AiiN — AI news, hourly

Two Hacker News posts on harnesses and agent.md files reveal where real agent quality now lives.

Enterprises favor cheaper, faster models over Anthropic's flagship, signaling price beats benchmarks in production.

Oxford Economics ties a widening US-Europe capex gap to the AI infrastructure buildout, not just cyclical spending.

Local opposition to AI data center construction is turning into a real electoral liability nationwide.

The chipmaker's potential bet follows Perplexity crossing $750M in annualized revenue, per Techmeme.

A second departure in a year shows how fluid research talent has become at the frontier AI labs.

A new arXiv framework scores AI agency delegation by stack layer, from app permissions to chip-level control.

New research quantifies how splitting compute between generation and critique changes self-refinement outcomes.

The OpenAI CEO's warning about AI centralization is a cue to avoid single-vendor lock-in.

A reported $13B sale talk for Hugging Face raises questions about open-source AI's future.

Multi-agent AI systems fail from missing management structure, not weak models, a new analysis argues.

An unclaimed model called Ox Alpha is topping benchmarks, and nobody will say who built it.

Enterprise card data suggests spend follows price more than hype, and Fable 5 is losing share to Opus 5.

Courts keep splitting on AI training lawsuits — here's what that means for anyone building on these models.

The AI needed a human to point out its own rules before it would actually act on them.

Agentic token usage on OpenRouter jumped 14x, signaling autonomous AI is moving from demos to production.

A global memory chip crunch is adding roughly 15% to the cost of Nvidia-based AI servers.

A new study complicates the AI-productivity story in research, with real stakes for R&D ROI calls.

Resellers are undercutting Anthropic's official pricing, revealing how much unmet demand exists in China.

A protocol standardizing tool access and an agent that watches live code both attack the same problem.

Vera Rubin and Grace Blackwell pricing is set to jump in 2027, reshaping multi-year AI infrastructure budgets.

A DeepMind-alumni startup claims its agent outperforms rivals at reproducing scientific results.

Every token you send to a foreign LLM API is a line item someone else prices — and can reprice.

The chipmaker is now backing data center construction directly, not just shipping GPUs into finished buildings.

Bloomberg finds cheap, capable Chinese models pulling enterprise buyers away from US API providers on cost alone.

A lab known for fighting AI regulation is now asking California to make its safety bill tougher.

None of the top AI labs has published a plan for containing a model that slips control.

One base agent, cloned into a full office cast, exposes what real multi-agent orchestration actually requires.

An unnamed lab's model went viral on OpenRouter with claims of 1M-token context and 100T tokens a day.

No camera, no computer vision — just live text overlays, at a lower price and privacy cost.

A new study maps why structured skills boost AI agents, and where the approach quietly fails.

A little-known data supplier is minting AI-boom revenue faster than most software companies ever have.

New research finds that predicting human actions requires modeling what they believe, not just what they do.

Netflix is testing a language model to replace decades of hand-tuned recommendation logic.

SemiAnalysis finds the open-closed AI gap now halves every model era, not just every year.

New research shows psychological techniques bypass AI safety filters that standard red-teaming never catches.

TTS synthesis just got fast enough to disappear from the conversation — shifting the real bottleneck elsewhere.

After a purge of senior executives, Brockman now controls both product and scaling at OpenAI.

Meta's new Mac app lets Meta AI act inside other applications via voice or text commands.

Ramp built an internal router that picks the cheapest capable AI model for each task, no vendor lock-in.