Drew Breunig 点破了过去三年的隐性补贴：工程上的糙活由下一代模型免费兜底，所以打磨 harness 和上下文策略显得像过度优化。这个循环一断，精细化工程第一次有了明确回报期——它从可选项变回了竞争壁垒。

在 LLM 时代，这是对“上下文窗口”重要性最精辟的注解。Agent 不具备人类的隐性知识和环境感知能力，因此显式的上下文（如 context.json）成为了其行动的基石。这提醒我们，在设计 AI 辅助系统时，构建高质量的上下文生成机制往往比优化模型本身更为关键。

相关工作「上下文拖拽」（Contextual Drag）的存在，说明这个研究方向正在快速形成：不只是「无关上下文缩短推理」，还有「错误上下文拖拽推理方向」。两篇论文合在一起暗示了一个新的研究领域：「上下文污染对推理模型的系统性影响」。对 AI Agent 的工程实践者而言，这意味着上下文管理策略（截断、摘要、过滤）将成为保障推理质量的核心工程能力，而非仅仅是 token 节省手段。

Wordcraft Writers Workshop by Andy Coenen - PAIR, Daphne Ippolito - Brain Research Ann Yuan - PAIR, Sehmon Burnam - Magenta

cross reference: ChatGPT

Many people treat their blogging practice as an experimental thought space. They try out new ideas, explore a small space, attempt to come to understanding, connect new ideas to their existing ideas.

Ton Zylstra coins/uses the phrase "metablogging" to think about his blogging practice as an evolving thought space.

How can we better distill down these sorts of longer ideas and use them to create more collisions between ideas to create new an innovative ideas? What forms might this take?

The personal zettelkasten is a more concentrated form of this and blogging is certainly within the space as are the somewhat more nascent digital gardens. What would some intermediary "idea crucible" between these forms look like in public that has a simple but compelling interface. How much storytelling and contextualization is needed or not needed to make such points?

Is there a better space for progressive summarization here so that an idea can be more fully laid out and explored? Then once the actual structure is built, the scaffolding can be pulled down and only the idea remains.

Reminiscences of scaffolding can be helpful for creating context.

Consider the pyramids of Giza and the need to reverse engineer how they were built. Once the scaffolding has been taken down and history forgets the methods, it's not always obvious what the original context for objects were, how they were made, what they were used for. Progressive summarization may potentially fall prey to these effects as well.

How might we create a "contextual medium" which is more permanently attached to ideas or objects to help prevent context collapse?

How would this be applied in reverse to better understand sites like Stonehenge or the hundreds of other stone circles, wood circles, and standing stones we see throughout history.

Oded Shenkar's three step process for effective problem solving using imitation: - Step 1. Specify your problem and identify an analogous problem that has been successfully solved. - Step 2. Analyze why the solution was successful. - Step 3. Identify how your problem and circumstances differ from the example problem and figure out how to best and most appropriately adapt the original solution to the new context.

The last step may be the most difficult.

The IndieWeb broadly uses the idea of imitation to work on and solve a variety of different web design problems. By focusing on imitation they dramatically decrease the work and effort involved in building a website. The work involved in creating new innovative solutions even in their space has been much harder, but there, they imitate others in breaking the problems down into the smallest constituent parts and getting things working there.

Link this to the idea of "leading by example".

Link to "reinventing the wheel" -- the difficulty of innovation can be more clearly seen in the process of people reinventing the wheel for themselves when they might have simply imitated a more refined idea. Searching the state space of potential solutions can be an arduous task.

Link to "paving cow paths", which is a part of formalizing or crystalizing pre-tested solutions.

Hoffman, R., Mueller, S., Klein, G., & Litman, J. (2021). Measuring Trust in the XAI Context. PsyArXiv. https://doi.org/10.31234/osf.io/e3kv9

This is a valid concern that needs to be addressed. While some engineers certainly do use IRB for their projects, it is not nearly as common as it should be.