Four new cards from three operators. Two threads connect from independent angles: where the human layer must hold even as AI handles more of the work, and where organizational structure either distributes or removes judgment from the people closest to the problem.
Theme 1, Where AI falls short and the human layer must hold
Tom Tunguz published on AI ghostwriting on August 3 and Hamel Husain spoke on eval failures at the Arize AI Engineer Summit on July 30. Neither was writing about the same problem, but both point at the same failure pattern: the AI layer gets blamed for an output problem that was created in the human layer upstream.
AI fails as a ghostwriter because every user shares the same underlying voice; it succeeds as an editor that suppresses itself makes the voice-convergence case. Because every writer using the same model draws from the same trained corpus, AI-generated prose converges to a shared style. Tunguz calls the result "the age of slop." The fix is not avoiding AI but sequencing the work: the author embeds idiosyncratic markers first, then uses AI as an editor that suppresses itself rather than a ghostwriter that blends its voice in. The practical signals are specific.
"The problem with AI as a ghost writer: everyone uses the same ghost writer."
Most LLM evaluation failures trace to product design problems upstream of the model, not model capability runs the same argument one layer down. When LLM evaluations fail, teams adjust prompts, swap models, and rewrite rubrics. Husain argues the correct first question is whether the task was solvable in the first place. If the user's input is ambiguous and the product does not enforce context, the LLM never had a chance. The most important AI engineering skill, in his framing, is looking at raw input-output pairs before attributing failure to the model.
"I get into eval reviews all the time where it's really obvious that the product is not built correctly. Stop. This is a product issue."
- Tom Tunguz, AI fails as a ghostwriter because every user shares the same underlying voice; it succeeds as an editor that suppresses itself. AI prose converges across users who share the same model; writing distinctively requires embedding idiosyncratic markers before AI refines the draft.
- Hamel Husain, Most LLM evaluation failures trace to product design problems upstream of the model, not model capability. Most LLM eval failures are product design failures upstream of the model; the first intervention is reading data, not adjusting prompts.
Theme 2, GPU supply pressure and the tier routing response
Tom Tunguz's August 4 post on what happens if GPU prices double is the most data-dense card this week. AI labs sustain consumption growth under supply constraints by segmenting models into tiers so workloads route to cheaper alternatives rather than stopping traces the economic mechanism by which AI labs sustain Jevons' Paradox even as frontier model prices rise under supply constraints. The response is tier segmentation: premium models cost 13x more than value-tier for only 20% additional capability; mid-market delivers 96% of frontier performance at 40% of cost; value reaches 84% at 1 to 5% of cost. Workloads that can no longer afford frontier pricing route down rather than stopping.
Three hyperscaler CEOs confirmed supply constraints in Q2 2026 earnings. Sundar Pichai: "We continue to be supply constrained, a sign of momentum and rapid adoption." Jensen Huang: "Cloud GPUs are sold out." The router is the layer Tunguz identifies as the key asset: the intelligence layer that routes requests to the right tier becomes the margin-capturing position as price sensitivity grows. Companies building on AI inference need to know which of their workloads are truly frontier-only and which can be rerouted down without quality loss.
- Tom Tunguz, AI labs sustain consumption growth under supply constraints by segmenting models into tiers so workloads route to cheaper alternatives rather than stopping. Supply-constrained AI labs sustain total GPU-hour consumption through three-tier model segmentation; workloads route to cheaper tiers rather than being abandoned, preserving Jevons' Paradox as frontier prices rise.
Theme 3, Judgment at the team level
Tom Verrilli's August 2 Lenny podcast episode is a clean structural argument. Fixed PM pods produce alignment theater; rotating problem-centric DRI assignments restore judgment to the people doing the work makes the case that permanent PM-to-team assignments are self-reinforcing: engineers and designers learn not to decide because someone always interprets the work for them. Whatnot replaced this with six-month rotating DRI assignments tied to problems rather than teams. The DRI can be a PM, an engineer, or a designer, depending on which discipline is best suited to own that problem. Whatnot runs the entire marketplace with 21 to 22 PMs under this model.
"Hiring so many PMs infantilizes the engineers and the designers who are perfectly capable of making good decisions but just never had to because there was always a PM to babysit them."
The structural read is that PM headcount growth is a signal, not a target. When decision authority sits above the people doing the work, the team optimizes for coordination rather than judgment. Rotating DRI ownership forces that judgment to distribute.
- Tom Verrilli, Fixed PM pods produce alignment theater; rotating problem-centric DRI assignments restore judgment to the people doing the work. Permanent PM assignments remove decision ownership from engineers and designers; six-month rotating DRI roles tied to problems restore judgment to the people closest to the work.