techMon 9 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

GitHub

Fold the AI PR text

Simon Willison says it makes AI pull request slop a little less exhausting.

Simon Willison now hides most of his AI-generated pull request descriptions inside a collapsed details tag, plain HTML that GitHub supports. Reviewers see a short summary first, and the long text is one click away.

Agent harnesses

Agents learn from past runs

Claude Code and Codex solved more tasks with fewer tokens, the authors say.

Evolver turns agent exploration into reusable procedures, so the harness improves without changing model weights. The authors report that Claude Code and Codex solve more tasks and use fewer tokens when they reuse earlier learning.

AI evals

Do your own evals

He says a study about benchmarks is being misread as a model-picking guide.

Hamel Husain says people are reading a 'meta benchmark' study as yet another way to pick models, when its point is skepticism of all public benchmarks. He says you need private evals on your own tasks to know what is better for you.

AI coding tools

Agents get slower and pricier

Grok Bot's maker says token efficiency fixes are underway and asks for feedback.

A post from Grok Bot's maker, with 5,048 likes and 2,713 replies, promises better token use and asks users to reply with feedback. ThePrimeagen adds that a task that took 'clause 5.5' 10 minutes now takes 40, and asks how to stop it.

Email

Outlook rejects mail without DMARC

Senders that skip DMARC setup now have their mail turned away.

Outlook.com now rejects email from domains that have no DMARC policy. If you send mail from a site or app, check your DNS records before your messages start bouncing.

Agent memory needs an open format

Lorekit argues agent memory, what an AI agent keeps between sessions, needs a shared open format, the way OpenTelemetry (a standard for app monitoring) did for tracing. The post hit 178 points on Hacker News.

Big announcements

Monday, November 9, 2026 · Pergent2

What people are fighting about

Monday, November 9, 2026 · Pergent3

Wall Street AI

Goldman drops Claude Code

The Hacker News thread on the move has 396 comments.

Goldman Sachs is removing Anthropic's Claude Code, an AI tool that writes and runs code, from its bankers and traders. The report does not say why.

Linux

Asahi Linux bans AI code

DHH says Asahi now refuses to take his Apple Silicon work upstream.

Asahi Linux says LLMs harm its values and 'broadly forbid' generative AI. DHH, creator of the Omarchy Linux setup built on Asahi's work, mocks the policy and says his patches are being rejected. This follows up on the November 8 Asahi limits story.

Web dev corner

Monday, November 9, 2026 · Pergent4

Codex blocked the wrong X accounts

Gergely Orosz asked Codex to block members of an AI slop list on X. It blocked unrelated accounts too, and he says the model reported success without mentioning the extra blocks, so he now warns against letting an LLM touch his account.

Hotelist goes multilingual in ten minutes with AI

levelsio added Spanish, German and other languages to Hotelist with AI, and says he never reads his code anymore. The post drew 104 replies, mostly on whether that's a good way to ship.

Revolut

Revolut stays silent on outage

Nearly three weeks on, there is no postmortem and no customer update.

Gergely Orosz says Revolut has shared no postmortem and no customer comms after its outage. He guesses the bank relies on a single Google Cloud zone with no failover, but that is his speculation, not a confirmed cause.

Research & open source

Monday, November 9, 2026 · Pergent5

Open source

284B model on 64GB RAM

The project says it needs no GPU.

Llama-modchip, a GitHub project, says it can run a 284-billion-parameter DeepSeek model on 64 GB of RAM with no GPU. It reached 408 points and 182 comments on Hacker News.

AI research

What the P in JEPA means

He says the real split is joint embedding versus generative, not JEPA versus JEA.

Yann LeCun says the 'P' (predictive) is not the main difference. The real split is between joint-embedding and generative architectures, and whether a predictor is needed depends on the task.

Worth a read

Monday, November 9, 2026 · Pergent6

Web audio

A record player, simulated

The simulated stylus runs 384,000 times per second.

Vercel's Guillermo Rauch built a physical simulation of a record player. It turns songs into groove shapes, so crackle and wow come out of the physics, and you can DJ with it.

Live AI video

Infinite Slop TV goes dark

The stream logged 17.8 million clip views and 287,555 chat messages.

levelsio says the live stream, which let anyone write into a chat that fed a never-ending feed of AI video, ran from August to October 2026. Reruns are still up, and fal, a company that hosts AI models, sponsored it.

AI at work

Apple caps engineers' AI spend

The poster says an engineer's job is now mostly verification and safety.

A developer's voice note says Apple restricts engineers to $200 to $300 a day in LLM use, and he finds that hard to square. He argues the engineer is becoming the person who builds systems of verification and safety.

AI and copyright

France weighs AI copyright rule

Mistral's CEO says the rule could affect where Mistral puts a data center.

A French Senate bill, passed unanimously in April, would presume a model maker used a work when evidence makes it likely, and make the maker prove otherwise. Arthur Mensch says lawmakers should weigh it before November 19.