Two Belgians shipped a 16MB Python library called Semble that beat grep at the agent’s own job. Their static-embedding index hits 94% recall at a 2k-token budget where grep needs 100k tokens to crawl to 85%, indexes a full repo in 263ms, and installs with one pip install.
It lands at the end of a year of static-embedding research from MinishLab in Antwerp, the shop that’s been arguing 16-million-parameter models can match 137-million transformers at retrieval. If they’re right, every usage-priced agent built on grep just inherited a budget problem.
In today’s indie hacker news:
- 🛠️ Semble: a CPU-only code-search MCP that undercuts grep
- 🧠 KPMG architect demolishes the AI-shrinks-your-timeline myth
- 🇬🇧 Mozilla begs UK not to age-gate VPNs as the Lords already voted to
- 📟 Doogee U10 tablet boots Debian from an SD card, stock Android intact
- ✉️ MailMark: BYO-domain cold email, no rented-inbox tax
TOP STORIES
GREP’S TOKEN TAX EXPIRED
🛠️ Semble cuts Claude Code’s grep tax 98% in 16MB of Python
The story: MinishLab shipped Semble, an MCP code-search server that runs on CPU and trades a fraction of grep’s input-token bill for the same retrieval quality. Maintainer Thomas van Dongen (Stephan Tulkens left MinishLab in September 2025) posted the Show HN yesterday: “We kept running into the same problem while using Claude Code on large codebases: when the agent can’t find something directly, it falls back to grep, reading full files or launching subagents. This uses a lot of tokens, and often still misses the relevant code.” 1.4k stars in under a day, top of Show HN.
The details:
- Stack is static Model2Vec embeddings from MinishLab’s own 16M-parameter potion-code-16M, plus BM25, Reciprocal Rank Fusion, and code-aware rerank signals like identifier stems and definition boosts
- Self-reported speed: 1.5ms p50 query, 218x faster indexing than CodeRankEmbed and 11x faster queries on their 1,250-pair benchmark across 63 repos and 19 languages
- Retrieval quality lands at 0.854 NDCG@10 vs 0.862 for CodeRankEmbed Hybrid, so you keep 99% of the transformer’s accuracy at the CPU price
- Wires into Claude Code with
claude mcp add semble -s user -- uvx --from "semble[mcp]" semble; runs entirely local with no GPU and no external service - HN user aadishv ran a real test on OpenCode: 14.7% context and $0.282 with Semble vs 19.0% and $0.352 without, but wall clock came in 2x slower on that repo
Why builders care: If your agent’s monthly bill has a “grep + read” line item, this is the cheapest experiment you’ll run this week. Wire it in behind a feature flag and measure the input-token delta yourself. Honest HN caveat: agents trained on grep often distrust probabilistic retrievers and re-read anyway, so the 98% is the ceiling, not the floor.
Work from any WiFi like it's your home network. NordVPN's Meshnet runs a free private mesh between your laptop, dev box, and home server. SSH from a café without exposing a port, the way you'd use Tailscale. The paid VPN on top lets you test geo-fenced Stripe checkouts or feature flags from any country.
We get a cut if you sign up. Only added for tools we use ourselves.
TYPING WAS NEVER THE BOTTLENECK
🧠 KPMG architect: AI can’t shrink your timeline because typing was never the slow part
The story: Frederick Vanbrabant, Manager and Enterprise Architect at KPMG Belgium, published a post on May 15 that hit #1 on Hacker News with 515 points and 365 comments. His thesis is borrowed from Eli Goldratt’s The Goal: the visible bottleneck isn’t always the real one, and AI is being sold as a fix for a step that was rarely the slow part. “You can’t make projects go faster just by typing faster. If that were the case we would all be taking typing lessons.”
The details:
- His Gantt example: a real project runs 10 days of feature exploration, 25 days of dev exploration, 70 days of build. Execs imagine AI crushes that build phase to three days; a realistic version still needs 40 days of detailed docs to brief the agent plus 40 days of AI-assisted build
- The example ticket he tears apart is “send mail to user once sale is completed”, which any agent will half-build wrong without clarification on content, error handling, and what “completion” even means
- His closing line, quoted across the HN thread: “This exact thing is what software developers have been begging for since the beginning of the profession,” meaning the AI-spec PMs now write is the same brief devs always wanted
- Top HN comment from pron: “It’s 2026 and the idea that even with detailed-enough requirements you can one-shot even a workable solution also needs to die”
Why builders care: Hand this post to the non-technical co-founder asking for an “AI 10x” timeline. The work doesn’t disappear, it shifts upstream onto whoever writes the PRD. Audit which step actually eats your calendar before adding more agents. If it’s clarification cycles with PMs and clients, more agents just moves the bottleneck, not the deadline.
1984 AS DIGITAL ROADMAP
🇬🇧 Mozilla begs UK not to age-gate VPNs as the Lords already voted to do it
The story: Mozilla policy lead Svea Windwehr filed a formal response on May 15 to the UK’s “Growing up in the online world” consultation, which closes May 26 and considers age-assurance rules for VPN providers themselves. Days earlier, the House of Lords passed an amendment to the Children’s Wellbeing and Schools Bill that would force VPNs to age-check every UK user, 207 to 159 along party lines. Mozilla’s brief argues activists, journalists, and dissidents rely on the same VPNs the government wants gated, and that the better lever is to “hold platforms to account, encourage responsible parental controls use, and invest in digital skills.”
The details:
- Backdrop: when the Online Safety Act took effect on July 25 2025, Proton VPN saw a 1,400% hourly signup surge sustained at +1,800% for several days, and five VPN apps hit the top 10 of the UK Apple App Store the Monday after
- Pornhub UK daily visitors fell from 3.2M to 2M (-38%) in early August 2025, while UK Google searches for VPN terms went up sevenfold
- Ofcom is scheduled to publish a June 2026 report on what’s blocked age-assurance so far, and DSIT’s government response on the VPN amendment is due summer 2026
- The Register’s coverage noted Russia restricted 400+ VPN services by January 2026, a sharp jump from autumn, as the comparison case
- Top HN comment from robotswantdata: “1984 was meant to be a warning, not the UK’s digital infrastructure roadmap”
Why builders care: If you sell to UK consumers, age-assurance just stopped being a porn-site problem. Plan for jurisdiction-agnostic hosting in your pitch, stop trusting IP-based geofencing for tax and compliance, and price in whatever your fraud vendor decides to do with UK VPN traffic this summer.
E-WASTE TO EDGE COMPUTE
📟 $80 Doogee tablet boots Debian from SD, no bootloader unlock needed
The story: A solo hacker posting as tech4bot pushed rk3562deb to GitHub on May 14 and the Show HN hit 280 points in 14 hours. Target hardware is the Doogee U10, an $80 Android tablet with a Rockchip RK3562 quad-core A53 at 2.0GHz, 4GB RAM, 128GB eMMC, Mali-G52 GPU, and a 1 TOPS NPU. His build system produces a Debian 12 Bookworm image you flash to SD: BootROM checks the card first and falls back to eMMC, so insert the card and Debian boots, pull it and stock Android comes back. Skips the bootloader unlock entirely and leaves the device reversible.
The details:
- He reverse-engineered the device without any BSP, kernel source, or vendor docs, working only from a DTB extracted from the stock Android firmware and the Firefly rk356x U-Boot tree
- Working today: display, 10-point multitouch, Wi-Fi 6, Bluetooth, audio in and out, accelerometer, battery, USB OTG, SD boot, flashlight. Partial: Panfrost 3D acceleration, NPU inference, cameras
- For Wi-Fi drivers he emailed the Chinese chip factory, got no English reply, rewrote the email in Chinese using AI translation, and got the drivers
- His own quote on the workflow: “We are not yet at the point where you give AI a tablet and it magically returns a working image. AI helped a lot, but it also introduced bugs more than once”
- Companion YouTube walkthrough covers the UART wiring and DTB adaptation
Why builders care: An NPU-equipped touchscreen Linux box at consumer-Android pricing is now a viable target for kiosks, ambient dashboards, edge-AI demos, or homelab nodes. The cheap-edge-compute math just got a real benchmark, and AI assistants now compress months of solo kernel-bringup work into weeks, even for builders who openly say they aren’t kernel devs.
OWN THE DOMAIN, KEEP THE INBOXES
✉️ Power-plant engineer ships $10/mo cold email where you own the mailboxes
The story: Debasish Barai, an electrical engineer at a power plant in India, shipped MailMark and posted it to Show HN on May 17. Bring your own domain and mailboxes, MailMark walks you through MX, SPF, DKIM, and DMARC, runs a 28-day warming flow, then hands you a unified Gmail-style inbox across domains plus a Resend-style developer API. Founder line from the HN post: “I literally own so many domains which are all related to my products. Maintaining all these domains and their email Ids honestly speaking was difficult.”
The details:
- Starter is $10/mo for 1 domain, 3 mailboxes, 1k emails. Pro is $50/mo with 5 domains, unlimited mailboxes you bring yourself, 25k emails. Business runs $100/mo for unlimited domains and 100k emails
- Instantly’s headline price is $47/mo, but rented inboxes push real TCO to $150 to $400, and Smartlead’s Unlimited Smart tier sits at $174/mo. The BYO escape hatch is whatever you pay your registrar
- Tech stack is Convex for backend, Clerk for auth, Vercel for hosting, with sender-reputation monitoring on SPF, DKIM, DMARC, and blacklist status
- The Feb 2024 Gmail and Yahoo bulk-sender rules and Nov 2025 hard-reject 550s killed shared-IP cold email. Owning the sending domain means a reputation you can move
- Prior project was a side-hustle Android app called Railify. No funding, no team, started writing Android in 2015 in Java pre-Kotlin
Why builders care: Cold email is the rare $1B niche where the incumbents priced themselves toward the BYO escape hatch. Same energy as Plausible vs GA4, Resend vs Mailchimp, and Cal.com vs Calendly. The lesson for indie SaaS: when the floor price of an incumbent quietly triples on add-ons, the bring-your-own version writes itself.
TRENDING TODAY
🧱 Anti-AI sentiment hit the front page twice in one day. Pew’s March 2026 reading says US adults more concerned than excited about AI runs 50% vs 10% more excited (up from 37% concerned in 2021), and 53% say AI worsens creative thinking. Add Vanbrabant’s blog at #1 the same morning and the vibe shift looks operational, not ethical. Then Chrome Web Store rolled out a fourth competitor in the “strip em dashes from AI output” niche this week alone (DashAway, undash, Em-Dash Replacer, plus today’s stack pick). Builders are watching their feeds get crammed with output and quietly shipping the filters.
🪶 Agent runtimes raced to the bottom of the hardware spec this weekend. Two more Show HNs alongside Semble pushed the same thesis: Agetor is a local-first kanban that orchestrates multiple coding-agent CLIs in parallel via git worktrees (TypeScript, MIT, v0.0.2 shipped May 17 on Electrobun), and Cheap-IM replicates Thinking Machines’ Interaction Models demo on a CPU laptop by stitching Silero VAD, Kroko ASR, YOLO11 pose, Piper TTS, and a remote LLM into a single asyncio loop. The frontier moved this week: smallest hardware footprint wins the launch post.
🔧 Weekend tinkerers swept the HN front page. Beyond tech4bot’s Debian tablet, grid2poster generates print-ready electrical-grid posters from OpenStreetMap data using GeoPandas and Matplotlib (82 stars, any country or continent in PNG, SVG, or PDF), and IEEE Spectrum’s VoIP payphone story about Beta Technologies engineer Patrick Schlott installing six coinless payphones across rural Vermont towns hit 129 HN points ahead of Vermont’s September 2026 school-phone ban. Hobby ships do well on Sundays when the rest of the internet starts feeling fake.
FIRST DOLLAR
PLAYBOOK YES, MATH MAYBE
💸 GojiberryAI’s @pierreeliottlall posts a 5-step growth playbook with one inflated stat
@pierreeliottlall dropped a 5-step playbook claiming “$0 to 2,000+ paying customers in 9 months with no outside funding” for his B2B intent-data SaaS GojiberryAI. Two corrections first: the 2,000 figure looks like the pre-pivot waitlist (founder-told versions land closer to 100 paying customers in the first 60 days post-pivot and roughly $1.4M ARR), and GojiberryAI is YC P26-backed, so “no funding” is overstating. The playbook itself still holds: 7-slide deck before any product, manual high-intent lead lists, 180 LinkedIn invites and 1,000 cold emails a week, then Reddit value-comments stacked with three customer-story posts a week. The post ends “comment GOJI for a trial,” so read it as a lead-magnet ad with a real log buried inside.
STACK OF THE DAY
🧹 Dashbuster
A Chrome extension that strips the AI tell, every em dash, from any page you visit. Show HN’d at 1:48 UTC this morning. Swaps the dash for a period, comma, or hyphen depending on the surrounding clause, so AI-generated copy reads like a human typed it. 30-second install, no account or tracking. Solo builder shipped it free to the Chrome Web Store. Bonus pattern: keep it on for a week and you get a passive AI-content detector for the pages you visit, the dashier the page, the higher the chance you’re reading model output.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
🌞 Tesla Solar Roof quietly on life support, pivots to regular panels - Electrek’s postmortem on the 2016 product Elon promised would be cheaper than a normal roof, with 192 HN points and 187 comments arguing the install math never worked.
🦴 Prolog Coding Horror - Markus Triska’s walkthrough comparing what Prolog could already do in 1992 against what most modern languages still can’t, 71 HN points and a thread full of people remembering why declarative programming briefly felt like the future.
⚡ A nicer voltmeter clock - lcamtuf shipped a weekend desk clock that drives two analog meters as a digital-to-analog display, with a homemade PCB and a microcontroller doing the heavy lifting. The kind of “I built this Sunday” post HN still rewards.
Curated by AI, built by a human. Subscribe for daily indie hacker news.