AI Daily Brief — 22 April 2025

EU AI Office set the August enforcement clock with preliminary GPAI guidelines. Google’s antitrust trial inadvertently confirmed the Motorola–Perplexity–Microsoft tie-up. And Cursor’s AI support bot got caught inventing a policy that didn’t exist — a viral cautionary tale about agent-only customer support.

Top stories

• EU AI Office publishes preliminary GPAI guidelines. European Commission’s AI Office clarified the scope of obligations for general-purpose AI providers, including a 10^22 FLOP training-compute threshold for presumed GPAI status ahead of the Aug 2, 2025 enforcement deadline. Stakeholder consultation open until May 22. via EU AI Act tracker
• Google VP confirms Motorola+Perplexity+Microsoft AI in antitrust trial. During Google’s antitrust remedies trial, a Google VP testified that upcoming Motorola phones would ship with AI features from Perplexity, Microsoft and Google — the first public confirmation of Perplexity’s Motorola integration before its formal Apr 24 announcement. via Bloomberg
• Cursor AI support bot caught fabricating a device-limit policy. Cursor’s customer-support agent ‘Sam’ invented a one-device-per-subscription rule and emailed it to a frustrated user; the post went viral on Reddit/HN, triggering subscription cancellations and a public apology from co-founder Michael Truell. via WinBuzzer
• USC publishes ‘Tina: Tiny Reasoning Models via LoRA’ on arXiv. A 1.5B base model fine-tuned with LoRA during RL matches or surpasses state-of-the-art reasoning models at ~$9 post-training cost — >20% reasoning gain and 43.33% AIME24 Pass@1. via arXiv
• Apple and Google mark Earth Day. Apple highlighted Apple 2030 carbon-neutral roadmap (60%+ reduction in global GHG emissions vs 2015) and a 10%-off accessory promo for device recycling. Google’s homepage Doodle featured satellite imagery (Maldives atolls, French Alps, Quebec, Utah canyons, NSW Australia) under the ‘Our Power, Our Planet’ theme.

Who shipped

No frontier-lab model launches. EU AI Office ran the policy headline. Cursor ran the cautionary tale.

Open-source pulse

The Tina paper sharpens the small-model reasoning thesis — $9 post-training for a tiny model matching frontier reasoning is the strongest evidence yet that the reasoning-as-RL recipe scales down dramatically.

Money, infra & hardware

No major standalone funding rounds dated Apr 22.

Quiet corners

Frontier labs (OpenAI, Anthropic, Google DeepMind, Meta, xAI) quiet on the model side. Chinese labs quiet ahead of Qwen3.

By the numbers

• 10^22 FLOPs — EU GPAI presumption threshold
• Aug 2, 2025 — GPAI enforcement deadline
• May 22 — EU consultation close
• $9 / 43.33% — Tina total post-training cost / AIME24 Pass@1
• 1 — fabricated device-limit policy that took Cursor down
• Most-mentioned entity: EU AI Office

Compiled by AI Feed’s editor from verified web sources for 22 April 2025.