So you want to use OpenRouter? (606 pts)

The top story today is a brutally honest field report from the operator of Olly, an iMessage AI assistant that has routed over 18 million messages — roughly a third through OpenRouter on open models. The post catalogs ten hard-won lessons that amount to one thesis: the abstraction OpenRouter sells you is leaky, and the leaks will bite you in production.

The most damning finding: the same model weights (DeepSeek V4 Flash 0731) score 90% GPQA and 81% TAU-Bench through first-party DeepSeek, but drop to 75% and 58% through DigitalOcean. That’s a 23-point swing on tool-calling reliability — the metric that actually matters for agents. Vision models are worse: DeepInfra’s Qwen endpoint read a letter K as R and called red blue, while Venice and Together silently swallowed image inputs entirely, returning 200 OK with no content. The reasoning.effort knob, which should control compute allocation, is ignored by several providers including DigitalOcean and Venice.

The post also demonstrates that declared quantization (fp4 vs fp8) is a poor proxy for quality — fp4 hosts land in the middle of the fp8 pack — and that pinning providers is fragile. The author pinned three “reliable” providers and still hit total outage when Baidu started rate-limiting, Cloudflare dropped the model, and Alibaba followed with 429s. The lesson: filter on benchmark boards, not on metadata claims. Test from production infra, not your laptop. And parse tool calls on your end because provider-side parsers will silently fail.

Source: mmoustafa.com

Claude is only available to people over 18 years (454 pts)

Anthropic updated its support documentation to clarify that Claude’s consumer product requires users to be 18 or older. The system uses automated detection to flag accounts suspected of belonging to minors, then requires age verification through Yoti — a third-party platform offering three methods: facial age estimation via selfie, government ID upload, or a pre-verified Digital ID attribute.

Anthropic claims it never sees the ID or image data — only a pass/fail result from Yoti, which is SOC2-audited and deletes verification data immediately after processing. The HN discussion (505 comments) predictably split between those seeing this as responsible AI deployment and those viewing it as surveillance creep dressed up as safety policy.

The real tension here is regulatory. The UK’s Online Safety Act, the EU’s Digital Services Act, and various US state-level age verification laws are all pushing platforms toward exactly this kind of gatekeeping. Anthropic is getting ahead of enforcement, but the approach raises the same questions that have plagued age verification across the internet: facial estimation accuracy varies by demographics, ID verification excludes people without government documents, and the “pass/fail only” architecture is only as trustworthy as Yoti’s audit trail. For a tool increasingly used as a productivity workhorse, locking out anyone under 18 is a meaningful access restriction — especially in educational contexts.

Source: support.claude.com

Google will buy half the electricity from one of Finland’s nuclear power plants (268 pts)

Google announced a €13 billion investment in Finland — its largest single investment in Europe — to build three new data centers and expand an existing site in Hamina. The headline grab: a 22-year power purchase agreement with Fortum for up to 50% of the output from the Loviisa nuclear power plant, which currently generates about 10% of Finland’s electricity. Construction is slated for 2027-2028, and Google claims it will support 37,000 jobs during construction and add €3.6 billion annually to Finnish GDP.

Finland checks every box for data center siting: cold climate (reduced cooling costs), low-carbon grid, uncongested power infrastructure, and stable governance. TikTok announced a $1 billion Finnish data center investment the same week. The Nordic countries are quietly becoming the compute backbone of European AI infrastructure.

The nuclear angle is the real story. Tech companies are signing ever-longer power agreements directly with generation assets rather than relying on grid electricity or even renewable energy certificates. A 22-year commitment to a nuclear plant is effectively co-investing in the plant’s lifetime extension — Fortum explicitly said Google’s commitment supports their investment program to extend Loviisa’s operating life and increase capacity. This is infrastructure-level vertical integration: the hyperscalers aren’t just buying compute, they’re buying the atoms that power it.

Source: BBC

Cherenkov Radiation (205 pts)

The IAEA published an explainer on Cherenkov radiation — the blue glow visible in nuclear reactor cooling pools — timed to Pavel Cherenkov’s birthday anniversary. The physics: when charged particles (typically electrons) travel faster than the local speed of light in a medium like water, they emit a shockwave of photons analogous to a sonic boom but in the visible spectrum. The high energies produce short wavelengths that skew blue and violet, with significant ultraviolet output invisible to the naked eye.

Nothing exotic — this is standard nuclear physics education. What makes it worth reading is the practical application section. IAEA safeguards inspectors use next-generation Cherenkov viewing devices (XCVDs and DCVDs) to verify spent fuel inventories at nuclear facilities. By measuring the Cherenkov glow from spent fuel ponds, inspectors can cross-check whether declared fuel quantities match reality — a direct tool for detecting nuclear material diversion.

It’s a useful reminder that nuclear verification depends on fundamental physics, not just paperwork. The IAEA’s mandate under the NPT relies on these kinds of ground-truth measurements to maintain the integrity of the non-proliferation regime. In an era where arms control agreements are fraying, the unglamorous work of safeguards inspection matters more than ever.

Source: IAEA

Throughline

Three of today’s four stories orbit the same gravitational center: the physical infrastructure tax that AI is levying on the real world. The OpenRouter post exposes how fragile the model-serving layer is — dozens of providers running the same weights with wildly inconsistent quality, silently failing tool calls, and ignoring parameters. Google’s nuclear deal is the supply-side response: lock in generation capacity at the atomic level because the compute demand is non-negotiable. And Anthropic’s age gate is the governance layer catching up — once AI tools become critical enough to warrant identity verification, they’ve crossed from novelty to infrastructure. Cherenkov radiation, sitting quietly at the bottom, is a reminder that the physics underneath all of this hasn’t changed: charged particles still obey the same rules they did in 1934, and the blue glow is still the most honest signal in the room.