Самое крупное обновление GOOGLE

Scale of Google’s AI usage (tokens / compute)

• Google’s AI “token” processing is described as a unit of account for AI workload.
• In 2024, it was cited as ~10B tokens/month across platforms, while Google is currently processing about 3.2 quadrillion tokens/month.
• Token consumption is described as growing roughly 7× per year.
• Google has 13 products with 1B+ users each (examples mentioned: YouTube and Gemini).

Gemini rollout across Google products (product-wide integration)

• Google plans to embed Gemini into “almost every” service, aiming to make it helpful “everywhere and in everything.”
• Google Maps example: users can ask for practical tasks (e.g., where to buy a dress within a time window). Gemini interprets and uses relevant data, including payment/card-related details.
• YouTube “ask-first” interface (beyond recommendations):
  • Instead of only searching, you ask questions, and YouTube selects the best video.
  • Claimed advantage: YouTube has transcripts, so the system can understand video topics in advance.
  • The UI is described as a chat/LLM-like experience, including:
    • a mini-transcript
    • a breakdown by key points
    • “Click the video inside the chat” to open it at that moment
  • Context/memory claim: YouTube can remember the ongoing conversation.

Availability / regions

• The new YouTube + Gemini chat mode is described as rolling out this summer in the US.

Search upgrades

• Housing search demo: users describe requirements; Google searches across social networks and rental sites, not just a single site.
• Proactive reminders: the system can notify users when something new appears.
• Plans to add interactive “graph/visual” style answers similar to an Anthropic-style feature (interactive, informative visuals vs plain text).
• Example themes mentioned for interactive visuals: learning/science, including gravitational waves.

Google Docs / productivity via chat

• Gemini in docs enables users to:
  • talk live while Gemini drafts documents
  • use data from:
    • Google Drive
    • Email
    • additional online sources
• Gemini can rewrite in simpler language and add structure such as tables.
• “Create without touching the keyboard” capability is expected this summer.

New and upgraded Gemini models / multimodal generation

• “Gemini Omni” is described as generating “anything” from input data (compared to Gemini content-generation tooling).
• Referenced multimodal/video/world tools:
  • Veo (video generation)
  • Nanobana (image generation/editing)
  • a model for virtual worlds (walk/fly)
• Demo concept: prompt like “Make a clay animation explaining protein folding.”
• Video editing: users can upload a video and ask to add/remove/change elements (analogous to image editing workflows).

Detecting AI-generated / edited content (Synthetic ID)

• Google introduced/expanded “Sint ID” (“SYN ID”) to label/augment search results with content provenance metadata.
• Example use:
  • identify whether a photo was edited via Nanobana or processed through Google Photos.
• Search can then answer questions like “was it generated?” and add context (e.g., whether it circulated online).
• The approach is suggested to become more common due to industry adoption (mentions Nvidia, Open, and “Elon Laps” as parties implementing similar IDs).

Model performance

• Gemini Flash 3.5 is described as a fast model (“thinks less”) but with stronger benchmark performance than a slower 3.1 Pro.
• It’s positioned as very fast, and a review is promised (creator notes upcoming content).

“Antigravity” / agent tooling and developer features

• Antigravity 2.0 updates include:
  • voice input
  • improved agent dialogue/orchestration
  • splitting complex tasks into sub-tasks handled by multiple agents
  • a new Antigravity CLI (not to be confused with “GMI CLI”)
• Integration mentioned with:
  • Android
  • Firebase
  • Google Studio
• Additional reference: using the engine inside Google Search for dynamic behavior and code execution.

Autonomous agent: Gemini Spark

• Gemini Spark is described as a personal AI agent running on dedicated Google servers.
• It uses Antigravity and MCP to connect to third-party services (framed as “OpenAI-like only in Google’s cloud”).
• Availability: only for AI Ultra subscribers.
• Additional mentions:
  • Spark added to Chrome so it can browse/act for the user
  • A Universal Card / single basket concept: users collect items from YouTube / Gemini / mail into one shared basket that checks price and stock in the background
  • Spark/agent capabilities later integrated into the Gemini app and other Google surfaces

Gemini app UX and settings

• Gemini interface redesign: “more attractive” with updated themes (a “black” style mentioned).
• Voice mode: choosing different English dialects.
• A daily “brief” feature: Gemini reviews links/data (email, calendar, task book) and reminds you what to do.

Mac app and multimodal editing tools

• Gemini app for Mac: select multiple files and ask (via a keyboard function key or voice) to draft something like a letter; Spark integrated for actions.
• A separate Google Pix/Photos tool:
  • hover/select parts of an image
  • delete/drag/add text
  • create banners/posters from photos

Flow updates (video generation/editing)

• Flow can now:
  • upload a photo and ask it to choose best angles
  • return multiple video options (e.g., 16)
  • regenerate those options with additional prompts (e.g., “from morning to night” phrasing)
• Flow Tools: apply presets to videos/photos or create from ready-made templates without typing prompts.
• Music generation is mentioned as improving instruments/presets.

Audio-focused smart glasses

• Wearable described as audio smart glasses (no reliance on text on a display; audio only).
• Demo includes:
  • using Maps navigation
  • asking Gemini/Spark to do tasks in the background on Google servers (e.g., “Order some coffee”)
• Claim: Nanobana-type image processing works via glasses as well (take photo → ask to process).
• Collaboration mentioned with Samsung and cross-OS support (iOS and Android).
• Multiple models expected “in the fall.”

Security and research AI tooling

• Cybersecurity agent “Code…/Codenr” (name partially garbled):
  • a dedicated Google agent for protecting codebases
  • automatically finding and fixing critical vulnerabilities
• “Jmy for Science” toolkit for accelerating research:
  • finding relevant preprints
  • converting research into working code
  • proposing hypotheses
  • running simulations
• Science examples:
  • Alpha Earth / digital twin concepts for deforestation and food security
  • weather forecasting model predicting a hurricane path 3 days early
  • future plans for virtual cell simulations
  • Isomorphic Labs using AI to model molecular interactions to develop drugs faster, mentioned as working in preclinical cancer/immune disorder areas

Main speakers/sources (as referenced in the subtitles)

• “Google” / Google presenters at the main Google presentation
• Hasabis (DeepMind/Google leadership; referenced while speaking about security)
• The video creator/host (narrator guiding viewers, referencing their Telegram channel and promising reviews)