Skip to content
Source · Daily Brief

AI Daily Brief — 20 May 2025

The largest single-vendor AI announcement bundle of 2025. Google I/O shipped or previewed everything from a research-mode reasoning model to a watermark-detection tool, a 1,000-2,000 tok/s text-diffusion model, a 3D telepresence platform with HP, and a $249.99/mo VIP subscription. The Gemini app crossed 400M MAU.

Top stories

  • Sundar Pichai opens I/O 2025 — Gemini 2.5 Pro tops LMArena across categories; Gemini app 400M MAU. Elo scores up more than 300 points since the first Gemini Pro. via Google
  • Gemini 2.5 Pro ‘Deep Think’ enhanced reasoning mode. Explores multiple hypotheses in parallel before responding. Scored highly on 2025 USAMO, led LiveCodeBench for competition-level coding, hit 84.0% on MMMU. Initially offered to trusted testers via the Gemini API on Vertex AI. via DeepMind
  • Veo 3 — first major video model with native synchronized audio. 4-8 second clips up to 4K at 24fps with joint audio-visual generation — synchronized dialogue, sound effects and ambient noise produced in a single pass, lip-sync within ~120ms. Available to US AI Ultra ($249.99/mo) subscribers and on Vertex AI. via CNBC
  • Imagen 4 — sharper text, 2K resolution, fast mode 10x faster than Imagen 3. Sharper text rendering, photorealistic detail, output up to 2K. All outputs carry invisible SynthID watermarks. Available via Gemini API, AI Studio, Vertex AI and Workspace. via TechCrunch
  • AI Mode in Search rolls out to all US users. No Labs sign-up. Dedicated AI Mode tab in Search and the Google app. Under the hood: ‘query fan-out’ — breaking a question into subtopics with multiple parallel searches. Deep Search and Project Mariner-powered agentic capabilities follow in Labs. via TechCrunch
  • Project Astra becomes agentic — controls Android, comes to Search and Gemini Live. ‘Universal AI assistant’ with improved memory, computer-use control of an Android phone, and enhanced voice output. Real-time video and screen-sharing to all Gemini Live users. via TechCrunch
  • Project Mariner expands — runs in the cloud, 10 parallel tasks. Gemini-powered agent that browses and uses websites. Runs on cloud VMs (no longer pinned to a local Chrome tab) and can handle up to ~10 simultaneous tasks. Ships first to US AI Ultra subscribers; capabilities exposed in the Gemini API and Vertex AI. via TechCrunch
  • Gemini Diffusion — text-diffusion model at 1,000-2,000 tok/s. Generates text and code by denoising rather than left-to-right token prediction. 4-5x faster than Google’s fastest public LLM while matching its coding performance. Experimental demo with a waitlist. via DeepMind
  • Project Starline becomes Google Beam — 3D video calls with HP. AI volumetric video model turns standard 2D streams into realistic 3D via a light-field display and six cameras. Ships to early customers (Deloitte, Salesforce, Citadel, NEC, Duolingo) later in 2025 through HP. Google Meet also gains near-real-time speech translation (English-Spanish first). via Google
  • Google AI Ultra at $249.99/mo. Bundles Gemini 2.5 Pro Deep Think, Veo 3, Flow, NotebookLM (highest limits), Project Mariner access, YouTube Premium and 30TB cloud storage. New subscribers got 50% off for three months. Gemini Advanced rebranded Google AI Pro at $19.99/mo. via CNBC

Who shipped

Google ran the day end-to-end across model + product + media + subscription tiers. HP, Duolingo, Salesforce, NEC, Citadel and Deloitte got launch-customer billing on Beam. Microsoft Build Day 2 ran in parallel with Windows AI Foundry and Microsoft Discovery deeper detail.

Open-source pulse

Gemma 3n preview — open mobile-first multimodal model handling audio/text/image/video inputs, 140+ languages, up to 32K context, ~1.5x faster on mobile than Gemma 3 4B with smaller memory footprint via Per-Layer Embeddings (PLE). 4B model nests a 2B submodel for on-the-fly quality/speed tradeoff. Open weights, commercial use allowed. Lyria 2 + Lyria RealTime open Google’s music models to developers and YouTube Shorts.

Money, infra & hardware

Jules — Google’s autonomous coding agent powered by Gemini 2.5 Pro — enters public beta inside GitHub. Asynchronous: given a task, Jules clones the repo into an isolated cloud VM, plans, edits, runs tests and returns a PR. Free during beta with usage limits. Google Flow AI filmmaking tool ships exclusively to US AI Pro ($20/mo) and AI Ultra ($249.99/mo) subscribers.

Quiet corners

SynthID Detector launched as a portal that scans images, video, audio and text for invisible SynthID watermarks. Google said 10B+ pieces of content have been SynthID-watermarked across Gemini, Imagen, Lyria and Veo. Stitch — Google Labs UI design tool from the Galileo AI acquisition — generates mobile/web designs and exports production-ready code in seven frameworks. No major OpenAI, Anthropic, Meta or xAI launch on the date — Google’s I/O blitz was the story.

By the numbers

  • 400M MAU — Gemini app monthly active users
  • +300 Elo — Gemini 2.5 Pro improvement since first-gen Gemini Pro
  • 84.0% — Deep Think MMMU
  • 4-8 sec / 4K / 24 fps / ~120ms — Veo 3 clip duration / resolution / frame rate / lip-sync
  • 2K / 10x — Imagen 4 max resolution / fast-mode speedup
  • 10 — Mariner parallel cloud tasks
  • 1,000-2,000 tok/s — Gemini Diffusion speed
  • $249.99 / $19.99 per month — AI Ultra / AI Pro
  • 10B+ — SynthID-watermarked content items
  • Most-mentioned company: Google

Compiled by AI Feed’s editor from verified web sources for 20 May 2025.