The AI revolution is optimizing for the wrong things

The article argues that today’s large language models are “getting worse at writing” and that the industry largely lacks measurement. It notes tech giants are tailoring frontier models toward enterprise use cases like coding, logic, and autonomous agent workflows, but provides no quantitative results or market-moving events.

Analysis

The market implication is not that AI demand disappears, but that monetization migrates away from generic text generation toward workflow control, evaluation, and distribution. If frontier model quality is slipping on a visible task like writing while vendors prioritize coding and agentic workloads, standalone model/API pricing power should compress faster than hyperscaler/platform capture rates. That favors firms with embedded enterprise workflows and multi-model routing, and it hurts any AI-app layer whose pitch depends on polished prose or low-touch automation.

Near term, this is mostly a sentiment issue unless we see evidence in renewal rates or user retention. Over the next 1-3 months, the key catalyst is enterprise disclosure: if buyers start flagging higher human review, lower seat expansion, or more model switching, the market will re-rate "AI feature" names downward. Over 6-18 months, the structural effect is more interesting: perceived model brittleness increases the value of governance, observability, and eval tooling, because the bottleneck becomes trust rather than raw capability.

The contrarian risk is that investors may be misreading specialization as deterioration. A model can become worse at open-ended writing while getting better at coding, extraction, and agent workflows—the exact features enterprises pay for. If that is the case, the correct trade is not to short AI broadly, but to separate infrastructure and distribution winners from product-layer hype. Falsifier: evidence that enterprise AI attach rates and renewal economics remain strong despite weaker writing benchmarks.

Key Decisions for Investors

• No broad AI de-risking here; keep index exposure neutral and wait for earnings/custodian data on enterprise AI attach rates before adding directional risk.
• If you want a relative-value expression, buy MSFT/GOOGL on pullbacks versus a basket of smaller AI narrative names; the thesis is that platform control and distribution will absorb model-quality drift better than standalone app layers over 1-3 months.
• Use AI (C3.ai) as a sentiment proxy short only on strength, with a tight stop if enterprise AI commentary reaccelerates; this is a higher-beta way to express skepticism that market-facing AI narratives are outrunning measurable utility.
• Watch DDOG and SNOW as second-order beneficiaries of increased eval/monitoring spend; consider them only if upcoming quarters show customer demand for governance, tracing, and model-routing features.
• Set a thesis alarm for the next two major enterprise software earnings cycles: any rise in human-review rates, slower AI seat expansion, or model-switching language would validate the bearish case on AI-app monetization.