How to Disable the AI Features in Gmail and Google DocsNew AI toolbars and prompts are showing up in Google Docs and Gmail. If you don’t want Gemini’s help in writing documents and emails, here’s how to turn that stuff off.
Show HN: Alyph, a manual transmission for LLM contextSince I started using LLMs, I've always had this feeling of wanting to go back, change the history, and then go down another branch.The lightbulb moment for me was when Google AI Studio started to let you actually edit the AI's responses. I realized, YES, I need a multiverse type of situation to branch all my thought streams and variants of the same thread. I wanted to try Gemini and ChatGPT in parallel on the same thought.# IdeaThe original concept for Alyph goes back to 2024. I had some early concepts that were a lot more sci-fi (if you look at the codebase, you’ll see names like Stargate, Hermes, etc.).I even thought about making it 3D, like a holo deck, where you visit "planets" of topics and content. But the execution now is much more, eh, pragmatic.# Alyph todayAlyph is a 2D canvas where you can ask LLMs stuff in parallel. When you go down a rabbit hole and finally get an answer (be it code or text), you can pull that answer up and delete all the unnecessary back-and-forth in bet
Show HN: Llmem – Local persistent memory for AI coding, no embeddingsHi folks. This part of AI coding always felt really dumb: after working on many projects it repeats the same mistakes, uses awkward patterns or looks up stuff over and over. It just does not remember. Skills exists, but I don't like maintaining lots of markdown files. There were other solutions for this and I'm sure there are even more now, but I simply did not fancy those for one reason or another.So I wrote my own good-enough persistent memory which easily integrates with all AI coding tools I use. Lightweight, local, 1 Go binary + 1 SQLite file. Speaks MCP + good old REST API I can curl.After some testing I settled on BM25 lexical search with optional WordNet synonym expansions - no embeddings, no vector DB, no external API calls. A downside is English memories only, but that's OK for me.Memory_context pulls priority-tagged memories at session start, auto-consolidation merges near-duplicates to prevent bloat. Later I added optional memory types, project scopes and hashtags.I've been
Show HN: Sidebar – a private ESP32 intercom for five kids, no accountsHello all - Inspired by many who build/assemble for their daughters/sons, I took cue and assembled a two-way communication between my son and his friends. Nothing spectacular. Just basic stuff to talk.He is eleven and doesn't have a phone. I wanted him to talk to his friends without an account, a feed, or an algorithm in between, so I built five ESP32-S3 touchscreen units, one per kid. Tap a friend's name, their unit rings. Tap Everyone, they all ring.Each unit dials outbound to a relay I run, so no house opens a port. Audio is uncompressed 16 kHz PCM in 20 ms frames, AES-256-GCM with a fresh key per call, over UDP with an encrypted-WebSocket fallback. To be precise: encrypted to my relay, not end-to-end. No rooms, no recordings, no free-text contacts; the roster lives in a server-side allow-list.I'm a PM, engineer by degree but not by profession - let alone a firmware engineer. I assembled this with Claude as the engineering partner, which is why the write-up includes what went wrong:
Show HN: Capy – A Git-style platform for managing your team's secretsHello HN!We're all spending more and more time making stuff with agents, but I've noticed that one of the things that is the most disconnected from my daily engineering workflow is dealing with secrets and credentials. It often involves a lot of click-ops, copying/pasting, and collaboration, and none of the secrets management products out there have truly scratched that itch.I built Capy to solve this. It's a secrets manager whose entire frontend is a developer CLI, and I find it extremely ergonomic to use by hand. You also don't need to leave the CLI to sign up for and use it! You can install and authenticate without leaving your terminal (or agent) session.It also has very powerful version management with git-like branching and conflict resolution. You can push a version manifest to source control and have collaborators pull a specific version of the secrets at any point in time.The platform itself is extremely secure. It encrypts your local .env files so they can't be read directly.
Show HN: Webklip – ephemeral online clipboard with live sync and E2E encryptionI've build an online clipboard on sterorids.Or rather, a shareable temporary workspace where I can paste text and files, setup an expire timer, or configure read&burn. For really private stuff use the E2E encryption, which encrypts the text in the browser before persisting in the DB (files don't get encrypted). It also exposes an API that one can use to integrate it in automatic workflows.Too often I've found myself using WhatsApp or Slack as a makeshift clipboard to share bits of data with myself (or colleagues): configs, notes, passwords, sometimes files. It's a quick & dirty solution and it does the job but over time my chats become cluttered with things I no longer need.There are also times when I need to access that data from devices where I can't use WhatsApp or Slack. For those situations, I've relied on online clipboard services like cl1p.net or pastebin but I kept running into limitations that got in the way.Another use case is automation. Sometimes I need to make arbitrary da
Ask HN: How do you correct spatial reasoning of LLMs?I'm working on a complex multi-part 3D-printed product, and I needed some engineering input from LLMs. I use highest Gemini Pro reasoning levels, Fable, and K3 Max with everything tuned to highest.The problem I'm facing is mounting a sensor on a barbell sleeve. The sensor is a semi-hollow 100mm cylinder, 60mm diameter, with a 50mm diameter 50mm deep tube cutout for the sleeve. (It's a bit more complex than that.)I'm trying to brainstorm different clamping mechanisms, and evaluate durability and manufacturability of several designs.No matter what I try, the output of all LLMs is absolutely wild, especially if you ask to make a technical drawing. Anthropic reliably been the worst, and Gemini, surprisingly, the best.But none understand what goes where in any detail, besides spewing paragraph and paragraph of wild ideas including rotational cam clamps over a recessed ring TPU collet (apparently, a "brilliant pivot") and stuff like this.Is this the "unconquered frontier"? Do I need to blow
Ask HN: Dear Anthropic, can we please have thought traces back?Dear Anthropic, can we please have thought traces back?I can't verify whether or not the LLM is arriving at the conclusion from cheating, or if it's fudging or making stuff up.Opus 4.6 remains the best model because of this.
Show HN: Blender for AI AgentsCreating 3D is hard. LLMs seem to be getting better at tool use and spatial understanding. While MCPs have proved to be a good way to use these tools- the current methods have these challenges:- Access to scene graph and core C modules of Blender- Lack of parallelism, only way is to run blender headless- Lack of deterministic and fast verification layer- Inference stack- only way to use inference is to hook another MCPWe're building Mixar, think Cursor for 3D. One access point to all generative inference, an agent to build scenes/blockouts, do boring stuff like UVs and export standard formats(.glb/gltf/obj/gbx/usd etc.)Try it: https://mixar.app - first week is free. Would love to get your feedback.
Show HN: Jotchi, a smart scratchpad for the busy brainI love the idea of projects like Obsidian, but since I already have difficulty dealing with one brain, the idea of managing a second has never really worked for me.Since I'm always working on a bunch of stuff, I never really know if a particular piece of information will be important later or not. Ideally, I wanted a place that I could just dump notes into, and let the notepad itself organize everything in a clear and fast to navigate way.Ultimately, that lead to Jotchi, which became a sort of cell-based scratchpad. It's keyboard navigable, supports inline arithmetic (with units and re-calculation on edits,) connects to existing cloud storage providers, has Markdown support, and more.While there is an API and MCP server, Jotchi itself uses various heuristics to attempt to automatically organize your information. For example, typing similar notes in quick succession will automatically group them together, or typing a hex code will display it as a color. This happens way faster than an L
Show HN: Zizq is a fast, single-binary job queue that fits into any stackI’ve been quietly working on this in my free time since around March and while there is still a lot to do, I’m feeling really happy with where things are at and how well it all works. Zizq is a single binary, zero-dependency job queue that sits in any stack and can enqueue and process jobs across programming language boundaries. It’s designed to be really easy to get up and running with, and it’s designed to make it easy to manage jobs that have already been enqueued.I’ve been having a lot of fun thinking hard, designing and implementing this, and it solves some real problems I’ve come across building software at non-trivial scale. My favourite feature to this point has probably been batched jobs, where successive enqueues of jobs of the same type are folded/merged into one another to be processed as a larger batch when eventually picked up by a worker.As of now, Zizq can already:* Do all the basic FIFO / backoff + retry / concurrent worker stuff you just expect * Efficiently schedule
Show HN: A little physical breakout cloneIdea: How about a 5 minute game I can play in my browser and it's actually fun to play.Controls: Mouse or keyboard (left/right, left alt/right alt or A/D), space and esc are pause toggle. Page up / Page Down are next or previous level (if already unlocked).Cheats: Z toggles zapper, M toggles magnets.Debug stuff: I, P, E keys are toggles for various stuff (E lets you change some aspects of physics engine).Game is maybe playable on mobile devices, there is a menu on top right, and then click the circle to select type of control.Physical engine is decoupled from refresh rate and runs at 240 Hz.Fully vibe coded (deepseek, chatgpt), levels are human made with some amount of love.
Show HN: Local text, image, video, music and 3D from one CLI, no PythonHi HN! I'm the author of mere.run a local first inference runtime built around an installable CLI. I believe that whenever possible we should use the stuff we already own (like our Mac laptops, decent machines gathering dust, our gaming PC) and the limited electrical power we have easy access to, like the socket in the wall next to most of us. We shouldn't have to send our data to the cloud hoping some T&C will prevent it from being used in a way that we'd regret. Most of the local AI solutions are technical, involved, and land a curious body in some package hell. People are optimizing for one system and not another, the fun stuff is on PC if you own a Mac, and on Mac if you own a PC.So mere.run was my choice to begin to patch many of those things that I saw as problematic. Text, chat, code, image gen, speech tts+asr, vision (caption/ground/segment/track/pose/depth/face/OCR/etc), music, sfx, video, 2d->3d, persistent worlds, lora training plus a few more things I am probably forgetting
Show HN: Nurb – Agentic CAD for 3D printingI've had a love/hate relationship with Fusion over the years. I can model stuff there but it's just always slow and painful and Fusion has a deep desire to crash on me.I started dabbling with the Fusion MCP a couple of months ago and found that to be a much more pleasant way to interact with it. But then I realized I was just using Fusion as a rendering engine, which was terribly overkill.So I built Nurb.You just describe what you want ("Make an adapter that connects my shop vac hose to the dust port on my table saw") and it asks for clarification, does web searches for part measurements/details and then generates an interactive viewer as well as optimized STL files to print.The whole system hangs on the convention that a part is a function, and its keyword defaults are its parameters. That concept drives the CLI, the viewer’s sliders, the tests, and the agent’s interface. There’s no schema, no project file, no init step. A project is any directory with a parts/ folder.Under the hood i
Show HN: I built an accountability companion that lives in your textsProblem:You have a million things going on and stuff slips. Passport renewal, car registration, kid's doc appt, why you started going to the gym in the first place. You text yourself reminders on the go or dump it in your notes app and never look again.Solution:Faro. Text it what you need to do, and every morning it texts you the list of what's still open.Why it works:1. It texts you every morning. Not a one-time notification, not a note you never open. Stays top of mind.2. Done with something? Just tell it. No editing notes, no re-texting yourself a new list.3. Zero friction. No calendar menus, no alarm settings. You remember something on the go, you text it, done.4. Reminds with context! One user prepping a conference talk doesn't just get nagged, Faro reminds her who's in that room waiting to get value from it.Try it today (or just text it lmao, that's the whole onboarding): "hey faro" to +1 (646) 879-0459 or check out: https://www.justtextfaro.com/
Ask HN: What have you been curious about?If you are anything like me, you will have about five answers at any given point in time for "What are you curious about?".Arguably, a more interesting, revealing, and kinder question is; "What have you been curious about?".I feel a looking-back-to-curiosity exercise has power to shed light on where to look ahead (and why to do it). More so, I feel it has power to grant permission to that "insatiably curious thirteen year old self" to step out and play. Because it probably already has been up to something, while your attention was elsewhere.So, here is a prompt (see what I just did there? ;-))Um... Except for LLM stuff, if you please. Just for this thread. Because these days the plurality of HN discussion appears to be about the curiosities of LLMs.(cf. More about this in a blog post I published a while back https://www.evalapply.org/posts/what-have-you-been-curious-about/ )
Show HN: MandoCode Desktop – Native Windows AI Coding Assistant on Ollama .NETQuick backstory on how this version came to be: the original MandoCode is a CLI, and its UI is built with RazorConsole — Blazor components rendered into a terminal. Genuinely clever tech! I wanted tabs, though, and when I attended MS Build this year I got a good look at what WinUI apps could do and saw the potential to give the original agent some actual windows and tabs. Having up to 4 agents running in one window is pretty cool. Aside from that — tabs! I feel like everything is a chat these days, and for some reason my brain keeps screaming AGENTS, not chats, at me. I love the tabbed-browser experience a lot moreSome of the stuff I got to toy with once the plumbing worked:- Context snapshots — closing an agent archives the conversation instead of deleting it, and you can snapshot mid-conversation to get an AI-written recap you can hand to a different model or a fresh agent and have it pick up the thread. You pick the LLM from a dropdown right when you create the snapshot.- Transcript
Ask HN: Do we need a stronger process for vetting code?Now that writing code is largely automated, I've seen Github being filled to the brims with slop, and it feels like the overall quality of everything produced with code (either the tooling or the software itself) has been getting worse. Pre-AI, times were a bit simpler: You looked at the stars, looked at the issues, poked around the code, and gave the repository a chance. Today there's just way too much stuff being produced. I don't see this changing in the future, so I was thinking that I'd be cool to make some spin out of it: My idea is to create a data aggregator based on https://github.com/explore that is targeted towards developers. Think of it as the "Rotten Tomatoes" of code. Here are a few ideas that I've been bouncing around: Repositories having different scores and being organized by tech stack, language, etc. Certified classics by category and technology.Each repository has a scorecard. - its dependencies (too much - potential bloat) - issue progression over time (number of
SOTA on the hardest AI memory benchmark (BEAM, 10M tokens), with a smaller modelHey HN. I'm Johnny, founder of Exabase. We just hit the highest reported score on BEAM, the hardest AI memory benchmark, at every scale up to 10 million tokens. We also ran our evaluation using Gemini 3 Flash, when all previous leaders depended on a much larger model (Gemini 3 Pro).At 10M tokens, the scale is vastly larger than any model's context window, so context stuffing isn't an option (aside from the fact that only about half of a large window can be effectively utilised without degradation). The only way to score well is recall that fundamentally works.Our system (M-1) scored 76.9% at 100K, 75.0% at 1M, and 68.0% at 10M. Previous leaders were Hindsight (73.4%, 73.9%, 64.1%) and Honcho (63.0%, 63.1%, 40.6%), both using Gemini 3 Pro, while we used Flash.We saw the competitive gap get wider at scale: 3.5 points ahead of Hindsight at 100K, 3.9 at 10M. The gap with Honcho goes from 13.9 to 27.4 points. As the corpus gets bigger, it filters out effective recall vs. brute-forcing / mod
Show HN: We killed our AI stock analysis and kept the news feed (API and MCP)I run AlphaAI, it is a news feed for trading agents, REST plus MCP. Before it was "AI stock analysis" site. That version demoed great and made no money. i was paying for market data, users was not paying me. Nobody buys prompt output, everyone already has own ChatGPT or Claude and own prompts. The only thing people were watching was the news feed with ratings, so I built a new product around this feature and it turned out to be really useful.What's under the hood: GDELT parsing, SEC EDGAR, google public RSS and several other sources. Everything goes through a deduplication and enrichment pipeline. Only a small part of the really useful stuff gets into the feed, it's the filtering that's valuable to the user. Question to people who wire news into agents: what metadata you missing usually? For me this is main thing i want to learn from this post.