techThu 12 Nov, 07:30

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

Frontend Focus

Frontend tools for real screens

Ahmad Shadeed's new free tool previews layouts inside split views, docked DevTools and in-app browsers.

Ahmad Shadeed's free Site Seeing tool shows your layouts with on-screen keyboards, split views and docked DevTools, the real conditions users have.

Playwright screenshots give agents a visual feedback loop

A Mainmatter post shows using Playwright's toHaveScreenshot so coding agents can check their work visually and locally.

Rowboats lets agents work together in one space

Rowboats is a free shared workspace where Hermes, Claude Code, Codex or OpenClaw agents get a shared goal, plan, write docs and hand off work.

Lock your GitHub Actions to exact versions

Hugo van Kemenade explains how to pin GitHub Actions to exact commits so a changed action can't sneak into your builds. It's a short, practical fix for supply chain attacks.

This Hue Shall Pass fixes failing color pairs

Paste a palette and get every text and background pair graded for WCAG, with nudges to pass while keeping the hue.

Atuin Threads: all your agent sessions on every machine

Atuin's new Threads syncs your agent sessions across machines, so you can find past work from any laptop.

Big announcements

Thursday, November 12, 2026 · Pergent2

Google

Google ships Gemini Decisions API

It is much faster and cheaper than a full LLM, and LangChain has day-one support.

Google's new API lets you hand a model text, images, audio, video or PDFs and get back a structured pick from options you define. Google says it is much faster and cheaper than a full chat model, and the output is reliable. It landed in LangChain on day one.

NVIDIA research

NVIDIA's tiny model matches giants

The 117M-parameter model matches or beats models up to 10 times larger on five benchmarks.

Déjà View takes photos from different angles and rebuilds 3D shape and camera positions. It reuses one transformer block in loops, and you can turn the loop count up or down to fit your compute budget. It is a NeurIPS 2026 oral paper, and the code is out.

What people are fighting about

Thursday, November 12, 2026 · Pergent3

Gergely Orosz: frontend jobs are shrinking

Orosz says fresh data confirms native mobile and frontend roles are dying. He argues you now need to be a general engineer who uses AI to stay employable. The post drew 152 replies and many pushed back.

Raycast

Raycast skips code review

Raycast says it ships straight to main with no code reviews and still passes SOC 2 compliance checks.

Gergely Orosz says Raycast pushes to main without code review and is still SOC 2 compliant, because it validates changes in other ways. Mitchell Hashimoto's reply on the same topic, saying small teams move fast without compliance, drew a strong reaction. The argument: modern practice and compliance do not have to conflict.

Web dev corner

Thursday, November 12, 2026 · Pergent4

Wes Bos: an MCP server for all your apps

Wes Bos is trying RhysSullivan's Executor to connect apps once and reach them over one self-hosted MCP connection.

Node.js ships its November security updates

This follows the heads-up from earlier: Node.js 22, 24 and 26 now have updates for llhttp, undici and sqlite.

Web security

Apache server has a critical RCE bug

The flaw, CVE-2026-119618, lets an attacker run code on a server running Apache HTTP Server.

Wiz published details of CVE-2026-119618, a critical remote code execution flaw in Apache HTTP Server. If you run it, check your version and update. The Hacker News thread has 59 comments on the fix.

AWS

AWS releases an emulator

The AWS emulator lets you run AWS services on your own machine for testing, and it is on the HN front page.

AWS has put out an emulator that lets developers test against AWS services locally, without touching the real cloud. A companion post asks what an engineer's job is when agents write the code.

TLS trust

A certificate authority leaked a key

eMudhra turned off FIPS mode and copied out a certificate private key that should never leave the hardware.

A report says eMudhra, a certificate authority, turned off FIPS mode on its hardware security module and copied out a private key. That key should never leave the device, and the company is leaving the TLS business as a result. It raises the question: how can we tell if other certificate authorities behave the same way?

From the newsletters

Thursday, November 12, 2026 · Pergent5

Tinkerfont inspects and swaps fonts on any page

Tinkerfont is a Chrome and Firefox extension to inspect page fonts and try alternatives directly on the page.

System colors can fail contrast together

Aaron Gustafson found that HighlightText and Highlight reach only 1.98:1 contrast in Chrome on Android dark mode.

Decision models in the browser

Agustin Barrientos looks at small decision models that pick from options you define, including Chrome's experimental on-device Decisions API.

Research & open source

Thursday, November 12, 2026 · Pergent6

Agent research

Agents fail to learn over time

Even Hermes and OpenClaw could not memorize or generalize on a new test, while a small trained 9B model could.

Let's Unify AI's new Continual-ARC benchmark says OpenClaw, Hermes Agent and Prime Agent memorized nothing useful and could not generalize. Training a small 9B model's weights did both. The post drew 515 replies.

Formal proofs

AI proves Signal's security code

Gemini 4 Carbon agents wrote about 6.5 million lines of Lean proofs, and five people finished in four weeks.

Google DeepMind used agents to prove key security properties of Signal's libsignal code, down to compiled machine code. Lean checks every proof. For comparison, verifying the seL4 microkernel took twenty person-years.

Worth a read

Thursday, November 12, 2026 · Pergent7

Travel and AI

United bars booking agents

The ban covers AI agents booking flights on behalf of travelers.

United says AI agents can no longer book flights for travelers. The Hacker News thread has over 500 comments on whether agents should be allowed to act for people.

Agent economy

Agents now make most web traffic

HUMAN Security saw AI agent web traffic grow 8x from August to October.

a16z's Olivia and Justine Moore argue that agents now make up most web traffic and write most code pushed to production. They say businesses must treat agents as customers, not bots to block.