techThu 22 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

Node.js ships security fixes, and Vercel blocks the attack

New Node.js releases for versions 22, 24 and 26 fix several bugs, some high severity. Vercel put firewall rules in front of all its deployments for today's Node RCE (remote code execution), and will update Node for its managed sites. If you self-host Node behind Vercel, update today. This follows up on the heads-up that these releases were coming.

Lighthouse for Claude Code

Enter a URL and get a CLAUDE.md fix plan built from a PageSpeed Insights audit.

Chrome finishes deprecating the unload event

Unload handlers no longer fire by default. Use pagehide or visibilitychange instead.

pour: a free accessibility checker

It runs as a browser extension or bookmarklet and sorts violations by severity.

Claude Code 2.1.304 ships 76 CLI changes

The release includes new features and a long list of command-line fixes.

A debugger for scroll-triggered animations

A Chrome DevTools extension draws the active ranges of a scroll animation on a minimap and lets you edit them.

Big announcements

Thursday, October 22, 2026 · Pergent2

Omnara: an open-source alternative to Claude Managed Agents

A new open-source project aims to do what Anthropic's managed agent service does, but on your own setup.

Anthropic opens Claude Duet, a proactive personal agent

Anthropic opened a research preview of Claude Duet, a separate app where Claude works as an agent. It shows you what needs attention, keeps working when you step away, and checks with you before it acts for you. It runs on Opus 5.5 and is on a US waitlist for Pro and Max plans. Early users are enthusiastic: Dan Shipper says he now splits his time roughly 50/50 with ChatGPT.

Flint: Recursive's open coding model runs on your laptop

Recursive released Flint, a model with 40 billion total parameters that uses only 2.5 billion per word, so it is fast. It runs at 125+ tokens per second on a MacBook Pro, and its open weights are on Hugging Face under Apache 2.0. The team says its own automated research system helped build nearly every part of it in under four months.

Spacesheep: a shared home for agent work

Spacesheep (YC F26) lets agents put their output in a space your team can read and edit together.

Claude Security now re-scans your code as it changes

Claude Security connects to your repos once, runs a deep first scan, then re-scans as changes land. It is in public beta on Claude Enterprise, and eligible orgs get $25K a month in usage credits through December 31.

Mistral Large 4 is out

Mistral released Large 4, its newest large model, and it is a top story on Hacker News. My sources only have the headline and the crowd reaction, so the announcement post has the specs.

OpenAI models are coming to Google Cloud

OpenAI's frontier models and Codex agent workflows will appear in Google Cloud's Model Garden. They already run on Azure and AWS, so teams can use them on the cloud they already use.

NVIDIA's GTC Berlin: agent guardrails, open models and a robot duck

At GTC Berlin, NVIDIA added an Agent Knowledge Foundation to its Agent Toolkit, plus OpenShell, a reference design that controls what agents can touch. It also released more open models for European teams. Its Nemotron 3.5 Parse model took the top spot on the ParseBench document test.

Hermes Agent builds a training course for a robot duck on stage

Jensen Huang closed his keynote with Nous Research's Hermes Agent. It planned an obstacle course, had other tools build it in simulation, and trained thousands of virtual ducks until a control policy worked for Hugging Face's Microduck robot.

What people are fighting about

Thursday, October 22, 2026 · Pergent3

Are AI benchmarks just marketing now?

Reducto's Adit Abraham says benchmarks have become marketing. He says one published benchmark's provider gets much worse than rivals once documents are blurred or tilted a little. Hamel Husain agrees, calling it BenchmarkGate and saying we need evals on our evals. Why it matters: test models on your own documents before you trust a leaderboard.

Are AI bug hunters outpacing the people who patch?

The claim: publicly logged security bugs (CVEs) went from about 4,000 a month to about 15,000 since April, and investor @saranormous says that is not business as usual. Gergely Orosz says breaches that once took three or four months now seem to happen in a day. The pushback is that the PHP bug is less scary in practice: one widely shared post argues there is no easy exploit for it. Why it matters: how fast you can update is now the real race.

Did Duet kill the proactive agent startups?

Some timeline posts say Anthropic 'destroyed all proactive agent startups.' omarsar0 disagrees, saying personal agents will come in many flavors, and that trust and personality will matter more than speed. Alex Finn calls Duet the one he is most excited about. The real fight is not whether personal agents win, but which kind of personal agent people trust.

Are you 'qualified' to build public software yet?

Matt Pocock's three steps say you should build personal software, then internal software, then public software. His post drew 2,218 likes and 109 replies. Many readers pushed back on the word 'qualified,' so he changed it to 'confident.' The point stands: the stakes and the testing should go up as your audience grows.

Does bad architecture kill startups?

Gergely Orosz claims he has never seen a startup fail because of its architecture. He says they fail to find users, and that bad tech can be fixed once traction arrives. The post drew 173 replies, which is a sign that people are split. Why it matters: it argues against polishing code before you have users.

Web dev corner

Thursday, October 22, 2026 · Pergent4

Railway got hacked, and deploys stayed off for hours

Railway, a hosting platform, published a security disclosure after a breach. Andrew Curran says attackers went after AI keys stored there. Theo says deploys were still off about 17 hours later, and that switching hosts would cost him about three times as much. Greg Isenberg's advice: rotate every AI key tonight and set a hard monthly spend cap.

Motherload imports data files into TypeScript with types

Colin Hacks' library imports .md, .toml, .yaml, .csv and more with full type information, built on TypeScript 7.1 beta content mappers.

LiteLLM 1.107.0 on PyPI was compromised

A bad version of LiteLLM, a popular Python library for talking to many AI model APIs, was published on PyPI. It installs a credential stealer and a backdoor. If you installed it, check your version, rotate your keys, and audit the machine.

An AI found the PHP bug, and sites are going dark

OpenAI's GPT-5.6 Sol helped find critical PHP flaws, part of its Patch the Planet effort to fix them before public disclosure. Cloudflare says it saw thousands of exploit attempts before the announcement, tens of thousands in 45 minutes after, and now millions. The NSA told PHP users to update or take systems offline, and the Dutch cyber centre's own site went down because it runs on PHP. This follows the earlier PHP release story: update PHP now, and treat firewall rules as a stopgap.

From the newsletters

Thursday, October 22, 2026 · Pergent5

Fit an element's height to its background pattern

Ahmad Shadeed uses calc-size() with round() to snap an auto height to a multiple of the tile height.

Databricks' MemCurrent gives agents a memory that lasts

MemCurrent is a memory layer that lets agents keep knowledge across sessions. In an internal test, task success rose from 22% to 68%, and on three public benchmarks it matched giving the agent every past chat while using 12x fewer tokens. It is in Genie now, and Agent Bricks Memory is in beta.

Amazon's Kiro improves itself, with a referee

Amazon's coding agent Kiro now learns from its own runs. A tracing system proposes changes, a deterministic check decides which ones work, and a human approves. In week one, seven of eight proposed changes shipped, and one raised long-task success from 82% to 92%.

Safari blocks 'identity resolution' trackers

Cross-site cookies are gone in Safari, so trackers moved to turning the email you type into a cross-device ID. Safari now blocks that infrastructure at the domain level.

An MIT paper says AI is getting cheap at theory research

The share of theory problems where AI matched the human authors' conclusions rose from 3% (o1) to 35% (GPT-5.5) in sixteen months. Compute cost per correct answer fell from about $170 to $15, and the researchers say AI costs under 0.3% of a human expert.

Research & open source

Thursday, October 22, 2026 · Pergent6

Qwen3-Coder writes less secure code for a 'US persona'

Hirundo found that Qwen3-Coder produces less secure code when it is told the user is a US persona. The cause is not explained in the summary, so read the post before drawing conclusions.

OO-Harness: specialized harnesses that evolve themselves

A Google Cloud AI Research paper builds a class hierarchy of specialized harnesses from past runs. With the same Gemini 3.5 Flash model, it lifts AutomationBench from 61.6% to 77.4% and beats Meta-Harness on all three benchmarks tested.

Perplexity's JOLT: one model as teacher and student

JOLT trains one model to be both teacher and student, with the teacher learning from its own attempts while staying close to what the student can use. On MATH it matched GRPO's accuracy with 6.5x fewer training tokens. NVIDIA co-authored the work.

Mirendil claims pretraining without a pretraining team

Mirendil, formally announced less than four months ago, says its mission is to solve recursive self-improvement and use it to democratize frontier AI. Its post says pretraining at scale works without a dedicated team. The feed copy is cut off, so read the thread for numbers.

Worth a read

Thursday, October 22, 2026 · Pergent7

AI tools cite big sites, and small sites rarely get quoted

Tibo ran an experiment asking ChatGPT, Perplexity, Gemini and Claude the same questions, then checked 1,329 cited sources. The median domain rating was 86, and only 7% of citations came from sites under 30. Trust is built slowly, so small sites are at a disadvantage.

Hassabis says AGI is two to three years away

Demis Hassabis told the WSJ his AGI timeline has shortened from three to five years to two or three years.