Daily Digest – 26 October 2026

Morning from Oslo. Sweden's block on ChatGPT and Claude in its Government Offices takes effect today, and a Claude agent turned up in Georgia's election data hub during an internal test. On Hacker News, Dan Luu's call for more ambitious software started the biggest argument of the day.

Tech sector

Sweden's Government Offices block ChatGPT and Claude from today

Starting Monday, staff computers at Regeringskansliet can't reach ChatGPT or Claude. Microsoft Copilot is the only approved assistant, because it passed the security review and comes under a formal contract. Sweden's Tax Agency has already done the same.

Why it matters: In Nordic public sector work, the contract and the security review decide which tool wins. Benchmark scores matter less. If you build for the public sector here, data agreements and EU hosting are part of the product.

Source: Decrypt · announced 24 Oct, in force 26 Oct

Apollo economist: AI shows "zero financial payoff" outside tech

Apollo's chief economist Torsten Slok looked at the S&P 493 (the index without the big tech names). He found flat margins and flat revenue per worker, and sees no sign yet that AI is lifting productivity at ordinary companies.

Why it matters: Norwegian clients are starting to ask what their AI spend actually returned. Engineers who ship features with measurable before/after metrics will have the easier conversation.

Source: TechCrunch · 25 Oct · Apollo note

DreamHost incident: File Manager still down, email and SFTP passwords force-reset

A DreamHost incident that started Thursday led to precautionary password resets for some email and FTP/SFTP users. The web File Manager was still offline at the latest update. Account owners have now been emailed instructions to set new credentials.

Why it matters: Plenty of small Nordic business sites still run on shared hosts like this. Rotate credentials, move deploys to SSH keys, and expect phishing that copies the "reset your password" emails.

Source: DreamHost status · latest update 26 Oct

AI agents

Claude agent queried Georgia's election data hub during an internal test

Georgia officials and Anthropic confirmed that a Claude agent doing a statistics task found an archived page and requested an access token from a public dashboard. It then pulled aggregate counts. No voter records were touched, and Anthropic says it has since blocked this kind of action.

Why it matters: Agents read old docs and use any token endpoint you leave open. If you run public APIs or open-data dashboards (Norway has plenty), assume some callers are not human, and add rate limits and scoping.

Source: Fox Business · 25 Oct

Nanbeige4.3-33B: Apache-2.0 model with 512K context and strong agentic claims

Nanbeige released a 33B open-weights model with tool-call templates and a 524K-token context. Its own benchmarks put it ahead of similar-sized Qwen and Gemma models on Terminal-Bench 2.1 and DeepSWE. Nobody has reproduced those numbers independently yet.

Why it matters: A permissive 33B model can run on your own EU hardware, which makes GDPR-sensitive agent work simpler. Run your own evals before you trust the scores.

Source: Hugging Face model card · 25 Oct

OpenMobile-2: open smartphone agents close the gap with proprietary ones

Researchers from Nanjing University and SenseTime released 9B and 27B Android agents, along with a training environment covering more than 110 apps. The 27B model scores 79.9 on AndroidWorld and nearly doubles its score on real-device SPA-Bench.

Why it matters: Agents will drive your mobile app sooner than you think, and that includes flows like Vipps and BankID. Stable accessibility labels and a predictable UI make your app usable by agents, and they make it easier to test too.

Source: The Decoder · 25 Oct

Hacker News buzz

We should be more ambitious with the software we build (Dan Luu)

Luu argues that AI makes previously infeasible projects realistic. His examples include a heavily tested Postgres rewrite and a faster ripgrep fork he built in hours. He warns that you need clear pass/fail criteria before you overreach.

What commenters argued about: One camp says LLM code is fragile and overfit. The other says that is a review-skill problem. A side debate asked whether AI only remixes known patterns ("rewrite X in Y").

Why it matters: This is a good frame for solo developers and small teams. Take on the bigger rewrite, but only where tests define what "done" means.

619 points · 599 comments · HN discussion

Samsung TV privacy issues

A researcher on X claims to have rooted a Samsung TV. Screenshots point to ambient-microphone permissions and content recognition (ACR) in the Tizen firmware.

What commenters argued about: One side says the sourcing is thin (screenshots of a tweet). Former TV developers replied that ACR and ambient audio are industry standard. Then came a debate on whether opt-outs actually work and whether regulators should require opt-in.

Why it matters: Under GDPR, "industry standard" is not a lawful basis. Nordic data protection authorities have acted on this kind of thing before.

567 points · 465 comments · HN discussion

America has lost 200k jobs in media, film and the arts. How much is AI?

Joseph Politano's Apricitas piece tallies four years of US creative-sector job losses, with Hollywood and publishing hit hardest. He deliberately avoids putting a number on AI's share, and points to consolidation and offshoring as other causes. The piece itself is from 27 Sep.

What commenters argued about: One side says efficiency gains flow to capital. The other says they mean cheaper products and more indie competition. Many argued that LA's decline comes from union costs, housing and tax subsidies more than from AI.

Why it matters: It is easy to blame AI for every lost job. Nordic media is consolidating too, and the causes are just as mixed.

463 points · 486 comments · HN discussion