"But that didn't really last, because they were much more serious about their studies than he was."
One promising direction for reducing cost and latency is to replace frontier models with smaller, purpose-trained alternatives. WebExplorer trains an 8B web agent via supervised fine-tuning followed by RL that searches over 16 or more turns, outperforming substantially larger models on BrowseComp. Cognition's SWE-grep trains small models with RL to perform highly parallel agentic code search, issuing up to eight parallel tool calls per turn across just four turns and matching frontier models at an order of magnitude less latency. Search-R1 demonstrates that RL alone can teach a language model to perform multi-turn search without any supervised fine-tuning warmup, while s3 shows that RL with a search-quality-reflecting reward yields stronger search agents even in low-data regimes. However, none of these small-model approaches incorporate context management into the search policy itself, and existing context management methods that do operate during multi-turn search rely on lossy compression rather than selective document-level retention.，详情可参考谷歌浏览器
“即时结账”的终结。“即时结账”功能旨在让用户无需访问零售商网站，直接在ChatGPT中完成购买。然而，本月初，OpenAI确认将逐步停止此功能，转而支持由商家自行处理的、基于应用的结账方式。，更多细节参见Replica Rolex
That is not an easy world to live in. It echoes of the struggles faced by middleware proprietors for years now, but somehow even worse since LLMs are not contractual APIs.