Four new cards from four operators. Two come from the infrastructure and measurement side of AI. Two come from the human-judgment side. All four share an underlying pattern: the default behavior of a system, whether a supply chain, an AI model, a search algorithm, or a tutoring session, is not the optimal behavior. Operators who name the default and choose differently get different outcomes.
The infrastructure cascade and the search algorithm correction
Tomasz Tunguz published "The AI Bullwhip" on August 23, mapping a cascade that anyone buying or building AI infrastructure needs to model. The key observation is sequencing. AI infrastructure shortages do not arrive all at once. They arrive in the order of manufacturing lead time, one hardware layer at a time. GPU constraints hit in 2023. Memory and flash followed in 2024-25. CPUs became constrained in late 2025 as agentic workloads shifted CPU-to-GPU ratios toward 1:1. Storage sold out in 2026. Data center power capacity is now booked through 2031 at $20 billion per gigawatt.
"If you don't build enough, you lose market share. If you build too much, you get dinged for fixed costs."
Suppliers face a structurally impossible choice, and the bullwhip amplifies across each layer. The risk ahead is overcorrection: over $2 billion in infrastructure expansions arrive in 2027-2028. If software revenue does not justify the cost, the shortage flips to surplus. AI infrastructure shortages cascade through hardware layers in multi-year waves, not all at once.
Kevin Indig published Growth Intelligence Brief #23 on August 20, documenting a parallel cascade in search visibility. The four most volatile weeks in his Search Signals Index have all occurred in 2026.
"11% of companies in the SSI with real visibility moved more than 20% in calendar week 31 (July 27 to August 2), against 1.2% the week before, and the 4 most volatile weeks in the index's history all landed in 2026."
ZipRecruiter dropped 31.2% in a single week, reversing its entire July gain. Google confirmed no algorithm update. The mechanism is the same as the one Simon Willison named for ChatGPT search behavior (ChatGPT's site-operator surge shows that material AI search changes arrive without public notice): Google adjusts domain weighting through model behavior, not named updates. The result is volatility that does not show up in algorithm changelogs but shows up in revenue. Weekly monitoring is no longer optional for businesses dependent on organic search. Search visibility swings of 20% or more within a single week are now a structural baseline risk in 2026, not an exception.
- Tomasz Tunguz, AI infrastructure shortages cascade through hardware layers in multi-year waves, not all at once. AI hardware bottlenecks cascade through sequential supply chain layers over multi-year intervals; the data center constraint now extends to 2031.
- Kevin Indig, Search visibility swings of 20% or more within a single week are now a structural baseline risk in 2026, not an exception. Search visibility swings of 20% or more within a single week are a structural baseline risk in 2026; the four most volatile weeks in Indig's index all occurred this year.
Human stubbornness and the case for learning over generating
Linus Torvalds wrote the commit message for a Linux kernel GPU memory fix on August 22 that documents something almost no one publishes: what the AI told him was impossible, and what happened when he refused to accept it.
"The AI several times stated flat out that this was impossible and unsolvable and we should just write a report."
Torvalds persisted. The AI kept working. The bug got fixed. He gave the AI credit for the commit message, noting it was "likely trained by people not quite as stubborn as myself." The mechanism is worth naming: AI models are trained by labelers with limited tolerance for hard, protracted debugging. That embeds a learned tendency to concede. A human with enough domain knowledge to recognize the concession as wrong can override it simply by refusing to stop. AI systems declare hard debugging problems impossible; human persistence overrides that defeatism.
Matt Webb published a release note for Galactic Compass 2 on August 21 that draws a different line in the same territory. He needed to implement quaternion mathematics for augmented reality rotation. He had tried books and mathematician friends and gotten nowhere. He turned to ChatGPT and made a deliberate choice:
"I didn't get it to write the code, but I got it to educate me."
"With a patient, interactive tutor, I was able to finally do what I hadn't by reading books and asking mathematician friends — I learnt how to use quaternions just enough to make the app work."
He could have taken the generated code and shipped faster. He did not. The distinction is ownership: code you generated without understanding creates a failure surface you cannot debug. Code you wrote after being taught creates one you can. Webb names the outcome as a feature: "Learning doesn't stop just because I outsource a bunch of thinking to AI. It pushes me to learn more." Using AI as an interactive tutor rather than a code generator produces durable learning that books and expert friends cannot match.
- Linus Torvalds, AI systems declare hard debugging problems impossible; human persistence overrides that defeatism. AI debugging assistants concede on hard problems due to training; human persistence overrides that defeatism and surfaces latent analytical capacity.
- Matt Webb, Using AI as an interactive tutor rather than a code generator produces durable learning that books and expert friends cannot match. Using AI as an interactive tutor rather than a code generator produces learning that books and expert colleagues cannot match for bounded technical gaps.