techFri 13 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

Vercel UI SDK

Vercel and shadcn ship UI SDK

It has 50+ behaviors like popups, selection and focus, with components for React, Vue, Svelte, Solid and Web Components.

shadcn says agents are now both the builders and users of interfaces, so the logic of a select box, its focus and keyboard handling, is split out and tested as reusable pieces. Vercel's Guillermo Rauch says agents can build with it, and every app exposes a WebMCP for agents to use.

Split big desktop bundles with lazy imports

Byte Binding's Kettanaito cut load time by breaking huge chunks into smaller lazy-loaded files, a simple checklist worth copying.

Node.js

Node ships security fixes

Seven bugs are rated high, and the worst lets a remote client read other connections' memory.

Node 26.13.1, 24.22.3 and 22.23.5 patch HTTP/2, TLS, zlib and permission-model flaws. Five bugs were reported with help from Claude and Anthropic Research, and one came from Trail of Bits.

Your JS error can lose its message before it is logged

Gábor Koós explains why some errors log as an empty object and what gets lost along the way.

Supply chain

Hijacked npm release passed checks

The backdoored version was published with valid provenance, so the usual trust signal looked fine.

Socket walks through how the Ethersphere bee-js release process was hijacked. It shows that trusted publishing alone does not stop a bad release.

Firefox adds streaming uploads with the Fetch API

Apps can now send request bodies in pieces from a ReadableStream instead of buffering the whole file first.

Big announcements

Friday, November 13, 2026 · Pergent2

Perplexity open-sources Decider v1.3 for free on OpenRouter

The new locate feature finds text and returns exact character positions, and the models are free on OpenRouter for a week.

Open-weight AI

Inkling 2 Small goes open

The model has 594B total parameters, with 17B active for each token.

Thinking Machines released the trained model files on Hugging Face under Apache 2.0, a license that lets anyone use and change them. It takes text and image input and runs on vLLM from day zero, and Cloudflare already offers it on Workers AI.

World models

World model Atlas opens up

Users can build 3D scenes with an agent, then have Atlas generate the final pixels.

Fei-Fei Li's team released Atlas to all users, with a Chisel web app and Codex and Claude Code integrations. The agent sketches coarse 3D geometry, and Atlas handles the polished look.

Edge0-145B promises frontier-level AI on your own computer

A new open model aims to run near frontier quality on a home machine, and the HN thread is active.

Google Gemini

Gemini 4 Hydrogen arrives

It scores 38.2% on Cognition's FrontierCode 2.0, just behind GPT-6.1 Sol.

Google's newest Gemini 4 model is out, following Carbon from yesterday. Simon Willison finds the pricing confusing: Hydrogen is only 25% cheaper than Carbon and Argon, and those two double in price on December 1. ThePrimeagen jokes that the element names are a mess.

Anthropic

Claude joins your video calls

It starts in beta with Google Meet, on Team and Enterprise plans.

Claude sits in on the call and builds docs, whiteboards and other artifacts as you talk. Afterward, the notes, action items and transcript are saved on the meeting's page.

Apple's Foundation Models framework gets a page

Apple published a developer page for its on-device AI framework, and Hacker News picked it up with 194 points.

vLLM hits 58x GB300 throughput on Vera Rubin

vLLM reports DeepSeek-V4.1-Flash at 1.7 million tokens per second per GPU on NVIDIA's Vera Rubin chips, with a score on SemiAnalysis AgentX.

What people are fighting about

Friday, November 13, 2026 · Pergent3

Open source AI

Will open AI survive Nvidia?

dhh's post on the news got 7,550 likes, and Hugging Face says it will keep fighting.

dhh told people to back up open-weight models and some gear to run them, saying the world just lost one of free AI's biggest backers. Clement Delangue of Hugging Face replied that open-source AI will be fought for harder than ever. Sebastian Raschka hopes the Nemotron model work survives, since open data and training recipes are harder to replace than weights.

X and AI hype

Paid hype on AI launches

Two developers say marketers pay for nice posts about new agents.

Peter Steinberger warned that marketers offer money for positive posts about a new agent launching this week. Wes Bos says AI tools are among the best-paying ad slots, and that a paid quote-tweet scene has grown up around them. Both want readers to watch for honest reviews.

AI chips

Claim: banned chips trained China AI

The evidence is a weight-pattern fingerprint, and the files show no outside check of it.

Will Depue says DeepSeek, Moonshot, MiniMax and Ant Group built their newest models on Nvidia Blackwell chips, which US export rules block from China. His method: a new model's starting weights follow a pattern tied to the exact GPU and seed used. The post drew 31,905 likes and 1,358 replies, and the HN thread is long and skeptical.

Web dev corner

Friday, November 13, 2026 · Pergent4

Neon AI Gateway cut frontier-model token use by 73% in one test

The MubitLabs team reports a 73% token cut and 40% speedup on Neon's Lakebase Postgres with no accuracy loss.

Cloudflare rate limits can now weigh requests

The Rate Limiting binding accepts a delta in limit(), so an expensive database search can spend more of the allowance than a cheap lookup.

Turso Cloud BYOC runs inside your own AWS account

Turso's bring-your-own-cloud option keeps the database and storage in your account, with Turso's side holding only billing and metrics.

GraphQL over HTTP spec v1 is out

The GraphQL project released the first version of its spec for sending GraphQL over HTTP.

Durable Objects relocate about 45% faster

Cloudflare's Ashley Peacock cut p50 and p99 latency on cache misses for relocated Durable Objects by about 45%.

Postgres security

Old Postgres bug turned into RCE

Researcher Mehmet Ince chained a type-confusion flaw into full code execution, CVE-2026-6460.

Type confusion means the database treats one kind of data as another, and an attacker can abuse that. The writeup walks through how the bug became full remote code execution, meaning an attacker runs their own code on the server.

Postgres releases

Postgres patches 52 bugs

Neon says its customers were running patched versions within 30 minutes.

New minor releases 18.7, 17.12, 16.16, 15.20 and 14.25 are out, plus 19 RC 2. Always-on Neon computes pick up the update on next restart.

Microsoft outage

Microsoft 365 down for hours

Outlook.com has been out for about 12 hours, and Microsoft has not yet found the cause.

Outlook.com and Microsoft 365 have been down globally for about 12 hours. Gergely Orosz reports that servers keep dying when they come back online. He suspects a thundering herd, where many clients retry at once and flatten a server that just returned. The HN thread has 691 comments.

Nintendo's DMCA sweep takes down emulators and mod-chip code on GitHub

Nintendo removed Switch emulators and firmware from GitHub in one big takedown, with 230 comments on HN.

Research & open source

Friday, November 13, 2026 · Pergent5

Agents may manage their own memory with linked docs

SoMem lets agents keep ordinary linked documents instead of a fixed schema, which the author says is close to how he already works.

Chain-of-thought monitoring is at risk if labs optimize the reasoning text

Sebastian Raschka notes that training on reasoning text can hide what a model is really doing, which makes it harder to check.

Document parsing

ParseBench-2 grades parser errors

It covers 4.3k pages and 330+ error types, and the best parser scores 92 on charts.

LlamaIndex's new benchmark gives each error a severity, so one wrong number costs more than a harmless format change. The severities are open, so anyone can change them.

SGLang gets a consortium of big chip and cloud names

Nvidia, AMD, Google, Meta, Intel, Cloudflare and Modal joined the SGLang maintainers to steer the open inference project.

From the newsletters

Friday, November 13, 2026 · Pergent6

Qwen4-Max is in preview for coding and knowledge work

Alibaba's new model is built for complex coding and knowledge tasks and is now in preview.

RunThruVideo shows that an agent's pull request works

RunThruVideo records proof that a PR from an agent actually runs.

Postgres.js 4.1 adds sql.batch() and query_timeout

The release also restores Node 18 support.

Node Web Worker API is on by default in main

The browser-style Web Worker API is enabled by default on Node's main branch, so expect it in a Node 27 alpha.

An open PR proposes a built-in node:yaml module

The PR would let Node parse YAML without a package.

Grok Build

Grok apps leak data

Claroc checked 37,571 live apps and 94.7% opened without a login.

The publish screen preselected access for anyone with the link. SpaceXSI took down 24 public apps built by its own staff after the report.

vlt 1.7 adds deploy and pins git dependencies

The package manager can copy a workspace and production deps to a standalone folder and now pins git deps to a commit.

Netlify Server is in early access

It is a long-running Node backend for WebSockets and in-memory state.

Node.js

Node tests green threads

A thousand workers use about 1GB, versus 6.7GB for regular worker threads.

James Snell opened a PR for a lightweight worker that shares a small pool of OS threads, with each worker getting its own V8 isolate but no Node APIs. It is early days.

Forges: one typed client for GitHub, GitLab and more

Nuxt's lead maintainer Daniel Roe released one client for issues, PRs, checks and webhooks across many Git hosts.

Hackers reuse AI playbooks across targets

Profero counted over 250 named targets in at least nine countries, mostly for reconnaissance with no data taken.

Super runs 70 agents to review player sessions

Super's agents review more than 100 player recordings a week, up from five to ten by hand.

ElevenLabs adds customer memory and follow-up for agents

The new feature lets AI agents remember customers and automate follow-ups.

Singapore Airlines lets you describe your trip

Its flight recommender reads plain-language requests and returns options, and usage nearly tripled in a year.

MiniMax H3-Image brings 4K generation in about 10 seconds

MiniMax's image model is pitched for fast 4K output, which vLLM-Omni also supports.

ChatGPT's Veo-style Grok Bot joins video meetings

Grok Bot joins video meetings to take notes and answer in chat.

DaVita staff have built over 10,000 agents

Clinical, legal and security staff review use cases, and the company retires agents that duplicate others.

Runway releases its robot-control model for download

The model can be downloaded and used commercially, Superpower Daily reports.

An online shop's support agent replies in about six minutes

After five days of drafts, the agent started sending replies, answering about four in ten conversations in full.

Scribe v3 is ElevenLabs' most accurate speech-to-text

Scribe v3 and Scribe v3 Live are ElevenLabs' newest speech-to-text models.

Fujifilm tests an AI contract reviewer in a sandbox

Fujifilm's legal team is testing an agent that checks contracts against standards and suggests edits, aiming for production in early December.

Worth a read

Friday, November 13, 2026 · Pergent7

Apple

AI floods Apple app review

Synth developers say review has turned into a nightmare.

Synth developers say AI-made apps are making Apple's review slow and hard for everyone.

An old essay: math alone won't make software correct

Ron Pressler's 2016 essay argues that correctness is a practical problem, not just a math one.

Running websites makes one want to quit

A personal post on the tiring side of running your own site drew 56 comments on HN.