techWed 21 Oct, 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 makes any agent harness durable

Vercel's durable-harness runs Pi, OpenCode, Codex and Claude Code so they resume after crashes and can run for days. It ships adapters for the popular harnesses and works self-hosted too.

Node.js security releases land tomorrow

Node 26.x, 24.x and 22.x get security fixes, the highest rated high.

Kent C. Dodds on why legacy codebases need verification

His view is that the problem is confidence, not speed, and the fix is letting agents verify their work like humans do: run it, use the UI, seed data.

TypeScript 7.1 beta adds content mappers

Content mappers let TypeScript read files like .md or .vue and give them static types, so you can import typed frontmatter directly. Colin Hacks calls it a game changer.

Hugging Face's Agent Manager orchestrates many agents

It runs dozens of agents across Claude Code, Codex and OpenCode, and keeps them going on HF Spaces.

Big announcements

Wednesday, October 21, 2026 · Pergent2

Perplexity's agent can now run fully on your Mac

Perplexity Computer runs its model and agent locally on a Mac, with no cloud and no WiFi needed. Sensitive local files never have to leave the machine.

ChatGPT adds Locked Chats behind a PIN or Face ID

You can move sensitive chats into a protected space. New messages there do not update the memory used in your normal chats. It rolls out to most plans starting today.

DeepSeek open-sources DeepInfer, its own serving engine

DeepSeek released DeepInfer, the software that runs its V4.1 Flash model at scale, under a free Apache 2.0 license. It is built mostly on vLLM and it hit number one on Hacker News with 2,200 points.

Mistral releases Voxtral 3

Mistral's new speech model, Voxtral 3, went to the top of Hacker News with almost 600 points. Read the launch post for the details.

vLLM open-sources Humming, one kernel library for every format

Humming is a single library of fast kernels, the low-level GPU code that does the math, for all quantization formats and GPUs. vLLM and Ant Group say it gives up to 5x kernel speedups and 14 to 43 percent more MoE throughput.

What people are fighting about

Wednesday, October 21, 2026 · Pergent3

Did Opus 5.5 get dumber, or is it a skill issue?

Users say Claude Opus 5.5 feels worse. Theo's Nerf Bench says it is actually 5.8 percent smarter than at launch. Riley Brown adds that only Claude and GPT feel usable to him, which gets a lot of replies.

Is harness engineering dead? Omar Sarar says no

Some recent papers say models are absorbing the harness, the code around them. Omar Sarar calls that a claim built on weak tests on benchmarks that models already ace, and says a stable minimal harness still matters.

Hugging Face's CEO says a Netflix doc got the AI attack wrong

Clem Delangue says a Netflix documentary about a cyberattack on Hugging Face used too much doom and skipped how the team defended itself. Kevin Roose and Arvind Narayanan's circle say the 'AI is normal technology' view looks odd after a rogue-agent attack.

Web dev corner

Wednesday, October 21, 2026 · Pergent4

Next.js patches two critical RCE flaws

Two remote code execution bugs were patched: one in use cache functions from a React Server Components bug, and one in the image optimizer's SVG handling. Upgrade to 16.4.1 or 15.5.28 because a public proof-of-concept already exists.

The future of CSS, through Interop 2027

Josh Comeau walks through what browser makers plan to make consistent next year. It is a practical look at what can be used without hacks.

tsrs: TypeScript's compiler ported to Rust, up to 8x faster

Cloudflare and VoidZero ported the Go-based TypeScript 7 compiler to Rust. It runs as a drop-in replacement, is experimental, and can run in WebAssembly. Steinberger says it made his OpenClaw setup about 5x faster.

eMudhra's intermediate certificate keys were tiny numbers

eMudhra, a certificate authority that issues the trust that makes padlocks work, leaked intermediate private keys that anyone could recover in seconds. With them, a scammer could mint a fully trusted certificate for any site, such as irs.gov, and sit in the middle.

OpenClaw's frontend moved from Lit to Solid 2.0 in about 600 PRs

Peter Steinberger ran an agent loop for two weeks. The result is a smaller JS and CSS footprint and finer update control, with checks every other day.

Google Cloud outage takes down European banks and Revolut

A Google Cloud outage hit many banks in Europe and the UK, including HSBC, Revolut and Bank of Scotland. Revolut had a 7-plus hour incident, and Gergely Orosz notes that Monzo has a stand-in bank for exactly this case.

From the newsletters

Wednesday, October 21, 2026 · Pergent5

OpenCode Console breach exposes AI keys

A breach of the OpenCode console exposed AI API keys and led to unauthorized charges.

Mercury 3: a diffusion model that writes many tokens at once

Inception's Mercury 3 generates several tokens in parallel, reaching about 1,000 tokens a second. It matches Haiku and GPT-6 Luna class models on quality and costs 10 cents in and 35 cents out per million tokens during its launch promotion.

AI agents probed 1,200 sites and broke into more than 100

Reuters reports one hacker used autonomous agents against over 1,200 websites and compromised over 100 companies' sites, including Deere and Cummins. How much data was stolen is still unclear.

Research & open source

Wednesday, October 21, 2026 · Pergent6

Anthropic finds MLP neurons are easier to read than assumed

Across 13 open transcoders, many late-layer features are close copies of single MLP neurons, and only those keep a stable link to the underlying network.

DeepSeek V4.1 Flash beats GPT-6 Sol on an offensive cyber task

Vals AI says DeepSeek's 1017 model beat GPT-6 Sol, which was tested via OpenAI's trusted access program, on a multi-step cyber task. Full results come in the next weeks.

Worth a read

Wednesday, October 21, 2026 · Pergent7

Publish fake stats, get cited by LLMs

Peec.ai ran a test where made-up statistics spread across the web and got cited by AI models. It is a strong example of how AI search can be gamed, which is also what agencies are chasing.