techMon 26 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

AI coding

How one dev builds apps

Each feature gets a screen recording before the pull request merges.

Nutlope lists six steps: record a long voice prompt, add visual inspiration, let the agent ask questions, review a first version with voice feedback, check each feature on video, and iterate until happy. The post has 571 likes and 83 replies.

opencode adds background compaction

The coding agent now has an experimental setting that tidies its context in the background, and its weekly active users are climbing fast.

Proc-CLI manages processes and ports for agents

An agent-friendly command-line tool for managing processes and ports.

Why your app shows black bars on iPhone Duo

iOS letterboxes apps built before the new folding Duo existed, because the old developer tools never knew that screen shape.

AI agents

Is the CLI agent done?

Omar Sar says apps like the Codex app now do the job better.

@omarsar0 says the command line is no longer the best way to work with agents. He has been a heavy CLI user and says it hurts to admit it. He argues newer apps make it easier to scale agent work and manage the whole dev cycle. The post has 85 replies on 228 likes, a sign the argument is live.

Server security

One dev's server lockdown

Each server is isolated on Tailscale and cannot reach the laptop or phone without approval.

levelsio asked Claude Code for a full security audit of his sites and servers. He already had automatic security updates and reboots. He also put each VPS (a rented virtual server) on Tailscale, a private network, with rules that stop any one server from reaching the others or his devices.

Big announcements

Monday, October 26, 2026 · Pergent2

AI pricing

AI prices fall up to 80%

Both labs cut prices within a few months, and a long blog post asks why.

A blog post asks why OpenAI and Anthropic cut their LLM prices by up to 80% in just a few months. It hit Hacker News with 278 comments, and it follows the OpenAI discount news from the October 25 digest.

Kimi K3.1 Pro

Kimi resets quotas after complaints

Every plan gets a one-time reset worth 25% of the monthly quota.

Moonshot says its K3.1 Pro plan hit users with rate limits and wasted quota, and admits it should have fixed this sooner. Every subscriber gets a one-time reset, and the team is cutting unneeded subagent calls and shorter outputs so quotas go further.

What people are fighting about

Monday, October 26, 2026 · Pergent3

Anthropic

Claude's cyber refusals

One creator says his video workflow stops mid-task on a cybersecurity video.

Users report Claude refusing reverse engineering and malware removal on their own devices. Andrew Curran says small wording changes can flip a refusal, and that OpenAI calls its similar Codex refusals accidental and is fixing them. He expects Anthropic to do the same.

Code review

Auto-merge for safe changes

Theo says if you still merge most pull requests by hand, you are behind.

Pocock thinks a tool could check, with strict rules, whether a pull request (a proposed code change) is safe to merge without a human. The rules: it only moves code, doesn't change the module graph, and doesn't lower test coverage. He admits it would not work for new features and might fail for bug fixes.

AI models

Gemini's agent problem

In one run it quoted line 173 of a 72-line file.

Theo says Gemini is badly unreliable at agentic work, with made-up line numbers, random IP addresses, invented script arguments and fake function exports. The same week he calls Opus 5.5 the greatest model of all time. The replies (109) are not in the files, so the other side is not shown.

Programming languages

The next agent language

Pocock broke Bend's online playground with three or four prompts.

Matt Pocock says the next great language for agents should be as fast as Rust, have as little design space as Go, and have the strictest type checks ever written. He calls Bend an interesting first try, then says it is not ready. The post has 4,276 likes and 359 replies, so the idea is catching on even where the verdict is harsh.

Web dev corner

Monday, October 26, 2026 · Pergent4

Bun

Dev tools get cheap to clone

Bun's Jarred Sumner says a linter or formatter now takes a day or two to build.

Bun's creator says the whole category of linters, formatters, package managers, type checkers, test runners and bundlers used to be hard to build, and now it can be prompted in a day or two. He says part of Bun's bet was that bundling everything in one tool was too much work for others, and that is no longer true. The post has 1,252 likes.

Web performance

Dropping Next.js for a blog

The blog now scores 100 in every Lighthouse category.

A developer removed Next.js from their blog and got a perfect score in all Lighthouse categories. The post hit 108 points and 62 comments on Hacker News, so it touched a nerve about framework weight.

Security

Dreamhost forces password resets

Dreamhost's status page reports an active incident, and Hacker News posted it with 141 points.

Web hosting company Dreamhost has an active incident that is forcing users to reset passwords. It lands in a week when many devs are rotating passwords and adding two-factor login, after a wave of security scares.

Research & open source

Monday, October 26, 2026 · Pergent5

Google DeepMind

Teaching models to argue ethics

The trained model stopped flipping its answer when the options were reordered.

The setup uses real dilemmas from classic novels and trains the model with RL, where it rates pairs of its own arguments. Its ratings agree with human annotators. Sycophancy and reframing still moved its judgments.

Medical AI

Old CNN beats new LLMs

The Hacker News thread on it has 230 comments.

A lab's post says a 10-year-old convolutional network (CNN) outperforms frontier LLMs at diagnosing skin cancer. It is a reminder that a narrow model built for one job can beat a huge general one.

Worth a read

Monday, October 26, 2026 · Pergent6

AI evals

Benchmarks are shaky

Hamel Husain says the grader fails on many examples, so check them by hand.

Husain looked at popular public benchmarks and found poor quality, next to how much the industry relies on them. He points to Epoch AI's benchmark site, which checks the data, and says to read the examples the grader fails.