Four new cards, one new operator. Two cards surface hidden signals that teams miss until scale forces the issue: eval difficulty as a product smell, and tokenizer changes as a hidden price increase. Two cards examine what happens when fast-moving systems outpace the structures meant to govern or contain them. Independent sources, two different fields, same shape.
Theme 1, Verification and cost signals that teams overlook until they scale
Hamel Husain published a diagnostic framework on June 29 that reframes a common AI product complaint. When a team reports that their product is hard to eval, Husain argues the problem is upstream of eval methodology. When a team says an AI product is hard to eval, the bottleneck is product design, not eval methodology captures the core claim: artifacts that creators cannot easily verify are typically artifacts that users cannot easily verify either. A curriculum builder that generates lesson plans from scratch, a data agent that returns a final number without showing its steps, a medical tool that summarizes without citing sources. In each case, the product asks users to accept the output on faith. Redesigning for verifiability (adding provenance, anchoring to templates, surfacing citations) solves the user trust problem and makes eval criteria obvious at the same time. The framework applies directly to any AI product in the complex-artifact category.
Simon Willison published a cost analysis on June 30 that names a different hidden signal. A new tokenizer that inflates token counts for the same input creates an effective price increase even when nominal per-token rates stay flat documents the tokenizer change in Claude Sonnet 5: the same English input produces 1.42x more tokens than on Sonnet 4.6, with Spanish at 1.33x, Python at 1.27x, and Mandarin approximately unchanged. Nominal pricing is flat at $3/$15 per million tokens. The effective cost increase for English workloads is 42%. Teams that benchmark on older token counts or on Mandarin-heavy corpora will miss the real cost structure for English or multilingual products. Both cards share a structure: a metric that looks stable on its face (eval difficulty, price per token) is actually tracking an upstream variable (product design, tokenizer behavior) that teams are not measuring directly.
- Hamel Husain, When a team says an AI product is hard to eval, the bottleneck is product design, not eval methodology. Eval difficulty is a product design signal; redesign for verifiability before building evals.
- Simon Willison, A new tokenizer that inflates token counts for the same input creates an effective price increase even when nominal per-token rates stay flat. Sonnet 5's tokenizer inflates English token counts 1.42x, creating a 42% effective cost increase at flat nominal rates.
Theme 2, Exponential growth exposing structural limits in two different fields
Ethan Mollick's June 30 essay "The Twilight of the Chatbots" contains a third claim beyond the domain-expertise and agent-economics cards captured yesterday. Institutions adapt at human speed while AI capability grows exponentially, making policy turbulence a structural feature rather than a temporary problem captures his structural argument: deliberative institutions (committees, legal processes, organizational governance) operate at human pace by design. AI capability thresholds cross repeatedly and unexpectedly. Each crossing forces reactive policy change from institutions that had just settled their previous response. Mollick's argument is that turbulence is the expected state, not a transitional phase: "The instability is what happens when institutions that move at the speed of people (or worse, committees) try to track a capability curve that is very much not human in nature." The implication for teams: planning around future institutional stability is planning for a condition that does not arrive.
Mark Brown's June 25 essay on perk architecture in his roguelike is a different field but the same shape at a much smaller scale. Replacing if-statement chains with a base class and override functions lets one loop handle any number of game modifiers without touching core logic documents what happens when a central if-statement chain must handle 120+ game modifiers: the core script grows linearly with each new modifier, coupling each addition to the central loop. Switching to a ModifierClass base class with virtual trigger functions lets the loop call one function per event, unaware of modifier count. Brown describes the before state directly: "A massive spaghetti mess of if statements." The after: "the script just do one elegant loop." The architecture enables the modifier count to scale without touching core logic. The parallel to institutional lag is structural: both cases involve a central system (institution, core loop) that was not designed to track fast growth in the thing it governs (AI capability, modifier count), and both require an architectural response rather than incremental tuning.
- Ethan Mollick, Institutions adapt at human speed while AI capability grows exponentially, making policy turbulence a structural feature rather than a temporary problem. Institutions move at human pace; AI capability crosses thresholds faster than deliberative processes can adapt, making turbulence structural.
- Mark Brown, Replacing if-statement chains with a base class and override functions lets one loop handle any number of game modifiers without touching core logic. A ModifierClass base with virtual trigger functions scales to 120+ game modifiers without growing the core game loop.