File TXT tidak ditemukan.
Transcript
Qd67wF7vtko • Google Gemini 3.5: The AI Upgrade That Changes Everything (AGI Closer Than We Think?)
/home/itcorpmy/itcorp.my.id/harry/yt_channel/out/BitBiasedAI/.shards/text-0001.zst#text/0264_Qd67wF7vtko.txt
Kind: captions
Language: en
You're probably typing these long,
overly detailed prompts into Gemini,
thinking that more words equals better
results. Maybe you're even copying those
perfect prompt templates from Reddit,
stuffing them with formatting rules and
instructions.
Well, I've spent months testing Google's
new Gemini 3 model, and here's what
surprised me. All that extra fluff
you're adding, it's actually making your
results worse. The model is so advanced
now that it wants you to communicate
like a human, not a robot. So, in this
video, I'm going to show you how to
actually get the best results from
Gemini 3 in 2026. We're talking about
prompting strategies that work with the
model's intelligence instead of against
it. Real examples across coding, content
creation, research, and productivity,
plus the hidden features most people
don't even know exist.
First up, let me show you the single
biggest shift in how you should be
writing prompts now. And trust me, it's
going to feel almost too simple.
The new prompting reality.
Here's the thing about Gemini 3 that
Google quietly released in November
2025. This model is wildly different
from anything that came before. We're
talking about over 650 million people
using this every month, and most of them
are still prompting like it's 2023. The
old way of prompting was all about being
hyper specific. You'd write something
like, "Please analyze the attached
report and extract the following fields
using JSON format with double quotes
around keys for revenue, expenses, and
margin, ensuring that numerical values
are not quoted." That's exhausting just
to read, right? But here's where it gets
interesting. Gemini 3 has been trained
to understand context and structure
naturally.
So now that same prompt becomes
extract Q3 2025 revenue expenses and net
profit margin from the report
return as JSON.
That's it. Two sentences
and Gemini will give you perfectly
formatted JSON without you having to
explain what JSON even looks like. The
model can handle up to a million tokens
of context.
To put that in perspective, you could
paste an entire novel, a research paper,
multiple spreadsheets, images, even 2
hours of video content, and Gemini will
synthesize all of it. It's not just
reading words anymore. It's
understanding images, interpreting
charts, analyzing code, and connecting
dots across different types of media
simultaneously.
And the accuracy
on reasoning and coding benchmarks, it's
hitting state-of-the-art scores. For
factual Q&A, it's sitting around 72%
accuracy, which means about 7 out of 10
times it's nailing the answer.
That's massive progress, but we'll talk
about that other 28% later because it
matters.
The clarity first framework. Let me show
you the framework that's actually
working in 2026.
Forget everything you learned about
prompt engineering from a year ago. The
new approach is what Google calls
clarity first instructions. And it's
honestly a gamecher. Think of your
prompts in three simple parts. Task,
input, and output. That's the mental
model.
You're telling Gemini what you want it
to do, giving it the materials to work
with, and specifying how you want the
result delivered.
Here's a real example. Let's say you
have a document about project risks the
old way. I am providing you with a
document. Please carefully read through
it. I need you to identify any risks
that are mentioned. For each risk, I
also need you to find if there's a
mitigation strategy discussed.
Please format this as a table with
proper columns and rows. The columns
should be labeled risk and mitigation.
Use markdown formatting for the table.
The new way task list the top five risks
mentioned in the document.
Output markdown table with columns risk
and mitigation. See the difference?
You're treating Gemini like an
intelligent collaborator, not like it
needs step-by-step instructions for
breathing. The model will automatically
create a properly formatted markdown
table. It knows what that means. It's
inferred thousands of table structures
during training.
Now, if you want to give Gemini
additional context or perspective, you
absolutely can.
You might start with something like
you're a financial project manager or as
a cyber security expert.
This gives the model a lens to view the
problem through. But here's the key.
You're setting context, not writing a
novel.
one sentence about the role or
perspective, then move directly into
what you need. The same goes for
constraints. If you need British
spelling, say use British English. If
you want a casual tone, say use casual
tone. If there's a word limit, state it
max 150 words.
Gemini respects these constraints
without you having to repeat them or
create complex rules around them.
Multimodal magic.
But wait until you see what happens when
you start combining text with images,
videos, and documents.
This is where Gemini 3 really pulls
ahead of everything else. Let's say you
have a complex chart from a quarterly
report. You upload the image and instead
of describing what's in the chart, you
just write,
"What does the red segment represent in
this chart?" Explain in two sentences.
Gemini looks at the image, identifies
visual elements, interprets the data,
and gives you a precise answer.
It's genuinely analyzing the visual
information, not just looking at
metadata or alt text. The same principle
applies to video. If you're on the pro
tier, you can upload up to 2 hours of
video content. Imagine you have a
recorded meeting or a lecture. You can
reference specific timestamps.
summarize the discussion between 15:30
and 22 SAR in this video. Gemini will
process that section, extract the key
points, and deliver a summary. It's like
having someone watch the video for you
and report back on exactly what you need
to know. For PDFs and long documents,
the million token context window means
you're not limited to summaries or
excerpts anymore. You can feed Gemini an
entire research paper, a contract, a
technical manual, whatever you need
analyzed, and ask specific questions
about any part of it.
What methodology did they use in section
3 or list all the legal obligations in
this contract as a numbered list? It
handles the full document without
breaking a sweat. There's even this new
feature called Gemini Canvas where you
can ask it to create visuals.
You might say, "Generate an infographic
showing the water cycle or create a
poster design for a community event."
The model will produce actual visual
assets that you can use and iterate on
real world use cases.
Let me walk you through how this
actually plays out across different
domains because this is where theory
meets practice. Coding and development.
If you're a developer or you work with
code at all, Gemini 3 is probably going
to become your favorite tool. The model
topped the webdev arena benchmark with a
score of 1487 ELO, which in the coding
world is basically saying it's better
than most human programmers at certain
tasks. Here's the practical reality. You
can describe what you need in plain
English and Gemini will write the code.
something like write a Python function
called calculate discount that takes a
price and discount percentage, returns
the final price, and includes a dock
string with type hints.
Within seconds, you get a fully formed
function with proper documentation, type
annotations, everything formatted
correctly.
But it goes deeper than just writing
snippets. You can ask Gemini to refactor
existing code, explain complex
algorithms, debug errors, even generate
entire web interfaces.
There's a demo where someone simply
asked for a landing page for a retro
dance night. And Gemini produced
complete HTML, CSS, and JavaScript for a
functioning web page, not a template,
not a framework.
Actual custom code based on that
description. The Agentic coding features
take this even further.
Gemini can now execute code in a real
development environment, run tests, use
version control.
Google's anti-gravity platform
demonstrated this by having Gemini solve
complex problems like balancing a
simulated cartpole by writing and
running code iteratively until it
worked.
content creation and creative work. For
writers, marketers, anyone creating
content, Gemini handles everything from
blog posts to social media copy to video
scripts. The key is being specific about
tone and format without overexplaining.
You might prompt, "Write a 200word blog
post about sustainable travel tips. Use
a friendly, conversational tone."
That's all you need.
Gemini will draft something engaging
that matches the style you requested.
If you're not satisfied, you can
iterate, make it more professional, or
add a section about carbon offsets. One
of my favorite examples is when someone
asked Gemini to explain nuclear fusion,
and the model responded with both
working code to visualize the process
and a poem about fusion. That's the kind
of creative range we're talking about.
It's not just generating text. It's
thinking across mediums and finding
interesting ways to communicate ideas.
For visual content, you can request
infographics, diagrams, illustrations.
Create an infographic showing the
benefits of remote work will get you an
actual visual asset. It's not perfect
every time, and you'll need to refine,
but having a first draft appear in
seconds changes the entire creative
workflow.
Research and learning.
This is where that massive context
window becomes crucial. You can upload
entire textbooks, research papers,
lecture transcripts, whatever learning
materials you're working with, and
Gemini will process all of it. A
practical scenario, you're studying for
an exam, and you have 200 pages of
lecture notes. You upload them and ask,
"Create a study guide covering the five
main topics with key concepts and
definitions under each." Gemini reads
through everything, identifies the main
themes, extracts the important details,
and organizes it into a study guide
format. Or let's say you're doing
research and you come across a dense
academic paper. Instead of spending an
hour reading it, you can ask, "Summarize
the methodology and key findings in
three bullet points.
What are the limitations mentioned by
the authors?"
You get the essentials in seconds. And
if you need to dive deeper into specific
sections, you can ask follow-up
questions in Google searches AI mode.
There's even functionality where Gemini
can create interactive simulations on
the fly. Someone asked how RNA
polymerase works. And instead of just
describing it, Gemini coded a mini
simulation showing the process visually.
That's learning taken to a different
level.
Productivity and professional tasks
in day-to-day work. Gemini becomes like
having a highly capable assistant who
never gets tired. Email drafting,
meeting summaries, task prioritization,
data analysis, it handles all of it. For
emails, you can be super brief.
Compose an email inviting the team to a
planning meeting next Tuesday at 2 p.m.
Keep it friendly but professional.
Gemini drafts it, you review, maybe
tweak a line or two, and send. What used
to take 10 minutes now takes one. Travel
planning is surprisingly robust. If you
say, "Plan a 3-day trip to Rome with a
$1,000 budget, including flights,
accommodation, and activities," Gemini
will come back with a complete
itinerary, cost breakdowns, attraction
recommendations, even timing for each
day. There's a visual itinerary mode
that looks like a travel magazine spread
with maps and images.
The Gemini agent features which are
still rolling out can actually connect
to your Gmail and calendar. You can tell
it summarize all the emails from the
last week about the marketing campaign
and highlight any action items and it
will triage your inbox, pull out the
relevant threads and give you a
synthesized report. It can even schedule
meetings by looking at everyone's
availability, though it always asks for
confirmation before taking any action.
For data analysis, if you have a
spreadsheet of sales data, you can paste
it and ask, "What are the trends in
quarterly revenue? Are there any
products underperforming?"
Gemini will analyze the numbers,
identify patterns, and present insights
in plain language. You're not doing
formulas or pivot tables. You're just
asking questions and getting answers.
Expert strategies. Now, let me share
some strategies that the AI community
has figured out through extensive
testing because there are definitely
ways to make Gemini work even better.
First, context first prompting.
This is where you set the stage before
asking the question.
As an expert data scientist, given this
attached sales report from Q4 2024,
you're priming Gemini to approach the
task from a specific perspective. It's
not mandatory, but for complex or
specialized tasks, it significantly
improves the quality of responses.
Multi-step instructions work really
well. Instead of cramming everything
into one long sentence, break it into
numbered steps.
Task: Analyze this code. Step one,
explain what it does. Step two, identify
any bugs or inefficiencies.
Step three, suggest improvements.
Gemini loves structure like this. It
processes each step sequentially and
gives you organized output. For complex
reasoning tasks, explicitly asking
Gemini to show its work can improve
accuracy.
Phrases like, "Explain your reasoning
step by step or walk me through how you
arrived at that answer trigger the model
to be more deliberate." The chain of
thought approach, as it's called, helps
prevent shallow or incorrect answers.
There's also this interesting technique
some people use called simulating a
debate. Before making a recommendation,
you can prompt,
discuss the pros and cons of option A
versus option B before giving your final
recommendation.
This forces Gemini to consider multiple
perspectives, which usually leads to
more balanced and thoughtful responses.
One thing that's really powerful, Gemini
maintains conversation context across
multiple turns.
So, if you give it a detailed document
or set of instructions once, you can ask
follow-up questions without repeating
everything. Based on the report I shared
earlier, what about the section on
customer retention? Gemini remembers,
you're building a conversation, not
starting from scratch each time. And
here's a tip that seems obvious, but
makes a huge difference. Iterate on the
answers you get. If the first response
isn't quite right, don't give up. Ask
Gemini to refine it. Make this more
concise. Change the tone to be more
formal. Add specific examples.
The second or third pass is often
significantly better than the first.
Warnings and limitations.
Before we wrap up, I need to talk about
the limitations because Gemini is
powerful, but it's not magic, and you
need to know where the edges are.
Hallucinations are still a thing. Even
though Gemini 3 scores around 72% on
factual accuracy benchmarks, that means
roughly one in four answers could
contain errors or madeup information.
If you're asking about obscure facts,
niche topics, or anything that could
have multiple interpretations, always
verify the response with trusted
sources.
Don't treat Gemini's output as gospel,
especially for critical decisions. There
are biases baked into the model. Just
like with any AI trained on internet
data, Google has done extensive safety
testing, but the model can still exhibit
biases or misunderstand sensitive
queries. Be thoughtful about how you
phrase questions around controversial
topics. And be aware that the responses
might reflect certain perspectives more
than others. Privacy is a real
consideration. Anything you upload to
Gemini, whether it's text, images,
documents, could potentially be stored
or used in model training unless you're
using specific enterprise settings that
guarantee data isolation.
Don't upload confidential business
documents, personal information, or
anything you wouldn't want on a server
somewhere. Read the terms of service for
whatever tier you're using. If you're
using the agent features or any
functionality that connects to your
Google Workspace, be careful. These
tools can perform actions like
organizing emails, scheduling meetings,
or creating documents. They always ask
for confirmation before doing anything
significant, but still supervise them.
Don't just blindly approve actions
without understanding what's happening.
The free version of Gemini has usage
limits. You might hit monthly caps on
the number of queries or the length of
conversations.
The paid pro and ultra tiers give you
more capacity and speed, plus access to
features like deep think mode and
extended video processing.
If you're planning to use Gemini
heavily, you'll probably need a
subscription. And finally, remember,
we're early in the Gemini 3 era.
Features are rolling out gradually. Some
capabilities mentioned here might be in
beta or limited to certain regions or
account types. Google is updating the
platform constantly, so what's true
today might change in a month. Keep an
eye on official announcements. So that's
Gemini 3 in 2026.
The biggest shift is moving away from
overengineered prompts and toward clear
conversational instructions.
Think of it as talking to a highly
capable colleague who understands
context and can work across text,
images, code, and video simultaneously.
Use the strategies we covered. Keep
prompts simple. Structure them with task
and output. Leverage the multimodal
capabilities and iterate on responses
until they're exactly what you need.
Whether you're coding, creating content,
doing research, or managing daily tasks,
Gemini can genuinely save you hours of
work if you know how to guide it
properly. What are you most excited to
try with Gemini? Drop a comment and let
me know if this guide helped clarify how
to actually use this tool effectively.
Hit the like button and subscribe for
more AI breakdowns. I'm constantly
testing these models and sharing what
actually works versus what's just hype.
Thanks for watching and I'll see you in
the next one.