, and legacy systems — remains largely unchanged.