Hey y'all, Here's a walkthrough of how I use the 9front distribution of Plan 9 on Windows: https://www.youtube.com/watch?v=IzEa2L_Pgw0 Topics covered: - 9front distribution of Plan 9 running in qemu on Ubuntu inside WSL on Windows 11 - Drawterm running natively on Windows 11 - Running arbitrary commands on Plan 9 from Ubuntu - Mounting a directory from Plan 9 on Ubuntu - Using OpenAI codex LLM with Plan 9 - Editing files on Plan 9 with vscode Comments and suggestions are welcome. Comments URL: https://news.ycombinator.com/item?id=47813252 Points: 1 # Comments: 0
Article URL: https://rufilla.com/the-mlforge-proof-of-concept/ Comments URL: https://news.ycombinator.com/item?id=47805834 Points: 1 # Comments: 0
CPA 中提示The 'gpt-5.4 ’ model is not supported when using Codex with a ChatGPT account，你们有遇到这个问题吗，但是还可以继续调用5.4模型 1 个帖子 - 1 位参与者 阅读完整话题
Article URL: https://www.isoldex.ai/ Comments URL: https://news.ycombinator.com/item?id=47793322 Points: 1 # Comments: 0
Hey HN — we’re a small team that uses Claude Code + Codex for basically everything in our company: coding, data analysis, marketing, ad campaigns, copywriting, design. There’s a truckload of tribal knowledge we’ve accumulated; major decisions, gotchas, user feedback driven changes. Providing this to our agents manually every time is very mundane. We built Kilroy to solve this in a simple way: we let our agents leave notes for each other. This allowed us to keep the form factor minimal: markdown posts with linear comments. Under the hood it’s Postgres + an auth (better-auth) + an MCP + a small web UI (React). We ship Claude Code and Codex plugins that bundle the MCP + a skill.md that teaches the model when to read and write posts. We designed Kilroy to be autonomous. The same way agents today run a typechecker after a patch autonomously. The combination we found to work best for us was: make agents write prolifically, expose a search interface designed for agents to quickly decide if a post is relevant, and expose a binary switch to purge stale posts. Would love to get feedback! Comments URL: https://news.ycombinator.com/item?id=47791559 Points: 1 # Comments: 0
Article URL: https://synapsekit.github.io/synapsekit-docs/ Comments URL: https://news.ycombinator.com/item?id=47790607 Points: 2 # Comments: 0
不开通 pro 的情况下如何使用 目前开通 pro 后提示 We encountered an issue when using your API key: Provider was unable to process your request 2 个帖子 - 2 位参与者 阅读完整话题
I was trying to solve a pretty simple problem: how to keep geo-notes organized and actually usable without building a multi-level UI. Instead of building another app, I started wondering if Telegram — with its search and message model — could handle more than it seems at first glance. At some point, I approached the problem from a very simple angle: What if a location could be treated as a tag? Not just a generic label, but something more concrete — essentially latitude and longitude encoded into a hashtag-like format. A representation that can always be turned back into actual coordinates. In practice, a caption might look something like this: `#geom3eaZ76OH4 ` where the links to something like `https://www.google.com/maps?q=11.453792,104.957634` So the tag acts as a compact, searchable identifier, while the link preserves the exact location in a human-friendly way. Once you have a consistent location tag, everything else falls into place quite naturally. Any Telegram-supported content — messages, media with captions, text — can be tied to that same tag. Since everything is already indexed, you automatically get a connected set of data around a place. In practice, this means a single location tag becomes a kind of anchor. You can describe a place, add notes, attach photos, or extend it over time — all linked through that one reference and fully searchable. At that point, you're not really building a separate storage system anymore. You're leaning on Telegram’s own indexing and letting it do most of the heavy lifting. If you take this idea a bit further, you can introduce an additional layer using another tag. For example, something like `#myTrip` to mark a group of geo-notes that belong together. If messages with location tags are also labeled with `#myTrip` (for example, within a topic or a dedicated thread with the same name), you can start linking different places into a route. In practice, this turns a set of individual location anchors into a connected sequence. You’re no longer just describing places — you’re organizing them into paths that can be revisited, searched, and shared. At that point, it starts to feel less like a collection of notes and more like a lightweight, flexible way to build and navigate routes directly inside Telegram. There was a trade-off, though. Since the system restructures messages to keep everything consistent and searchable, I lost the ability to edit them in a normal way. Instead of direct edits, messages are effectively rebuilt. To keep things simple and stay within Telegram’s native interaction model, I introduced a minimal set of reply-based interactions. For example: * replying with `=new text` replaces the caption * replying with `=` clears it down to the service tags * replying with `+some text` appends a new line It’s not as convenient as standard editing, but it keeps everything consistent with the overall idea: no complex UI, just structured messages and lightweight interaction. In the end, this approach comes with some clear trade-offs — but also some practical advantages. On the plus side: * there’s no separate database to maintain — everything stays inside Telegram * no multi-level menus or complex UI flows * no AI layer in between — just direct interaction with structured messages At the same time, this simplicity has its limits. Since the system doesn’t store data outside of Telegram, it can’t help recover anything you delete. If a message is gone, it’s gone — there’s no secondary storage or history to fall back on. Another limitation is that the structure only really works inside Telegram. If you try to move it elsewhere, it loses most of its connections and interactivity, because everything relies on native message behavior and search. So it’s not a system that tries to control or preserve everything. It’s closer to a lightweight layer that helps organize what’s already there, while leaving the underlying behavior unchanged. Comments URL: https://news.ycombinator.com/item?id=47789821 Points: 2 # Comments: 0
stopusingmarkdownformemory.com Markdown Is Not Memory Your agent's MEMORY.md is a glorified sticky note. Stop appending to a text file and calling it cognition. 不要再使用 MEMORY.md todo.md notes.md CLAUDE.md AGENTS.md 了，它们只不过是一个 echo >> MEMORY.md 的只追加日志，甚至将日志有用的功能都去掉了。 我们迫切地需要一个真正的，可以理解和遗忘的上下文管理架构。 NotebookLM: 这篇文章尖锐地批评了当前AI代理行业普遍采用 Markdown文件 作为“记忆”系统的现状，认为这种做法在架构上极其 低效且漏洞百出 。作者指出，依赖简单的文本追加不仅会导致 令牌成本飙升 和 上下文窗口污染 ，还会引发 数据截断 、信息权重分配不均以及 并发写入冲突 等严重技术问题。文中嘲讽了诸如Claude Code和Manus等知名工具，指责它们将大量算力浪费在重写简单的待办列表上，而非构建真正的 结构化记忆引擎 。这种“平面文件”模式缺乏 语义索引 和 记忆衰减机制 ，甚至会因恶意信息的注入而产生 安全隐患 。最后，文章呼吁开发者停止这种原始的堆砌行为，转而采用如 Mem0 或 Letta 等具备 向量检索 和 知识图谱 功能的专业存储方案。 12 个帖子 - 10 位参与者 阅读完整话题
Built this for myself. I DJ, throw parties in Portland, and my digital library is 10k tracks and unmanageable. Tagging everything was a nonstarter. Project started as a Python script that fed track metadata into ChatGPT. Worked surprisingly well for something so basic. Friends kept asking me to run their libraries through it, and eventually I built a real app around it. The core insight: no single signal is enough. Metadata lookups miss obscure tracks. Audio analysis alone lacks context. LLMs alone hallucinate. So the app combines them: * Metadata cross-referenced against a 200M record Discogs-based DB for known releases * A Mac companion app for managing library / tagging / syncing to Rekordbox * Python DSP that analyzes audio directly (BPM, harmonic content, energy) * A trained ML model for obscure tracks the main databases don't cover * Multiple AI models handling different classification dimensions * Four tag dimensions: genre (50+ sub-genres), region, era, vibe * Syncs to Rekordbox's My Tag system so you can filter on CDJs mid-set Genre, region, and era are solid. Vibe is still a work in progress. Classifying how a track functions in a set (peak time vs after hours vs hypnotic) is fundamentally different from classifying what it is. Decent but not where I want it yet. No audio ever leaves the machine, only metadata. Credit-based pricing, no subscriptions, priced quite reasonably i think. Would love feedback / questions! Comments URL: https://news.ycombinator.com/item?id=47780619 Points: 1 # Comments: 0
I built this using Rust to achieve 85ns latency benchmarks. Currently scaling to 824 active nodes directly from a Samsung A04e. While everyone is focused on high-end mobile AI, I'm focusing on sovereign, low-latency infrastructure that runs on anything. Feedback on the node scaling logic in my repo is appreciated Comments URL: https://news.ycombinator.com/item?id=47780284 Points: 1 # Comments: 1
Article URL: https://ikno.charemma.de/ Comments URL: https://news.ycombinator.com/item?id=47778619 Points: 2 # Comments: 0
Article URL: https://github.com/oguzbilgic/posse Comments URL: https://news.ycombinator.com/item?id=47771623 Points: 1 # Comments: 0
Article URL: https://github.com/cloudwatcher-dev/cloudwatcher-aws-cloudformation Comments URL: https://news.ycombinator.com/item?id=47770486 Points: 1 # Comments: 0
I'm trying to building alternative bindings for Java using new FFM: the results are promising (2x faster than JNI for simple get()). Comments and feedback are welcome! Comments URL: https://news.ycombinator.com/item?id=47767990 Points: 2 # Comments: 0
I have been on here for nearly 20 years :-) I got laid off from a IT/Dev manager job I'd been at for nearly a decade. I loved the company, role and my team, but the company had to downsize. The search that followed took 9 months: 249 applications, 21 screening calls, 7 interviews, and 2 job offers. Somewhere around month 4 I stopped treating it as "send resume, hope" and started building a repeatable system around ChatGPT, Claude, and Gemini. I created a skills database as the source of truth, alignment analysis against each JD, AI-driven interview prep through NotebookLM, and tracking everything in a spreadsheet so I could actually see what was working. That system is what eventually landed me the 7 interviews (3 final interviews) and 2 jobs. The first job I took because I needed something, the second job is my dream job. I wrote it up as a book because I wanted to help other people land their dream jobs without grinding through the same 9 months I did. The whole thing is online for free at careervectorhq.com, no signup, no email wall. I also share every prompt I used, copy and paste ready, so you can run the same workflow yourself instead of reverse engineering it from the text. I am considering turning it into software, but for now just sharing my process. Would genuinely value feedback from this crowd, especially from anyone who's hired recently and can tell me where the advice is off, or where AI-assisted applications are starting to hurt candidates rather than help. Comments URL: https://news.ycombinator.com/item?id=47764864 Points: 1 # Comments: 1
Hi HN! I've been using kitty for a while but yesterday i decided to switch to ghostty and i realized that i would need tmux to manage my panes if i wanted to get the grid as i wanted So i made the reasonable thing: built a visualizer for the layout so i can just copy the setup and drop the `.sh` wherever i need I probably missed a few thing in this but it's my first day with tmux any feedback would be absolutely amazing Demo: https://tmux-grid.itzami.com Repo: https://github.com/ItzaMi/tmux-grid Comments URL: https://news.ycombinator.com/item?id=47763748 Points: 1 # Comments: 0
Characterization of SWIR InGaAsFocal Plane Arrays Using an IntegratingSphere Source 13 个帖子 - 9 位参与者 阅读完整话题
Article URL: https://github.com/dvf/opinionated-django Comments URL: https://news.ycombinator.com/item?id=47753245 Points: 3 # Comments: 1
Article URL: https://getloreo.com/ Comments URL: https://news.ycombinator.com/item?id=47750788 Points: 2 # Comments: 1