Daily Digest – 4 November 2026

Morning from Oslo. Today's biggest story is a European fintech showing that self-hosted coding models can work at scale. Apple also opened Siri to app developers, and Hacker News spent the day angry at Postman.

Tech sector

Revolut cuts AI coding costs ~90% by self-hosting GLM

Revolut moved most of its coding-agent traffic off Claude Code and Codex onto self-hosted GLM-5.3, running on 56 Nvidia B200s at roughly $0.30 per million tokens. It serves about 2,000 engineers and keeps the commercial models as a fallback.

Why it matters: This is a large European company showing that open-weight coding models hold up in production. If your team has to think about cost or where its data lives, this gives you a concrete example to point to.

Source: The Register · 3 Nov

AMD beats on data center, guides ~$17B next quarter

Q3 revenue rose 45% to $13.38B, and data center now makes up 64% of AMD's sales. Guidance came in well above analyst expectations, and the stock rose more than 5% after hours.

Why it matters: Stronger competition for Nvidia should make GPU capacity cheaper over time. That includes the hydro-powered Nordic data centers that are competing for AI workloads.

Source: SiliconANGLE · 3 Nov

Big Tech's less-visible AI debt reportedly hits $3.2T

Nikkei Asia reports that AI-related obligations at US tech giants have grown about 90% in six months. Separate reports say lenders are now charging higher premiums on this debt.

Why it matters: If the AI spending boom slows, cloud and AI API prices could change quickly. Products that can switch providers are better protected.

Source: Nikkei Asia · 4 Nov (paywalled)

Safari 27.3 beta adds Adaptive Dark Mode

When a user turns it on, Safari darkens pages that only have a light theme. Sites that declare color-scheme: light dark are left alone, and light only opts out.

Why it matters: If your site has no dark theme, Safari may now improvise one for you. A one-line CSS change keeps you in control.

Source: WebKit blog · 3 Nov

AI agents

Apple opens Siri AI to every developer via App Intents

The iOS and macOS 27.3 betas include a developer preview that lets Siri trigger an app's actions and chain them into multi-step "skills". Apps can be submitted to the App Store with this in the spring.

Why it matters: Apps with clear, narrowly scoped actions (transit, banking, delivery) will be easy for the phone's built-in agent to use. Now is a good time to define yours.

Source: Apple Developer · 3 Nov

Apple's annual filing lists rogue agents as a business risk

Apple's 10-K warns that autonomous agents could go beyond their permissions, ignore instructions, or cause harm that is hard to undo.

Why it matters: When one of the most cautious companies in tech lists this as a risk, it is time to treat narrow permissions, audit logs and undo as basic requirements. That also fits how the EU AI Act approaches risk.

Source: CNBC · 3 Nov

GPT-6 Astra designed a hardware device in under a week

OpenAI says the model produced the enclosure, circuit board and firmware for a small Codex companion device, and prototypes were made within a month. It needed several rounds of revisions because it missed manufacturing tolerances.

Why it matters: Agents are moving beyond writing code into CAD and firmware work. Hardware teams can prototype faster, but people still have to make sure the design can actually be manufactured.

Source: The Decoder · 3 Nov · OpenAI post

AI agents now run SETI's Allen Telescope Array

Agents built on Nvidia Holoscan point the telescopes and analyze radio data as it arrives, cutting some tasks from weeks to hours. The team reports detecting a pulsar in real time.

Why it matters: This is a rare example of agents running live scientific operations. It is directly relevant to research facilities like EISCAT 3D up north.

Source: The Decoder · 3 Nov

Hacker News buzz

Postman silently auto-installs AI plugins and tries to share API data

A GitHub issue reports that Postman added AI plugins without asking and tried to send workspace API data to them. It reached the top of HN in less than a day.

What commenters argued about: Most blamed bloat driven by investor pressure and recommended Bruno, Yaak or plain .http files instead. A smaller group argued that every paid tool eventually goes this way.

Why it matters: API collections contain tokens and internal endpoints, so this is a GDPR and security problem as much as an annoyance. Check what your dev tools are sending, and where.

1,186 points · 391 comments · HN discussion

Google will turn off Search for schools that reject its AI

Schools that disable Google's AI tools for students will also lose Google Search on those accounts.

What commenters argued about: Many called it coercion and compared it to the Internet Explorer antitrust case. A side debate asked whether AI summaries are actually more useful than ranked results, given that they sometimes contradict their own sources.

Why it matters: Many Norwegian schools run on Chromebooks and Google Workspace, so it is worth watching whether this policy reaches Europe.

601 points · 427 comments · HN discussion

"The Astra weight-reading post is misleading"

Matthew Honnibal argues that the viral post about an LLM "reading and writing weights" to improve Qwen overstated its result. In the Qwen experiment the model wrote programs that edited the weights, and reading raw weights directly was only shown on tiny networks. (I covered the original claim in yesterday's digest.)

What commenters argued about: One side said the framing was close to malpractice. The other said the paper itself is clear and only the title goes too far. Several pointed to earlier work like Tracr.

Why it matters: The original post got around 2.9K points. It is a good reminder to read the method before sharing the headline.

956 points · 199 comments · HN discussion · original post