🛠️ Hacker News Tools

[Other] Show HN: Bonsai 1.7B ternary model at 442T/s on M4 Max We took a recently released Bonsai 1.7B ternary model from PrismML (<a href="https:&#x2F;&#x2F;github.com&#x2F;PrismML-Eng&#x2F;Bonsai-demo" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;PrismML-Eng&#x2F;Bonsai-demo</a>) and ran our agentic evolution search on it for 6 hours to optimize the Metal kernels. The search was fully autonomous.<p>Measured against unmodified upstream llama.cpp at the same Bonsai&#x2F;Q2_0 commit, same M4 Max:<p>- tg128: 309.82 → 442.42 t&#x2F;s (+42.0%)<p>- pp512: 4250.32 → 4622.63 t&#x2F;s (+8.8%)

[Other] Show HN: Replacing spec-driven development with just facts I had a lot of issues with spec-driven approaches, agents are too readily producing fluff, large projects have so many specs agents start making mistakes maintaining them. There&#x27;s a constant consistency tax.<p>In the end every spec is just a bunch of facts, so I decided to leave that and throw away everything else while making it friendlier for agentic use.<p>Introducing facts - skills and CLI for agents to use facts-driven development. <a href="https:&#x2F;&#x2F;github.com&#x2F;av&#x2F;facts" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;av&#x2F;facts</a>

[Other] PyInfra 3.8.0 Is Out

[Monitoring/Observability] Setting up server monitoring for a Rails app on Hatchbox

[Other] DeepClaude – Claude Code agent loop with DeepSeek V4 Pro, 17x cheaper

[Other] How Kepler built verifiable AI for financial services with Claude

[Other] Show HN: Ableton Live MCP Ever wanted to control Ableton with just your voice? Me too! I made this MCP server so I could just ask Codex to do anything in Ableton Live for me, while I was nap-trapped by my baby.<p>The chat messages I sent to Codex to make this:<p><i>in ableton, make a self reflective song, with audio vocals (via macos say) and chip tunes and 80&#x27;s drum machines. should be a real edm banger<p>i want midi for everything but vocals please, with ableton devices. not prerendered audio for instruments<p>needs some fills<p>and should hit way harder after &quot;3-2-1 i become the sound&quot;<p>the vocals are squished too much (read too quickly), give them a little more length<p>add some dynamics, the song is basically one volume. and some pumping side chain<p>improve dynamics of the clap, seems a bit flat and indistinguished, want it harder after the 3-2-1 drop<p>introduce a new element on a new track after the 3-2-1 drop, that comes in but then recedes before the final exit<p>doesn&#x27;t seem like the new thing has any notes<p>the element is a bit muddy&#x2F;indistinct. perhaps it needs simplification and more space, different instrument choice, i dunno</i>

[Other] I recreated the Apple Lisa computer inside an FPGA [video]

[Other] Collaborative Editing in CodeMirror (2020)

[Other] Metal Gear Solid 2's source code has been leaked on 4chan

[Other] Show HN: Software Engineer to Novelist: Writing a Book Like Coding I just published my first book, Means and Motive. ( <a href="https:&#x2F;&#x2F;www.amazon.com&#x2F;dp&#x2F;B0GYCZJVGX" rel="nofollow">https:&#x2F;&#x2F;www.amazon.com&#x2F;dp&#x2F;B0GYCZJVGX</a> )<p>As a software engineer, I approached writing like a software project. I used familiar tools (Emacs and HTML) for the primary writing.<p>I built my own tool (EPublish) to transform the HTML manuscript into an .epub file, the source for the ebook version. And I wrote shell scripts to reliably and repeatably transform the .epub version into PDF files for the printed editions.<p>I wrote &#x27;design&#x27; and &#x27;architecture&#x27; docs, describing the world, key actors, and timelines. I kept a task list of chapters and key scenes that needed to be written, in priority order. Along the way, I kept my files version-controlled so I could see the progress of the novel and edit mercilessly, without worrying about keeping old text around in backup files should I want it back for some reason.<p>If you&#x27;ve thought about writing a book, I highly recommend it. There are many similarities to the software engineering process. You&#x27;ll also gain a newfound appreciation of the design, layout, and typesetting world, exactly how much work goes into each book you read.

[Testing] Testing macOS on the Apple Network Server 2.0 ROMs

[Other] Wiki Builder: Skill to Build LLM Knowledge Bases

[Monitoring/Observability] Alert-driven monitoring

[Other] Show HN: I built a tool that helps predict HN front page success Hey HN community,<p>I built a tool that helps optimize your post for hitting the first page of Show HN.<p>How it works: I used a Hugging Face dataset of all Hacker News posts from the past 3 years and trained a model that predicts how successful your post might be. There&#x27;s still a lot of randomness on HN, so nothing is guaranteed, but the tool helps optimize your post for higher odds.<p>A couple of interesting findings:<p>- GitHub repo links work x3 better than regular domains - Open-source tools have a steady virality rate (13.9% - one of the highest) - &quot;I built&quot; outperforms &quot;We built&quot; - Using parentheses and mentioning technologies (Lua, Postgres, Rust, etc.) helps a ton.<p>You can try the tool at wannalaunch.com or read the blog posts for more insights from the analysis. The model is also available as open source if you want to retrain it or look under the hood.<p>Happy to hear the feedback!

[Other] Show HN: I built a RISC-V emulator that runs DOOM Demo: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=f5uygzEmdLw" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=f5uygzEmdLw</a><p>Hi HN,<p>I built a RISC-V emulator that implements the RV32IM instruction set and a minimal syscall interface to run DOOM. A few weeks ago, I got my first output with a simple hello world assembly program.<p>Since then I have been working tirelessly to get DOOM to run.<p>I needed to figure out how to run C programs first, and came across newlib, which allows the underlying environment to implement the syscall stubs one by one until the programs run.<p>I have also added ELF loading, but currently only a single `PT_LOAD` segment is supported.<p>To port DOOM, I used doomgeneric, which was quite convenient to get working once the required stubs were in place.<p>DOOM renders to a fixed area in memory (0x705FDD = VRAM_START):<p><pre><code> 0x7FFFFF +-------------------------------------+ | | | QUEUE_SIZE (32 bytes) | | | 0x7FFFDF +-------------------------------------+ &lt;-- QUEUE_START 0x7FFFDE | QUEUE_READ_IDX | 0x7FFFDD | QUEUE_WRITE_IDX | +-------------------------------------+ | | | | | VRAM (1,024,000 bytes) | | | | | 0x705FDD +-------------------------------------+ &lt;-- STACK_START | Stack | | | | | v | | | | ^ | | | | | Program data + Heap | | | 0x000000 +-------------------------------------+ </code></pre> I made a small linker script so that the entry point of a C program is at _start and virtual address is always 0. That kept the ELF loader code simple.<p>Inputs are written to the queue by rvcore which are then intercepted by DOOM running inside it.

[Other] Show HN: Apple's SHARP running in the browser via ONNX runtime web Hi HN, author here. SHARP is Apple&#x27;s recent single-image 3D Gaussian splatting model (<a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2512.10685" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2512.10685</a>). Their reference code is PyTorch + a pretty heavy pipeline; I wanted to see if it could run in a browser with no server hop, so I exported the predictor to ONNX and ran it via onnxruntime-web with the WebGPU EP.<p>What works: drop in an image, get a .ply you can download or preview live, all on your machine — your image never leaves the tab. The model is large (~2.4 GB sidecar) so first load is slow on a cold cache, but inference itself is a few seconds on a recent Mac.<p>Caveats: SHARP&#x27;s released weights are research-use only (Apple&#x27;s model license, not the code&#x27;s). I host the exported ONNX on R2 so thedemo &quot;just works&quot;, but you can also export your own from the upstream Apple repo and upload locally.<p>Happy to talk about it in the comments :)

[Other] Kimi K2.6 just beat Claude, GPT-5.5, and Gemini in a coding challenge

[DevOps] Show HN: I'm running parallel Pi agents on a local sandbox I&#x27;ve been running Pi using SmolVM to build SmolVM!<p>SmolVM provides an abstraction over microVMs to easily create sandboxes for coding agents, OpenClaw, or just to build a custom harness.<p>To use it, install using: curl -sSL <a href="https:&#x2F;&#x2F;celesto.ai&#x2F;install.sh" rel="nofollow">https:&#x2F;&#x2F;celesto.ai&#x2F;install.sh</a> | bash<p>and then run: smolvm pi start

[Other] Windows API is Successful Cross-Platform API (2024)