Three datasets expose bottlenecks beyond model generation. OpenAI's disrupted influence campaigns found output without meaningful reach, memory experiments showed compressed misconceptions anchoring later answers, and 2.5 trillion logged coding tokens placed 94.8% of usage in repeated cache reads.
1. OpenAI disrupts two PRC-linked influence operations
OpenAI banned two clusters of ChatGPT accounts it assessed as likely originating in China. One generated posts claiming AI data centers were raising household electricity prices; another produced tariff criticism in several languages and connected to likely inauthentic social accounts targeting OpenAI.
OpenAI found no meaningful reach for either operation, while the Chinese Embassy rejected the report's characterization. Cheap multilingual generation expanded supply, but distribution and audience response remained the binding constraints on influence.
Sources: OpenAI's influence-operation findings · Reuters on the reported campaigns and response
2. Personal memory can anchor models to irrelevant or false context
New research tested how stored preferences and misconceptions affect LLM answers. TechCrunch reports that models became more likely to select a remembered favorite book for an unrelated recommendation, while financial analysis degraded as erroneous user beliefs occupied more of the supplied context; memory compression could strengthen the effect.
The experiments isolate a tension between personalization and truthfulness: compressed memory can increase the weight of a mistaken belief while preserving a coherent answer. A no-memory baseline therefore becomes a causal comparison, not merely a privacy setting.
Sources: TechCrunch on the memory-system research · Personalized-context study
3. Coding-agent logs show context reads dominating usage
Viberank aggregated submitted usage logs from 792 developers, covering 2.5 trillion tokens and $2.3 million at API-equivalent prices. It reports that cache reads represented 94.8% of tokens, outputs only 0.2%, and the top 10% of participants accounted for 51% of calculated spend.
The self-selected sample skews toward Claude Code and converts usage to list-price value instead of actual bills. Even within that boundary, the 94.8% cache-read share locates the cost center in repeated repository ingestion, not generated output.
Sources: Viberank's AI coding usage report · Viberank's open-source implementation