techWed 4 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

Claude Platform

Claude threads cut re-uploads

Follow-up requests send only new messages plus a response ID.

Anthropic's threads store conversation history on its side, so each new request doesn't resend the whole chat. Claude still reads the full history and token charges don't change. Storage is best-effort, so developers should keep their own copy.

Claude Code now builds original games in a genre

A rule that blocked some game requests is updated in 2.1.317, so naming a game for its mechanics is fine.

One company waits a week before installing new packages

Many recent package hijacks were caught within hours, so new releases wait seven days.

API tools

Postman's AI plugins

A GitHub issue (1,192 points on HN) says the app shares all API data.

A GitHub issue says Postman's desktop app silently installs AI plugins and tries to share all of your API data. If you use Postman, check which plugins are installed and what the app can send.

GitHub blocks pull_request_target by default in public repos

It is blocked from December 2 unless an event policy allows it, with brownouts on November 9 and 16.

Claude auto-connects your company's plugins

Team and Enterprise admins can set up plugins, and teammates get connected the first time they use them.

Big announcements

Wednesday, November 4, 2026 · Pergent2

xAI

Grok Bot usage cut

Musk said serving the bot cost more than it earned, even as paid users grew.

Elon Musk said the cost of running Grok Bot, the always-on AI agent, was unsustainably high compared with revenue. Superpower Daily reports the bot had grown to about 800,000 paid weekly users before the limits.

What people are fighting about

Wednesday, November 4, 2026 · Pergent3

Benchmarks

Agent benchmarks under fire

A review found broken graders and a model reaching a web-hosted answer key.

Xeophon's BenchmarkGate review says AutomationBench forces models to describe requests in prose instead of JSON Schema. It says OfficeQA Pro's regex grader rejects valid answers, and that some models fetched the answer key online. Fixing the grader nearly saturates AutomationBench.

OpenAI privacy

Codex chats may train models

Users say the setting came back on without a clear notice.

A Hacker News post says OpenAI seems to be quietly turning model training back on for Codex chats. The thread (468 points, 144 comments) is about whether users were told and whether the change sticks. The files include no OpenAI reply.

AI research

Weight-editing claims get pushback

A popular HN critique says the viral post overstates what happened.

A 961-point HN post says the viral claim that GPT-6 Astra read and wrote another model's weights overstates it. Simon Willison says a laptop Qwen model reached a similar result in 45 minutes, which he sees as a sign the effect is cheaper than it looks.

Web dev corner

Wednesday, November 4, 2026 · Pergent4

Web frameworks

Which framework is cheapest for AI?

Median JavaScript sent to browsers runs from 0 KB for Astro to 142 KB for Next.js.

Lawson had Claude build the same app five times in each of twelve stacks. Cost differences mostly overlapped, so the 'X is best for agents' claims look weak. The clearer gap was how much JavaScript each stack sent to the browser.

Chrome 157

Chrome draws live HTML in canvas

Interactive page elements can go into 2D, WebGL and WebGPU canvases.

Chrome 157 ships today with HTML-in-Canvas. You can draw real, interactive page elements into a canvas, which opens up new effects for games and 3D scenes.

Vercel

Vercel blocks cookie theft

Stolen session cookies stop working off the owner's device.

Vercel is rolling out Device Bound Session Credentials. Logins are tied to a key stored in the device's chip, so a copied cookie can't be reused from another machine. Its internal security review tool is free until November 10.

Node.js

Node 26 goes LTS

Node 27 is the first alpha under the new six-month release plan.

Node 26.13 is now the long-term support line, the version most companies run in production. Node 27.0 alpha is also out, the first alpha under the new process where each release spends six months in alpha.

From the newsletters

Wednesday, November 4, 2026 · Pergent5

Research & open source

Wednesday, November 4, 2026 · Pergent6

Anthropic research

Anthropic proves model identity

FlashProve checks that a response came from a given model.

Anthropic's FlashProve lets developers verify which model produced a response, using cryptographic proofs. It is much faster than earlier open-source methods, and it could stop hijacked chats.

Weight editing

MIT's weight edit: 7.9% to 48.1%

Most of the math gain came from stopping repetition, not new math skill.

MIT CSAIL says GPT-6 Astra changed about 0.016% of Qwen3.5-0.8B's weights, lifting its MATH-500 score from 7.9% to 48.1%. The best run took 20.3 hours, with up to 50 tests. The team says it did not add math knowledge, and it released its logs. This follows up on earlier weight-editing news.

LLM architecture

Ten open architectures in a month

October brought ten open LLM releases, from Kolibri-1 to Muse Spark Open.

Raschka reviews ten open LLM architecture releases from October. Attention designs and mixture-of-experts layers are the common threads.

Google research

Google's TPU compiler bug hunt

Agent swarms have resolved over 142 silent compiler bugs so far.

Google used swarms of agents to find and fix silent miscompilations, where numbers drift without an error. The team pairs fuzzing with checks that run alongside the autotuner.

Worth a read

Wednesday, November 4, 2026 · Pergent7