🛠️ All DevTools

A Commons of Software Productive Infrastructure, by and for Capital

Human migration has surged since 2000 – these maps reveal where people are going

Show HN: Claumon – forecasting Claude Code usage limits with a Gamma process Anthropic&#x27;s usage analytics dashboard is only available to Team and Enterprise org admins. On a Pro or Max plan all you get is &#x2F;usage and the claude.ai usage page, which show where you stand right now but not where you&#x27;re heading. I looked at various open-source projects but none quite matched what I had in mind: an all-round control panel for Claude Code that&#x27;s also a single binary, with no dependencies and no install steps.<p>The other thing I cared about was forecasting the usage limits. Existing tools do burn-rate projections (ccusage) or percentile heuristics (Claude-Code-Usage-Monitor), which felt too simplistic for what I wanted - I was after a calibrated statistical model with proper credible intervals.<p>I built claumon over the last few months in Go. It runs on Linux, macOS and Windows, with a Homebrew install. It has the usual consumption gauges, cost breakdowns and conversation history, plus two tabs for memory management: after a while I had memories scattered across several projects and wanted to see them in one place and prune the stale ones.<p>In the last few weeks I focused on the forecast model. I started from an empirical-Bayes linear regression with Brownian noise, but ended up with a Gamma process for the path noise: token usage can&#x27;t decrease over time, so Brownian motion, however mathematically convenient, was the wrong choice. The intervals are calibrated against your own recorded history, and there&#x27;s tooling that scores the forecasts out-of-sample, so the coverage is checked rather than assumed.<p>I wrote a formal, versioned spec for the model, and the implementation follows it: <a href="https:&#x2F;&#x2F;github.com&#x2F;fabioconcina&#x2F;claumon&#x2F;blob&#x2F;main&#x2F;internal&#x2F;forecast&#x2F;MODEL.pdf" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;fabioconcina&#x2F;claumon&#x2F;blob&#x2F;main&#x2F;internal&#x2F;f...</a><p>Everything runs locally - nothing leaves your machine. It&#x27;s open source, MIT. I&#x27;d welcome feedback on the model especially.

Web Browsers on Video Game Consoles

Show HN: I built a Red Flag Warning zone-check tool for the East Bay in 48h Hey HN. I&#x27;m a high schooler in Fremont, CA. Tuesday morning I got a county-wide AC Alert text telling everyone in Alameda County to prepare a go-bag for an East Bay Hills Red Flag Warning that starts tonight at 11 PM. The text went to ~half a million phones.<p>The actual NWS warning polygon only covers East Bay Hills (NWS zone CAZ515). Most people who got the text don&#x27;t need a go-bag tonight. Some in the hills don&#x27;t realize how close they are. So I built this tool - <a href="https:&#x2F;&#x2F;redflag-check.info&#x2F;" rel="nofollow">https:&#x2F;&#x2F;redflag-check.info&#x2F;</a> mit licensed public github - <a href="https:&#x2F;&#x2F;github.com&#x2F;vedant-f-is-ma&#x2F;redflag-check" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;vedant-f-is-ma&#x2F;redflag-check</a><p>It does a few things - tells people if they are in the flagged zone, and also provides a way to check if a buddy is in flagged zone and send them a text. Everything without installing an app.<p>I heard back from Oakland Firesafe Council director about a gap in my understanding (and the tool). To my surprise, and through feedback, I realized that you cannot assume that only the flagged area is at risk. Adjacent areas are at risk too! Fires do not follow zone boundaries! I fixed the tool.<p>I built this in 48 hours to close that specific gap: type your address, get a yes&#x2F;no on whether the NWS polygon covers it, your Genasys evacuation zone, tonight&#x27;s wind + humidity at your point, a plain-English action checklist, a per-school decision view for East Bay districts, and a one-tap iMessage buddy-check template for a hill-neighbor at 10:30 PM.

Teenagers Stayed Overnight at Their School and Found Hidden Ancient Roman Ruins

Inverse Rubric Optimization: A testbed for agent science

Are insecure code completions in PyCharm a vulnerability?

Lattice Triangles Are Rare <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2603.23928" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2603.23928</a><p><a href="https:&#x2F;&#x2F;github.com&#x2F;AxiomMath&#x2F;lattice-triangle" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;AxiomMath&#x2F;lattice-triangle</a>

Deficient executive control in transformer attention

Show HN: Llmbuffer – Python library for cache-optimized LLM conversation history I was not getting good cache utilization when including dynamic context in agent threads. After a lot of experimentation, I found a good pattern that minimizes how often long lived conversation history gets modified while still supporting dynamic context. It has flexible hooks for doing things like truncating or summarizing tool outputs when transitioning messages to the long term history. And I&#x27;m seeing &gt;&gt;90% of tokens hitting the cache for my agents despite including a lot of dynamic user context.<p>There are a wide range of agent prompting strategies so I&#x27;d love to hear where this library works well and where there are patterns that don&#x27;t fit well into the current API!

Show HN: Magenta Real-Time Music Generation Locally on iPhone, Without the GPU Last Thursday, Deepmind released Magenta Realtime 2 , an open source music generation model. They said it could run on Mac, but not iPhone.<p>As a v̵i̵b̵e̵ ̵c̵o̵d̵i̵n̵g̵ ̵a̵d̵d̵i̵c̵t̵ agentic AI maxxi and person who has melted iPhones before (link at bottom), I took that as a personal challenge and made it my weekend project.<p>On Saturday, I got it to run for 10min straight on an iPhone 12 Pro from 2020 without melting the phone or - shockingly - touching the GPU.<p>How? I chopped the model up into 5 pieces and set them each to run on different parts of Apple&#x27;s system on a chip (SoC).<p>My past experience taught me that if you can actually leverage it, the iPhone&#x27;s NPU is incredibly powerful, and power efficient. If you&#x27;re doing sustained real-time generation for long periods of time on a device without a fan, you gotta use the neural engine or else you will melt the device.<p>See: <a href="https:&#x2F;&#x2F;accelerateordie.com&#x2F;p&#x2F;we-melted-iphones-for-science" rel="nofollow">https:&#x2F;&#x2F;accelerateordie.com&#x2F;p&#x2F;we-melted-iphones-for-science</a><p>The Apple Neural Engine has a ton of constraints, the main one being that it only accepts fixed shape inputs, and only supports some architectures -- which is why I chopped the model up into pieces.<p>But it works! And I wrote zero lines of code by hand. Back when I was running VC-backed companies, I would have needed a small team of grumpy greybeard engineers to do this and it would have taken 2-6 weeks. Now I can feed my own nerd fetish and do this stuff myself.<p>Next up: I&#x27;m building an iPhone app that ties into your heart rate, movement data, location etc to generate a real-time soundtrack to you life.<p>What a time to be alive!<p>update: Demo video of it running on my iPhone 15 Pro: <a href="https:&#x2F;&#x2F;x.com&#x2F;mattmireles&#x2F;status&#x2F;2064849702156415214?s=46" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;mattmireles&#x2F;status&#x2F;2064849702156415214?s=46</a>

A Written Language for the Cherokee So Efficient It Was Thought to Be Magic

Free financial literacy platform for kids – 90 lessons, no paywall

Show HN: Kctx – A read-only Kubernetes context engine for SREs and AI Agents

Show HN: I am building a map of people who lived in the Roman Empire Driving home from work one day, I wanted to know how many people we knew the names of who lived during the Roman era. Searching around, I found lists of Consuls and officials, but nothing that covered ordinary people or even most people like freedmen and slaves. So I ended up building a pipeline to process the more than 500k Latin inscriptions in the Epigraphic Database Clauss-Slaby <a href="https:&#x2F;&#x2F;edcs.hist.uzh.ch&#x2F;en&#x2F;" rel="nofollow">https:&#x2F;&#x2F;edcs.hist.uzh.ch&#x2F;en&#x2F;</a> and extract the names of people (and attempt to cluster them, but this is a work in progress).<p>There are databases where Classicists have done this manually for specific regions, Trismegistos <a href="https:&#x2F;&#x2F;www.trismegistos.org&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.trismegistos.org&#x2F;</a> and Latin Inscriptions of the Roman Empire (LIRE) <a href="https:&#x2F;&#x2F;pure.au.dk&#x2F;portal&#x2F;en&#x2F;publications&#x2F;latin-inscriptions-of-the-roman-empire-lire&#x2F;" rel="nofollow">https:&#x2F;&#x2F;pure.au.dk&#x2F;portal&#x2F;en&#x2F;publications&#x2F;latin-inscriptions...</a> are two major efforts I found. But there doesn&#x27;t seem to be a project that did what I set out to do, although I have read in some places that it was believed to be possible.<p>I am not a classicist or a web developer, but I have Claude and Gemini and I can sort of read basic Latin - so I set to work. I used LIRE and another database as ground truth and built a pipeline to extract and process the inscriptions to recover the names. The process I developed uses a high end LLM like Sonnet or Gemini Pro to supervise the extraction and tuning process on a regional basis until the obvious error rate is reasonable. For this, so far, reasonable to me means less than 1-2% in the smaller initial samples of 100-500 and no observed systemic issues. The different regions often need different prompts, so this basically became an exercise in letting the higher level AI tune the prompt for the lower level AI. The extraction when measured against LIRE produces an F1 score between 0.64 and 0.87, but take this with a grain of salt.<p>Once I had done a few regions, I wanted to see the work, so I threw together a pretty crude website but as I am not a web developer, it was crude in how it accessed its data. It does look cool and I also added summarization, and machine translation to each entry. I wanted to eventually get feedback from an actual team of classicists and make the website work better, so I am rewriting it as we speak but it is broadly functional now with a few extra bugs but substantially improved performance compared to the old one. All entries link back to the proper sources, and the old web app linked to several additional sources where the data was present, but I haven&#x27;t gotten that working again just yet on the new one. (The old web interface is still available at <a href="https:&#x2F;&#x2F;roman-names.com" rel="nofollow">https:&#x2F;&#x2F;roman-names.com</a>, but I will warn you it is clunky and not mobile friendly at all)<p>Key findings so far:<p>AI supervised AI extraction saved me time. I was manually tuning things for a while and then the runbook became an idea that I feed my instructions in and let the big AI go with sparse oversight from me.<p>The extraction improved significantly (by about 10 F1 points) when I fed the model the raw text including the markers, vs a cleaned up version of the text.<p>I just thought it was a cool little project and wanted to share. If you happen to work in any adjacent space and there is something I could do better etc let me know.

Farmer donates land for a park, city sells it for $10M as data center land

The Dynamo and the Computer: The Modern Productivity Paradox (1989) [pdf] <a href="https:&#x2F;&#x2F;gwern.net&#x2F;doc&#x2F;economics&#x2F;automation&#x2F;1989-david.pdf" rel="nofollow">https:&#x2F;&#x2F;gwern.net&#x2F;doc&#x2F;economics&#x2F;automation&#x2F;1989-david.pdf</a>

I Hacked into the Worst E-Bike and Fixed It [video]

How JPL keeps the 13-year-old Curiosity rover doing science