techSun 20 Sept, 08:00

Sources93 X accounts · techtwitter.com · Hacker News · 8 newsletters
  1. Runs every day at 07:30 London time.
  2. It is shown its own papers of the last 3 days, to know what it has already said.
  3. Haiku 5.5 writes the paper from what was fetched, with no tools.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

93 X accounts

Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.

techtwitter.com

What tech Twitter is discussing, ranked by discussion.

Hacker News

The front page, plus stories over 80 points in the last 24 hours.

8 newsletters

Every new issue, each delivered once, looking back up to 7 days.

The tech daily written for one reader

Node.js got 0.8ms faster

Colin Hacks merged a PR improving Node startup time by seeding V8's random number generator from the OS instead of OpenSSL. Small win, big compounding effect on startup-heavy apps.

The senior engineer death spiral

Sunil Pai's post about a failure mode in engineering: you get stuck trying to do a month's work in a week, tell no one, hit burnout. It's been shared thousands of times. Managers should read it too. The counterintuitive solution: ask for help, or explicitly reduce scope. The post resonated because it's real.

"Jev" branding is worn out

CompleteSkeptic (probably from the Jev team) is tired of the branding. "Our brand is irreverent and insane, not jev"—says it was fun because the name itself was insane, but now it's exhausted.

Kody shows what agent-native tooling looks like

Dodds highlighted documented utilities in Kody: basically a skill + a tool that agents can search for and trust ("merge a pull request"). The goal is reliability and repeatability that plain skills don't guarantee.

RSA-896 factored with AI

The factoring record jumped from 829 bits to 896 in 16 days using AI assistance. Not a practical threat to current RSA-2048, but shows the pace of mathematical breakthroughs with computational help.

Big announcements

Sunday, September 20, 2026 · Pergent2

Muse hitting TV this weekend

Meta's Muse agent product is getting its first TV ad, running during the big game(s). Signs the product is real and has backing behind it.

Kody v2026.09.19 ships experiments flag

Kent C. Dodds' new tool now lets users opt into unfinished features early. Kody itself is about turning skills into code you can run from any agent on any device—or a custom app.

Tin brings full-text search to Postgres

PlanetScale launched Tin, a full-text search extension for Postgres. Clean, native integration without spinning up another service.

Omarchy phone port looks smooth

The Omarchy Linux team is designing a phone UX that sounds ambitious: "the smoothest, most customizable phone experience ever built."

What people are fighting about

Sunday, September 20, 2026 · Pergent3

Inference providers are lying about cache rates

Thdxr called out a pattern: some providers claim 99% cache hit rates, but only Deepseek actually achieves that. The scam is wrapping Deepseek, then later switching to a cheaper model while still claiming the high rates. Also watching out for "bad cache rates" making things look cheap when they're actually 3x the token cost.

Claude Code took way too long to ship AGENTS.md

Gergely Orosz pointing out that Claude Code took 16 months longer than OpenCode and Codex to add AGENTS.md support. The subtext: best-in-class models don't guarantee execution speed. (Context: AGENTS.md is an emerging standard for agent interop that Claude Code was slow to embrace.)

AI scraping is "the largest theft of labor in human history"

Microsoft director called it out in a legal brief. The argument: training on human-created work without consent or payment is theft at scale.

Agents distract when you need control

Matt Poccock: using an agent to plan a course section felt hard—it asked irrelevant questions, jumped to conclusions, created complex assets too fast. With pen and paper, he had control and moved at human speed. Agents seem bad for work where you need oversight of the whole output at once.

What do we even call AI?

Trump considering a new name for AI. Gergely Orosz saying Anthropic's reasoning that agent families aren't exchangeable doesn't hold—open models work fine with closed ones. Brands are in chaos.

Web dev corner

Sunday, September 20, 2026 · Pergent4

Jev as a practical tool

People are rebuilding workflows around Jev (the fast decision model). Venturetwins scanned 1000s of Zillow listings to classify by architecture, renovation, freeway proximity—1,500+ items in 20 seconds for $0.18. Omarsar0 built custom verifiers for agent goals using Jev, running them cheap enough to scale. The thread: it's faster and cheaper than GPT-5.6 on practical tasks.

How decision models differ from LLMs

Aarondfrancis' analogy: LLMs are your dad explaining why the sky is blue (10 minutes). Jev is your mom saying "yes" immediately. The speed comes from skipping reasoning and going straight to decisions.

Supabase docs markdown trick

You can append .md to Supabase docs URLs to get the raw markdown instead of HTML. Small UX win, useful for people syncing docs locally.

Research & open source

Sunday, September 20, 2026 · Pergent5

Decision models aren't new

Laya posted: "I built non-autoregressive decision models with RL a year ago." The HN thread is full of people saying they've done similar work. The Jev moment is less about invention and more about execution and timing.

ZK-JPEG: zero-knowledge image editing

A new paper on zero-knowledge image compression and editing—cryptographic proof of image modification without revealing the image.

Worth a read

Sunday, September 20, 2026 · Pergent6

How to get rich with taste when everything is slop

EXM7777: decision-making is now 100x more valuable than building, because machines can build in seconds. The money flows to people who have taste and can filter, not makers. A practical guide for the next decade.

You should almost never use AI to write

Erich Grunewald's argument: AI-generated writing lacks voice and judgment. It can draft, but it can't choose. Reading it hit differently than the usual "AI is slop" takes because it's specific about where it breaks.

AI-generated posters don't have to be horrible

John Hartnup showed event posters made with AI that are actually good. Not "good for AI," actually good. Sparked real discussion about aesthetics vs. automation.

If math is more than proof, we need to celebrate the rest

Terry Tao on what math research actually is: it's experiments, intuition, computation, and exploration, not just formal proof. A long essay that lands differently depending on whether you've felt trapped by the "proof is all" narrative.