← Master Index
Vol. 22 Module 22.1 Lecture

Google / Google DeepMind (Gemini, Gemma, NotebookLM, Veo, Imagen, AI Studio, Vertex AI)

Frontier LLM Providers

How This Lesson Fits the Module & Volume

OpenAI and Anthropic were mostly lab + API + consumer chat. Google / DeepMind is a full cloud + consumer + open-weight estate: closed Gemini, open Gemma, NotebookLM, stills (Imagen), video (Veo), prototype surface Google AI Studio, and production surface Vertex AI. The client was Vol. 18 Gemini SDK.

This is often the right vendor when the customer already lives in Google Cloud or Workspace—not because Gemini always wins a blind eval. Vol. 21 products still need HITL; NotebookLM is not a licensed lawyer or doctor. Module 22.3/22.2 will revisit Veo/Imagen as modality vendors; here they sit on the same procurement card as Gemini.

Learning Objectives

By the end of this lesson, students should be able to:

  • Distinguish Gemini (closed product) from Gemma (open weights under Gemma terms).
  • Place NotebookLM, Veo, Imagen, AI Studio, and Vertex AI on one Google estate map.
  • Write a qualitative Pricing / Strengths / Weaknesses comparison vs OpenAI and Anthropic.
  • Choose AI Studio (spike) vs Vertex (IAM, billing, enterprise controls) honestly.
  • Sketch a Gemini generate_content call and know when to use Vertex ADC instead of an AI Studio key.
  • Avoid inventing DeepMind architecture diagrams or fake multimodal leaderboards.
Definition

Google DeepMind (with Google product orgs) ships the Gemini closed multimodal family via consumer Gemini apps, Google AI Studio, and Vertex AI. Gemma is Google’s open-weight developer LLM line (related in branding, not a download of Gemini flagship weights). NotebookLM is a grounded notebook product over user-provided sources. Imagen and Veo are Google’s closed still and video generators. AI Studio is the developer playground + API-key path; Vertex AI is the GCP production platform (IAM, VPC-SC, logging, Model Garden including third-party models). This lecture catalogs the estate—not unpublished model internals.

Estate Map

SurfaceClosed / openTypical job
Gemini apps + APIClosedMultimodal assistant; Flash vs Pro-style latency/quality tiers
GemmaOpen weights (Gemma terms)Self-host / fine-tune small–mid models; education and on-prem spikes
NotebookLMClosed productSource-grounded notes/audio over uploads—not a generic chatbot
ImagenClosedStills; Vertex + consumer surfaces; SynthID watermarking story
VeoClosedVideo; Vertex / Gemini apps; native-audio SKUs vary—read current IDs
AI StudioDev consoleKeys, prompts, quick multimodal tests
Vertex AIGCP platformProduction: IAM, billing accounts, Model Garden, enterprise networking

Pricing / Strengths / Weaknesses

AxisPricing (qualitative)StrengthsWeaknesses
Gemini API / Vertex Token-metered; Flash-class SKUs cheaper/faster than Pro-class. Consumer Gemini Advanced / Workspace add-ons are seats, not the same as Vertex invoices. Grounding / context-caching / thinking modes add billable units—read the live meter. Native multimodal parts in one request; Google Search grounding options; Workspace/GCP gravity; long-context SKUs. Closed Gemini weights; Google-shaped ToS and data maps; SDK/API surface has historically churned (AI Studio vs Vertex clients).
Gemma Weights free to download under terms; you pay GPU/host or a third-party inference vendor—not Gemini token list prices. Open path from the same company; good teaching contrast with Gemini. Not a secret Gemini distill you can assume; license/terms still apply; capability ceiling below flagship Gemini.
NotebookLM Product/plan limits (sources, seats), not a public per-token RAG SKU. Excellent grounded-notebook UX for teams with document piles. Not your multi-tenant research-assistant product; citation/privacy still Vol. 21 + 20 work if you rebuild it.
Imagen / Veo Per-generation / resolution / duration / audio tiers; video typically dwarfs text tokens. Same cloud bill as Gemini; watermarking + IAM for enterprises. Closed; compare on Vol. 16 rubrics vs DALL·E, Sora, Midjourney, Runway—no fake quality scores.
AI Studio vs Vertex Studio keys are easy/dev; Vertex rides GCP committed use + project billing. Studio: fastest spike. Vertex: production controls and Model Garden (including some non-Google models). Shipping on a Studio API key as “enterprise” is a governance miss.

Pick Google when

  • GCP / Workspace is already the estate
  • You need Gemini + Imagen + Veo on one bill
  • NotebookLM covers internal knowledge notebooks
  • You want Gemma for a small open-weight track

Look elsewhere when

  • Azure/OpenAI or Anthropic is the procurement lock
  • You need Claude Code / GPT-only toolchain familiarity
  • Self-host flagship-class models (Google does not give you Gemini weights)

Curriculum links

Estate advantage

  • One vendor: text, vision, video, stills, notebooks, cloud IAM
  • Open + closed pairing (Gemma vs Gemini) is pedagogically clean
  • Vertex Model Garden can host third-party models later

Estate cost

  • Lock-in to Google identity, networking, and ToS
  • AI Studio ≠ production Vertex
  • Do not treat NotebookLM citations as Vol. 19 hallucination-proof

Practical API Sketch

Google Gen AI SDK. AI Studio uses an API key; Vertex uses Application Default Credentials. SKUs change.

# Gemini via Google Gen AI SDK (AI Studio key path). # pip install google-genai # export GEMINI_API_KEY=... # AI Studio; Vertex uses ADC + project instead from google import genai client = genai.Client() # reads GEMINI_API_KEY by default MODEL = "gemini-2.0-flash" # teaching stand-in; pin current Vertex/Studio ID resp = client.models.generate_content( model=MODEL, contents="In 4 bullets, contrast Gemini (closed) with Gemma (open weights) for a GCP shop.", ) print(resp.text) # usage / prompt_token_count live on response metadata — feed Vol. 18 monitoring # Multimodal: pass parts (text + inline/file image or PDF) — see Vol. 18 Gemini SDK. # Production: Vertex AI + IAM + VPC-SC, not a long-lived Studio key in a mobile app. # Gemma: load weights via HF / vLLM / Ollama; do not expect generate_content to download Gemini weights. # NotebookLM / Imagen / Veo: separate product or Vertex model IDs — re-read current docs.

Related Lectures

LectureRole
AnthropicPrevious closed lab
Gemini family / GemmaClosed vs open Google lines
Imagen / VeoStills and video catalogs
Gemini SDKgenerate_content + Vertex path
Research assistantsNotebookLM is a product cousin, not a replacement architecture
xAI (Grok)Next frontier lab
Common Misconception

Gemma weights are Gemini. They are a separate open-weight family under Gemma terms. Second: AI Studio keys are production-ready Vertex. Third: NotebookLM is a licensed research or legal product. Fourth: Veo/Imagen quality can be ranked with invented scores here. Fifth: Google has no open path (Gemma exists). Sixth: picking Google removes Vol. 20 data-map work.

Knowledge Check

  1. Short Answer: Contrast Gemini and Gemma in one sentence. Answer: Gemini is Google’s closed multimodal product/API; Gemma is an open-weight developer family under Gemma terms.
  2. True/False: Google AI Studio and Vertex AI are the same production control plane. Answer: False.
  3. Multiple Choice: NotebookLM is best described as: (a) a source-grounded notebook product, (b) an open-weight LLM, (c) Azure OpenAI. Answer: (a).
  4. Short Answer: Name Google’s closed still and video generators. Answer: Imagen (stills) and Veo (video).
  5. True/False: This lecture publishes fake dollar prices for Gemini tokens. Answer: False.
  6. Multiple Choice: Vertex AI is the path for: (a) GCP IAM / billing / Model Garden production, (b) only Discord bots, (c) downloading Gemini weights. Answer: (a).
  7. Short Answer: Which Vol. 18 lecture is the Gemini client? Answer: Gemini SDK.
  8. True/False: Flagship Gemini weights are generally downloadable like Gemma. Answer: False.
  9. Multiple Choice: A good reason to pick Google is: (a) existing GCP/Workspace estate + multimodal + Veo/Imagen, (b) OSI Gemini-4 weights, (c) skipping privacy reviews. Answer: (a).
  10. Short Answer: Which provider comes next in 22.1? Answer: xAI (Grok, API, Reasoning).

Key Takeaways

  • Google’s card is an estate: Gemini, Gemma, NotebookLM, Veo, Imagen, AI Studio, Vertex.
  • Gemini closed vs Gemma open; Studio for spikes, Vertex for production.
  • Qualitative pricing: Flash vs Pro tokens, seats vs GCP invoices, generation tiers for Veo/Imagen.
  • Cross-link Vol. 11.5, Vol. 16 Imagen/Veo, Vol. 18 Gemini SDK. NotebookLM ≠ your Vol. 21 research product.
  • Next: xAI Grok.
Trainer’s Guide

Lab: Students draw the Google estate (seven boxes) and mark which they would use for a Vol. 21 research assistant vs a consumer chatbot vs on-prem Gemma. One generate_content spike with a pinned Flash-class ID. Write the data-map: AI Studio key vs Vertex project.

Whiteboard: Three-column OpenAI | Anthropic | Google. Add a fourth empty column labeled xAI.

Recap: Google/DeepMind is Gemini + Gemma + NotebookLM + Veo + Imagen + AI Studio + Vertex. Next is xAI (Grok).