techFri 6 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

Claude Code

Claude Code shares your email

A GitHub issue against Anthropic's coding tool reports it, and the thread went near the top of Hacker News.

The issue says the tool sends your email address to sites you visit. The files only give the issue title, so read the thread for what is confirmed. If you use the tool on sensitive sites, it is worth checking the thread before you trust it.

Matt Pocock's rule: ask how to make a change verifiable

The prompt makes the agent sign up for services and click through the frontend. Quality goes up.

Stroq audits what your coding agent did and why

A local tool that records a coding agent's actions so you can review them later.

Simulate.ai tests agents on every pull request

Runs your agent against real-style scenarios with sandboxes and fake data, one click per PR.

Rune, a new project from OpenCode's team, is on GitHub

The author invites people to point an agent at it and build something with it.

Supabase shows how to give your app its own MCP server

A demo adds an MCP server with an OAuth 2.1 server in Supabase Auth, so customers can use it from agents.

Big announcements

Friday, November 6, 2026 · Pergent2

Perplexity

Perplexity launches Teammates

Perplexity says it has about 10,000 active agent teammates inside its own workspace.

Teammates are personal work agents. You give a goal, and they run jobs using your tasks, skills, memory and tools. They can message coworkers in Slack. Max users get them on the web first, with native apps coming soon.

Thinking Machines

Inkling 2 Small arrives

It scores 31 on the AA Intelligence Index, a big jump on its predecessor.

You can fine-tune and run it today on Tinker, the company's platform. Full weights are planned soon. The company calls it a big step up for a small model.

OpenAI

OpenAI ships GPT-6 Cyber

Select Daybreak Red customers get it first, and access widens over time.

A model for security research that pairs stronger skills with tighter alignment. Mozilla, Palo Alto Networks, Cisco and OpenAI's own team are already using it, including on Linux kernel work. OpenAI says it was trained to stay inside the scope the operator sets.

Anthropic

Claude gets Personal Finance

Claude can see your balances but cannot move or invest your money.

Connect the accounts you pick and ask money questions in plain language. It is in beta in the US on Pro, Max and Ultra plans, rolling out gradually. Anthropic says there are no ads and it is not built to give investment advice.

What people are fighting about

Friday, November 6, 2026 · Pergent3

Benchmarks

Can you trust AI scores?

Hamel Husain says some models have already memorized the answer key.

Husain has been making this case for a while. Today he says teams should run private evals, since public tests have flaws. Swyx wants trusted third parties to rebuild evals with fresh problems that look different. A new finance benchmark, DeepFinance, has 21 problems that take experts 20 to 40 hours each.

Browsers

Is Safari really slower?

A popular developer says Safari trails on speed, features and site support.

Peter Steinberger says Safari is slower, has fewer features and weird bugs, and some sites don't work at all. The post drew 418 likes and 93 replies, so the argument is live. Safari is the default on every iPhone, so its gaps reach a lot of users.

Web dev corner

Friday, November 6, 2026 · Pergent4

Vercel

Vercel adds Schedules

Agents and apps can set up their own recurring jobs, even in preview.

Define static jobs once in code, or create them at runtime. They run in any environment. Guillermo Rauch calls them cron on steroids, since they scale and sit outside the sandbox.

Node.js

Node 27 alpha breaks things

Under the new schedule, the final release is due next April.

The alpha is built from main with breaking changes, so it is for library authors and testing, not production. V8 is upgraded to 15.6. The default heap limit becomes half of physical memory, clamped between 256MB and 4GB, and counts Buffers. structuredClone and postMessage use a format older Nodes cannot read. npm 12 is bundled.

Desktop apps

Discord ditches Electron

The new app uses less RAM, according to The Verge.

Discord is moving its desktop app off Electron, the framework that wraps a web page and a copy of Chrome into an app. Electron is often blamed for memory use. The move drew 366 comments on Hacker News.

Security alert

Alpine SSH login bypass

Any Alpine, OpenWrt or musl box with password logins should update now.

Attackers can get past password login on the SSH server (OpenSSH). Alpine has shipped a fixed package. OpenSSH 10.6p2 includes a workaround for Alpine builds without PAM. Vercel's Guillermo Rauch says a reproduction used a wrong password, then a password over 256 characters, and then got root.

Research & open source

Friday, November 6, 2026 · Pergent5

Hugging Face

Free MoE training handbook

The team spent 800,000 GPU hours on experiments and wrote everything down.

Thom Wolf says 18 of the 20 best open models are Mixture-of-Experts, and dense models are basically over. The handbook covers questions like how many experts to use, why half a training step can be spent in communication, and why a run can be 25 times slower with no error.

Benchmarks

DeepFinance tests AI on deals

The 21 problems take human professionals 20 to 40 hours each.

Experts from top PE and investment banking firms built problems from real deal data. The benchmark asks whether an agent can do diligence on a $100 million-plus deal, not just answer simple questions.

Worth a read

Friday, November 6, 2026 · Pergent6

Cyber

Exploits now almost free

The longest cyber task AI can reliably finish went from 5.5 minutes in late 2024 to 4.8 hours by April 2026.

Attacks used to land about 63 days after disclosure in 2018 and 2019, and 7 days before in 2025. The median time to fully fix a known-exploited bug rose to 43 days. a16z expects a long worm era, hitting hospitals and small water systems hardest.

From the newsletters

Friday, November 6, 2026 · Pergent7

Express's October security roundup covers the Brotli flaw

It lists body-parser and multer DoS fixes, and says to upgrade Node or use the workaround.