techFri 30 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 security

Node2Shell: update now

A bug in the bundled Brotli decoder can allow remote code execution.

Apps that decompress untrusted Brotli data, for example through node:zlib, are at risk. Node's October security releases (26.12.1, 24.22.1 and 22.23.4) fix it along with nine other CVEs. This follows our Oct 28 note on the Node security release.

A skill fix stops Codex subagents from forking the parent context

Matt Pocock's fix stops most subagents from inheriting the parent's context window, which matters for skills that need a clean start.

AgentBaton: switch to a cheap model once, mid-task

An NVIDIA paper cuts API cost by 65.9% on average across 11 model pairs while largely keeping accuracy.

GitHub Copilot

GitHub's cheaper coding mode

GitHub says Copilot Fusion Max matches Claude Code with Opus 5.5 Max while using half the tokens.

Fusion starts tasks on simpler models and moves up to stronger ones only when needed. GitHub's COO says early enterprise users see 50 to 60 percent lower costs on complex work.

Kody adds 'profiles' with limited permissions for agents

A profile gives an agent a token with a subset of your account's permissions, which is safer and easier to audit, says Kent C. Dodds.

Build evals before you call anything self-improvement

Omar Sarar argues agents need concrete measures to hill-climb, and that agents write weak evals.

Big announcements

Friday, October 30, 2026 · Pergent2

Open weights

Mistral Large 4 goes open

The 1-trillion-parameter model ships with day-one support in vLLM and NVIDIA's stack.

Mistral released the weights of Le Chonk, its new mixture-of-experts model (a design that wakes only some of its parts for each word, to save compute). It ranks first among open models on Cognition's FrontierCode coding test and posts strong cyber-defense results.

Anthropic safety

Claude checks its own chats

The check runs right after each reply and has caught misuse that other safeguards missed.

Anthropic calls the method self-assessment. Claude gives a quick safety read of the chat after it responds, and because it reuses what the model already read, it is cheap and should improve as models do.

OpenAI

Codex limits double

Pro 100, 200 and 500 tiers and Business Premium get double usage from today.

OpenAI says the change follows user feedback, and Pro 200 keeps its current limits. The company has also started a sprint to make its app simpler and faster.

What people are fighting about

Friday, October 30, 2026 · Pergent3

Coding agents

Agent coding mania

One post drew 10,214 likes, and a second says most developers miss flow state with agents.

Anthropic's Thariq says there is a kind of mania with coding agents, where you feel busy but nothing real happens, often as avoidance. Gergely Orosz reports that most developers he asked miss flow state, except those building agent systems themselves.

Agentic dev

Why devs are wary

The Hacker News thread on it has 131 comments.

Martin Fowler's post asks why strong engineers resist agentic development. The question sparked one of the longer discussions on Hacker News today.

Git workflow

Kill the pull request?

The post got 796 likes and 133 replies, a sign of a real debate.

The post argues that pull requests only make sense as a CI trigger, and that teams can push to main with trunk-based development. Defenders of pull requests say they are needed for code review and knowledge sharing.

X and bots

Ban AI replies?

The post has 4,292 likes and 1,064 replies, a quarter as many replies as likes.

Developer levelsio wants a zero-tolerance rule that suspends any account that posts an AI reply, saying bots are an existential threat to the platform. The thread is hot because many people feel replies are already flooded with bots.

AI race

Is distillation bad for America?

Lambert says the answer is not cut and dry, while Xu argues distillation helps the US lead.

Distillation trains a small model on a big model's outputs. Kevin Xu argues it benefits consumers and locks in America's lead, and says more attention should go to labs that do not distill.

Web dev corner

Friday, October 30, 2026 · Pergent4

JavaScript runtimes

Deno heads to Cloudflare

A RedMonk analyst worries the runtime field is narrowing to Node, Bun and workerd.

Node Weekly reports Deno's runtime is set to wind down after its team's move to Cloudflare. Deno says it is working toward a foundation, and a fork called Ende has appeared. This follows our Oct 28 foundation news.

Supabase

Postgres gets failover

Private alpha is open to organizations on paid plans.

If a database fails, a standby copy takes over on its own, and data the app already saved is not lost, Supabase says.

Server security

PHP2Shell chain roots 25 servers

The chain reached US government servers, according to the write-up.

Researchers chained the PHP2Shell flaw into an exploit that rooted 25 servers, including some US government machines. It follows our Oct 28 note on Wiz tracking PHP2Shell attacks.

From the newsletters

Friday, October 30, 2026 · Pergent5

Gemini 4 Carbon

Carbon tops the index

Carbon scores 59, one point above Claude Opus 5.5.

Artificial Analysis says Carbon is the new number one on its index v5, ahead of Opus 5.5 (58) and GPT-6 Astra (57). It trails on OSWorld and FrontierSWE, and this follows our Oct 29 launch note.

GitHub

Agents start 40% of PRs

That is up from 17% in March, about 110 million agent-started pull requests in one month.

GitHub's annual report says most of these figures measure drafting involvement, not autonomous delivery or verified code quality.

Banking AI

UBS staff run 12,700 agents

The bank also reports 608 AI use cases in production.

The disclosure shows agents are widespread in the bank, but it does not say how many people use each agent or what they add to results.

Research & open source

Friday, October 30, 2026 · Pergent6

YC Labs

Agent swarms play Minecraft

Researchers saw bureaucracy, thousands of short messages, and OpenAI models that read the game's source code.

YC Labs and olam let agent swarms share long-running game worlds to see what dynamics emerged. OpenAI models retrieved the game's source code to beat Claude.

Worth a read

Friday, October 30, 2026 · Pergent7

AI assistants

Claude Duet feels useful

The HN thread has 115 comments.

The post argues Duet is the first personal-assistant harness, the wrapper that runs the model, that actually helps day to day.

Software design

Carmack wants agent-friendly libraries

The post drew 3,797 likes and 256 replies.

AI can write big projects from scratch, but that is rarely the fastest or safest path. He argues for a standard library shaped for agents, possibly an odd new design.