Today's haul: 10 tools · 7 resources · 21 reads · 10 numbers · 13 things that happened. Every tool, resource and read below has a working link. No link, no listing.
⚡ 60-Second Catch-Up
The "route cheap work to cheap models" thesis stopped being a thesis this week. AT&T is now sending roughly 40% of employee AI queries to open models and intends to push that to 60–70%, running about 45 billion tokens a day; coding costs fell 56% while quality dropped about 2% (The Information). Mark Austin, an AT&T vice president, said open models are "just as good or better" than older paid models for many jobs. Goldman Sachs' Jim Covello calls the model optimization layer the real bottleneck for enterprise AI — high-consequence requests to frontier models, routine ones to cheap models. → The 2% is the number to steal, not the 56%. A router only pays if you can define "good enough" on your own real work; without an eval you have not built a router, you have built a slot machine. Start by writing down what a bad answer looks like for your three most common tasks.
Apple rewrote what it costs to ship an app in the EU, and the new tiers are worth doing arithmetic on. From October 1, apps using Apple's in-app purchase pay 26% instead of 30%; alternative in-app payment processing pays 20%; linking out to an external checkout pays 15%; and apps distributed through alternative marketplaces or the web pay a flat 5% Core Technology Commission, replacing the old per-install Core Technology Fee. Most developers still qualify for halved rates — 15% and 10% — through the Small Business Program, Mini Apps and Video Partner programmes, and on auto-renewing subscriptions after year one (Apple, CNBC, TechCrunch). → If you are a small-business-programme developer with subscriptions, the link-out tier at 10% is now less than half what you were modelling. Re-run your unit economics before October, because the cheapest path changed and the answer is no longer "whatever Apple defaults you to."
Nvidia ran the same unusual play a third time, and founders should learn its shape. Poolside signed a non-exclusive $6bn licensing deal for the software it used to build its models, took a $1bn investment at a $12bn pre-money valuation, and watched 109 of its staff receive Nvidia job offers — while all three founders stayed and the company insists this is neither an acquisition nor an acquihire, with the $6bn earmarked for distribution to investors by the end of next year (Newcomer). Nvidia had already structured deals this way with Groq and with Enfabrica for about $900m. → This is becoming a real third option between "raise again" and "get bought": licence the machine, keep the company, pay the cap table. If you have built tooling that a giant would rather rent than rebuild, that is now a financeable outcome — and it does not trigger the regulatory review a purchase would.
🛠 TOOLS — 10
Build & ship
Bun 1.4 ⭐ — The biggest Node.js-compatibility jump since 1.0: +1,517 tests from the Node test suite and over 2,900 fixes, plus 5× lower idle CPU, up to 35% less memory and 50% faster startup on Linux. Adds Bun.Image, Bun.WebView, Bun.markdown, Bun.cron() and Bun.Terminal · Open source
Fig — A TypeScript re-implementation of React — fibers, lanes, hooks, streaming SSR, selective hydration — with data resources and asset declaration built into the runtime. For a minimal interactive surface it measures 92.5 kB minified against React 19.2.7's 194.1 kB, roughly half. Stable pre-1.0, TanStack Start integration works, not a drop-in React replacement · Open source (MIT)
Rust Glancer — An alternative Rust LSP built for low memory — targets under 100 MB on reasonable projects and skips re-indexing on editor restart, so an already-indexed project comes straight back. Author tested it on an 8 GB M1 MacBook Pro · Open source
AI & agents
Linear coding sessions ⭐ — Linear's agent now installs the toolchain, runs your app, drives it in a browser, and captures before-and-after screenshots to prove the change — fixing and re-testing on its own before handing work back. The part to actually copy: admin-set workspace and per-user spend limits that reset daily, weekly or monthly and hard-stop credit use when hit · Paid · usage priced on model + runtime
Slack Code — Temporary channels where people and coding agents work a single task together, then archive automatically while staying searchable. The point is the audience: non-engineers can watch what an agent is building, open previews and comment, without leaving Slack · Paid (Slack plans)
OzBrain — One shared knowledge layer that Claude, ChatGPT, Cursor and other agents all read and write, structured so each agent pulls only the slice it needs instead of re-reading everything. Aimed at teams tired of re-explaining the same context to every tool. Show HN this week · Paid · see site
NoBuzz /debuzz — A Claude Code skill that pipes Claude's last reply through the Gemini CLI and prints the translation verbatim, stripping the "here's where it gets interesting" register down to plain engineering English. Three modes — colleague, manager, director. Requires the Gemini CLI, authenticated · Free · open repo
Money & payments
Zoneless ⭐ — A self-hosted, Stripe-compatible payments stack for USDC — checkout sessions, payment links, subscriptions, connected accounts, identity verification and marketplace payouts, following Stripe's object shapes, webhook patterns and idempotency model. Built for and running in production at PromptBase, which reports more than 500,000 users. Useful mainly for onboarding sellers in regions Stripe Connect does not cover · Open source (Apache-2.0)
Read & search
Kagi — Exclude paywalled websites — A single setting that strips paywalled results out of search entirely, instead of blocking domains one at a time. Shipped August 21; the control sits in Search settings · Paid (Kagi plans)
Cobalt — An open-source app platform for Kobo e-readers: a launcher, a signed app store, a Rust SDK, and a runtime that sandboxes each app in its own unprivileged process. Install once over USB, then everything after installs and updates on the reader over Wi-Fi — and a reboot returns you to the stock Kobo reader. Not affiliated with Rakuten Kobo · Open source
📚 RESOURCES — 7
Steal the template
Firsthand: How I Built a Brand with Heart at Lovable — Nad Chishtie joined as fourth employee and first designer, prototyped the launch identity in 30 minutes, and then had to rebuild it for scale while the company picked up $10M ARR in 60 days. The useful part is the method he used for the rebuild: talking to every paying user he could to find what the product actually made them feel, and designing from that.
Raising Your Next Round Is Harder Than Your Last One — the arithmetic behind why a company that raised a seed easily gets Series A nos without changing anything about itself. An accelerator at $330K–$2.5M post only needs a $33M–$250M exit to work; a $5–10M seed needs belief in $500M–$1B; a Series A underwrites a path to roughly $100M a year, which is why $1M ARR is not the milestone you think.
How we raised mobile end-to-end test stability to 98% — Shopify's E2E suite had gotten flaky enough to be pulled from PR checks. They rebuilt it with a builder-style API that forces an assertion on every step, prefixes escape hatches with UNSAFE_, and uses computer vision to find elements the way a merchant sees them rather than by view-hierarchy node.
PagedAttention: Virtual Memory for the KV Cache — the OS virtual-memory trick applied to LLM serving: fixed-size physical blocks from a shared pool instead of contiguous allocation, which is where the throughput on concurrent requests comes from.
Inside a TanStack Router Navigation — what actually happens across route loading, caching and rendering when navigations and preloads overlap, and how they avoid stepping on each other.
Technical Ingredients, Revenue Ingredients — Rich Mironov on the unforced error of treating shipped code as shipped revenue, and what market-readiness work has to exist in parallel when development outpaces selling.
Thoughts on Taking OpenAI Foundation Funding — SecureBio on accepting a $17.2M OpenAI Foundation grant while its AI division continues to evaluate OpenAI's models, written with unusual candour about the conflict.
Why You Need to Subtract: The Vasa Effect — Sahil Bloom on a 1628 warship that sank a mile into its maiden voyage because every individual addition made sense, paired with Leidy Klotz's research on how badly people default to adding.
I'm becoming AI-blind — on the creeping inability to read anything without wondering whether a model wrote it.
Stop Making TUIs — a contrarian case against the terminal-UI revival, from someone who has clearly used a lot of them.
Pushing the Speed-Cost Frontier for Qwen3-TTS — how Nari Labs got a text-to-speech model to 50 ms time-to-first-audio, with the cost side of the tradeoff included rather than buried.
📚 Longer, worth it
Rise of the Borderless Founder — on international founders who keep a foot in both markets, and the structural advantages that come with it: knowing the strongest engineers at home before anyone else does, and landing large domestic customers as early proof points.
Forget about the ARR — the case that standard diligence cannot tell you whether an AI company built what it is selling, and that you have to score model dependency, data moat, talent concentration and substitution risk by opening the codebase.
Going freestanding — porting Go's standard library to run with no operating system underneath it, via a Go subset that translates to C.
SpacetimeDB: a short technical review — a sceptical read of the 2.0 launch that separates the genuinely novel design from the benchmark marketing around it.
56% — the cut in AT&T's coding costs from routing work to open models, against a quality drop of about 2% (The Information).
40% → 60–70% — the share of AT&T employee AI queries going to open models today, and where the company intends to take it. It runs about 45 billion tokens a day (The Information).
26% — Apple's new standard EU App Store commission from October 1, down from 30%; 20% for alternative in-app payment processing, 15% for linking out, and a flat 5% Core Technology Commission outside the App Store (Apple).
$6bn — Nvidia's non-exclusive licence for Poolside's model-building software, alongside a $1bn investment at a $12bn pre-money valuation and job offers to 109 staff (Newcomer).
39% — how far below its long-run trend US call-centre employment now sits, against 33% in Canada and 27% in Germany, in Goldman Sachs' August 19 study of where AI is already visible in labour markets (Goldman Sachs, CNBC).
$10M ARR in 60 days — what Lovable picked up after launching in November 2024, while its brand was being rebuilt by a design team of one (Nad Chishtie, First Round Review).
98% — the stability Shopify's rebuilt mobile end-to-end suite reached, after the old one got too flaky to keep in PR checks (Shopify Engineering).
+1,517 — Node.js test-suite tests added in Bun 1.4, alongside over 2,900 fixes, 5× lower idle CPU and up to 35% less memory (Bun).
$17.2M — the OpenAI Foundation grant to SecureBio's Detection division, disclosed alongside a public account of the conflicts it creates (SecureBio).
92.5 kB vs 194.1 kB — Fig's minified runtime against React 19.2.7's, for a comparable minimal interactive surface (Fig).
📰 WHAT HAPPENED — 13
Models & math
Bun 1.4 shipped its largest Node.js-compatibility jump since 1.0, adding image processing, a headless WebView, markdown parsing, cron and terminal primitives (Bun).
Nari Labs published how it serves Qwen3-TTS at 50 ms time-to-first-audio, with the cost curve alongside the latency (Nari Labs).
A researcher documented taking over part of the e164.arpa DNS tree and inadvertently logging hundreds of thousands of phone calls, including ones to military numbers (Lina, lina.sh).
Safety & policy
Apple set new EU App Store business terms effective October 1, replacing the per-install Core Technology Fee with a 5% Core Technology Commission and letting a single app offer both Apple in-app purchase and alternative payments for the first time (Apple).
SecureBio disclosed a $17.2M OpenAI Foundation grant funding its Detection division, and published its reasoning about independence given that its AI division evaluates OpenAI's models (SecureBio).
The NSA, CISA, the FBI, the Department of Energy and the EPA issued a joint advisory on August 19 warning that attackers are using AI to write exploit scripts against internet-exposed Siemens S7 programmable logic controllers in water, energy and manufacturing — scanning with tools like Censys and ZoomEye, then walking in on default or weak credentials. "This is not a theoretical risk — it is an active threat" (NSA, CISA, FBI, DOE, EPA).
Goldman Sachs found AI already weighing on employment across developed economies, concentrated in call centres, software publishing, management consulting and advertising, and hitting entry-level roles hardest (Goldman Sachs).
Business moved
Nvidia agreed to pay Poolside $6bn to license its model-building software, invest $1bn at a $12bn pre-money valuation and hire 109 of its staff, with all three founders staying and the company insisting it is not an acquisition — the third deal Nvidia has structured this way, after Groq and Enfabrica (Newcomer).
AT&T is routing roughly 40% of employee AI usage to open models, cutting coding costs 56% for about a 2% quality tradeoff (The Information).
Meta launched a three-month Startup School for early-stage consumer brands (Meta).
Platforms shifted
Salesforce launched Slack Code, putting coding agents into temporary, auto-archiving Slack channels that non-engineers can follow and review (Salesforce).
Linear's coding sessions gained configurable environments, in-browser testing with before-and-after screenshots, and admin-set spend limits that reset on a schedule (Linear).
Kagi added a single setting to strip paywalled links out of search results, replacing manual per-domain blocking (Kagi).