Real problems people complain about online, pulled every morning and scored out of 100. Build, validate, or skip. How scoring works
Hey HN, Guanming and Bill here from General Instinct (https://general-instinct.com/).After years of working in robotics, we kept running into the same problem: the best models never fit the hardware we actually had available.The models that performed best were usually designed around datacenter assumptions: large GPUs, lots of memory bandwidth, and reliable network access. But most physical systems have the opposite constraints.That led us down the path of figuring out how much of a frontier model could be preserved while still making it practical to run on edge hardware.As part of that work, we recently open sourced InstinctRazor (https://github.com/General-Instinct/InstinctRazor)One result we're excited about is compressing Qwen3.5-122B-A10B, a roughly 245 GB BF16 MoE model, into a 48 GiB GGUF. The resulting model is actually smaller than Gemma-4-26B-A4B while outperforming it on benchmarks like MMLU-Pro and GPQA-D etc. we preserve the parts that are always active (router, norms, Gated-DeltaNet/SSM layers, vision pathway, etc.) and quantize the routed experts much more aggressively. We then use on-policy distillation to recover capability lost during quantization.The model can also run in a "small GPU" configuration where experts are streamed from system RAM. With an 8k context window, peak VRAM usage is around 7.6–8 GB.If you're interested in the technical details, we wrote up the approach here (https://general-instinct.com/blog/frontier-moe-sub-4-bit)We're especially interested in hearing from people deploying models onto robots or other edge devices. What models are you trying to run locally today? What has been the biggest bottleneck in getting them into production?
Hacker News3mo agoToolAI
I got really tired, as a human, of parsing the standard marketing heavy web we have today. I've always loved the simplicity of gopher and gemini web.Recently I found myself manually adding `/llm.txt` to most websites I visit because I find the content for LLMs strait to the point and clear. The only annoyance is web browsers like chrome do not render the markdown.So could the AI revolution actually fix the web for humans as a side effect?Do you find yourself doing the same?
Hacker News3mo agoToolAI
Hi HN, we’re Nick and Drew, and we’re building boxes.dev – the first cloud-only agentic dev environment (ADE) that gives every Codex and Claude Code agent its own cloud computer.We’re two engineers who previously built Gem (co-founder/CTO and first hire), and we spent the last year coding almost exclusively using Codex and Claude Code. It’s been a huge change to how we code, and it’s been exhilarating seeing the models keep getting better – but we eventually realized that developing on localhost was holding us back:- Git worktrees are clunky to set up and use for parallelizing work - It’s 2026, but somehow everyone is still walking around with laptops cracked open or SSHing into mac minis in their garage so their agents don’t stop working. - Mobile is treated like an afterthought even though coding is just texting now We started hitting resource constraints when multiple parallel agents test their own work by running the full app locally. - We tried different products, but couldn’t find any that solved all of our pain points – so we pivoted and decided to just build the ADE we wanted for ourselves.Boxes.dev is a desktop and mobile app that lets you run Claude Code, Codex (using your subscription!), and the full dev environment for whatever you’re building, all on remote compute. It’s similar to Conductor or the Codex desktop app, except everything is in the cloud.We use coding agents to scan your local dev setup and port it to the cloud. Then every Claude Code/Codex thread starts from a snapshot of the full setup, with its own filesystem and compute. No more git worktrees, no more cracked-open laptops, and your coding agents can actually test their work end-to-end because they can run your full app in isolation.We’ve mirrored the Claude Code and Codex UX to feel natural to power users, and also have a fully-featured mobile app (no handoffs or remote control), plus scheduled automations and a Slack integration.We’re obviously biased, but we’ve been buildin
Hacker News3mo agoToolAI
You can get a 2h free trial by solving a proof-of-work captcha when topping up your account for the first time.If you'd like to learn more, an independent interview was posted a couple of weeks ago [1], and the FAQ [2] has a lot of information as well.For the source code sharing, we've talked with lawyers and are inclined to no longer require the NDA/NCC for privacy reasons shared with us before (signing requires identification), but instead use a source-available permissive license that doesn't allow competition, like PolyForm Shield [3] (we do still have about 6 months before finalising a decision, here).This does come with a lot more risks for us (it's harder to track down if someone publishes the code or uses it against the license), but given we've already passed 100 monthly active accounts, we're feeling more confident it's an acceptable risk.The plan is to give logged in accounts (who are 12 months old or more) a way to download a ZIP of the current code base that's in the server.Obviously there's no easy way to prove that's the case, but we're open to ideas/suggestions if someone here has them.[1]: https://theprivacydad.com/interview-with-the-engineer-of-uru...[2]: https://uruky.com/faq[3]: https://polyformproject.org/licenses/shield/1.0.0
Hacker News3mo agoToolAI
We launched Infracost on HN five years ago (https://news.ycombinator.com/item?id=26064588) where our CLI generated cost estimates for infra-as-code, e.g. "this Terraform PR adds $400/mo". The idea was to shift cloud costs (FinOps) left, so engineers get visibility of costs before deployment and make better decisions.Earlier this year we started seeing agent traffic in our logs and it looked like coding agents were calling our CLI. But that CLI wasn't designed with coding agents in mind. We went down a philosophical rabbit hole to see if a CLI is even needed anymore given that Claude, Copilot et al. already follow best practices. Ultimately we decided to create a new CLI from the ground up with coding agents in mind for two reasons:1. We optimized the CLI for agent callers and cut Claude's output token usage by up to 79% and API cost by up to 67% versus a bare-Claude baseline. We wrote a blog documenting our lessons on optimizing user token usage when designing a CLI, e.g. using predicate flags so the agent doesn't compose jq | python | wc pipelines, output format that strips JSON's redundant field names. The blog is here: https://www.infracost.io/resources/blog/we-cut-claude-s-toke...2. With cloud costs, precision matters. Telling a coding agent "make this Terraform cost-optimized" can be expensive and lossy. You burn tokens loading code and policy context into every conversation. Your agent could make up a price and you wouldn't know because it's difficult to verify that across the ~10M price points that AWS, Azure and Google have. The CLI runs static analysis on the code, uses the latest prices from cloud vendors, and passes that context to the coding agent.So that's what we're launching today - Cost.dev: https://cost.dev/.- It runs locally. Your code never leaves your machine, you get a fast feedback loop, and you're not burning API calls p
Hacker News3mo agoToolAI
Hacker News3mo agoToolAI
Hey HN, we’re Shalin & Kanyes, best friends who've been hacking together for 10+yrs, and now founders of Hyper (https://heyhyper.ai/). Hyper is a shared “company brain” that plugs into information flowing inside a company to make AI agents and automations better and ultimately save people time.Models have gotten good enough that they can (mostly) take on long-horizon, complex tasks. We believe the bottleneck now is that these smart-enough models often lack information about your company, which is scattered in people's heads, Slack threads, stale docs, and in back-and-forth convos with AI.MCP is useful for getting some info in front of an agent, but there are problems: (1) Once the session dies, so does the insight, so instead of copy-pasting a whole doc each time you're telling the agent to dig through Drive each time - not much of a win; (2) Even when MCP works, what it gathers isn't comprehensive, because people decide things on a whiteboard, brainstorm out loud, post a little in Slack, and scribble the rest in a doc, which leaves the agent working from partial information; (3) And even if it had everything, it doesn't do the meta-reasoning required to do a great job. If you paste in a Notion doc and it won't learn your design taste or your writing style unless you tell it to, and it won't know why a decision was made or when.As undergrads 5 years ago, we were into the tools-for-thought wave and became power users of Notion, Obsidian, Roam, Anki, real believers in building a second brain. After GPT-3.5 came out we started to realize how much more powerful that second brain could be if an AI could actually read it, because suddenly it would know our backstory, our taste, our preferences, and unlock genuinely new capabilities. That’s why we’re building Hyper.We know it’s not for everybody! But for people who do want to be on the cutting edge, this is a force multiplier that makes agents faster and better. It inc
Hacker News3mo agoToolAI
Hi all,I'm excited to show off Textile, a desktop app I recently built.Textile can combine bits of text using various inputs, such as commands on your computer, the contents of your clipboard, and hard-coded strings that you provide. It lets you carefully build up and modify a dynamic string, step by step, until it's exactly how you need it. The saved steps can then be executed on demand, with the click of a button or using a keyboard shortcut.I built Textile because I was often constructing complicated, dynamic URLs from various sources that all existed on my computer. I got tired of manually switching between different apps, copying and pasting various chunks of text, and assembling them all together somewhere. I've also found Textile to be quite useful as a kind of repository for obscure bits of static text, such as ½ and other fraction characters, when I can't be bothered to remember their built-in keyboard combinations.I also built Textile because I wanted to learn Electron, although I expect there will be some gnashing of teeth about this here. :) I think desktop development is quite interesting, in part because it doesn't require me, the developer, to pay for an API server and database in the cloud. The app itself is both the UI and the "server," and the local drive is effectively the "database." I knows this trades away syncing with the cloud but, on the other hand, there's something nice about knowing that your files are on your drive and not on somebody else's server.I realize that something like Textile may already exist, and may have much more functionality but, again, I wanted to learn. I must say that multi-sequence keyboard shortcuts are hard, and there are cases that don't work right in Textile. I feel vulnerable admitting that my approach has much room for improvement!For what it's worth, I did not use an LLM to write any code for Textile (although I did ask many questions of an LLM, a
Hacker News3mo agoToolAI
I built a terminal app that paces slow breathing at 6 breaths per minute for vagal tone training. It's a single Python file, stdlib only, no dependencies — just run breathe and follow the bar.I'm a cardiology patient (HFrEF). Slow breathing at resonance frequency is one of the few non-pharmacological interventions shown to improve cardiac vagal tone and baroreflex sensitivity (Bernardi et al., Circulation 2002; Lancet 1998). I wanted a frictionless daily habit tool — no app store, no account, no subscription, just open terminal and go.Design constraints, all grounded in the clinical literature:- No breath retention — Valsalva risk in cardiac patients- No rapid breathing — minimum 8-second cycles- Exhale ≤ 2x inhale — no evidence for extreme ratios- Immediate exit, always — q or Ctrl+C restores the terminal even on crashThe README includes a resonance frequency measurement protocol for anyone with a chest-strap HRV monitor who wants to find their individual optimum instead of using the 6 bpm default.macOS only (uses afplay for audio cues). MIT licensed.pip install breathe-cliorbrew tap marekkowalczyk/breathe && brew install breathe.
Hacker News3mo agoToolAI
Hi I'm Dan from Elodin, making an open source real-time capable flight software simulation.For AI Grand Prix contestants, the wait for the Round 1 virtual qualifier simulation has been grueling.If you’re competing, check out our simulation harness to tide you over, built to match the published competition constraints and message format. It runs against real Betaflight, which we learned requires at least 1000 sensor samples per second to run real-time correctly.The competition warranted introducing a new feature to generate the camera sensor directly in the simulation loop. Typically people connect to Unreal or similar game engine to create a camera sensor, which works well but is very heavy. For the simple needs of this challenge, creating sample directly in the loop is very handy and easy to use. Happy to hear your feedback on this! While it's not fancy looking currently, it uses the Rust Bevy game engine, which should allow us to improve the visual fidelity quickly.We all should easily be able to shift our implementation to the published competition sim once it lands. Hope you enjoy and good luck!
Hacker News3mo agoToolAI
Hi HN, we’re Rishi and Sahil. We’ve developed Rudus (https://www.rudus.ai/), an AI-powered takeoff and estimation platform built for concrete subcontractors.Takeoff is the process of measuring and quantifying materials from concrete plan sheets. Rudus identifies every concrete structure (footings, walls, columns, slabs), pulls in related details, and eliminates hours of manual quantity calculation. Here’s a demo: https://www.youtube.com/watch?v=PAMNDRWEdlI.The problem: Concrete subcontractors are the backbone of every building, but their estimating workflow hasn't changed in 20 years. Right now, a senior estimator opens a PDF, manually traces every footing and grade beam, then hand-builds an Excel spreadsheet with 300+ line items- volumes, formwork, rebar by bar size with lap splices and development lengths. Bids can take weeks and even months. Most firms have just a few estimators, meaning they physically cannot bid on most of the work available to them.The software incumbent in this trade hasn’t been updated since 2020. Beyond that, every AI takeoff tool on the market was built for GCs and treats concrete as one checkbox, rather than working around how concrete estimators actually price work. We’re building Rudus for this trade and only this trade.We started this when Sahil took a construction management class and realized how the estimation workflows hadn't changed in decades. We started cold calling, walking into offices with donuts, showing up at job sites, and everyone told us the same thing: slow estimation is the biggest bottleneck in growing their business, but every new product they've tried has failed. We quickly realized that the reason those tools failed is a lack of trust and frequent errors causing later problems. Estimators stake million to billion dollar bids on these numbers, and they are clear that they won’t trade their workflow for a black box. We took a different approach: software that intelligently
Hacker News3mo agoToolAI
Built an open JSON Schema for defining AI agent teams.Multi-agent systems are becoming a real deployment pattern — not single assistants, but teams with roles, handoffs, and human checkpoints. But there's no shared way to define one that travels across frameworks. Every implementation is scattered, locked to whichever tool you picked first. Built the schema to fix that.The schema lives at schema.openenvelope.org and is registered in SchemaStore, so if you drop a .envelope.json file in VS Code you get autocomplete and validation without installing anything. It's also on npm as @openenvelope/schema if you want to validate programmatically.The spec covers: agent definitions (role, prompt, model, access policy), supervisor/sub-agent hierarchy, human-in-the-loop gates, pipelines, schedules, and secrets/variables that get injected at deploy time. Access policies let you declare exactly which hosts each agent can call — the runtime enforces this at the network level, not in the prompt.The goal is a portable definition format — define a team once, any compatible runtime can execute it. Similar to how Dockerfiles describe a container without being tied to a specific host. There's a managed runtime at openenvelope.org but the schema is Apache 2.0 and anyone can implement it.Happy to answer questions on any part of the spec — especially interested in feedback from people who've built multi-agent systems and have opinions on what's missing.
Hacker News3mo agoToolAI
Hey HN, we’re Ismaeel, Eren, Yafet and Nikodem. We built Expanse (https://expanse.sh/) to increase the effective capacity of your HPC/GPU clusters running schedulers/orchestrators like Kubernetes and SLURM. We read the source code, job submission script, and the hardware a workload is about to run on to predict what the job actually needs before the cluster sees it. We also flag failures we think are about to happen and surface line-level optimisations the researcher can apply themselves.The problem: Datacenters run at roughly 30% to 40% effective utilisation. Users request more resources than what they actually need, because of asymmetric risk: while over-requesting is bad because it’s expensive and wastes capacity that someone else could have used, under-requesting kills your job mid-run and you lose days of work. So everyone over-requests by two to three times.We measured one national-scale HPC cluster for a month and from 122k jobs, 59% of the compute was wasted. At on-demand cloud rates for the same hardware, that’s roughly $8.5M of compute wasted in one month on one cluster. The pattern is similar in large scale compute industries as well, such as quant funds, AI labs, and manufacturing.The four of us ran HPC and GPU training workloads at the largest quant funds and HPC facilities. Ismaeel did research at EPCC (Edinburgh’s Parallel Computing Centre, the UK’s national HPC site) under Adrian Jackson, where he built the first multimodal HPC resource predictor: a model that ingests job source code, submission scripts, hardware telemetry and cluster metadata in order to figure out how much compute will actually be needed. On a dataset of real workloads on EPCC’s own clusters it scored 34% better than any other baseline, and outperformed frontier general-purpose LLMs prompted on the same prediction task by roughly 8x. These results convinced us the problem was solvable with software.Expanse installs on every node and hooks into SLURM (or the
Hacker News3mo agoToolAI
Fungible is a terminal-based personal finance app that fills the Mint-shaped void in my life. It runs using your own plaid credentials (optional) and has its own integrated chatbot (also optional and BYO key).You connect banks via Plaid or import CSVs. Transactions get auto-categorized by rules you define. On top of normal categories there's a flexibility layer (fixed / flexible / discretionary) so you can see at a glance what's actually controllable spending. There are also tags (also separate from categories) for isolating things like trips or hobbies.The financial health screen does savings rate, liquidity runway, and FIRE projection with adjustable assumptions. Probably overkill but I like it.It also has an MCP server so Claude/ChatGPT can talk to you about your finances, create rules/tags, etc. That’s always the most annoying thing for me when trying a new personal finance tool. Hopefully this brings down the barrier to usefulness. The agent/chatbot in the app has the same tools as the Claude/ChatGPT would have via the MCP.
Hacker News3mo agoToolAI
### What variant of Codex are you using? App ### What feature would you like to see? Many Windows users cannot easily install applications from the Microsoft Store due to system restrictions, corporate policies, offline environments, or personal preference for traditional installers. Currently the Codex Windows app appears to be distributed only through the Microsoft Store. It would be very helpful if a standalone installer such as `codex-setup.exe` or `.msi` were provided. Benefits of a standalone installer: - Allows installation without Microsoft Store - Works in offline or restricted enterprise environments - Lets users choose the installation directory - Easier automation and deployment (scripts, DevOps, package managers) - Better compatibility with portable setups and custom Windows environments Many developer tools provide both distribution methods (Store + standalone installer), which gives users more flexibility. A simple downloadable installer from the GitHub releases page or official website would solve this. Example: ``` codex-setup.exe codex-x64.msi ``` ### Additional information Other developer tools commonly provide standalone installers alongside store versions, which improves accessibility for developers using custom setups or environments where the Microsoft Store is disabled. Providing a traditional installer would make Codex easier to adopt across a wider range of developer environments.
GitHub6mo agoToolAI
Plug-in solar panels (no electrician needed) have just become legal in the UK and will go on sale soon. Helios estimates how much electricity a typical installation could generate at a given address and what that's worth against your tariff.It uses UK government LIDAR data to reflect the actual skyline, so it knows whether there's a building or a hill blocking the sun.Caveats: - Outside LIDAR coverage (most of Scotland and Wales) it falls back to a synthetic horizon (less accurate). - Trees and recent developments (post-2022 or so) may not be in the data, and some address placements could be off (geocoding via OSM).Feedback on the shading model especially welcome.
Hacker News3mo agoToolAI
Businesses replacing employees with AI face total chaos because there's no easy way to capture and feed all company-specific knowledge into the models as structured context.
X3mo agoToolAI
Hi, I’m Kenny, I’ve been building aislop. I starting working on this after using Claude Code, codex and opencode several times and noticing some slops. They aren’t syntax and passes most tests, they are patterns like empty catch blocks, useless comments, duplicated helpers, dead code and many more. So I built a tool to scan and check for these patterns and wired it into hooks so after each tool call, the agent checks for the slops.You can try it out with npx aislop scan.It’s all local and no code is transferred. Thank you.
Hacker News3mo agoToolAI
Hi HN, we’re open-sourcing ktx. It’s an executable context layer that makes agents reliable on your data stack.We built it after going through the experience of building production-grade data agents for dozens of companies. If you’ve also tried building them, or simply tried using Claude Code or Codex on your data warehouse, you’ll know that accuracy is the #1 issue. Agents are great at generating valid SQL, but it’s not always correct SQL.To cite a few examples of “agents gone wrong”:- Stale column + hidden business rule: when preparing a board report, a finance analyst asks Claude Code for “ARR by customer segment”, it derives ARR from multiple tables (subscriptions, plans, accounts), then groups by accounts.industry. But CC doesn’t know that this industry column was deprecated a few months prior, or that past board reports excluded paused subscriptions from the ARR calculation- Join fanout: a data analyst at a retailer uses their company’s internal agent to prep a product revenue deck for a QBR. The agent joins orders to order_items, then sums orders.total_amount_cents grouped by order_items.product_id. The SQL runs fine, but each order’s revenue is repeated once per line item, which most people will miss if most orders only have 1 item- Missing attribution logic: a marketing analyst asks Codex “Which campaigns drove the most revenue?” Codex joins marketing_touches to users to orders and groups by utm_campaign. But since each order can have multiple touches before purchase, the same order can be credited to first touch, last touch, every touch, or every campaign the user clicked before buying. If the agent chooses the method that doesn’t match the team’s attribution logic, they’ll make suboptimal decisionsTo solve this at first we gave the agent more context through skills + a wiki-style knowledge base. That gives it some useful extra context but still relies on it writing the SQL without incorrect assumptions.The next solution we explored was implementing a clas
Hacker News3mo agoToolAI