Latest

Compacting aggressively changes the cost calculus for long-running agents: short state summaries can outperform full transcripts on both accuracy and spend, since late-token generation runs up to 7x pricier. Builders running multi-step coding or tool-use agents can test end-of-conversation compaction before paying for million-token runs. The gains were measured on Claude Opus 4.5, so teams using other frontier models should verify the effect on their own stacks.

elvis@omarsar011 sourcesSource

The talk frames late 2025 model releases as the moment coding agents crossed from error-prone to dependable, a shift with practical consequences for how teams staff and structure engineering work. Practitioners evaluating agent workflows can use the annotated slides as a curated timeline of which models and tools mattered. Willison's pelican-bicycle test also offers a lightweight way to compare model capabilities hands-on.

FAB measures whether models can combine images with instruction manuals, a skill that transfers to appliance repair and equipment inspection rather than simple document recall. Engineers building visual inspection or step-by-step guidance tools can use the 80% vs 28% gap as a marker of how quickly multimodal spatial reasoning has matured. Chinese models Alibaba Qwen and Google Gemini were also noted for characteristic error patterns in the study.

斯坦福发了个AutoMem，给智能体装上可训练的记忆模块，让32B模型打平Claude Opus 4.5。省去调任务动作，光改记忆就提升2-4倍，搞长程任务的别错过。

elvis@omarsar02 sourcesSource

Simon 用五分钟讲清了 LLM 过去半年的关键转折——编码代理从玩具变成生产力工具，做 AI 开发或重度使用编程助手的团队值得花五分钟了解这个趋势，看完会对模型选择和工具策略有更清晰的判断。