Daily Digest – 17 October 2026

Morning from Oslo. Today's biggest story is close to home: Nscale's flagship AI data center in Glomfjord is six months behind schedule. PostgreSQL 19 also slipped after AI review found more bugs than humans could fix, and Perplexity open-sourced the sandbox it uses for agents.

Tech sector

Nscale's Norwegian flagship is six months late ahead of a $35B IPO

Bloomberg reports that the Glomfjord AI data center has only 6.4 of a planned 30 MW running, after cooling-system leaks, a transformer fire and permitting problems. Nscale owes penalties to Microsoft and ByteDance while it pitches investors on $103B of contracts.

Why it matters: Cheap hydro power made Norway the poster child for sovereign AI compute. This is the reality check. Getting power is easy, and building fast in a Nordic fjord is hard.

Source: DCD · 16 Oct · Bloomberg

PostgreSQL 19 slips to around 19 November after AI review surfaces a bug backlog

LLM-assisted review found a long tail of edge-case bugs, and 8–10 significant features were pushed to PG 20. Committer Peter Eisentraut says the quality bar went up, and that vibe-coded patches did not cause the delay.

Why it matters: Plan your upgrades around PG 18 for now. This is also a preview of what many teams will face soon: AI finds bugs faster than humans can fix them, so triage becomes the bottleneck.

Source: The Register · 16 Oct

Apple runs an 816B open model on four Mac Studios with MLX

Apple ML Research shows Muse Spark Open decoding at more than 36 tokens/s across four M5 Ultra Mac Studios. The machines are linked by RDMA over Thunderbolt 5 and use tensor parallelism.

Why it matters: Frontier-class inference on a desk, with no US cloud in the loop, is a real option for GDPR-sensitive teams in health, finance and the public sector.

Source: Apple ML Research · 15–16 Oct · HN discussion

Waymo announces Toronto, with first riders targeted for 2028

Waymo starts with manually driven mapping runs in Toronto. Local partner Element Fleet Management will run fleet operations. Paid robotaxi rides are about two years away.

Why it matters: Toronto means snow, slush and a cold-climate city. If Waymo handles it, Oslo and Stockholm become realistic markets.

Source: Waymo blog · 16 Oct

AI agents

Perplexity open-sources SPACE, its Firecracker sandbox for agents

SPACE gives each agent task its own microVM and injects credentials at the network layer, so secrets never enter the guest. It is written in Rust and licensed under Apache 2.0. In Perplexity's own red-team test, nine models with root access found no sandbox escape. A partial-network bypass worked 11 of 54 times before it was patched.

Why it matters: If you run coding agents in CI or for customers, this is a production-grade isolation layer you can host yourself.

Source: GitHub · 16 Oct · Red-team report

Four Gemini 4 Argon agents beat Minecraft from a random spawn

DeepMind researchers had four agents control Mineflayer bots and coordinate only through chat and shared memory. They killed the Ender Dragon in about 27 hours, wrote 2,783 JS programs, died 605 times and settled into roles on their own.

Why it matters: Agents splitting up roles on their own in a long task is the multi-agent pattern people are trying to build into real dev workflows.

Source: Paper (Hugging Face) · 16 Oct

StarSkirmish Bench: LLMs write StarCraft bots and mostly lose

Nine flagship models each get an hour to write a Brood War Protoss bot. GPT-6 Astra leads, but the best model wins only 7% of games against top human-written bots.

Why it matters: This balances the Minecraft result. Under real-time, adversarial pressure, code written by agents is still far behind code written by experts.

Source: StarSkirmish · 15–16 Oct · HN discussion

How the Claude Code team reviews code when Claude writes about 90% of it

The team layers pre-PR checks, blocking AI and security review, CI and feature flags, and scales human approval by risk tier. Merged PRs more than doubled from June to September while the revert rate fell by over 40%.

Why it matters: This is a concrete playbook for having at most one human per PR. A small product team can copy it without a dedicated platform team.

Source: claude.dev · 15–16 Oct · HN discussion

Hacker News buzz

Tesla Roadster

Tesla's Roadster demo showed the car jumping a semi using fan-based downforce and cold-gas thrusters. Production is slated for 2028.

What commenters argued about: Fans called it the coolest demo in years. Skeptics called the physics a gimmick and pointed to deposits taken in 2017 for a car promised for 2020.

Why it matters: Norway is Tesla's favourite EV market, so expect to hear a lot about this one here.

1,049 points · ~1,900 comments · HN discussion

90% of US working hours sit outside AI benchmarks

Mercor mapped benchmark prompts to job tasks and found that most work time has little or no benchmark coverage. High-wage jobs are overrepresented in what is covered.

What commenters argued about: Many said real white-collar work lacks tight feedback loops. Developers said their time goes to coordination and politics more than to typing code.

Why it matters: This is useful evidence the next time someone says AI will replace developers next quarter.

350 points · 302 comments · HN discussion

Claude traces where the NIST elliptic-curve seeds came from

Anthropic says Claude found that the seeds for P-192 through P-521 came from lrand48 seeded with a Unix timestamp, and were not hash outputs. This does not break the curves, and it makes an NSA backdoor in those seeds less likely.

What commenters argued about: Some called the backdoor question settled. Skeptics noted that timestamps could still have been searched for a weak curve. A side thread complained that model safety filters block legitimate security work.

Why it matters: P-256 sits under your TLS, BankID-style auth and JWTs. An AI agent helping solve a decades-old crypto mystery is worth reading about.

546 points · 235 comments · HN discussion