[SHARED ONLINE] - GDG Cloud Iasi - Build with AI - Sponsored by Qodea.pdf

Google - Build with AI
April 2025 Florian Blaga
Security and Compliance Lead

Our People & Places
Our 400+ best-in-class experts are based across
the EMEA region, with locations including:
Romania
Iași
The Netherlands
Utrecht
UK
London, Manchester, Swindon, and Edinburgh
France
Paris
Belgium
Brussels

And Google has put AI to work for 3B users in Google Workspace
Spelling & Grammar
Suggestions
Smart Reply
Noise Cancellation
Portrait Restore
Smart Compose
Malware Detection and Prevention
2017 2018 2019 2020 2021 2022 2023
2016
Explore
2015
Suggested Actions
Chart Suggestions
Answers
Grammar Suggestions
Formula Acceleration
Formula Suggestions
Smart Compose
Smart Compose
File Suggestions
Write & Re-Write
AI-generated Summaries
Sharing Suggestions
Practice sets
Originality reports Interactive questions for YT videos in
Classroom

Google keeps more people
safe online than anyone else
4 billion
Devices running
Chrome are protected
each day against
malware and social
engineering attacks
2.5 billion
Active Gmail users
protected against
phishing, malware,
and spam
5 billion
Google Safe Browsing
users devices protected
each day from malware
and social engineering .
Petabytes
Of cloud telemetry analyzed
each day by Chronicle and
Security Command Center
(CSPM) for threat detection
and response.

Generative AI is a type of artificial intelligence that
helps humans create content, such as text and
images.
It’ll enable a new era of work, as it can make humans
significantly more productive, collaborative, and
creative.

Generative AI can transform industries and roles by a
multiplying factor that is still unknown, and professionals
that don’t use it will likely be less productive than those
that do.
As a consequence, Google believes organizations that adopt
generative AI will have an opportunity to leapfrog those
who don't.

Generative AI has the potential to change the anatomy of
work, augmenting the capabilities of individual workers by
automating some of their individual activities. Current generative
AI and other technologies have the potential to automate work
activities that absorb 60 to 70 percent of employees’ time
today.”
McKinsey: The economic potential of generative AI: The next productivity frontier

[SHARED ONLINE] - GDG Cloud Iasi - Build with AI - Sponsored by Qodea.pdf

What is it Why it matters
Gemini 2.0 Flash
GA
First
production-ready
Gemini 2.0 model
Offers the same 1M context window as 1.5 Flash
but with expanded capabilities (e.g., sophisticated
integration and tool use) and better performance,
exceeding 1.5 Pro on key benchmarks at twice the
speed
Gemini 2.0
Flash-Lite
Preview
Low-cost,
high-performance
Gemini 2.0 model
Offers the same 1M context window as 1.5 Flash
but with improved performance and a lower price
for workloads that blend small- and large-context
inputs and outputs.
Gemini 2.5 Pro
Experimental
Superior intelligence
with a 2M context
window
Ideal for coding, reasoning, complex prompts, and
long context

3 Sizes | Improved Performance | New Output Capabilities

ChatBot Arena
#1,6,6
,10
Top models as
blindly evaluated by
users
SEAL Leaderboard
#1
Mode for Visual
Understanding while
being orders of
magnitude cheaper
SEAL Leaderboard
#4
Model for Coding
while being orders of
magnitude cheaper
Vectara
Hallucination
#1,2,4
,5,8
Lowest hallucination
Rate

Features Notes
Gemini 2.0 Flash Gemini 2.0 Flash (Model ID: gemini-2.0-flash-exp) is available as an experimental preview release through
AI Studio, the Gemini Developer API and the Vertex AI Gemini API. Gemini 2.0 Flash offers improvements in
speed, quality, and advanced reasoning capabilities including enhanced understanding, coding, and
instruction following. (Only available in us-central)
Gemini 2.0 Flash Thinking Gemini 2.0 Flash Thinking is an experimental model that's trained to generate the "thinking process" the
model goes through as part of its response.
Google Gen AI SDK The new Google Gen AI SDK provides programmatic access to Gemini 2.0 using both the Developer and
Vertex AI APIs in 4 languages: Python and Go with Java and Javascript coming in the future.
Speech generation
(early access/allowlist)
Gemini 2.0 supports a new multimodal generation capability: text to speech as a private experimental
release.
Image generation
(early access/allowlist)
Gemini 2.0 supports image generation as a private experimental release.
Multimodal Live API The Multimodal Live API enables low-latency, two-way interactions using text, audio, and video input, and
audio and text output.
Tools: Search as a tool Google Search is available as a tool. Using Grounding with Google Search, you can improve the accuracy
and recency of responses from the model.
Tools: Code execution The code execution feature enables the model to generate and run Python code and learn iteratively from
the results until it arrives at a ﬁnal output.

Gemini 2.0
Portfolio
Walkthrough

Comprehensive Table of Google Gemini AI Models
Feature
Gemini 2.5 Pro
Experimental Gemini 2.0 Flash Gemini 2.0 Flash-Lite Gemini 1.5 Flash Gemini 1.5 Flash-8B Gemini 1.5 Pro Imagen 3
Current release gemini-2.5-pro-exp-03-25
Latest (gemini-2.0-flash),
Stable (-001)
Latest
(gemini-2.0-flash-lite),
Stable (-001)
Latest stable
(gemini-1.5-flash), Stable
(-001, -002)
Latest stable
(gemini-1.5-flash-8b),
Stable (-001)
Latest stable
(gemini-1.5-pro), Stable
(-001, -002) imagen-3.0-generate-002
Current model gemini-2.5-pro-exp-03-25
gemini-2.0-flash, -001,
-exp, -thinking-exp gemini-2.0-flash-lite, -001
gemini-1.5-flash, -latest,
-001, -002
gemini-1.5-flash-8b, -latest,
-001
gemini-1.5-pro, -latest,
-001, -002 imagen-3.0-generate-002
Multimodal inputs Audio, images, video, text Audio, images, video, text Audio, images, video, text Audio, images, video, text Audio, images, video, text Audio, images, video, text Text
Text output Yes (limit 65,536) Yes (limit 8,192) Yes (limit 8,192) Yes (limit 8,192) Yes (limit 8,192) Yes (limit 8,192)
No (input for image
generation)
Image output No Experimental No
No (processes images as
input) No No Yes (up to 4 per prompt)
Audio output No Coming soon No No No No No
Multimodal Live API No Experimental No No No No No
Thinking model Yes
Experimental version
available No No No No No
Context window
Input: 1,048,576; Output:
65,536
8,192
8,192
8,192
8,192
8,192
Input: N/A; Output: Up to 4
Images
Controlled generationStructured outputs
Structured outputs,
Function calling Structured outputs
JSON mode, JSON
schema
System instructions, JSON
mode, JSON schema
System instructions, JSON
mode, JSON schema Details not available
Function calling Yes Yes No Yes Yes Yes No
Search as a tool Yes (Search grounding) Yes (Search) No Yes (Search) No No No
Code execution Yes Yes No Yes Yes Yes No
Batch prediction Information unavailable Supported on Vertex AI Supported on Vertex AI Supported on Vertex AI Information unavailable Supported on Vertex AI Information unavailable
Tuning No
Supported on Vertex AI
(-001)
(-001) Supported Supported
(-002)
(customization,
fine-tuning)

AI Model Feature Comparison Table
Feature GPT-4.5 GPT-4o
Gemini 2.5 Pro
Experimental Gemini 2.0 Flash Claude 3.7 Sonnet Grok-3 DeepSeek R1 Llama 3.2
Current Release Feb 2025
May 2024 (Image Gen
Mar 2025) Mar 2025 Feb 2025 Feb 2025 Feb 2025 Jan 2025 Sep 2024
Current Model Orion Omni
gemini-2.5-pro-exp-03-
25 gemini-2.0-flash-001
claude-3-7-sonnet-202
50219 Grok-3 671B MoE (37B active) Llama 3.2
Multimodal Inputs:
Text Yes Yes Yes Yes Yes Yes Yes Yes
Multimodal Inputs:
Image Yes Yes Yes Yes Yes Yes Yes Yes
Multimodal Inputs:
Audio No Yes Yes Yes No No Yes Yes
Text Output Yes Yes Yes Yes Yes Yes Yes Yes
Image Output No Yes Yes Yes No Yes (via Aurora) No Yes
Audio Output No Yes Yes Yes No No No Yes
Multimodal Live API No Yes No Yes No No No No
Thinking Model
Scaled Unsupervised
Learning
Generative Pre-trained
Transformer
Thinking Model
(Reasons Through
Thoughts)
Thinking Model
(Reasons Before
Answering) Hybrid Reasoning
Reasoning-Based
(Test-Time Computing)
Reasoning Model
(MoE)
Decoder-Only
Transformer
Context Window 128K 128K 1M (2M Soon) 1M 200K 128K 16K 128K
Controlled Generation Yes (Steerability) Yes Yes Yes
Yes (Adjustable
Reasoning Budget)
Yes (Think/Big Brain
Modes) Yes (Temperature) Yes
Function Calling Yes Yes
Yes (Excluding
Compositional) Yes Yes (via Tools) No Explicit Mention No Yes (Tool Calling)
Search as a Tool Yes Yes Yes Yes Yes Yes No Explicit Mention No Explicit Mention
Code Execution Yes Yes Yes Yes Yes (via Claude Code) Yes No Explicit Mention
Yes (Tool Calling
Implied)
Batch Prediction Yes Yes No Explicit Mention Yes Yes (Preview) No Explicit Mention Yes No Explicit Mention
Tuning Yes Yes Yes Yes No Explicit Mention Yes Yes Yes

Online Prediction Gemini 1.5 Flash Gemini 2.0 Flash-Lite Gemini 2.0 Flash
Performance
(Relative to Gemini 1.5 Flash)
Slightly better than 1.5 Flash Much better than 1.5
Flash
Input text / image / video $0.075
$0.15 (long context)
$0.075 $0.10
Input audio $0.075
$0.15 (long context)
$0.075 $1
Output text $0.3
$0.6 (long context)
$0.3 $0.40
Output image NA NA Coming Soon
Output audio NA NA Coming Soon
Search as a Tool free query per
day (QPD)
$35/1K
(standard search)
NA 1,500 QPD Free
(+ $35/1K overage)
Context caching (storage) $1.00 / 1,000,000
tokens per hour
Available March 31,
2025
How does pricing compare between models

The enhanced reasoning
model with 1M context window
improved performance and
explainability. Building on the
speed and performance of 2.0
Flash, 2.0 Flash thinking exp
excels in science and math while
revealing its thought process
to solve complex problems
Progress is quick in this space
Gemini Thinking: Now available in Gemini Advanced & Google AI Studio

Instruction Models
Eg. Gemini 2.0 Flash, GPT-4o, Claude 3.5, Llama 3…
Strengths:
● Faster
● Cost Effective
Weaknesses:
● Less capable at the most complex reasoning
Example Use Cases:
● Summarization & Extraction of Content
● Creative Writing
● RAG & Fact based question answering
Thinking Models
Eg. 2.0 Flash Thinking, O3-mini, O1, Deepseek R1…
Strengths:
● First Principles Thinking & Problem Solving
Weaknesses:
● Much Slower
● More expensive (larger outputs)
● Overkill for basic tasks
Example Use Cases:
● Multi Step Agentic Reasoning
● Novel Problem Solving (eg. Mathematics & Code)

The new ‘googling’ -
structuring your request
Persona
Task
Context
Format
You are a Sales Manager. Write a guide describing
3 sales techniques to use for the Sales Team when
selling a new car to 30 year olds. Limit to bullet
points with a description and conversation example.
From Google’s Prompting Guide

Iterative Prompting
● Limits: Gemini has limited output ~5,000
- 6,000 words of standard text, less with
tables.
● "Continue" Prompt: If the model reaches
the output token limit, simply reply with
"continue" to prompt it to generate the
next part of the response.
● Iteration: Repeat the "continue" prompt
as needed until the model has generated
the complete response

https://www.youtube.com/watch?v=V-r0WjXJhL8

Proprietary + Confidential
Next ‘25

[SHARED ONLINE] - GDG Cloud Iasi - Build with AI - Sponsored by Qodea.pdf

More Related Content

Similar to [SHARED ONLINE] - GDG Cloud Iasi - Build with AI - Sponsored by Qodea.pdf (20)

Recently uploaded (20)

[SHARED ONLINE] - GDG Cloud Iasi - Build with AI - Sponsored by Qodea.pdf