- Author
- Posted by emersonmacro, with replies from many commenters
- Published
- 1 October 2026
- Kind
- Hacker News discussion thread on an arXiv paper
- Original
- news.ycombinator.com/item?id=49922437
- Subject
- Harnesses and agent loops
What it is
The linked paper is titled "Context Language Models". The thread has 178 points and was posted by emersonmacro. Replies cover cache costs, context compaction and comparisons with related approaches.
What it says
- Commenter plastic-enjoyer quotes the paper: context is a file the model can freely update.
- Commenter Bolwin says the paper kept invalid cache suffixes and performance did not suffer.
- Commenter _jayhack_ warns that frequent prefix edits cut the cache hit rate sharply.
- Commenter dist-epoch says a stale cache will not crash but can surprise the model.
- Commenter nsingh2 says Codex is moving toward unreleased model-kept notes instead of summary compaction.
- Commenter bob1029 has not seen recursive harnesses work well past depth 1, the paper's limit.
What evidence it gives
The thread has one quoted passage from the paper and much reasoning from commenters. Some commenters describe their own handover-note workflows. No commenter reports reproducing the paper's results.
Interests and limits
The linked paper was not read. Its claims reach this page only through commenters. Commenters are self-selected. Several speculate about how the method works. One comment is flagged and two dismiss the paper without argument.