Google Gemini Explained: 80% of Google Gemini You Never Use (2026 Guide)

BitBiasedAI 首頁 每日列表 原始影片
實際影片長度：16:37.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:02.320
You've probably typed something into Gemini ,
0:02.320–0:03.010
got an answer ,
0:03.020–0:04.147
and closed the tab ,
0:04.147–0:06.552
the same way you'd use any other chatbot .
0:06.562–0:07.663
Here's the thing,
0:07.663–0:10.372
that's maybe 10% of what it actually does.
0:10.382–0:11.073
Use it right ,
0:11.073–0:13.835
and it can quietly take over almost half the busy
0:13.845–0:15.666
work you're still doing by hand .
0:15.676–0:18.077
I spent hours mapping every model ,
0:18.077–0:18.849
every mode ,
0:18.859–0:21.954
and every product Gemini is quietly wired into ,
0:21.964–0:23.983
and the number that stopped me was this .
0:23.993–0:28.315
Gemini's app alone has over 650 million monthly users ,
0:28.325–0:31.520
and that's before you count everyone using it inside search .
0:31.530–0:33.830
Most of them are using maybe 20% of it,
0:33.830–0:35.496
with no idea the rest even
0:35.506–0:36.469
exists .
0:36.479–0:39.025
Look at this data from the Census Bureau .
0:39.035–0:43.527
Only about one in five US businesses actually use AI in their operations
0:43.527–0:44.402
.
0:44.402–0:44.523
So ,
0:44.523–0:45.554
if you run a business ,
0:45.554–0:47.313
and you're even thinking about this
0:47.313–0:48.602
,
0:48.602–0:50.394
you're ahead of most of your competition .
0:50.404–0:53.876
What you might not know is that alongside covering AI news ,
0:53.886–0:56.477
we work with business owners to help them implement AI in their
0:56.487–0:57.958
business .
0:57.968–1:00.879
Our engineering team gets to know how your business runs ,
1:00.889–1:03.080
then builds the automation with you .
1:03.090–1:05.480
You'll find the link in the description below .
1:05.490–1:05.987
Click it ,
1:05.987–1:08.400
fill out a short form about your business ,
1:08.410–1:10.119
and we'll get in touch to set up a call .
1:10.129–1:10.302
So ,
1:10.302–1:11.253
in this video ,
1:11.253–1:14.279
I'm breaking down exactly what Gemini is as
1:14.289–1:15.428
of mid 2026 ,
1:15.428–1:17.580
every current model ,
1:17.580–1:18.719
every mode ,
1:18.729–1:21.518
and everywhere Google has quietly built it in .
1:21.528–1:22.114
By the end ,
1:22.114–1:25.118
you'll know exactly which Gemini tool to reach for
1:25.128–1:27.478
depending on what you're actually trying to do ,
1:27.488–1:30.557
instead of just typing into whichever box is in front of you .
1:30.567–1:30.909
First ,
1:30.909–1:33.237
let's clear up the biggest misconception .
1:33.247–1:35.477
Gemini isn't one product at all .
1:35.487–1:36.945
What Gemini actually is ,
1:36.945–1:39.277
here's the mental model you need before
1:39.287–1:40.717
any of this makes sense .
1:40.727–1:42.152
Gemini isn't a single AI ,
1:42.152–1:44.476
it's Google's umbrella name for a whole
1:44.486–1:45.060
platform ,
1:45.060–1:46.855
a family of models underneath ,
1:46.855–1:48.076
and a set of products
1:48.086–1:50.236
on top that let you actually talk to them .
1:50.246–1:51.795
Think of it in two layers .
1:51.805–1:53.884
The bottom layer is the models themselves,
1:53.884–1:54.835
things like Gemini
1:54.845–1:55.368
3 .
1:55.378–1:57.315
6 Flash or Gemini 3 .
1:57.325–1:58.195
1 Pro .
1:58.205–2:00.915
These are the engines tuned for different jobs .
2:00.925–2:02.209
Some built for speed ,
2:02.209–2:03.794
some for heavy reasoning ,
2:03.804–2:05.634
some for images or audio .
2:05.644–2:07.994
You never see these names unless you go looking .
2:08.004–2:10.434
The top layer is everything you actually click on .
2:10.444–2:11.520
The Gemini app ,
2:11.520–2:13.673
AI mode inside Google search ,
2:13.683–2:15.432
Gemini inside Gmail and Docs ,
2:15.432–2:17.473
the voice assistant on your phone
2:17.473–2:17.983
.
2:17.983–2:20.673
All of those are just different doors into the same underlying
2:20.683–2:21.433
models .
2:21.443–2:23.072
That's the whole point of this video .
2:23.082–2:25.752
Google isn't trying to build one great chatbot .
2:25.762–2:29.152
It's trying to put the same AI brain behind every product you already
2:29.162–2:29.832
use .
2:29.842–2:30.010
So ,
2:30.010–2:31.854
let's start with the brains ,
2:31.854–2:33.112
the actual models ,
2:33.122–2:35.751
because once you know what each one is built for ,
2:35.761–2:37.671
everything else clicks into place .
2:37.681–2:39.231
The current model lineup .
2:39.241–2:40.818
This is a demo checklist ,
2:40.818–2:42.631
so we're going model by model .
2:42.641–2:43.211
What it is ,
2:43.211–2:44.779
what it's actually good for ,
2:44.779–2:45.990
and where you can get
2:46.000–2:46.390
it .
2:46.400–2:47.350
Gemini 3 .
2:47.360–2:48.830
7 flash .
2:48.840–2:50.405
Launched on August 13th ,
2:50.405–2:50.718
2026 ,
2:50.718–2:52.595
this is Google's newest flash
2:52.605–2:55.119
model and its most capable workhorse yet .
2:55.129–2:58.885
It's built primarily for coding and AI agents with major improvements
2:58.895–3:00.371
in software engineering ,
3:00.371–3:01.355
web development ,
3:01.355–3:02.690
and complex multi step
3:02.700–3:03.892
workflows .
3:03.902–3:06.616
Google has already made it generally available through the Gemini
3:06.626–3:07.069
API ,
3:07.069–3:08.840
positioning 3 .
3:08.850–3:12.746
7 flash as the new go to model when you want strong intelligence
3:12.756–3:15.549
without giving up the speed and efficiency the flash lineup is
3:15.559–3:16.270
known for .
3:16.280–3:17.285
Gemini 3 .
3:17.295–3:18.354
6 flash .
3:18.364–3:20.140
This is Google's current flagship ,
3:20.140–3:21.599
announced in a company blog
3:21.609–3:23.621
post on July 21st ,
3:23.621–3:24.196
2026 .
3:24.206–3:25.750
It's built as a workhorse ,
3:25.750–3:26.831
strong at coding ,
3:26.841–3:27.874
knowledge work ,
3:27.874–3:29.305
and multimodal tasks .
3:29.315–3:31.223
And according to Google's own numbers ,
3:31.223–3:32.577
it does the job using about
3:32.587–3:36.167
17 fewer tokens on average than its predecessor .
3:36.177–3:39.919
Fewer tokens means faster answers and a lower bill if you're paying
3:39.929–3:41.515
for it through the API .
3:41.525–3:44.208
You can reach it through the Gemini API ,
3:44.208–3:45.465
through AI Studio ,
3:45.475–3:49.267
or simply by using the Gemini app and searches AI mode .
3:49.277–3:50.994
No extra setup required .
3:51.004–3:53.685
If you only remember one model name from this video ,
3:53.695–3:57.139
make it this one because it's what most of Gemini is quietly running
3:57.149–3:58.020
on right now .
3:58.030–3:59.061
Gemini 3 .
3:59.071–4:00.101
5 Flash .
4:00.111–4:04.342
This one launched back in May 2026 and it's the model that was
4:04.352–4:07.710
actually powering AI mode in search before 3 .
4:07.720–4:09.104
6 Flash took over .
4:09.114–4:12.465
Google described it as frontier level intelligence at exceptional
4:12.475–4:14.490
speed and by its own benchmarks ,
4:14.490–4:16.506
it pushed output throughput to
4:16.516–4:19.746
roughly four times faster than other top models at the time .
4:19.756–4:23.068
It's still very much in active use across the Gemini app ,
4:23.078–4:26.629
Google's anti gravity platform , and enterprise tools .
4:26.639–4:28.017
It's slightly behind 3 .
4:28.027–4:31.551
6 now , but it's the model quietly sitting behind a huge share
4:31.561–4:32.991
of what shipped this year .
4:33.001–4:33.918
Gemini 3 .
4:33.928–4:35.072
5 Flashlight .
4:35.082–4:38.073
Same July announcement , different job entirely .
4:38.083–4:40.834
This is the stripped down , high throughput sibling .
4:40.844–4:44.955
Google sites roughly 350 tokens per second , which is built for
4:44.965–4:46.476
volume , not depth .
4:46.486–4:48.716
You wouldn't use this for a hard reasoning problem .
4:48.726–4:52.198
You'd use it for background tasks and agent pipelines that need
4:52.208–4:54.198
to move fast and cheap at scale .
4:54.208–4:55.092
Gemini 3 .
4:55.102–4:55.999
1 Pro .
4:56.009–5:00.080
Released in February 2026 , this is the reasoning specialist .
5:00.090–5:03.041
Google's benchmarks claim roughly double the logic performance
5:03.051–5:04.922
of the earlier Gemini 3 Pro .
5:04.932–5:08.283
It's mostly gated behind preview access through the API ,
5:08.293–5:12.604
anti gravity , Vertex AI , and Pro or Ultra subscriptions in the
5:12.614–5:13.684
consumer app .
5:13.694–5:14.325
If 3 .
5:14.335–5:16.525
6 Flash is built for speed , 3 .
5:16.535–5:18.566
1 Pro is built for depth .
5:18.576–5:21.207
The model you'd want on a genuinely hard problem ,
5:21.217–5:22.247
not a quick one .
5:22.257–5:24.928
And if you're actually building on top of these through the API
5:24.938–5:27.929
, price is where the real world decision gets made .
5:27.939–5:30.570
According to Google's own published rates , 3 .
5:30.580–5:32.517
6 Flash runs about 1 .
5:32.527–5:35.185
50 per million input tokens and 7 .
5:35.195–5:37.092
50 per million output tokens .
5:37.102–5:40.933
For context , that's noticeably cheaper on the output side than
5:40.943–5:42.467
GPT 5 .
5:42.477–5:44.975
6 Luna's roughly 6 per million .
5:44.985–5:48.776
That gap is exactly why so many developers default to flash tier
5:48.786–5:51.176
models for anything running at volume .
5:51.186–5:54.898
Now , a handful of specialty models worth knowing by name ,
5:54.908–5:56.778
even if we don't dwell on each one .
5:56.788–6:00.700
Nano Banana 2 is Gemini's current image generation and editing
6:00.710–6:04.041
model , replacing the older Imagen line entirely .
6:04.051–6:07.351
And that's not a small detail because Imagen is actually shutting
6:07.361–6:10.362
down on August 17th , 2026 .
6:10.372–6:13.522
If you've had workflows built on Imagen , that clock is already
6:13.532–6:14.181
running .
6:14.191–6:17.694
There's also a Nano Banana 2 light variant built purely for speed
6:17.704–6:20.790
, trading a small amount of quality for much faster ,
6:20.800–6:22.753
cheaper output at high volume .
6:22.763–6:23.602
VIO 3 .
6:23.612–6:26.864
1 is Google's video generation model , still in beta ,
6:26.874–6:29.975
built to turn a text prompt into a short clip with matching audio
6:29.975–6:30.907
.
6:30.907–6:31.970
Gemini audio 3 .
6:31.980–6:36.217
5 live translate handles real time speech to speech translation
6:36.227–6:39.783
across more than 70 languages , already built into Google Meet
6:39.793–6:40.655
and Android .
6:40.665–6:43.468
And if you're curious about the more niche end of the lineup ,
6:43.478–6:46.137
there's a security focused variant called 3 .
6:46.147–6:50.185
5 flash cyber built to coordinate with vulnerability scanning tools
6:50.185–6:50.898
.
6:50.898–6:52.027
And Lyra 3 .
6:52.037–6:55.879
5 , Google's music model , which can now generate tracks up to
6:55.889–6:58.019
3 minutes long from a text prompt .
6:58.029–7:00.238
Here's the honest limitation worth naming .
7:00.248–7:04.041
Google ships a lot of these models fast , and the naming gets confusing
7:04.051–7:05.230
on purpose or not .
7:05.240–7:05.765
3.5,
7:05.775–7:06.905
3.6,
7:06.915–7:07.980
3.1 Pro,
7:07.990–7:10.820
Flash, Flash Cyber.
7:10.830–7:13.909
If you're not building on top of the API professionally ,
7:13.919–7:16.677
you genuinely don't need to memorize this list .
7:16.687–7:18.522
You just need to know the shape of it .
7:18.532–7:21.611
Fast and cheap , deep reasoning , and multimodal .
7:21.621–7:24.298
That's really three categories wearing a lot of different name
7:24.308–7:25.181
tags .
7:25.191–7:28.952
The modes you actually interact with , models are the engine .
7:28.962–7:30.476
Modes are the steering wheel .
7:30.486–7:33.324
Here's where things get useful for anyone who isn't a developer
7:33.324–7:34.754
.
7:34.754–7:37.656
AI mode inside Google Search turns your search bar into a conversation
7:37.656–7:39.188
.
7:39.188–7:40.386
As of I/O 2026, it's globally powered by Gemini 3.5 Flash
7:40.386–7:42.544
and instead of 10 blue links, you get a written answer
7:42.554–7:46.361
5 flash and instead of 10 blue links , you get a written answer
7:46.371–7:49.771
with follow up questions and sometimes an interactive widget built
7:49.781–7:50.573
on the fly .
7:50.583–7:52.178
Anyone with search can use it .
7:52.188–7:53.782
No subscription required .
7:53.792–7:56.706
Ask something like , What's a quick dinner with what's in my fridge
7:56.706–8:00.354
?
8:00.354–8:00.699
and it answers in full sentences , not a list of recipe blogs .
8:00.709–8:04.005
Deep Think is the extra effort version of the Gemini app .
8:04.015–8:07.773
It spends more compute per answer to reason through harder problems
8:07.783–8:08.815
step by step .
8:08.825–8:11.406
Google gates this one behind Google AI Ultra
8:11.406–8:15.468
and it's built for genuinely difficult science or engineering questions ,
8:15.478–8:16.831
not everyday chat .
8:16.841–8:19.927
Now , Deep Research is where this stops being a chatbot
8:19.927–8:21.641
and starts being an assistant .
8:21.651–8:22.815
You give it a topic
8:22.815–8:27.653
and instead of one reply , it plans a research strategy , opens web pages ,
8:27.663–8:28.420
reads them
8:28.420–8:32.622
and if you allow it , pulls from your own Gmail and Drive , too .
8:32.632–8:34.585
What comes back isn't a paragraph .
8:34.595–8:38.234
It's a full multi page report inside Gemini's canvas .
8:38.244–8:42.362
This is the part of Gemini that actually earns the word agent and
8:42.372–8:45.047
we're coming back to why that matters in a few minutes .
8:45.057–8:47.652
Gemini Live is the voice and camera mode .
8:47.662–8:51.540
Say , Hey Google , let's chat and you're talking to it hands free
8:51.550–8:54.226
with the option to point your camera at something and ask what
8:54.236–8:55.829
it's looking at live .
8:55.839–8:57.913
And Canvas is the workspace mode .
8:57.923–9:01.761
Type , Create a quiz app about planets and it writes the code ,
9:01.771–9:05.448
the interface and the content in one pass , right there for you
9:05.458–9:06.089
to edit .
9:06.099–9:09.416
One more worth a mention briefly because it's still early .
9:09.426–9:14.145
Gemini Spark , a personal agent announced at IO 2026 ,
9:14.155–9:17.031
meant to run continuously in the background handling things like
9:17.041–9:17.953
scheduling .
9:17.963–9:20.598
Right now , it's limited to early Ultra testers .
9:20.608–9:23.043
So , treat this one as coming , not here .
9:23.053–9:25.047
Quick gut check before we move on .
9:25.057–9:27.532
If all of that sounds like a lot of separate tools ,
9:27.542–9:28.374
that's fair .
9:28.384–9:29.696
But , notice the pattern .
9:29.706–9:32.676
Every single one of these modes is just Gemini 3 .
9:32.686–9:33.664
5 or 3 .
9:33.674–9:35.909
6 flash wearing a different job title .
9:35.919–9:39.316
You're not learning six different AIs , you're learning six different
9:39.326–9:41.520
ways to ask the same brain for help .
9:41.530–9:43.192
Multimodal in practice .
9:43.202–9:46.257
Let's talk about what multimodal actually means day to day ,
9:46.267–9:48.527
because it's more than a buzzword on a slide .
9:48.537–9:50.756
Gemini reads and writes text and code .
9:50.766–9:52.707
Obviously , that's the baseline .
9:52.717–9:56.170
But drop a photo into a chat and ask it to caption or edit it ,
9:56.180–9:57.842
and Nano Banana handles that .
9:57.852–10:01.284
Ask it to speak an answer out loud , and Gemini's audio models
10:01.294–10:05.089
generate that voice on the spot with actual control over tone and
10:05.099–10:05.810
pacing .
10:05.820–10:09.135
Ask for a short video and VO builds one from scratch .
10:09.145–10:12.259
Ask it to edit an existing clip , swap the sky ,
10:12.269–10:16.024
change the style , and that's a separate tool called Gemini Omni
10:16.034–10:18.668
doing frame by frame editing by voice command .
10:18.678–10:22.393
Inside Google Docs and Sheets , the same underlying models can
10:22.403–10:25.838
draft a document from your meeting notes or build a spreadsheet
10:25.848–10:29.843
out of a pile of invoices , complete with formulas and charts .
10:29.853–10:32.166
Not just raw numbers dumped into cells .
10:32.176–10:35.611
In Slides , hand it a list of bullet points and it can lay out
10:35.621–10:38.936
an actual deck , not just text on blank slides .
10:38.946–10:42.020
And through Gemini Live's camera mode , you can point your phone
10:42.030–10:45.479
at a menu in a language you don't speak and get a live translation
10:45.489–10:47.149
overlaid on what you're looking at .
10:47.159–10:50.229
Or ask it to identify an object it's looking at through the lens
10:50.229–10:51.979
.
10:51.979–10:52.059
No typing involved .
10:52.059–10:53.588
Here's the part worth remembering .
10:53.598–10:55.059
You never pick the model .
10:55.069–10:56.410
You just say what you want .
10:56.420–10:58.000
Make this an infographic .
10:58.010–10:59.233
Translate this .
10:59.243–11:00.624
Write this in Python .
11:00.634–11:04.026
And Gemini quietly roots the request to whichever model actually
11:04.036–11:05.188
does that job .
11:05.198–11:07.754
That's the design philosophy in one sentence .
11:07.764–11:11.141
One platform , and it decides the plumbing so you don't have to
11:11.141–11:12.546
.
11:12.546–11:13.165
Where Gemini actually lives .
11:13.175–11:15.570
This is the part that's easy to underestimate .
11:15.580–11:17.935
Gemini isn't confined to one app .
11:17.945–11:20.621
It's spread across nearly everything Google ships .
11:20.631–11:23.467
In Search , it's AI mode , already covered .
11:23.477–11:26.794
In Gmail , it's behind Smart Compose and auto reply suggestions
11:26.794–11:27.979
.
11:27.979–11:31.355
In Docs , Sheets and Slides , Ultra and Pro Pro Pro get Gemini
11:31.365–11:35.418
drafting text , building formulas , and designing slide layouts
11:35.428–11:38.454
, pulling context from your own files when you let it .
11:38.464–11:41.884
In Drive , it can find and summarize documents for you .
11:41.894–11:44.755
In Google Meet , it's doing live caption translation .
11:44.765–11:48.345
On Android , especially Pixel devices , it's baked straight into
11:48.355–11:49.501
the voice assistant .
11:49.511–11:52.213
And there's a Chrome extension that lets the browser send page
11:52.223–11:55.683
content straight to Gemini , so you can ask questions about whatever
11:55.693–11:56.640
tab you're on .
11:56.650–12:00.508
And for developers , all of it is exposed through Google AI Studio
12:00.518–12:04.560
and the Gemini API , plus a newer platform called Antigravity for
12:04.570–12:07.008
building multi agent workflows on top of it .
12:07.018–12:09.175
The strategic point here isn't subtle .
12:09.185–12:12.084
Google isn't trying to win the best standalone chatbot argument
12:12.084–12:13.727
.
12:13.727–12:15.073
It's trying to make sure you're never more than one product away
12:15.083–12:18.403
from Gemini , no matter what you're doing on a Google device or
12:18.413–12:19.526
in a Google app .
12:19.536–12:22.215
Agents : the part that actually matters .
12:22.225–12:25.224
Now , here's the shift I promised earlier , the one that actually
12:25.234–12:27.350
changes what this platform is for .
12:27.360–12:31.323
Everything so far has been ask a question , get an answer .
12:31.333–12:34.693
Agents are Google trying to move Gemini past that entirely .
12:34.703–12:37.863
Deep research is the clearest example already live ,
12:37.873–12:42.076
plan , browse , synthesize , write , without you babysitting every
12:42.086–12:42.757
step .
12:42.767–12:46.489
Spark is the early , still limited attempt at a persistent personal
12:46.499–12:49.178
agent running continuously in the background .
12:49.188–12:50.493
And on the developer side ,
12:50.493–12:52.980
Antigravity lets companies build coordinated
12:52.990–12:54.420
teams of sub agents .
12:54.430–12:57.859
Google's own blog post gave an example of businesses running parallel
12:57.869–12:59.698
agents to analyze data at scale ,
12:59.698–13:02.020
rather than one model doing everything
13:02.030–13:02.980
sequentially .
13:02.990–13:04.820
Picture the difference in practice .
13:04.830–13:05.510
The old way ,
13:05.510–13:06.416
you ask Gemini ,
13:06.416–13:08.380
What should I know before a trip
13:08.390–13:09.260
to Japan ?
13:09.270–13:10.660
and it gives you a paragraph .
13:10.670–13:11.767
The agent way ,
13:11.767–13:12.365
you say ,
13:12.365–13:14.060
Plan my trip to Japan .
13:14.070–13:15.448
and it checks flights ,
13:15.448–13:16.980
compares hotel options ,
13:16.990–13:19.076
and comes back with an actual itinerary ,
13:19.076–13:20.340
pausing to confirm with
13:20.350–13:21.900
you before it books anything .
13:21.910–13:23.525
That's the same underlying model ,
13:23.525–13:25.020
just given permission to take
13:25.030–13:27.980
more than one step before handing control back to you .
13:27.990–13:29.925
None of this is science fiction anymore ,
13:29.925–13:30.980
and none of it is fully
13:30.990–13:31.980
finished , either .
13:31.990–13:33.220
That's the honest read .
13:33.230–13:35.580
Deep Research genuinely works today .
13:35.590–13:37.240
Spark is still in early testing ,
13:37.240–13:39.080
but the direction is unmistakable.
13:39.080–13:41.157
.
13:41.157–13:41.237
break it into steps ,
13:41.237–13:42.420
and execute most of them without you typing
13:42.430–13:46.020
a follow up for every single one .
13:46.030–13:48.020
What actually makes Gemini different ?
13:48.030–13:48.114
So ,
13:48.114–13:50.140
how does this stack up against everyone else building the
13:50.150–13:52.781
same kind of thing ?
13:52.791–13:53.258
Let's be balanced here ,
13:53.258–13:54.061
because Google's advantages are real ,
13:54.071–13:57.542
but so are its weak spots .
13:57.552–13:59.183
The clearest edge is data .
13:59.193–14:00.652
Gemini can pull from live search results ,
14:00.652–14:00.824
Maps ,
14:00.834–14:01.133
Gmail ,
14:01.133–14:03.825
and Drive in ways that a closed sandbox chatbot simply
14:03.835–14:08.587
can't match without plugins bolted on .
14:08.597–14:10.868
The second edge is reach .
14:10.878–14:12.468
Every Android phone is a potential Gemini client ,
14:12.478–14:15.470
and every Workspace business account already has it available .
14:15.480–14:18.951
No competitor has that kind of built in distribution .
14:18.961–14:19.966
And on raw benchmarks ,
14:19.966–14:22.032
Gemini 3 Pro topped the LM Arena leaderboard
14:22.032–14:24.511
,
14:24.511–14:24.591
which ,
14:24.591–14:26.034
regardless of how much weight you put on any single leaderboard
14:26.034–14:29.045
,
14:29.045–14:29.815
says Google's infrastructure and DeepMind's research are producing
14:29.825–14:30.362
real ,
14:30.362–14:32.241
top tier results ,
14:32.241–14:33.717
not just hype .
14:33.727–14:33.893
But ,
14:33.893–14:35.439
and this matters for credibility ,
14:35.439–14:36.598
Google is genuinely more
14:36.608–14:40.359
conservative about rollout than some competitors .
14:40.369–14:42.867
Deep Think and Spark are still gated behind ultra subscriptions
14:42.877–14:43.826
or limited testing ,
14:43.826–14:45.960
while some rivals ship new capabilities to
14:45.970–14:49.160
everyone at once .
14:49.170–14:50.360
And the tier structure itself , free , pro , ultra ,
14:50.370–14:54.840
API pricing , can be genuinely confusing next to a simpler flat
14:54.850–14:58.799
subscription from a competitor .
14:58.809–15:00.440
If you've ever opened the Gemini pricing page and closed it 5 minutes
15:00.450–15:03.525
later still unsure which plan you need , that's not just you .
15:03.535–15:07.029
Where this is actually headed , a few things are confirmed and
15:07.039–15:09.930
a few are still rumor , and it's worth keeping those separate .
15:09.940–15:12.870
Confirmed , Gemini 3 .
15:12.880–15:14.690
5 Pro is currently in partner testing with a public release expected
15:14.700–15:18.067
soon , and Google has already started training on Gemini 4 ,
15:18.077–15:21.532
soon , and Google has already started training on Gemini 4 ,
15:21.542–15:24.795
according to its own July 2026 announcement, though there's no
15:24.805–15:26.406
public timeline for that yet.
15:26.416–15:30.475
Workspace AI rollout continues expanding, and Gemini Live's regional
15:30.485–15:32.127
language support keeps growing.
15:32.137–15:35.309
Speculative and worth labeling clearly as such,
15:35.319–15:38.613
there's talk of a dedicated on-device AI chip for future Pixel
15:38.623–15:42.529
phones, and some experimental DeepMind research around 3D avatars
15:42.539–15:45.597
and world simulation that hasn't shipped as a product.
15:45.607–15:48.785
Treat both of those as possible, not coming.
15:48.795–15:50.976
Nothing official has confirmed either one.
15:50.986–15:51.773
The verdict.
15:51.783–15:53.367
So, where does that leave things?
15:53.377–15:57.033
Gemini in 2026 isn't a chatbot you occasionally open.
15:57.043–15:59.982
It's an AI layer Google has threaded through search,
15:59.992–16:04.285
Gmail, your documents, and increasingly your phone itself.
16:04.295–16:07.752
The models handle the thinking, the modes handle how you ask,
16:07.762–16:10.661
and agents like Deep Research are the clearest sign of where all
16:10.671–16:12.055
of it is actually heading.
16:12.065–16:15.124
If there's one thing worth trying this week, it's Deep Research
16:15.134–16:17.773
on something you'd normally spend an evening looking into yourself
16:17.783–16:20.822
, and actually watching it work instead of just reading the final
16:20.832–16:21.539
report.
16:21.549–16:24.328
Drop a comment with which piece of this surprised you most,
16:24.338–16:27.715
the model lineup, the agent side, or just how much of this you
16:27.725–16:29.628
were already using without realizing it.
16:29.638–16:32.417
I'll be back soon with a deeper breakdown on how Deep Research
16:32.427–16:35.127
actually performs against a real research task.
16:35.137–16:37.000
Thanks for watching, and I'll see you in the next one.
0:00.000–0:02.320
你可能已經在 Gemini 輸入過一些內容，
0:02.320–0:03.010
得到答案，
0:03.020–0:04.147
然後關閉分頁，
0:04.147–0:06.552
就像使用其他任何聊天機器人一樣。
0:06.562–0:07.663
問題在於，
0:07.663–0:10.372
這可能只佔到它實際功能的 10%。
0:10.382–0:11.073
正確使用它，
0:11.073–0:13.835
它就能悄悄接管你目前仍需手動處理的
0:13.845–0:15.666
近一半繁瑣工作。
0:15.676–0:18.077
我花了數小時梳理每個模型、
0:18.077–0:18.849
每個模式，
0:18.859–0:21.954
以及 Gemini 悄悄整合進去的每個產品，
0:21.964–0:23.983
而讓我停下來的一個數字是這個。
0:23.993–0:28.315
僅 Gemini 應用程式就有超過 6.5 億月活躍用戶，
0:28.325–0:31.520
這還不包括在搜尋中使用它的所有人。
0:31.530–0:33.830
大多數人可能只使用了它的 20%，
0:33.830–0:35.496
根本不知道其餘功能
0:35.506–0:36.469
甚至存在。
0:36.479–0:39.025
看看來自美國人口普查局的數據。
0:39.035–0:43.527
美國企業中，實際上只有約五分之一在其營運中使用 AI
0:43.527–0:44.402
。
0:44.402–0:44.523
所以，
0:44.523–0:45.554
如果你經營企業，
0:45.554–0:47.313
並且你甚至正在考慮這一點
0:47.313–0:48.602
，
0:48.602–0:50.394
你已經領先於大多數競爭對手。
0:50.404–0:53.876
你可能不知道的是，除了報導 AI 新聞外，
0:53.886–0:56.477
我們還與企業主合作，幫助他們在
0:56.487–0:57.958
企業中實施 AI。
0:57.968–1:00.879
我們的工程團隊會了解你的企業如何運作，
1:00.889–1:03.080
然後與你一起建立自動化系統。
1:03.090–1:05.480
你會在下方描述中找到連結。
1:05.490–1:05.987
點擊它，
1:05.987–1:08.400
填寫一份關於你企業的簡短表格，
1:08.410–1:10.119
我們會聯繫你安排通話。
1:10.129–1:10.302
所以，
1:10.302–1:11.253
在這部影片中，
1:11.253–1:14.279
我正在詳細拆解 Gemini 作為
1:14.289–1:15.428
2026 年中，
1:15.428–1:17.580
每個目前的模型，
1:17.580–1:18.719
所有模式，
1:18.729–1:21.518
以及 Google 悄悄內建的所有地方。
1:21.528–1:22.114
到了最後，
1:22.114–1:25.118
你會清楚知道該使用哪個 Gemini 工具，
1:25.128–1:27.478
取決於你實際上想做的事，
1:27.488–1:30.557
而不是隨便對著眼前的輸入框打字。
1:30.567–1:30.909
首先，
1:30.909–1:33.237
讓我們釐清最大的誤解。
1:33.247–1:35.477
Gemini 根本就不是單一產品。
1:35.487–1:36.945
Gemini 真正是什麼，
1:36.945–1:39.277
這裡有一個你需要在
1:39.287–1:40.717
理解這一切之前建立的思維模型。
1:40.727–1:42.152
Gemini 不是單一的人工智慧，
1:42.152–1:44.476
它是 Google 對整個
1:44.486–1:45.060
平台的總稱，
1:45.060–1:46.855
其下包含一系列模型，
1:46.855–1:48.076
以及頂層的一組產品，
1:48.086–1:50.236
讓你能夠實際與它們互動。
1:50.246–1:51.795
我們可以從兩個層面來理解。
1:51.805–1:53.884
底層是模型本身
1:53.884–1:54.835
像是 Gemini
1:54.845–1:55.368
3.0
1:55.378–1:57.315
6 Flash 或 Gemini 3.0
1:57.325–1:58.195
[未翻譯]
1:58.205–2:00.915
這些是針對不同任務進行優化的引擎
2:00.925–2:02.209
有些專為速度打造，
2:02.209–2:03.794
有些專為複雜推理設計，
2:03.804–2:05.634
有些則專為圖像或音訊處理。
2:05.644–2:07.994
除非你刻意去尋找，否則你不會看到這些名稱。
2:08.004–2:10.434
上層是你實際點擊的所有介面。
2:10.444–2:11.520
Gemini 應用程式、
2:11.520–2:13.673
Google 搜尋中的 AI 模式、
2:13.683–2:15.432
Gmail 和 Docs 中的 Gemini、
2:15.432–2:17.473
手機上的語音助手
2:17.473–2:17.983
。
2:17.983–2:20.673
所有這些都只是通往相同底層
2:20.683–2:21.433
模型的不同門戶。
2:21.443–2:23.072
這就是本影片的重點。
2:23.082–2:25.752
Google 並非試圖打造一個偉大的聊天機器人。
2:25.762–2:29.152
它試圖在你已經使用的每個產品背後，植入相同的 AI 大腦
2:29.162–2:29.832
使用。
2:29.842–2:30.010
所以，
2:30.010–2:31.854
讓我們從大腦開始，
2:31.854–2:33.112
也就是實際的模型，
2:33.122–2:35.751
因為一旦你知道每個模型的設計用途，
2:35.761–2:37.671
其他一切就都對上了。
2:37.681–2:39.231
目前的模型陣容。
2:39.241–2:40.818
這是一個示範清單，
2:40.818–2:42.631
所以我們將一個模型一個模型地介紹。
2:42.641–2:43.211
它是什麼，
2:43.211–2:44.779
它實際上擅長什麼，
2:44.779–2:45.990
以及你可以在哪裡取得
2:46.000–2:46.390
它。
2:46.400–2:47.350
[未翻譯]
2:47.360–2:48.830
7閃光燈
2:48.840–2:50.405
於 2026 年 8 月 13 日推出，
2:50.405–2:50.718
這是 Google 最新的 flash
2:50.718–2:52.595
模型，也是迄今為止最強大的主力模型。
2:52.605–2:55.119
它主要為程式設計和 AI 代理而設計，並在
2:55.129–2:58.885
它主要為編碼和 AI 代理而設計，並在軟體工程方面有重大改進，
2:58.895–3:00.371
軟體工程方面，
3:00.371–3:01.355
網頁開發、
3:01.355–3:02.690
以及複雜的多步驟
3:02.700–3:03.892
工作流程中。
3:03.902–3:06.616
Google 已透過 Gemini
3:06.626–3:07.069
API 提供一般可用性，
3:07.069–3:08.840
定位 3。
3:08.850–3:12.746
7 Flash 定位為當您需要強大智慧
3:12.756–3:15.549
同時不犧牲 flash 系列所具備的速度和效率。
3:15.559–3:16.270
所聞名的速度與效率時的首選模型。
3:16.280–3:17.285
[未翻譯]
3:17.295–3:18.354
閃電六型
3:18.364–3:20.140
這是 Google 目前的旗艦產品，
3:20.140–3:21.599
於 2026 年 7 月 21 日
3:21.609–3:23.621
的公司部落格文章中宣布。
3:23.621–3:24.196
它被設計為主力模型，
3:24.206–3:25.750
擅長程式碼編寫、
3:25.750–3:26.831
知識工作、
3:26.841–3:27.874
以及多模態任務。
3:27.874–3:29.305
以及多模態任務。
3:29.315–3:31.223
根據谷歌自己的數據，
3:31.223–3:32.577
它完成任務所需的
3:32.587–3:36.167
平均比其前身少約 17 個 token。
3:36.177–3:39.919
更少的 token 意味著更快的回答速度，以及如果你透過 API
3:39.929–3:41.515
付費的話，費用也會更低。
3:41.525–3:44.208
你可以透過 Gemini API 使用它，
3:44.208–3:45.465
透過 AI Studio，
3:45.475–3:49.267
或者只是簡單地使用 Gemini 應用程式並搜尋 AI 模式。
3:49.277–3:50.994
不需要額外的設定。
3:51.004–3:53.685
如果你只記得這部影片中的一個模型名稱，
3:53.695–3:57.139
那就記住這個，因為目前大多數的 Gemini 都在悄悄運行
3:57.149–3:58.020
這個模型。
3:58.030–3:59.061
[未翻譯]
3:59.071–4:00.101
第五代閃電模型
4:00.111–4:04.342
這個模型於 2026 年 5 月推出，它是
4:04.352–4:07.710
實際上在 3 之前為搜尋中的 AI 模式提供支援。
4:07.720–4:09.104
6 Flash 接管之前，實際驅動搜尋 AI 模式的模型。
4:09.114–4:12.465
Google 將其描述為具備卓越
4:12.475–4:14.490
速度的前沿級智能，並且根據其自身的基準測試，
4:14.490–4:16.506
它將輸出吞吐量推升至
4:16.516–4:19.746
速度大約是當時其他頂級模型的 4 倍。
4:19.756–4:23.068
它仍在 Gemini 應用程式、
4:23.078–4:26.629
Google 的 Antigravity 平台以及企業工具中廣泛使用。
4:26.639–4:28.017
它目前略遜於 3.
4:28.027–4:31.551
6，但它是今年出貨量中佔據巨大份額的幕後模型。
4:31.561–4:32.991
今年出貨量中的份額。
4:33.001–4:33.918
[未翻譯]
4:33.928–4:35.072
同樣是七月宣布，但任務完全不同。
4:35.082–4:38.073
這是經過簡化、高吞吐量的兄弟版本。
4:38.083–4:40.834
Google 標示每秒約 350 個 token，這是為
4:40.844–4:44.955
大量處理而非深度推理而設計的。
4:44.965–4:46.476
你不會用它來解決複雜的推理問題。
4:46.486–4:48.716
你會用它來處理背景任務和需要
4:48.726–4:52.198
快速且低成本大規模運作的代理管道。
4:52.208–4:54.198
以規模快速且便宜地進行。
4:54.208–4:55.092
1 Pro。
4:55.102–4:55.999
於 2026 年 2 月發布，這是推理專家。
4:56.009–5:00.080
Google 的基準測試聲稱其邏輯性能
5:00.090–5:03.041
大約是早期 Gemini 3 Pro 的兩倍。
5:03.051–5:04.922
較早的 Gemini 3 Pro。
5:04.932–5:08.283
它主要透過 API 的預覽存取來限制使用，
5:08.293–5:12.604
以及 Vertex AI 和消費者應用程式中的 Pro 或 Ultra 訂閱方案。
5:12.614–5:13.684
消費者應用程式。
5:13.694–5:14.325
如果 3.
5:14.335–5:16.525
6 Flash 是為速度而設計，3.
5:16.535–5:18.566
1 Pro 是為深度而設計。
5:18.576–5:21.207
你希望在真正困難的問題上使用的模型，
5:21.217–5:22.247
而不是快速的問題，你會想要使用的模型。
5:22.257–5:24.928
如果你確實是透過 API 在上面進行開發，
5:24.938–5:27.929
價格是做出現實世界決策的關鍵。
5:27.939–5:30.570
根據 Google 自己公布的費率，3.
5:30.580–5:32.517
6 Flash 每百萬輸入記號約為 1.
5:32.527–5:35.185
50，每百萬輸出記號約為 7.
5:35.195–5:37.092
每百萬輸出記號 50 美元。
5:37.102–5:40.933
作為參考，這在輸出端的價格明顯比
5:40.943–5:42.467
GPT 5 明顯便宜。
5:42.477–5:44.975
6 Luna 約為每百萬 6。
5:44.985–5:48.776
這個差距正是為什麼許多開發者對於任何大量運行的應用，預設選擇 flash 層級
5:48.786–5:51.176
任何大量運行的模型都要便宜。
5:51.186–5:54.898
現在，少數值得以名稱了解的專業模型，
5:54.908–5:56.778
即使我們不對每一個模型深入探討。
5:56.788–6:00.700
Nano Banana 2 是 Gemini 目前用於影像生成與編輯的模型，
6:00.710–6:04.041
完全取代了舊版的 Imagen 系列。
6:04.051–6:07.351
這可不是小細節，因為 Imagen 實際上將於 2026 年 8 月 17 日停止服務。
6:07.361–6:10.362
如果你已經有基於 Imagen 的工作流程，那麼倒數計時已經開始。
6:10.372–6:13.522
此外，還有一個純粹為速度而設計的 Nano Banana 2 輕量版，
6:13.532–6:14.181
以犧牲少量品質為代價，換取在大量產出時更快、更便宜的結果。
6:14.191–6:17.694
此外還有一款專為速度打造的 Nano Banana 2 輕量版
6:17.704–6:20.790
1 是 Google 的影片生成模型，目前仍處於測試階段，
6:20.800–6:22.753
旨在將文字提示轉換為帶有同步音訊的短片。
6:22.763–6:23.602
[未翻譯]
6:23.612–6:26.864
5 live translate 負責處理超過 70 種語言的即時語音對語音翻譯，
6:26.874–6:29.975
目前已內建於 Google Meet 和 Android 中。
6:29.975–6:30.907
如果你對產品線中較為小眾的部分感到好奇，
6:30.907–6:31.970
則有一款專注於安全的變體，名為 3 .
6:31.980–6:36.217
5 即時翻譯功能可處理超過 70 種語言的即時語音對語音翻譯
6:36.227–6:39.783
該功能已內建於 Google Meet 和 Android 中
6:39.793–6:40.655
以及 Android 系統
6:40.665–6:43.468
如果你對產品線中較為小眾的選項感興趣
6:43.478–6:46.137
則有一款專注於安全的 3 版本
6:46.147–6:50.185
5 Flash Cyber 旨在與漏洞掃描工具協調運作
6:50.185–6:50.898
。
6:50.898–6:52.027
以及 Lyra 3.5，
6:52.037–6:55.879
Google 的音樂模型，現在可以根據文字提示生成長達
6:55.889–6:58.019
從文字提示生成僅需 3 分鐘
6:58.029–7:00.238
這裡有一個值得指出的誠實限制。
7:00.248–7:04.041
Google 快速推出許多這些模型，無論是否出於刻意，命名方式變得令人困惑
7:04.051–7:05.230
。
7:05.240–7:05.765
[未翻譯]
7:05.775–7:06.905
[未翻譯]
7:06.915–7:07.980
[未翻譯]
7:07.990–7:10.820
閃電、閃電網絡
7:10.830–7:13.909
如果你不是專業地在 API 之上進行開發，
7:13.919–7:16.677
你實際上不需要記住這個列表。
7:16.687–7:18.522
你只需要了解其架構。
7:18.532–7:21.611
快速且便宜、深度推理，以及多模態。
7:21.621–7:24.298
這真的是三個類別，披著許多不同名稱
7:24.308–7:25.181
標籤。
7:25.191–7:28.952
你實際互動的模式，模型是引擎。
7:28.962–7:30.476
模式是方向盤。
7:30.486–7:33.324
這裡是對於非開發者來說最有用的地方
7:33.324–7:34.754
。
7:34.754–7:37.656
Google Search 內的 AI 模式會將你的搜尋列變成對話
7:37.656–7:39.188
。
7:39.188–7:40.386
截至 I/O 2026，它由 Gemini 3.5 Flash 全球驅動
7:40.386–7:42.544
並且不再是 10 個藍色連結，你會得到書面答案
7:42.554–7:46.361
5 flash 並且不再是 10 個藍色連結，你會得到書面答案
7:46.371–7:49.771
帶有後續問題，有時還會即時建構互動小工具
7:49.781–7:50.573
。
7:50.583–7:52.178
任何會搜尋的人都能使用它。
7:52.188–7:53.782
不需要訂閱
7:53.792–7:56.706
詢問類似「冰箱裡有什麼可以做的快速晚餐」
7:56.706–8:00.354
？
8:00.354–8:00.699
它會以完整句子回答，而不是食譜部落格清單。
8:00.709–8:04.005
Deep Think 是 Gemini 應用程式的額外努力版本。
8:04.015–8:07.773
它會為每個答案投入更多運算資源，逐步推理較難的問題
8:07.783–8:08.815
一步一步。
8:08.825–8:11.406
Google 將此功能限制在 Google AI Ultra 之後
8:11.406–8:15.468
並且它是為真正困難的科學或工程問題而設計，
8:15.478–8:16.831
而非日常聊天。
8:16.841–8:19.927
現在，Deep Research 讓這裡不再只是聊天機器人
8:19.927–8:21.641
而是開始成為一位助手。
8:21.651–8:22.815
你給它一個主題
8:22.815–8:27.653
它不會只回覆一次，而是規劃研究策略、開啟網頁，
8:27.663–8:28.420
閱讀內容
8:28.420–8:32.622
如果你允許，它還會從你自己的 Gmail 和 Drive 提取資料。
8:32.632–8:34.585
回傳的不是一段文字。
8:34.595–8:38.234
而是 Gemini 畫布內一份完整的多頁報告。
8:38.244–8:42.362
這部分是 Gemini 真正配得上「代理（agent）」一詞的地方，
8:42.372–8:45.047
我們稍後會再回來討論為什麼這很重要。
8:45.057–8:47.652
Gemini Live 是語音和攝影機模式。
8:47.662–8:51.540
說聲「Hey Google，我們來聊天」，你就可以解放雙手與它對話
8:51.550–8:54.226
並可選擇將攝影機指向某物，詢問它
8:54.236–8:55.829
即時看到些什麼。
8:55.839–8:57.913
而 Canvas 是工作空間模式。
8:57.923–9:01.761
輸入「建立一個關於行星的測驗應用程式」，它會一次寫出程式碼、
9:01.771–9:05.448
介面與內容一次呈現，直接擺在你面前
9:05.458–9:06.089
供你編輯。
9:06.099–9:09.416
還有一個值得簡短提及的功能，因為它目前還處於早期階段。
9:09.426–9:14.145
Gemini Spark，一款在 IO 2026 上宣布的個人代理，
9:14.155–9:17.031
旨在於背景中持續運行，處理諸如
9:17.041–9:17.953
排程等事項。
9:17.963–9:20.598
目前，它僅限於早期 Ultra 測試人員使用。
9:20.608–9:23.043
因此，請將此視為即將推出，而非現已可用。
9:23.053–9:25.047
在繼續之前，快速進行直覺檢查。
9:25.057–9:27.532
如果上述內容聽起來像是許多獨立工具，
9:27.542–9:28.374
這很合理。
9:28.384–9:29.696
但請注意這個模式。
9:29.706–9:32.676
這些模式中的每一個都只是 Gemini 3 .
9:32.686–9:33.664
5 或 3 .
9:33.674–9:35.909
6 flash 換了不同的職位頭銜。
9:35.919–9:39.316
你並非在學習六種不同的 AI，而是在學習六種不同的
9:39.326–9:41.520
向同一個大腦尋求幫助的方式。
9:41.530–9:43.192
實務上的多模態。
9:43.202–9:46.257
讓我們談談多模態在日常生活中的實際意義，
9:46.267–9:48.527
因為它不僅是投影片上的流行詞。
9:48.537–9:50.756
Gemini 能讀寫文字和程式碼。
9:50.766–9:52.707
顯然，這是基本功能。
9:52.717–9:56.170
但將照片放入對話中，並要求它為照片加註說明或編輯，
9:56.180–9:57.842
Nano Banana 就能處理這些任務。
9:57.852–10:01.284
要求它大聲說出答案，Gemini 的音訊模型
10:01.294–10:05.089
即時生成該語音，並能實際控制語調和
10:05.099–10:05.810
節奏。
10:05.820–10:09.135
要求製作一段短片，VO 會從頭開始建立。
10:09.145–10:12.259
要求它編輯現有片段、替換天空、
10:12.269–10:16.024
改變風格，這是由另一個名為 Gemini Omni 的工具
10:16.034–10:18.668
透過語音指令逐幀編輯。
10:18.678–10:22.393
在 Google 文件和試算表中，相同的底層模型可以
10:22.403–10:25.838
根據您的會議記錄起草文件，或從一堆發票中建立試算表，
10:25.848–10:29.843
從一堆發票中，包含公式與圖表。
10:29.853–10:32.166
不只是將原始數字轉儲到儲存格中。
10:32.176–10:35.611
在投影片簡報中，提供它一份項目清單，它可以佈局
10:35.621–10:38.936
實際的簡報，而不僅僅是空白投影片上的文字。
10:38.946–10:42.020
透過 Gemini Live 的相機模式，您可以將手機指向
10:42.030–10:45.479
您不懂的語言的菜單，並獲得疊加在您所見內容上的即時翻譯。
10:45.489–10:47.149
或者要求它識別鏡頭所拍攝的物體
10:47.159–10:50.229
。
10:50.229–10:51.979
不需要打字。
10:51.979–10:52.059
這是值得記住的部分。
10:52.059–10:53.588
這部分值得記住。
10:53.598–10:55.059
你永遠不需要選擇模型。
10:55.069–10:56.410
你只需要說出你想要什麼。
10:56.420–10:58.000
將這個做成資訊圖表。
10:58.010–10:59.233
翻譯這個。
10:59.243–11:00.624
用 Python 寫這段程式。
11:00.634–11:04.026
然後 Gemini 會靜默地將請求路由到真正執行該工作的模型。
11:04.036–11:05.188
來完成那項工作。
11:05.198–11:07.754
這就是設計哲學的簡短總結。
11:07.764–11:11.141
單一平台，它負責處理底層架構，所以你不必操心。
11:11.141–11:12.546
。
11:12.546–11:13.165
Gemini 實際運行的地方。
11:13.175–11:15.570
這部分是很容易被低估的。
11:15.580–11:17.935
Gemini 並不限於單一應用程式。
11:17.945–11:20.621
它遍布於 Google 推出的幾乎所有產品中。
11:20.631–11:23.467
在搜尋中，它是 AI 模式，前面已經介紹過。
11:23.477–11:26.794
在 Gmail 中，它位於智慧撰寫和自動回覆建議背後。
11:26.794–11:27.979
。
11:27.979–11:31.355
在 Docs、Sheets 和 Slides 中，Ultra 和 Pro 版讓 Gemini
11:31.365–11:35.418
起草文字、建立公式，並設計投影片版面。
11:35.428–11:38.454
，在你允許的情況下，從你自己的檔案中提取內容。
11:38.464–11:41.884
在 Drive 中，它可以為您尋找並摘要文件。
11:41.894–11:44.755
在 Google Meet 中，它正在進行即時字幕翻譯。
11:44.765–11:48.345
在 Android 上，特別是 Pixel 裝置，它已直接內建於
11:48.355–11:49.501
語音助理中。
11:49.511–11:52.213
還有一個 Chrome 擴充功能，讓瀏覽器能將頁面
11:52.223–11:55.683
內容直接傳送給 Gemini，因此您可以針對目前
11:55.693–11:56.640
的分頁提出問題。
11:56.650–12:00.508
對於開發者而言，所有功能都透過 Google AI Studio
12:00.518–12:04.560
和 Gemini API 公開，以及一個名為 Antigravity 的新平台，用於
12:04.570–12:07.008
在其上建構多代理工作流程。
12:07.018–12:09.175
這裡的戰略重點並不隱晦。
12:09.185–12:12.084
Google 並非試圖贏得最佳獨立聊天機器人
12:12.084–12:13.727
的爭論。
12:13.727–12:15.073
它正試圖確保無論您在 Google 裝置上
12:15.083–12:18.403
或 Google 應用程式中進行什麼操作，您與
12:18.413–12:19.526
Gemini 之間永遠只隔一個產品。
12:19.536–12:22.215
代理：真正重要的部分。
12:22.225–12:25.224
現在，這裡有我之前承諾的轉變，那個真正
12:25.234–12:27.350
改變這個平台用途的轉變。
12:27.360–12:31.323
到目前為止的一切都是提問、獲得答案。
12:31.333–12:34.693
代理程式是 Google 試圖讓 Gemini 完全超越此階段的做法。
12:34.703–12:37.863
深度研究（Deep research）是目前已上線最清楚的範例，
12:37.873–12:42.076
它能規劃、瀏覽、綜合資訊並撰寫內容，
12:42.086–12:42.757
全程無需你一步步監督。
12:42.767–12:46.489
Spark 是早期且功能仍受限的嘗試，
12:46.499–12:49.178
旨在讓個人代理程式在背景持續運作。
12:49.188–12:50.493
而在開發者方面，
12:50.493–12:52.980
Antigravity 讓企業能夠建構
12:52.990–12:54.420
由子代理程式組成的協作團隊。
12:54.430–12:57.859
Google 自己的部落格文章舉例說明，企業可運行平行
12:57.869–12:59.698
代理程式來進行大規模資料分析，
12:59.698–13:02.020
而不是一個模型處理所有事情
13:02.030–13:02.980
依序執行所有任務。
13:02.990–13:04.820
想像一下實際應用上的差異。
13:04.830–13:05.510
舊有的方式是，
13:05.510–13:06.416
你詢問 Gemini，
13:06.416–13:08.380
旅行前我應該知道什麼
13:08.390–13:09.260
我該知道什麼？」
13:09.270–13:10.660
它會給你一段文字說明。
13:10.670–13:11.767
而代理程式的方式是，
13:11.767–13:12.365
你說，
13:12.365–13:14.060
幫我規劃去日本的行程。
13:14.070–13:15.448
然後它會檢查航班，
13:15.448–13:16.980
比較飯店選項，
13:16.990–13:19.076
並回傳實際的行程表，
13:19.076–13:20.340
在預訂任何東西之前暫停並與
13:20.350–13:21.900
在預訂任何東西之前先詢問你
13:21.910–13:23.525
這是相同的基礎模型，
13:23.525–13:25.020
只是被賦予了權限，
13:25.030–13:27.980
可以在將控制權交還給你之前採取多於一個步驟。
13:27.990–13:29.925
這些已不再是科幻小說，
13:29.925–13:30.980
而且也沒有任何一項是
13:30.990–13:31.980
完全完成的。
13:31.990–13:33.220
這是誠實的評估。
13:33.230–13:35.580
Deep Research 在今天確實有效。
13:35.590–13:37.240
Spark 仍在早期測試中，
13:37.240–13:39.080
但方向是明確無疑的。
13:39.080–13:41.157
。
13:41.157–13:41.237
將其分解為步驟，
13:41.237–13:42.420
並在沒有你輸入的情況下執行大部分步驟
13:42.430–13:46.020
針對每一項都進行後續追蹤。
13:46.030–13:48.020
究竟是什麼讓 Gemini 有所不同？
13:48.030–13:48.114
所以，
13:48.114–13:50.140
這與其他所有正在構建的人相比如何
13:50.150–13:52.781
相同類型事物的產品相比如何？
13:52.791–13:53.258
讓我們在這裡保持客觀，
13:53.258–13:54.061
因為 Google 的優勢是真實存在的，
13:54.071–13:57.542
但它的弱點也是如此。
13:57.552–13:59.183
最明顯的優勢在於數據。
13:59.193–14:00.652
Gemini 可以從即時搜尋結果中獲取資訊，
14:00.652–14:00.824
地圖，
14:00.834–14:01.133
[未翻譯]
14:01.133–14:03.825
以及 Drive，這些是封閉的沙盒聊天機器人
14:03.835–14:08.587
僅靠外接外掛就無法匹敵的。
14:08.597–14:10.868
第二個優勢在於觸及範圍。
14:10.878–14:12.468
每一支 Android 手機都是潛在的 Gemini 客戶，
14:12.478–14:15.470
而且每個 Workspace 企業帳戶都已提供此服務。
14:15.480–14:18.951
沒有任何競爭對手擁有這種內建的分發能力。
14:18.961–14:19.966
而在原始基準測試中，
14:19.966–14:22.032
Gemini 3 Pro 登上了 LM Arena 排行榜的首位
14:22.032–14:24.511
，
14:24.511–14:24.591
這意味著
14:24.591–14:26.034
無論你對任何單一排行榜賦予多少權重
14:26.034–14:29.045
，
14:29.045–14:29.815
都顯示 Google 的基礎設施與 DeepMind 的研究正在產生
14:29.825–14:30.362
真實的、
14:30.362–14:32.241
頂級的成果，
14:32.241–14:33.717
而不只是炒作。
14:33.727–14:33.893
不過，
14:33.893–14:35.439
這對於可信度很重要，
14:35.439–14:36.598
Google 在產品推出方面確實比某些競爭對手
14:36.608–14:40.359
在推出方面比一些競爭對手更為保守
14:40.369–14:42.867
Deep Think 和 Spark 仍受限於超級訂閱
14:42.877–14:43.826
或有限測試，
14:43.826–14:45.960
而一些競爭對手則會將新功能
14:45.970–14:49.160
一次性提供給所有人。
14:49.170–14:50.360
而且層級結構本身，免費、專業、超級，
14:50.370–14:54.840
加上 API 定價，與競爭對手較簡單的單一
14:54.850–14:58.799
訂閱方案相比，可能會讓人真正感到困惑。
14:58.809–15:00.440
如果你曾經打開過 Gemini 的定價頁面，五分鐘後就關閉
15:00.450–15:03.525
之後仍然不確定你需要哪個方案，這不只是你一個人的問題。
15:03.535–15:07.029
這裡實際上的走向是，有些事項已經確認，而有些仍是謠言，將這兩者區分開來是值得的。
15:07.039–15:09.930
有些事項已經確認，而有些仍是謠言，將這兩者區分開來是值得的。
15:09.940–15:12.870
已確認的是 Gemini 3。
15:12.880–15:14.690
5 Pro 目前正處於合作夥伴測試階段，預計很快就會公開發布。
15:14.700–15:18.067
不久後，Google 已經開始針對 Gemini 4 進行訓練，
15:18.077–15:21.532
不久後，Google 已經開始針對 Gemini 4 進行訓練，
15:21.542–15:24.795
根據其 2026 年 7 月的公告，雖然目前還沒有
15:24.805–15:26.406
公開的時間表。
15:26.416–15:30.475
Workspace AI 的推出持續擴大，而 Gemini Live 的地區
15:30.485–15:32.127
語言支援也不斷增加。
15:32.137–15:35.309
屬於猜測性質，值得明確標示為猜測，
15:35.319–15:38.613
有傳言稱未來 Pixel 將配備專用設備端 AI 晶片
15:38.623–15:42.529
手機將配備專屬的裝置端 AI 晶片，以及 DeepMind 針對 3D 虛擬角色
15:42.539–15:45.597
和世界模擬的一些實驗性研究，但這些尚未作為產品推出。
15:45.607–15:48.785
將這兩者視為可能發生，但並非必然到來。
15:48.795–15:50.976
沒有任何官方事項確認了其中任何一項。
15:50.986–15:51.773
結論。
15:51.783–15:53.367
那麼，這讓情況處於什麼狀態呢？
15:53.377–15:57.033
2026 年的 Gemini 不是一個你偶爾才會開啟的聊天機器人。
15:57.043–15:59.982
這是 Google 貫穿搜尋、
15:59.992–16:04.285
Gmail、你的文件，以及日益深入地嵌入你手機本身的 AI 層。
16:04.295–16:07.752
模型負責思考，模式負責你提問的方式，
16:07.762–16:10.661
而像 Deep Research 這樣的代理程式，則是最清楚地顯示出所有這些
16:10.671–16:12.055
最終將走向何方的跡象。
16:12.065–16:15.124
如果這週只推薦嘗試一件事，那就是 Deep Research
16:15.134–16:17.773
是對你通常會花一個晚上自行調查的事物進行 Deep Research，
16:17.783–16:20.822
並實際觀看它的運作過程，而不僅僅是閱讀最終
16:20.832–16:21.539
報告。
16:21.549–16:24.328
在留言區告訴我，這其中哪一部分最讓你驚訝，
16:24.338–16:27.715
模型陣容、代理端，還是你其實早就在無意識中使用的這些功能
16:27.725–16:29.628
還是你其實已經在無意識中使用這麼多功能。
16:29.638–16:32.417
我很快會回來，深入剖析 Deep Research
16:32.427–16:35.127
在面對真實研究任務時的實際表現。
16:35.137–16:37.000
感謝觀看，我們下一部影片見。
0:00.000–0:02.320
You've probably typed something into Gemini ,
你可能已經在 Gemini 輸入過一些內容，
0:02.320–0:03.010
got an answer ,
得到答案，
0:03.020–0:04.147
and closed the tab ,
然後關閉分頁，
0:04.147–0:06.552
the same way you'd use any other chatbot .
就像使用其他任何聊天機器人一樣。
0:06.562–0:07.663
Here's the thing,
問題在於，
0:07.663–0:10.372
that's maybe 10% of what it actually does.
這可能只佔到它實際功能的 10%。
0:10.382–0:11.073
Use it right ,
正確使用它，
0:11.073–0:13.835
and it can quietly take over almost half the busy
它就能悄悄接管你目前仍需手動處理的
0:13.845–0:15.666
work you're still doing by hand .
近一半繁瑣工作。
0:15.676–0:18.077
I spent hours mapping every model ,
我花了數小時梳理每個模型、
0:18.077–0:18.849
every mode ,
每個模式，
0:18.859–0:21.954
and every product Gemini is quietly wired into ,
以及 Gemini 悄悄整合進去的每個產品，
0:21.964–0:23.983
and the number that stopped me was this .
而讓我停下來的一個數字是這個。
0:23.993–0:28.315
Gemini's app alone has over 650 million monthly users ,
僅 Gemini 應用程式就有超過 6.5 億月活躍用戶，
0:28.325–0:31.520
and that's before you count everyone using it inside search .
這還不包括在搜尋中使用它的所有人。
0:31.530–0:33.830
Most of them are using maybe 20% of it,
大多數人可能只使用了它的 20%，
0:33.830–0:35.496
with no idea the rest even
根本不知道其餘功能
0:35.506–0:36.469
exists .
甚至存在。
0:36.479–0:39.025
Look at this data from the Census Bureau .
看看來自美國人口普查局的數據。
0:39.035–0:43.527
Only about one in five US businesses actually use AI in their operations
美國企業中，實際上只有約五分之一在其營運中使用 AI
0:43.527–0:44.402
.
。
0:44.402–0:44.523
So ,
所以，
0:44.523–0:45.554
if you run a business ,
如果你經營企業，
0:45.554–0:47.313
and you're even thinking about this
並且你甚至正在考慮這一點
0:47.313–0:48.602
,
，
0:48.602–0:50.394
you're ahead of most of your competition .
你已經領先於大多數競爭對手。
0:50.404–0:53.876
What you might not know is that alongside covering AI news ,
你可能不知道的是，除了報導 AI 新聞外，
0:53.886–0:56.477
we work with business owners to help them implement AI in their
我們還與企業主合作，幫助他們在
0:56.487–0:57.958
business .
企業中實施 AI。
0:57.968–1:00.879
Our engineering team gets to know how your business runs ,
我們的工程團隊會了解你的企業如何運作，
1:00.889–1:03.080
then builds the automation with you .
然後與你一起建立自動化系統。
1:03.090–1:05.480
You'll find the link in the description below .
你會在下方描述中找到連結。
1:05.490–1:05.987
Click it ,
點擊它，
1:05.987–1:08.400
fill out a short form about your business ,
填寫一份關於你企業的簡短表格，
1:08.410–1:10.119
and we'll get in touch to set up a call .
我們會聯繫你安排通話。
1:10.129–1:10.302
So ,
所以，
1:10.302–1:11.253
in this video ,
在這部影片中，
1:11.253–1:14.279
I'm breaking down exactly what Gemini is as
我正在詳細拆解 Gemini 作為
1:14.289–1:15.428
of mid 2026 ,
2026 年中，
1:15.428–1:17.580
every current model ,
每個目前的模型，
1:17.580–1:18.719
every mode ,
所有模式，
1:18.729–1:21.518
and everywhere Google has quietly built it in .
以及 Google 悄悄內建的所有地方。
1:21.528–1:22.114
By the end ,
到了最後，
1:22.114–1:25.118
you'll know exactly which Gemini tool to reach for
你會清楚知道該使用哪個 Gemini 工具，
1:25.128–1:27.478
depending on what you're actually trying to do ,
取決於你實際上想做的事，
1:27.488–1:30.557
instead of just typing into whichever box is in front of you .
而不是隨便對著眼前的輸入框打字。
1:30.567–1:30.909
First ,
首先，
1:30.909–1:33.237
let's clear up the biggest misconception .
讓我們釐清最大的誤解。
1:33.247–1:35.477
Gemini isn't one product at all .
Gemini 根本就不是單一產品。
1:35.487–1:36.945
What Gemini actually is ,
Gemini 真正是什麼，
1:36.945–1:39.277
here's the mental model you need before
這裡有一個你需要在
1:39.287–1:40.717
any of this makes sense .
理解這一切之前建立的思維模型。
1:40.727–1:42.152
Gemini isn't a single AI ,
Gemini 不是單一的人工智慧，
1:42.152–1:44.476
it's Google's umbrella name for a whole
它是 Google 對整個
1:44.486–1:45.060
platform ,
平台的總稱，
1:45.060–1:46.855
a family of models underneath ,
其下包含一系列模型，
1:46.855–1:48.076
and a set of products
以及頂層的一組產品，
1:48.086–1:50.236
on top that let you actually talk to them .
讓你能夠實際與它們互動。
1:50.246–1:51.795
Think of it in two layers .
我們可以從兩個層面來理解。
1:51.805–1:53.884
The bottom layer is the models themselves,
底層是模型本身
1:53.884–1:54.835
things like Gemini
像是 Gemini
1:54.845–1:55.368
3 .
3.0
1:55.378–1:57.315
6 Flash or Gemini 3 .
6 Flash 或 Gemini 3.0
1:57.325–1:58.195
1 Pro .
[未翻譯]
1:58.205–2:00.915
These are the engines tuned for different jobs .
這些是針對不同任務進行優化的引擎
2:00.925–2:02.209
Some built for speed ,
有些專為速度打造，
2:02.209–2:03.794
some for heavy reasoning ,
有些專為複雜推理設計，
2:03.804–2:05.634
some for images or audio .
有些則專為圖像或音訊處理。
2:05.644–2:07.994
You never see these names unless you go looking .
除非你刻意去尋找，否則你不會看到這些名稱。
2:08.004–2:10.434
The top layer is everything you actually click on .
上層是你實際點擊的所有介面。
2:10.444–2:11.520
The Gemini app ,
Gemini 應用程式、
2:11.520–2:13.673
AI mode inside Google search ,
Google 搜尋中的 AI 模式、
2:13.683–2:15.432
Gemini inside Gmail and Docs ,
Gmail 和 Docs 中的 Gemini、
2:15.432–2:17.473
the voice assistant on your phone
手機上的語音助手
2:17.473–2:17.983
.
。
2:17.983–2:20.673
All of those are just different doors into the same underlying
所有這些都只是通往相同底層
2:20.683–2:21.433
models .
模型的不同門戶。
2:21.443–2:23.072
That's the whole point of this video .
這就是本影片的重點。
2:23.082–2:25.752
Google isn't trying to build one great chatbot .
Google 並非試圖打造一個偉大的聊天機器人。
2:25.762–2:29.152
It's trying to put the same AI brain behind every product you already
它試圖在你已經使用的每個產品背後，植入相同的 AI 大腦
2:29.162–2:29.832
use .
使用。
2:29.842–2:30.010
So ,
所以，
2:30.010–2:31.854
let's start with the brains ,
讓我們從大腦開始，
2:31.854–2:33.112
the actual models ,
也就是實際的模型，
2:33.122–2:35.751
because once you know what each one is built for ,
因為一旦你知道每個模型的設計用途，
2:35.761–2:37.671
everything else clicks into place .
其他一切就都對上了。
2:37.681–2:39.231
The current model lineup .
目前的模型陣容。
2:39.241–2:40.818
This is a demo checklist ,
這是一個示範清單，
2:40.818–2:42.631
so we're going model by model .
所以我們將一個模型一個模型地介紹。
2:42.641–2:43.211
What it is ,
它是什麼，
2:43.211–2:44.779
what it's actually good for ,
它實際上擅長什麼，
2:44.779–2:45.990
and where you can get
以及你可以在哪裡取得
2:46.000–2:46.390
it .
它。
2:46.400–2:47.350
Gemini 3 .
[未翻譯]
2:47.360–2:48.830
7 flash .
7閃光燈
2:48.840–2:50.405
Launched on August 13th ,
於 2026 年 8 月 13 日推出，
2:50.405–2:50.718
2026 ,
這是 Google 最新的 flash
2:50.718–2:52.595
this is Google's newest flash
模型，也是迄今為止最強大的主力模型。
2:52.605–2:55.119
model and its most capable workhorse yet .
它主要為程式設計和 AI 代理而設計，並在
2:55.129–2:58.885
It's built primarily for coding and AI agents with major improvements
它主要為編碼和 AI 代理而設計，並在軟體工程方面有重大改進，
2:58.895–3:00.371
in software engineering ,
軟體工程方面，
3:00.371–3:01.355
web development ,
網頁開發、
3:01.355–3:02.690
and complex multi step
以及複雜的多步驟
3:02.700–3:03.892
workflows .
工作流程中。
3:03.902–3:06.616
Google has already made it generally available through the Gemini
Google 已透過 Gemini
3:06.626–3:07.069
API ,
API 提供一般可用性，
3:07.069–3:08.840
positioning 3 .
定位 3。
3:08.850–3:12.746
7 flash as the new go to model when you want strong intelligence
7 Flash 定位為當您需要強大智慧
3:12.756–3:15.549
without giving up the speed and efficiency the flash lineup is
同時不犧牲 flash 系列所具備的速度和效率。
3:15.559–3:16.270
known for .
所聞名的速度與效率時的首選模型。
3:16.280–3:17.285
Gemini 3 .
[未翻譯]
3:17.295–3:18.354
6 flash .
閃電六型
3:18.364–3:20.140
This is Google's current flagship ,
這是 Google 目前的旗艦產品，
3:20.140–3:21.599
announced in a company blog
於 2026 年 7 月 21 日
3:21.609–3:23.621
post on July 21st ,
的公司部落格文章中宣布。
3:23.621–3:24.196
2026 .
它被設計為主力模型，
3:24.206–3:25.750
It's built as a workhorse ,
擅長程式碼編寫、
3:25.750–3:26.831
strong at coding ,
知識工作、
3:26.841–3:27.874
knowledge work ,
以及多模態任務。
3:27.874–3:29.305
and multimodal tasks .
以及多模態任務。
3:29.315–3:31.223
And according to Google's own numbers ,
根據谷歌自己的數據，
3:31.223–3:32.577
it does the job using about
它完成任務所需的
3:32.587–3:36.167
17 fewer tokens on average than its predecessor .
平均比其前身少約 17 個 token。
3:36.177–3:39.919
Fewer tokens means faster answers and a lower bill if you're paying
更少的 token 意味著更快的回答速度，以及如果你透過 API
3:39.929–3:41.515
for it through the API .
付費的話，費用也會更低。
3:41.525–3:44.208
You can reach it through the Gemini API ,
你可以透過 Gemini API 使用它，
3:44.208–3:45.465
through AI Studio ,
透過 AI Studio，
3:45.475–3:49.267
or simply by using the Gemini app and searches AI mode .
或者只是簡單地使用 Gemini 應用程式並搜尋 AI 模式。
3:49.277–3:50.994
No extra setup required .
不需要額外的設定。
3:51.004–3:53.685
If you only remember one model name from this video ,
如果你只記得這部影片中的一個模型名稱，
3:53.695–3:57.139
make it this one because it's what most of Gemini is quietly running
那就記住這個，因為目前大多數的 Gemini 都在悄悄運行
3:57.149–3:58.020
on right now .
這個模型。
3:58.030–3:59.061
Gemini 3 .
[未翻譯]
3:59.071–4:00.101
5 Flash .
第五代閃電模型
4:00.111–4:04.342
This one launched back in May 2026 and it's the model that was
這個模型於 2026 年 5 月推出，它是
4:04.352–4:07.710
actually powering AI mode in search before 3 .
實際上在 3 之前為搜尋中的 AI 模式提供支援。
4:07.720–4:09.104
6 Flash took over .
6 Flash 接管之前，實際驅動搜尋 AI 模式的模型。
4:09.114–4:12.465
Google described it as frontier level intelligence at exceptional
Google 將其描述為具備卓越
4:12.475–4:14.490
speed and by its own benchmarks ,
速度的前沿級智能，並且根據其自身的基準測試，
4:14.490–4:16.506
it pushed output throughput to
它將輸出吞吐量推升至
4:16.516–4:19.746
roughly four times faster than other top models at the time .
速度大約是當時其他頂級模型的 4 倍。
4:19.756–4:23.068
It's still very much in active use across the Gemini app ,
它仍在 Gemini 應用程式、
4:23.078–4:26.629
Google's anti gravity platform , and enterprise tools .
Google 的 Antigravity 平台以及企業工具中廣泛使用。
4:26.639–4:28.017
It's slightly behind 3 .
它目前略遜於 3.
4:28.027–4:31.551
6 now , but it's the model quietly sitting behind a huge share
6，但它是今年出貨量中佔據巨大份額的幕後模型。
4:31.561–4:32.991
of what shipped this year .
今年出貨量中的份額。
4:33.001–4:33.918
Gemini 3 .
[未翻譯]
4:33.928–4:35.072
5 Flashlight .
同樣是七月宣布，但任務完全不同。
4:35.082–4:38.073
Same July announcement , different job entirely .
這是經過簡化、高吞吐量的兄弟版本。
4:38.083–4:40.834
This is the stripped down , high throughput sibling .
Google 標示每秒約 350 個 token，這是為
4:40.844–4:44.955
Google sites roughly 350 tokens per second , which is built for
大量處理而非深度推理而設計的。
4:44.965–4:46.476
volume , not depth .
你不會用它來解決複雜的推理問題。
4:46.486–4:48.716
You wouldn't use this for a hard reasoning problem .
你會用它來處理背景任務和需要
4:48.726–4:52.198
You'd use it for background tasks and agent pipelines that need
快速且低成本大規模運作的代理管道。
4:52.208–4:54.198
to move fast and cheap at scale .
以規模快速且便宜地進行。
4:54.208–4:55.092
Gemini 3 .
1 Pro。
4:55.102–4:55.999
1 Pro .
於 2026 年 2 月發布，這是推理專家。
4:56.009–5:00.080
Released in February 2026 , this is the reasoning specialist .
Google 的基準測試聲稱其邏輯性能
5:00.090–5:03.041
Google's benchmarks claim roughly double the logic performance
大約是早期 Gemini 3 Pro 的兩倍。
5:03.051–5:04.922
of the earlier Gemini 3 Pro .
較早的 Gemini 3 Pro。
5:04.932–5:08.283
It's mostly gated behind preview access through the API ,
它主要透過 API 的預覽存取來限制使用，
5:08.293–5:12.604
anti gravity , Vertex AI , and Pro or Ultra subscriptions in the
以及 Vertex AI 和消費者應用程式中的 Pro 或 Ultra 訂閱方案。
5:12.614–5:13.684
consumer app .
消費者應用程式。
5:13.694–5:14.325
If 3 .
如果 3.
5:14.335–5:16.525
6 Flash is built for speed , 3 .
6 Flash 是為速度而設計，3.
5:16.535–5:18.566
1 Pro is built for depth .
1 Pro 是為深度而設計。
5:18.576–5:21.207
The model you'd want on a genuinely hard problem ,
你希望在真正困難的問題上使用的模型，
5:21.217–5:22.247
not a quick one .
而不是快速的問題，你會想要使用的模型。
5:22.257–5:24.928
And if you're actually building on top of these through the API
如果你確實是透過 API 在上面進行開發，
5:24.938–5:27.929
, price is where the real world decision gets made .
價格是做出現實世界決策的關鍵。
5:27.939–5:30.570
According to Google's own published rates , 3 .
根據 Google 自己公布的費率，3.
5:30.580–5:32.517
6 Flash runs about 1 .
6 Flash 每百萬輸入記號約為 1.
5:32.527–5:35.185
50 per million input tokens and 7 .
50，每百萬輸出記號約為 7.
5:35.195–5:37.092
50 per million output tokens .
每百萬輸出記號 50 美元。
5:37.102–5:40.933
For context , that's noticeably cheaper on the output side than
作為參考，這在輸出端的價格明顯比
5:40.943–5:42.467
GPT 5 .
GPT 5 明顯便宜。
5:42.477–5:44.975
6 Luna's roughly 6 per million .
6 Luna 約為每百萬 6。
5:44.985–5:48.776
That gap is exactly why so many developers default to flash tier
這個差距正是為什麼許多開發者對於任何大量運行的應用，預設選擇 flash 層級
5:48.786–5:51.176
models for anything running at volume .
任何大量運行的模型都要便宜。
5:51.186–5:54.898
Now , a handful of specialty models worth knowing by name ,
現在，少數值得以名稱了解的專業模型，
5:54.908–5:56.778
even if we don't dwell on each one .
即使我們不對每一個模型深入探討。
5:56.788–6:00.700
Nano Banana 2 is Gemini's current image generation and editing
Nano Banana 2 是 Gemini 目前用於影像生成與編輯的模型，
6:00.710–6:04.041
model , replacing the older Imagen line entirely .
完全取代了舊版的 Imagen 系列。
6:04.051–6:07.351
And that's not a small detail because Imagen is actually shutting
這可不是小細節，因為 Imagen 實際上將於 2026 年 8 月 17 日停止服務。
6:07.361–6:10.362
down on August 17th , 2026 .
如果你已經有基於 Imagen 的工作流程，那麼倒數計時已經開始。
6:10.372–6:13.522
If you've had workflows built on Imagen , that clock is already
此外，還有一個純粹為速度而設計的 Nano Banana 2 輕量版，
6:13.532–6:14.181
running .
以犧牲少量品質為代價，換取在大量產出時更快、更便宜的結果。
6:14.191–6:17.694
There's also a Nano Banana 2 light variant built purely for speed
此外還有一款專為速度打造的 Nano Banana 2 輕量版
6:17.704–6:20.790
, trading a small amount of quality for much faster ,
1 是 Google 的影片生成模型，目前仍處於測試階段，
6:20.800–6:22.753
cheaper output at high volume .
旨在將文字提示轉換為帶有同步音訊的短片。
6:22.763–6:23.602
VIO 3 .
[未翻譯]
6:23.612–6:26.864
1 is Google's video generation model , still in beta ,
5 live translate 負責處理超過 70 種語言的即時語音對語音翻譯，
6:26.874–6:29.975
built to turn a text prompt into a short clip with matching audio
目前已內建於 Google Meet 和 Android 中。
6:29.975–6:30.907
.
如果你對產品線中較為小眾的部分感到好奇，
6:30.907–6:31.970
Gemini audio 3 .
則有一款專注於安全的變體，名為 3 .
6:31.980–6:36.217
5 live translate handles real time speech to speech translation
5 即時翻譯功能可處理超過 70 種語言的即時語音對語音翻譯
6:36.227–6:39.783
across more than 70 languages , already built into Google Meet
該功能已內建於 Google Meet 和 Android 中
6:39.793–6:40.655
and Android .
以及 Android 系統
6:40.665–6:43.468
And if you're curious about the more niche end of the lineup ,
如果你對產品線中較為小眾的選項感興趣
6:43.478–6:46.137
there's a security focused variant called 3 .
則有一款專注於安全的 3 版本
6:46.147–6:50.185
5 flash cyber built to coordinate with vulnerability scanning tools
5 Flash Cyber 旨在與漏洞掃描工具協調運作
6:50.185–6:50.898
.
。
6:50.898–6:52.027
And Lyra 3 .
以及 Lyra 3.5，
6:52.037–6:55.879
5 , Google's music model , which can now generate tracks up to
Google 的音樂模型，現在可以根據文字提示生成長達
6:55.889–6:58.019
3 minutes long from a text prompt .
從文字提示生成僅需 3 分鐘
6:58.029–7:00.238
Here's the honest limitation worth naming .
這裡有一個值得指出的誠實限制。
7:00.248–7:04.041
Google ships a lot of these models fast , and the naming gets confusing
Google 快速推出許多這些模型，無論是否出於刻意，命名方式變得令人困惑
7:04.051–7:05.230
on purpose or not .
。
7:05.240–7:05.765
3.5,
[未翻譯]
7:05.775–7:06.905
3.6,
[未翻譯]
7:06.915–7:07.980
3.1 Pro,
[未翻譯]
7:07.990–7:10.820
Flash, Flash Cyber.
閃電、閃電網絡
7:10.830–7:13.909
If you're not building on top of the API professionally ,
如果你不是專業地在 API 之上進行開發，
7:13.919–7:16.677
you genuinely don't need to memorize this list .
你實際上不需要記住這個列表。
7:16.687–7:18.522
You just need to know the shape of it .
你只需要了解其架構。
7:18.532–7:21.611
Fast and cheap , deep reasoning , and multimodal .
快速且便宜、深度推理，以及多模態。
7:21.621–7:24.298
That's really three categories wearing a lot of different name
這真的是三個類別，披著許多不同名稱
7:24.308–7:25.181
tags .
標籤。
7:25.191–7:28.952
The modes you actually interact with , models are the engine .
你實際互動的模式，模型是引擎。
7:28.962–7:30.476
Modes are the steering wheel .
模式是方向盤。
7:30.486–7:33.324
Here's where things get useful for anyone who isn't a developer
這裡是對於非開發者來說最有用的地方
7:33.324–7:34.754
.
。
7:34.754–7:37.656
AI mode inside Google Search turns your search bar into a conversation
Google Search 內的 AI 模式會將你的搜尋列變成對話
7:37.656–7:39.188
.
。
7:39.188–7:40.386
As of I/O 2026, it's globally powered by Gemini 3.5 Flash
截至 I/O 2026，它由 Gemini 3.5 Flash 全球驅動
7:40.386–7:42.544
and instead of 10 blue links, you get a written answer
並且不再是 10 個藍色連結，你會得到書面答案
7:42.554–7:46.361
5 flash and instead of 10 blue links , you get a written answer
5 flash 並且不再是 10 個藍色連結，你會得到書面答案
7:46.371–7:49.771
with follow up questions and sometimes an interactive widget built
帶有後續問題，有時還會即時建構互動小工具
7:49.781–7:50.573
on the fly .
。
7:50.583–7:52.178
Anyone with search can use it .
任何會搜尋的人都能使用它。
7:52.188–7:53.782
No subscription required .
不需要訂閱
7:53.792–7:56.706
Ask something like , What's a quick dinner with what's in my fridge
詢問類似「冰箱裡有什麼可以做的快速晚餐」
7:56.706–8:00.354
?
？
8:00.354–8:00.699
and it answers in full sentences , not a list of recipe blogs .
它會以完整句子回答，而不是食譜部落格清單。
8:00.709–8:04.005
Deep Think is the extra effort version of the Gemini app .
Deep Think 是 Gemini 應用程式的額外努力版本。
8:04.015–8:07.773
It spends more compute per answer to reason through harder problems
它會為每個答案投入更多運算資源，逐步推理較難的問題
8:07.783–8:08.815
step by step .
一步一步。
8:08.825–8:11.406
Google gates this one behind Google AI Ultra
Google 將此功能限制在 Google AI Ultra 之後
8:11.406–8:15.468
and it's built for genuinely difficult science or engineering questions ,
並且它是為真正困難的科學或工程問題而設計，
8:15.478–8:16.831
not everyday chat .
而非日常聊天。
8:16.841–8:19.927
Now , Deep Research is where this stops being a chatbot
現在，Deep Research 讓這裡不再只是聊天機器人
8:19.927–8:21.641
and starts being an assistant .
而是開始成為一位助手。
8:21.651–8:22.815
You give it a topic
你給它一個主題
8:22.815–8:27.653
and instead of one reply , it plans a research strategy , opens web pages ,
它不會只回覆一次，而是規劃研究策略、開啟網頁，
8:27.663–8:28.420
reads them
閱讀內容
8:28.420–8:32.622
and if you allow it , pulls from your own Gmail and Drive , too .
如果你允許，它還會從你自己的 Gmail 和 Drive 提取資料。
8:32.632–8:34.585
What comes back isn't a paragraph .
回傳的不是一段文字。
8:34.595–8:38.234
It's a full multi page report inside Gemini's canvas .
而是 Gemini 畫布內一份完整的多頁報告。
8:38.244–8:42.362
This is the part of Gemini that actually earns the word agent and
這部分是 Gemini 真正配得上「代理（agent）」一詞的地方，
8:42.372–8:45.047
we're coming back to why that matters in a few minutes .
我們稍後會再回來討論為什麼這很重要。
8:45.057–8:47.652
Gemini Live is the voice and camera mode .
Gemini Live 是語音和攝影機模式。
8:47.662–8:51.540
Say , Hey Google , let's chat and you're talking to it hands free
說聲「Hey Google，我們來聊天」，你就可以解放雙手與它對話
8:51.550–8:54.226
with the option to point your camera at something and ask what
並可選擇將攝影機指向某物，詢問它
8:54.236–8:55.829
it's looking at live .
即時看到些什麼。
8:55.839–8:57.913
And Canvas is the workspace mode .
而 Canvas 是工作空間模式。
8:57.923–9:01.761
Type , Create a quiz app about planets and it writes the code ,
輸入「建立一個關於行星的測驗應用程式」，它會一次寫出程式碼、
9:01.771–9:05.448
the interface and the content in one pass , right there for you
介面與內容一次呈現，直接擺在你面前
9:05.458–9:06.089
to edit .
供你編輯。
9:06.099–9:09.416
One more worth a mention briefly because it's still early .
還有一個值得簡短提及的功能，因為它目前還處於早期階段。
9:09.426–9:14.145
Gemini Spark , a personal agent announced at IO 2026 ,
Gemini Spark，一款在 IO 2026 上宣布的個人代理，
9:14.155–9:17.031
meant to run continuously in the background handling things like
旨在於背景中持續運行，處理諸如
9:17.041–9:17.953
scheduling .
排程等事項。
9:17.963–9:20.598
Right now , it's limited to early Ultra testers .
目前，它僅限於早期 Ultra 測試人員使用。
9:20.608–9:23.043
So , treat this one as coming , not here .
因此，請將此視為即將推出，而非現已可用。
9:23.053–9:25.047
Quick gut check before we move on .
在繼續之前，快速進行直覺檢查。
9:25.057–9:27.532
If all of that sounds like a lot of separate tools ,
如果上述內容聽起來像是許多獨立工具，
9:27.542–9:28.374
that's fair .
這很合理。
9:28.384–9:29.696
But , notice the pattern .
但請注意這個模式。
9:29.706–9:32.676
Every single one of these modes is just Gemini 3 .
這些模式中的每一個都只是 Gemini 3 .
9:32.686–9:33.664
5 or 3 .
5 或 3 .
9:33.674–9:35.909
6 flash wearing a different job title .
6 flash 換了不同的職位頭銜。
9:35.919–9:39.316
You're not learning six different AIs , you're learning six different
你並非在學習六種不同的 AI，而是在學習六種不同的
9:39.326–9:41.520
ways to ask the same brain for help .
向同一個大腦尋求幫助的方式。
9:41.530–9:43.192
Multimodal in practice .
實務上的多模態。
9:43.202–9:46.257
Let's talk about what multimodal actually means day to day ,
讓我們談談多模態在日常生活中的實際意義，
9:46.267–9:48.527
because it's more than a buzzword on a slide .
因為它不僅是投影片上的流行詞。
9:48.537–9:50.756
Gemini reads and writes text and code .
Gemini 能讀寫文字和程式碼。
9:50.766–9:52.707
Obviously , that's the baseline .
顯然，這是基本功能。
9:52.717–9:56.170
But drop a photo into a chat and ask it to caption or edit it ,
但將照片放入對話中，並要求它為照片加註說明或編輯，
9:56.180–9:57.842
and Nano Banana handles that .
Nano Banana 就能處理這些任務。
9:57.852–10:01.284
Ask it to speak an answer out loud , and Gemini's audio models
要求它大聲說出答案，Gemini 的音訊模型
10:01.294–10:05.089
generate that voice on the spot with actual control over tone and
即時生成該語音，並能實際控制語調和
10:05.099–10:05.810
pacing .
節奏。
10:05.820–10:09.135
Ask for a short video and VO builds one from scratch .
要求製作一段短片，VO 會從頭開始建立。
10:09.145–10:12.259
Ask it to edit an existing clip , swap the sky ,
要求它編輯現有片段、替換天空、
10:12.269–10:16.024
change the style , and that's a separate tool called Gemini Omni
改變風格，這是由另一個名為 Gemini Omni 的工具
10:16.034–10:18.668
doing frame by frame editing by voice command .
透過語音指令逐幀編輯。
10:18.678–10:22.393
Inside Google Docs and Sheets , the same underlying models can
在 Google 文件和試算表中，相同的底層模型可以
10:22.403–10:25.838
draft a document from your meeting notes or build a spreadsheet
根據您的會議記錄起草文件，或從一堆發票中建立試算表，
10:25.848–10:29.843
out of a pile of invoices , complete with formulas and charts .
從一堆發票中，包含公式與圖表。
10:29.853–10:32.166
Not just raw numbers dumped into cells .
不只是將原始數字轉儲到儲存格中。
10:32.176–10:35.611
In Slides , hand it a list of bullet points and it can lay out
在投影片簡報中，提供它一份項目清單，它可以佈局
10:35.621–10:38.936
an actual deck , not just text on blank slides .
實際的簡報，而不僅僅是空白投影片上的文字。
10:38.946–10:42.020
And through Gemini Live's camera mode , you can point your phone
透過 Gemini Live 的相機模式，您可以將手機指向
10:42.030–10:45.479
at a menu in a language you don't speak and get a live translation
您不懂的語言的菜單，並獲得疊加在您所見內容上的即時翻譯。
10:45.489–10:47.149
overlaid on what you're looking at .
或者要求它識別鏡頭所拍攝的物體
10:47.159–10:50.229
Or ask it to identify an object it's looking at through the lens
。
10:50.229–10:51.979
.
不需要打字。
10:51.979–10:52.059
No typing involved .
這是值得記住的部分。
10:52.059–10:53.588
Here's the part worth remembering .
這部分值得記住。
10:53.598–10:55.059
You never pick the model .
你永遠不需要選擇模型。
10:55.069–10:56.410
You just say what you want .
你只需要說出你想要什麼。
10:56.420–10:58.000
Make this an infographic .
將這個做成資訊圖表。
10:58.010–10:59.233
Translate this .
翻譯這個。
10:59.243–11:00.624
Write this in Python .
用 Python 寫這段程式。
11:00.634–11:04.026
And Gemini quietly roots the request to whichever model actually
然後 Gemini 會靜默地將請求路由到真正執行該工作的模型。
11:04.036–11:05.188
does that job .
來完成那項工作。
11:05.198–11:07.754
That's the design philosophy in one sentence .
這就是設計哲學的簡短總結。
11:07.764–11:11.141
One platform , and it decides the plumbing so you don't have to
單一平台，它負責處理底層架構，所以你不必操心。
11:11.141–11:12.546
.
。
11:12.546–11:13.165
Where Gemini actually lives .
Gemini 實際運行的地方。
11:13.175–11:15.570
This is the part that's easy to underestimate .
這部分是很容易被低估的。
11:15.580–11:17.935
Gemini isn't confined to one app .
Gemini 並不限於單一應用程式。
11:17.945–11:20.621
It's spread across nearly everything Google ships .
它遍布於 Google 推出的幾乎所有產品中。
11:20.631–11:23.467
In Search , it's AI mode , already covered .
在搜尋中，它是 AI 模式，前面已經介紹過。
11:23.477–11:26.794
In Gmail , it's behind Smart Compose and auto reply suggestions
在 Gmail 中，它位於智慧撰寫和自動回覆建議背後。
11:26.794–11:27.979
.
。
11:27.979–11:31.355
In Docs , Sheets and Slides , Ultra and Pro Pro Pro get Gemini
在 Docs、Sheets 和 Slides 中，Ultra 和 Pro 版讓 Gemini
11:31.365–11:35.418
drafting text , building formulas , and designing slide layouts
起草文字、建立公式，並設計投影片版面。
11:35.428–11:38.454
, pulling context from your own files when you let it .
，在你允許的情況下，從你自己的檔案中提取內容。
11:38.464–11:41.884
In Drive , it can find and summarize documents for you .
在 Drive 中，它可以為您尋找並摘要文件。
11:41.894–11:44.755
In Google Meet , it's doing live caption translation .
在 Google Meet 中，它正在進行即時字幕翻譯。
11:44.765–11:48.345
On Android , especially Pixel devices , it's baked straight into
在 Android 上，特別是 Pixel 裝置，它已直接內建於
11:48.355–11:49.501
the voice assistant .
語音助理中。
11:49.511–11:52.213
And there's a Chrome extension that lets the browser send page
還有一個 Chrome 擴充功能，讓瀏覽器能將頁面
11:52.223–11:55.683
content straight to Gemini , so you can ask questions about whatever
內容直接傳送給 Gemini，因此您可以針對目前
11:55.693–11:56.640
tab you're on .
的分頁提出問題。
11:56.650–12:00.508
And for developers , all of it is exposed through Google AI Studio
對於開發者而言，所有功能都透過 Google AI Studio
12:00.518–12:04.560
and the Gemini API , plus a newer platform called Antigravity for
和 Gemini API 公開，以及一個名為 Antigravity 的新平台，用於
12:04.570–12:07.008
building multi agent workflows on top of it .
在其上建構多代理工作流程。
12:07.018–12:09.175
The strategic point here isn't subtle .
這裡的戰略重點並不隱晦。
12:09.185–12:12.084
Google isn't trying to win the best standalone chatbot argument
Google 並非試圖贏得最佳獨立聊天機器人
12:12.084–12:13.727
.
的爭論。
12:13.727–12:15.073
It's trying to make sure you're never more than one product away
它正試圖確保無論您在 Google 裝置上
12:15.083–12:18.403
from Gemini , no matter what you're doing on a Google device or
或 Google 應用程式中進行什麼操作，您與
12:18.413–12:19.526
in a Google app .
Gemini 之間永遠只隔一個產品。
12:19.536–12:22.215
Agents : the part that actually matters .
代理：真正重要的部分。
12:22.225–12:25.224
Now , here's the shift I promised earlier , the one that actually
現在，這裡有我之前承諾的轉變，那個真正
12:25.234–12:27.350
changes what this platform is for .
改變這個平台用途的轉變。
12:27.360–12:31.323
Everything so far has been ask a question , get an answer .
到目前為止的一切都是提問、獲得答案。
12:31.333–12:34.693
Agents are Google trying to move Gemini past that entirely .
代理程式是 Google 試圖讓 Gemini 完全超越此階段的做法。
12:34.703–12:37.863
Deep research is the clearest example already live ,
深度研究（Deep research）是目前已上線最清楚的範例，
12:37.873–12:42.076
plan , browse , synthesize , write , without you babysitting every
它能規劃、瀏覽、綜合資訊並撰寫內容，
12:42.086–12:42.757
step .
全程無需你一步步監督。
12:42.767–12:46.489
Spark is the early , still limited attempt at a persistent personal
Spark 是早期且功能仍受限的嘗試，
12:46.499–12:49.178
agent running continuously in the background .
旨在讓個人代理程式在背景持續運作。
12:49.188–12:50.493
And on the developer side ,
而在開發者方面，
12:50.493–12:52.980
Antigravity lets companies build coordinated
Antigravity 讓企業能夠建構
12:52.990–12:54.420
teams of sub agents .
由子代理程式組成的協作團隊。
12:54.430–12:57.859
Google's own blog post gave an example of businesses running parallel
Google 自己的部落格文章舉例說明，企業可運行平行
12:57.869–12:59.698
agents to analyze data at scale ,
代理程式來進行大規模資料分析，
12:59.698–13:02.020
rather than one model doing everything
而不是一個模型處理所有事情
13:02.030–13:02.980
sequentially .
依序執行所有任務。
13:02.990–13:04.820
Picture the difference in practice .
想像一下實際應用上的差異。
13:04.830–13:05.510
The old way ,
舊有的方式是，
13:05.510–13:06.416
you ask Gemini ,
你詢問 Gemini，
13:06.416–13:08.380
What should I know before a trip
旅行前我應該知道什麼
13:08.390–13:09.260
to Japan ?
我該知道什麼？」
13:09.270–13:10.660
and it gives you a paragraph .
它會給你一段文字說明。
13:10.670–13:11.767
The agent way ,
而代理程式的方式是，
13:11.767–13:12.365
you say ,
你說，
13:12.365–13:14.060
Plan my trip to Japan .
幫我規劃去日本的行程。
13:14.070–13:15.448
and it checks flights ,
然後它會檢查航班，
13:15.448–13:16.980
compares hotel options ,
比較飯店選項，
13:16.990–13:19.076
and comes back with an actual itinerary ,
並回傳實際的行程表，
13:19.076–13:20.340
pausing to confirm with
在預訂任何東西之前暫停並與
13:20.350–13:21.900
you before it books anything .
在預訂任何東西之前先詢問你
13:21.910–13:23.525
That's the same underlying model ,
這是相同的基礎模型，
13:23.525–13:25.020
just given permission to take
只是被賦予了權限，
13:25.030–13:27.980
more than one step before handing control back to you .
可以在將控制權交還給你之前採取多於一個步驟。
13:27.990–13:29.925
None of this is science fiction anymore ,
這些已不再是科幻小說，
13:29.925–13:30.980
and none of it is fully
而且也沒有任何一項是
13:30.990–13:31.980
finished , either .
完全完成的。
13:31.990–13:33.220
That's the honest read .
這是誠實的評估。
13:33.230–13:35.580
Deep Research genuinely works today .
Deep Research 在今天確實有效。
13:35.590–13:37.240
Spark is still in early testing ,
Spark 仍在早期測試中，
13:37.240–13:39.080
but the direction is unmistakable.
但方向是明確無疑的。
13:39.080–13:41.157
.
。
13:41.157–13:41.237
break it into steps ,
將其分解為步驟，
13:41.237–13:42.420
and execute most of them without you typing
並在沒有你輸入的情況下執行大部分步驟
13:42.430–13:46.020
a follow up for every single one .
針對每一項都進行後續追蹤。
13:46.030–13:48.020
What actually makes Gemini different ?
究竟是什麼讓 Gemini 有所不同？
13:48.030–13:48.114
So ,
所以，
13:48.114–13:50.140
how does this stack up against everyone else building the
這與其他所有正在構建的人相比如何
13:50.150–13:52.781
same kind of thing ?
相同類型事物的產品相比如何？
13:52.791–13:53.258
Let's be balanced here ,
讓我們在這裡保持客觀，
13:53.258–13:54.061
because Google's advantages are real ,
因為 Google 的優勢是真實存在的，
13:54.071–13:57.542
but so are its weak spots .
但它的弱點也是如此。
13:57.552–13:59.183
The clearest edge is data .
最明顯的優勢在於數據。
13:59.193–14:00.652
Gemini can pull from live search results ,
Gemini 可以從即時搜尋結果中獲取資訊，
14:00.652–14:00.824
Maps ,
地圖，
14:00.834–14:01.133
Gmail ,
[未翻譯]
14:01.133–14:03.825
and Drive in ways that a closed sandbox chatbot simply
以及 Drive，這些是封閉的沙盒聊天機器人
14:03.835–14:08.587
can't match without plugins bolted on .
僅靠外接外掛就無法匹敵的。
14:08.597–14:10.868
The second edge is reach .
第二個優勢在於觸及範圍。
14:10.878–14:12.468
Every Android phone is a potential Gemini client ,
每一支 Android 手機都是潛在的 Gemini 客戶，
14:12.478–14:15.470
and every Workspace business account already has it available .
而且每個 Workspace 企業帳戶都已提供此服務。
14:15.480–14:18.951
No competitor has that kind of built in distribution .
沒有任何競爭對手擁有這種內建的分發能力。
14:18.961–14:19.966
And on raw benchmarks ,
而在原始基準測試中，
14:19.966–14:22.032
Gemini 3 Pro topped the LM Arena leaderboard
Gemini 3 Pro 登上了 LM Arena 排行榜的首位
14:22.032–14:24.511
,
，
14:24.511–14:24.591
which ,
這意味著
14:24.591–14:26.034
regardless of how much weight you put on any single leaderboard
無論你對任何單一排行榜賦予多少權重
14:26.034–14:29.045
,
，
14:29.045–14:29.815
says Google's infrastructure and DeepMind's research are producing
都顯示 Google 的基礎設施與 DeepMind 的研究正在產生
14:29.825–14:30.362
real ,
真實的、
14:30.362–14:32.241
top tier results ,
頂級的成果，
14:32.241–14:33.717
not just hype .
而不只是炒作。
14:33.727–14:33.893
But ,
不過，
14:33.893–14:35.439
and this matters for credibility ,
這對於可信度很重要，
14:35.439–14:36.598
Google is genuinely more
Google 在產品推出方面確實比某些競爭對手
14:36.608–14:40.359
conservative about rollout than some competitors .
在推出方面比一些競爭對手更為保守
14:40.369–14:42.867
Deep Think and Spark are still gated behind ultra subscriptions
Deep Think 和 Spark 仍受限於超級訂閱
14:42.877–14:43.826
or limited testing ,
或有限測試，
14:43.826–14:45.960
while some rivals ship new capabilities to
而一些競爭對手則會將新功能
14:45.970–14:49.160
everyone at once .
一次性提供給所有人。
14:49.170–14:50.360
And the tier structure itself , free , pro , ultra ,
而且層級結構本身，免費、專業、超級，
14:50.370–14:54.840
API pricing , can be genuinely confusing next to a simpler flat
加上 API 定價，與競爭對手較簡單的單一
14:54.850–14:58.799
subscription from a competitor .
訂閱方案相比，可能會讓人真正感到困惑。
14:58.809–15:00.440
If you've ever opened the Gemini pricing page and closed it 5 minutes
如果你曾經打開過 Gemini 的定價頁面，五分鐘後就關閉
15:00.450–15:03.525
later still unsure which plan you need , that's not just you .
之後仍然不確定你需要哪個方案，這不只是你一個人的問題。
15:03.535–15:07.029
Where this is actually headed , a few things are confirmed and
這裡實際上的走向是，有些事項已經確認，而有些仍是謠言，將這兩者區分開來是值得的。
15:07.039–15:09.930
a few are still rumor , and it's worth keeping those separate .
有些事項已經確認，而有些仍是謠言，將這兩者區分開來是值得的。
15:09.940–15:12.870
Confirmed , Gemini 3 .
已確認的是 Gemini 3。
15:12.880–15:14.690
5 Pro is currently in partner testing with a public release expected
5 Pro 目前正處於合作夥伴測試階段，預計很快就會公開發布。
15:14.700–15:18.067
soon , and Google has already started training on Gemini 4 ,
不久後，Google 已經開始針對 Gemini 4 進行訓練，
15:18.077–15:21.532
soon , and Google has already started training on Gemini 4 ,
不久後，Google 已經開始針對 Gemini 4 進行訓練，
15:21.542–15:24.795
according to its own July 2026 announcement, though there's no
根據其 2026 年 7 月的公告，雖然目前還沒有
15:24.805–15:26.406
public timeline for that yet.
公開的時間表。
15:26.416–15:30.475
Workspace AI rollout continues expanding, and Gemini Live's regional
Workspace AI 的推出持續擴大，而 Gemini Live 的地區
15:30.485–15:32.127
language support keeps growing.
語言支援也不斷增加。
15:32.137–15:35.309
Speculative and worth labeling clearly as such,
屬於猜測性質，值得明確標示為猜測，
15:35.319–15:38.613
there's talk of a dedicated on-device AI chip for future Pixel
有傳言稱未來 Pixel 將配備專用設備端 AI 晶片
15:38.623–15:42.529
phones, and some experimental DeepMind research around 3D avatars
手機將配備專屬的裝置端 AI 晶片，以及 DeepMind 針對 3D 虛擬角色
15:42.539–15:45.597
and world simulation that hasn't shipped as a product.
和世界模擬的一些實驗性研究，但這些尚未作為產品推出。
15:45.607–15:48.785
Treat both of those as possible, not coming.
將這兩者視為可能發生，但並非必然到來。
15:48.795–15:50.976
Nothing official has confirmed either one.
沒有任何官方事項確認了其中任何一項。
15:50.986–15:51.773
The verdict.
結論。
15:51.783–15:53.367
So, where does that leave things?
那麼，這讓情況處於什麼狀態呢？
15:53.377–15:57.033
Gemini in 2026 isn't a chatbot you occasionally open.
2026 年的 Gemini 不是一個你偶爾才會開啟的聊天機器人。
15:57.043–15:59.982
It's an AI layer Google has threaded through search,
這是 Google 貫穿搜尋、
15:59.992–16:04.285
Gmail, your documents, and increasingly your phone itself.
Gmail、你的文件，以及日益深入地嵌入你手機本身的 AI 層。
16:04.295–16:07.752
The models handle the thinking, the modes handle how you ask,
模型負責思考，模式負責你提問的方式，
16:07.762–16:10.661
and agents like Deep Research are the clearest sign of where all
而像 Deep Research 這樣的代理程式，則是最清楚地顯示出所有這些
16:10.671–16:12.055
of it is actually heading.
最終將走向何方的跡象。
16:12.065–16:15.124
If there's one thing worth trying this week, it's Deep Research
如果這週只推薦嘗試一件事，那就是 Deep Research
16:15.134–16:17.773
on something you'd normally spend an evening looking into yourself
是對你通常會花一個晚上自行調查的事物進行 Deep Research，
16:17.783–16:20.822
, and actually watching it work instead of just reading the final
並實際觀看它的運作過程，而不僅僅是閱讀最終
16:20.832–16:21.539
report.
報告。
16:21.549–16:24.328
Drop a comment with which piece of this surprised you most,
在留言區告訴我，這其中哪一部分最讓你驚訝，
16:24.338–16:27.715
the model lineup, the agent side, or just how much of this you
模型陣容、代理端，還是你其實早就在無意識中使用的這些功能
16:27.725–16:29.628
were already using without realizing it.
還是你其實已經在無意識中使用這麼多功能。
16:29.638–16:32.417
I'll be back soon with a deeper breakdown on how Deep Research
我很快會回來，深入剖析 Deep Research
16:32.427–16:35.127
actually performs against a real research task.
在面對真實研究任務時的實際表現。
16:35.137–16:37.000
Thanks for watching, and I'll see you in the next one.
感謝觀看，我們下一部影片見。

影片筆記：Google Gemini Explained: 80% of Google Gemini You Never Use (2026 Guide)

一句話總結

Google Gemini 在 2026 年已從單一聊天機器人演變為覆蓋底層模型與上層應用介面的完整生態平台，透過深度整合 Android、Google Workspace 及搜尋服務，並結合 Deep Research 與 Spark 等 Agent 功能，旨在取代傳統工作流並提供全域 AI 體驗。

核心重點

  1. 平台而非單一產品：Gemini 是 Google 的「傘狀名稱」，包含底層模型（Models）與上層互動介面/模式（Modes）。
  2. 模型分層策略：
  • Flash 系列：強調速度、編碼與高吞吐量（如 3.7 Flash, 3.6 Flash, 3.5 Flash）。
  • Pro 系列：強調深度推理與邏輯效能（如 3.1 Pro, 3.5 Pro）。
  • 專用模型：針對影像（Nano Banana 2）、影片（VIO 3.1）、音樂（Lyra 3.5）及即時翻譯（Gemini Audio）進行優化。
  1. 多模態與自動路由：使用者無需手動選擇底層模型，系統會根據意圖自動路由至適當模型處理文字、程式碼、影像、影片及音訊。
  2. 生態系深度整合：Gemini 已內建於 Google 搜尋（AI Mode）、Gmail、Docs、Sheets、Slides、Chrome 及 Android 語音助理中，目標是讓用戶在任何 Google 產品中都能觸及 AI。
  3. Agent 功能演進：從單純問答轉向執行多步驟任務，代表功能包括 Deep Research（深度研究）、Deep Think（深度思考）及 Gemini Spark（背景個人智能體）。

詳細大綱

1. Gemini 的核心概念與市場現狀

  • 架構分層：
  • 底層：模型本身（如 Gemini 3.6 Flash, 3.1 Pro），針對不同任務優化。
  • 頂層：使用者實際點擊的產品介面（Gemini App, 搜尋 AI Mode, Gmail/Docs 內建功能）。
  • 市場數據：
  • Gemini App 月活躍用戶超過 6.5 億。
  • 美國僅約 1/5 的企業在營運中使用 AI，使用 Gemini 的企業已領先競爭對手。

2. 當前模型陣容（Model Lineup）詳解

  • Gemini 3.7 Flash：
  • 發布日期：2026年8月13日。
  • 定位：最新 Flash 模型，編碼與 AI Agent 的主力。
  • 優勢：軟體工程、網頁開發、複雜多步驟工作流。
  • Gemini 3.6 Flash：
  • 發布日期：2026年7月21日。
  • 定位：目前旗艦（Flagship），工作馬（Workhorse）。
  • 優勢：編碼、知識工作、多模態任務。
  • 效率：平均比前代少使用 17 fewer tokens（註：原文表述，疑點見後）。
  • Gemini 3.5 Flash：
  • 發布日期：2026年5月。
  • 定位：搜尋 AI Mode 驅動者，極致速度。
  • 優勢：輸出吞吐量約為當時其他頂級模型的 4 倍。
  • Gemini 3.5 Flashlight：
  • 發布日期：2026年7月。
  • 定位：高吞吐量、簡化版（stripped down）。
  • 效能：約 350 tokens/秒，用於背景任務與 Agent 管線。
  • Gemini 3.1 Pro：
  • 發布日期：2026年2月。
  • 定位：推理專家（Reasoning Specialist）。
  • 優勢：邏輯效能約為早期 Gemini 3 Pro 的兩倍。
  • 定價參考（API）：
  • 3.6 Flash：輸入約 $1.50/百萬 tokens，輸出約 $7.50/百萬 tokens。
  • 對比：文中稱其比 GPT 5.6 Luna 的輸出價格（約 $6/百萬 tokens）更具競爭力（註：數值邏輯疑點見後）。

3. 專用模型（Specialty Models）

  • Nano Banana 2：影像生成與編輯，取代 Imagen 系列（Imagen 將於 2026年8月17日關閉）。有 Light 變體追求速度與低成本。
  • VIO 3.1：影片生成模型（Beta 階段），文字轉短片（含音訊）。
  • Gemini Audio 3.5 Live Translate：即時語音對語音翻譯，支援 70+ 語言，內建於 Google Meet 與 Android。
  • 3.5 Flash Cyber：安全專用變體，配合漏洞掃描工具。
  • Lyra 3.5：音樂模型，可生成長達 3 分鐘的曲目。

4. 使用者互動模式（Modes）

  • AI Mode (Google Search)：由 Gemini 3.5 Flash 驅動，將搜尋欄轉為對話，提供完整句子答案及即時互動小工具，無需訂閱。
  • Deep Think：Gemini App 的「額外努力」版本，分配更多運算進行逐步推理，限 Google AI Ultra 訂閱。
  • Deep Research：從聊天機器人轉變為助理，規劃研究策略、開啟網頁、閱讀內容，若授權可讀取 Gmail 和 Drive，輸出為 Gemini Canvas 中的多頁完整報告。
  • Gemini Live：語音與相機模式，免持對話、相機即時識別與提問。
  • Canvas：工作區模式，一次性生成程式碼、介面與內容並可直接編輯。
  • Gemini Spark：個人 Agent（I/O 2026 宣布），背景持續運行（如排程），僅限早期 Ultra 測試者。

5. 多模態實務應用（Multimodal in Practice）

  • 核心哲學：使用者只需表達意圖，系統自動路由至適當模型。
  • 具體應用場景：
  • 影像：Nano Banana 處理圖片標題或編輯。
  • 音訊：Gemini Audio 模型即時生成帶有語調控制的語音。
  • 影片：VIO 生成短片；Gemini Omni 進行逐幀編輯（如換天空、改風格）。
  • 文書處理：Docs/Sheets 從會議筆記起草文件或建立試算表；Slides 將要點轉換為簡報版面。
  • 即時翻譯：透過 Live 相機模式，對菜單進行即時翻譯疊加。

6. 生態系整合與 Agent 戰略

  • 全域整合：
  • Drive：搜尋並總結文件。
  • Google Meet：進行即時字幕翻譯。
  • Android (特別是 Pixel)：內建於語音助手中。
  • Chrome：透過擴充功能將頁面內容直接發送給 Gemini 提問。
  • 開發者工具：
  • 透過 Google AI Studio 與 Gemini API 開放。
  • 新增平台 Antigravity 用於構建多智能體工作流。
  • 核心戰略：不爭奪「最佳獨立聊天機器人」的論述，目標是讓用戶在任何 Google 裝置或應用中，距離 Gemini 不超過一個產品的距離。
  • Agent 功能演進：
  • Deep Research：已上線，具備規劃、瀏覽、綜合、撰寫能力。
  • Spark：早期且受限的嘗試，旨在作為在背景持續運行的個人智能體。
  • Antigravity：讓企業構建協調的子智能體團隊。
  • 實際場景對比：從「詢問去日本前該知道什麼」轉向「規劃我的日本行程」，檢查航班、比較酒店、生成行程表。

7. 競爭優勢與未來發展

  • 優勢 (Edges)：
  1. 數據整合能力：可從即時搜尋結果、Maps、Gmail、Drive 提取數據。
  2. 分發範圍 (Reach)：每部 Android 手機都是潛在客戶，每個 Workspace 商業帳戶已內建。
  3. 基準測試表現：Gemini 3 Pro 登上 LM Arena 排行榜頂端。
  • 劣勢/爭議點：
  1. 發布策略保守：Deep Think 與 Spark 仍受限於 Ultra 訂閱或有限測試。
  2. 定價結構複雜：免費、Pro、Ultra、API 定價結構可能比競爭對手更令人困惑。
  • 未來發展：
  • 確認事項：Gemini 3.5 Pro 處於合作夥伴測試階段；Gemini 4 已開始訓練（2026年7月公告）；Workspace AI 持續擴展。
  • 猜測事項：針對未來 Pixel 手機的專用裝置端 AI 晶片；DeepMind 關於 3D 化身與世界模擬的實驗性研究。

工具 / 模型 / 名詞整理

  • Gemini 系列模型：
  • Gemini 3.7 Flash
  • Gemini 3.6 Flash
  • Gemini 3.5 Flash
  • Gemini 3.5 Flashlight
  • Gemini 3.1 Pro
  • Gemini 3 Pro (提及為早期版本)
  • 3.5 Flash Cyber (安全專用變體)
  • 專用模型：
  • Nano Banana 2 (含 Light 變體)
  • Imagen (即將關閉)
  • VIO 3.1
  • Gemini Audio 3.5 Live Translate
  • Lyra 3.5
  • Gemini Omni (用於影片逐幀編輯)
  • 產品與功能介面：
  • Gemini App
  • AI Mode (Google Search)
  • Deep Think
  • Deep Research
  • Gemini Canvas
  • Gemini Live
  • Gemini Spark
  • Google AI Studio
  • Gemini API
  • Anti Gravity Platform (或 Antigravity)
  • Vertex AI
  • Google Meet
  • Android
  • Gmail (Smart Compose)
  • Google Docs
  • Google Sheets
  • Google Slides
  • Chrome
  • Google Drive
  • Maps
  • Workspace
  • 競爭對手產品：
  • GPT 5.6 Luna

操作流程整理

  1. 日常搜尋與對話：
  • 在 Google 搜尋中使用 AI Mode，系統自動路由至 Gemini 3.5 Flash 提供對話式答案。
  1. 深度工作與研究：
  • 訂閱 Google AI Ultra 後，在 Gemini App 中使用 Deep Think 進行複雜推理。
  • 使用 Deep Research 規劃研究策略，系統自動瀏覽網頁、閱讀內容（若授權可讀取 Gmail/Drive），並在 Canvas 中生成多頁報告。
  1. 多模態內容生成：
  • 輸入意圖（如生成影像、影片或音樂），系統自動路由至對應專用模型（如 Nano Banana 2, VIO 3.1, Lyra 3.5）。
  • 使用 Gemini Live 進行語音對話或透過相機進行即時翻譯與識別。
  1. 生態系整合應用：
  • 在 Gmail 中使用 Smart Compose 撰寫郵件。
  • 在 Docs/Sheets/Slides 中利用 Gemini 起草文字、建立公式或設計簡報。
  • 在 Chrome 中透過擴充功能將頁面內容發送給 Gemini 提問。
  • 在 Android/Pixel 手機上使用內建語音助理呼叫 Gemini。
  1. 開發者與企業應用：
  • 透過 Gemini API 或 Google AI Studio 接入模型。
  • 企業使用 Antigravity 平台構建多智能體工作流，協調子智能體團隊執行並行任務。

值得注意的限制或風險

  1. 訂閱與發布限制：部分高級功能（如 Deep Think, Spark）僅限 Google AI Ultra 訂閱者或早期測試者，發布策略相對保守。
  2. 定價結構複雜：免費、Pro、Ultra 及 API 定價結構可能比競爭對手的單一訂閱更令人困惑。
  3. 模型依賴與路由：使用者雖無需手動選擇模型，但依賴系統自動路由，若路由錯誤可能影響體驗（儘管影片強調系統能自動處理）。
  4. 授權與隱私：Deep Research 等功能若需讀取 Gmail 和 Drive，需用戶明確授權，涉及隱私與數據存取權限問題。
  5. 技術成熟度：部分專用模型（如 VIO 3.1）仍處於 Beta 階段；Spark 僅限早期測試，尚未全面開放。

逐字稿辨識疑點

  • 定價邏輯矛盾：逐字稿提到 3.6 Flash 輸出價格為 $7.50/百萬 tokens，並稱其比 GPT 5.6 Luna 的 $6/百萬 tokens "noticeably cheaper"（明顯更便宜）。數值上 7.50 大於 6，與 "cheaper" 描述矛盾，需查證原文數據或語意。
  • Token 數量描述：原文提到 3.6 Flash "does the job using about 17 fewer tokens"。未明確說明是絕對數量減少 17 個（對於複雜任務不合理）還是百分比減少（如 17%），或為聽寫錯誤。
  • 產品名稱拼寫：
  • "Nano Banana"：此名稱在現有 AI 領域較不常見，疑似為 "Nano Banana" 的正確名稱或聽寫錯誤（可能為 "Nano Banana" 或其他代號），需查證。
  • "Anti gravity platform" / "Antigravity"：Google 是否有名為 "Anti gravity" 或 "Antigravity" 的平台？疑似為 "Android" 或其他產品名的聽寫錯誤，或為特定內部代號。
  • "VIO 3.1"：Google 是否有名為 "VIO" 的影片生成模型？疑似為 "Video" 或其他產品名的聽寫錯誤。
  • 日期與版本：文中多次提及 2026 年的發布日期（如 3.7 Flash 於 2026年8月13日發布），若當前時間早於 2026 年，則為未來預測或影片設定在未來時間點，需確認影片發布時間背景。
  • GPT 5.6 Luna：Google 提及的競爭對手 "GPT 5.6 Luna" 名稱需查證是否為真實存在的產品名稱，或為聽寫錯誤（如 GPT-4o, GPT-5 等）。
  • Gemini 3 Pro：逐字稿提及此名稱並稱其登上 LM Arena 排行榜頂端，需查證該版本名稱是否準確或為聽寫錯誤。
  • LM Arena：需查證該排行榜的完整名稱或準確拼寫。

可延伸追問

  1. Gemini 3.6 Flash 的 "17 fewer tokens" 具體是指絕對數量還是百分比？這對成本計算有何影響？
  2. "Nano Banana 2" 和 "VIO 3.1" 的正式產品名稱為何？它們與現有的 Imagen 和 Video AI 模型有何具體區別？
  3. "Antigravity" 平台目前對開發者的開放程度如何？它與 Vertex AI 的關係為何？
  4. Gemini Spark 作為背景個人智能體，其數據隱私保護機制為何？它如何處理用戶的敏感資訊？
  5. 隨著 Gemini 4 的訓練開始，預計會對現有的 Flash 和 Pro 系列模型產生何種替代或升級影響？

生字列表

生字讀音類型中文
chatbot/ˈtʃætˌbɒt/聊天機器人
tab/tæb/noun分頁
workhorse/ˈwɜːrkˌhɔːrs/noun主力；耐用的工具
token/ˈtoʊkən/noun記號；權杖（AI術語：文字單位）
throughput/ˈθruːˌpʊt/noun吞吐量；產量
gated/ɡeɪtɪd/adjective受限制的；需許可的
buzzword/ˈbʌzˌwɜːrd/noun流行詞；口頭禪
pipeline/ˈpaɪpˌlaɪn/noun管線；流程
benchmark/ˈbentʃˌmɑːrk/noun基準；標準測試
genuine/ˈdʒɛnjuɪn/adjective真正的；真誠的
default/dɪˈfɔːlt/verb預設；默認
layout/ˈleɪˌaʊt/noun佈局；版面設計
rollout/ˈroʊˌlaʊt/noun推出；發布
hype/haɪp/noun炒作；誇大宣傳
thread/θrɛd/verb穿過；貫穿
verdict/ˈvɜːrdɪkt/noun結論；裁決
underestimate/ˌʌndərˈɛstɪmeɪt/verb低估
confined/kənˈfaɪnd/adjective受限的；局限的

生字解說

chatbot /ˈtʃætˌbɒt/

· B2

意思：聊天機器人

解說：指透過文字或語音與人互動的軟體程式。在原文中，作者指出 Gemini 不只是另一個「聊天機器人」，而是更複雜的平台。

影片原句
the same way you'd use any other chatbot .
就像使用其他任何聊天機器人一樣。
延伸例句
Many companies are replacing customer service chatbots with human agents.
許多公司正以人工客服取代客戶服務聊天機器人。

tab /tæb/

noun · A2

意思：分頁

解說：瀏覽器中的標籤頁。原文提到用戶通常輸入內容後就關閉分頁，暗示了淺層使用。

影片原句
and closed the tab ,
然後關閉分頁，
延伸例句
Please close the extra tab to free up memory.
請關閉額外的分頁以釋放記憶體。

workhorse /ˈwɜːrkˌhɔːrs/

noun · C1

意思：主力；耐用的工具

解說：原意為拉車的馬，引申為在特定領域中承擔主要、繁重工作的核心工具或模型。原文用來形容 Gemini 3.6 Flash 是主力模型。

影片原句
model and its most capable workhorse yet .
模型，也是迄今為止最強大的主力模型。
延伸例句
This old truck is a reliable workhorse for our delivery business.
這輛舊卡車是我們送貨業務中可靠的得力助手。

token /ˈtoʊkən/

noun · C1

意思：記號；權杖（AI術語：文字單位）

解說：在大型語言模型中，token 是處理文字的基本單位（通常為詞或字的一部分）。原文指出使用更少的 token 意味著更快的速度和更低的費用。

影片原句
it does the job using about 17 fewer tokens on average than its predecessor .
它完成任務所需的平均比其前身少約 17 個 token。
延伸例句
The API charges based on the number of input and output tokens.
該 API 根據輸入和輸出記號的數量收費。

throughput /ˈθruːˌpʊt/

noun · C1

意思：吞吐量；產量

解說：指系統在單位時間內處理的工作量。原文提到該模型將輸出吞吐量推升至其他頂級模型的 4 倍。

影片原句
it pushed output throughput to roughly four times faster than other top models at the time .
它將輸出吞吐量推升至速度大約是當時其他頂級模型的 4 倍。
延伸例句
The new server significantly improves network throughput.
新伺服器顯著提升了網路吞吐量。

gated /ɡeɪtɪd/

adjective · C1

意思：受限制的；需許可的

解說：指某些功能或資源僅對特定群體（如付費用戶或測試人員）開放。原文提到 3.1 Pro 主要透過預覽存取限制使用。

影片原句
It's mostly gated behind preview access through the API ,
它主要透過 API 的預覽存取來限制使用，
延伸例句
The beta features are gated to select users only.
測試版功能僅對選定用戶開放。

buzzword /ˈbʌzˌwɜːrd/

noun · B2

意思：流行詞；口頭禪

解說：指在特定時期被過度使用、缺乏實質意義的流行用語。原文指出「多模態」不僅是投影片上的流行詞，而是實際功能。

影片原句
because it's more than a buzzword on a slide .
因為它不僅是投影片上的流行詞。
延伸例句
Synergy is often just a buzzword in corporate meetings.
協同效應在企業會議中往往只是一個流行詞。

pipeline /ˈpaɪpˌlaɪn/

noun · C1

意思：管線；流程

解說：指一系列連續的步驟或程序。原文提到該模型適合用於需要快速且低成本大規模運作的代理管線。

影片原句
You'd use it for background tasks and agent pipelines that need to move fast and cheap at scale .
你會用它來處理背景任務和需要快速且低成本大規模運作的代理管道。
延伸例句
We need to optimize the data processing pipeline.
我們需要優化數據處理流程。

benchmark /ˈbentʃˌmɑːrk/

noun · C1

意思：基準；標準測試

解說：用於評估系統性能的標準測試。原文提到根據其自身的基準測試，該模型速度極快。

影片原句
by its own benchmarks , it pushed output throughput to roughly four times faster than other top models at the time .
並且根據其自身的基準測試，它將輸出吞吐量推升至速度大約是當時其他頂級模型的 4 倍。
延伸例句
The new chip sets a new benchmark for performance.
新晶片為性能設定了新的基準。

genuine /ˈdʒɛnjuɪn/

adjective · B2

意思：真正的；真誠的

解說：指真實存在、非虛構的。原文用來形容「真正困難的問題」，強調其複雜性。

影片原句
The model you'd want on a genuinely hard problem ,
你希望在真正困難的問題上使用的模型，
延伸例句
She showed genuine concern for her colleagues.
她對同事表現出真誠的關心。

default /dɪˈfɔːlt/

verb · B2

意思：預設；默認

解說：指在沒有特別選擇時自動採取的行動或狀態。原文指出開發者對於大量運用的應用，預設選擇 flash 層級。

影片原句
That gap is exactly why so many developers default to flash tier models for anything running at volume .
這個差距正是為什麼許多開發者對於任何大量運行的應用，預設選擇 flash 層級模型。
延伸例句
If you don't specify a language, the system defaults to English.
如果你未指定語言，系統將預設為英文。

layout /ˈleɪˌaʊt/

noun · B2

意思：佈局；版面設計

解說：指頁面或空間的排列方式。原文提到 Gemini 可以佈局實際的簡報，而不僅僅是文字。

影片原句
in Slides , hand it a list of bullet points and it can lay out an actual deck , not just text on blank slides .
在投影片簡報中，提供它一份項目清單，它可以佈局實際的簡報，而不僅僅是空白投影片上的文字。
延伸例句
The architect presented a new layout for the office.
建築師提出了辦公室的新的佈局方案。

rollout /ˈroʊˌlaʊt/

noun · C1

意思：推出；發布

解說：指產品或功能逐步向公眾發布的過程。原文指出 Google 在產品推出方面比競爭對手更為保守。

影片原句
Google is genuinely more conservative about rollout than some competitors .
Google 在產品推出方面確實比某些競爭對手更為保守
延伸例句
The rollout of the new software was delayed due to bugs.
新軟件的推出因錯誤而延遲。

hype /haɪp/

noun · B2

意思：炒作；誇大宣傳

解說：指過度或不實的宣传。原文強調 Google 的成果是真實的頂級結果，而不只是炒作。

影片原句
top tier results , not just hype .
頂級的成果，而不只是炒作。
延伸例句
The movie was good, but the hype was excessive.
這部電影不錯，但炒作過度了。

thread /θrɛd/

verb · C1

意思：穿過；貫穿

解說：原意為穿線，此處比喻將 AI 能力嵌入到各種產品中。原文說 Google 將 AI 層貫穿於搜尋、Gmail 等產品。

影片原句
It's an AI layer Google has threaded through search,
這是 Google 貫穿搜尋、
延伸例句
She threaded the needle with ease.
她輕鬆地將線穿過針眼。

verdict /ˈvɜːrdɪkt/

noun · B2

意思：結論；裁決

解說：指經過考慮後得出的最終判斷。原文用此詞引出對 Gemini 整體狀況的總結。

影片原句
The verdict.
結論。
延伸例句
The jury reached a verdict after hours of deliberation.
陪審團經過數小時的審議後達成了裁決。

underestimate /ˌʌndərˈɛstɪmeɪt/

verb · B2

意思：低估

解說：指對某事物的價值或重要性估計不足。原文指出 Gemini 實際運行的地方容易被低估。

影片原句
This is the part that's easy to underestimate .
這部分是很容易被低估的。
延伸例句
Don't underestimate the power of social media.
不要低估社交媒體的力量。

confined /kənˈfaɪnd/

adjective · B2

意思：受限的；局限的

解說：指被限制在特定範圍內。原文指出 Gemini 不限於單一應用程式，而是遍布多個產品。

影片原句
Gemini isn't confined to one app .
Gemini 並不限於單一應用程式。
延伸例句
The patient was confined to bed for a week.
病人被限制臥床一週。

句型解說（含實例）

not just A, but B

意思：不只是 A，而是 B

接續：not just + [Noun/Clause], but + [Noun/Clause]

解說：用於強調後者比前者更重要或更準確。原文用來區分 Gemini 不僅是聊天機器人，更是平台。

影片原句
It's not a single AI , it's Google's umbrella name for a whole platform ,
Gemini 不是單一的人工智慧，它是 Google 對整個平台的總稱，
實例
  1. It's not just a toy, but a powerful learning tool.
    這不僅僅是一個玩具，而是一個強大的學習工具。
  2. The solution is not just fixing the bug, but preventing it from happening again.
    解決方案不僅是修復錯誤，而是防止它再次發生。

the same way you'd...

意思：就像你會...一樣

接續：the same way + [Subject + would/could + Verb]

解說：用於比較兩種行為的相似性。原文指出用戶使用 Gemini 的方式與使用其他聊天機器人相同。

影片原句
the same way you'd use any other chatbot .
就像使用其他任何聊天機器人一樣。
實例
  1. You can solve this problem the same way you'd fix a flat tire.
    你可以用修補爆胎的方式來解決這個問題。
  2. He treats his employees the same way he'd want to be treated.
    他對待員工的方式就像他希望自己被對待一樣。

Here's the thing, that's...

意思：問題在於，那是...

接續：Here's the thing, + [Clause explaining the reality]

解說：用於引入一個關鍵的、可能與預期不同的事實或轉折。原文用來指出用戶只使用了 Gemini 功能的 10%。

影片原句
Here's the thing, that's maybe 10% of what it actually does.
問題在於，這可能只佔到它實際功能的 10%。
實例
  1. Here's the thing, we don't have enough budget for this project.
    問題在於，我們沒有足夠的預算來進行這個項目。
  2. Here's the thing, the meeting has been cancelled.
    問題在於，會議已經取消了。

depending on what...

意思：取決於你...

接續：depending on + [Wh-clause]

解說：用於表示結果或選擇隨條件變化。原文指出應使用哪個工具取決於你實際上想做的事。

影片原句
depending on what you're actually trying to do ,
取決於你實際上想做的事，
實例
  1. You can choose a route depending on the traffic conditions.
    你可以根據交通狀況選擇路線。
  2. The price varies depending on the size of the order.
    價格根據訂單大小而變化。

Think of it in two layers.

意思：我們可以從兩個層面來理解。

接續：Think of it + [Prepositional Phrase describing structure]

解說：用於引導聽眾從特定結構或角度理解複雜概念。原文引導觀眾將 Gemini 理解為兩個層面。

影片原句
Think of it in two layers .
我們可以從兩個層面來理解。
實例
  1. Think of it in three stages: planning, execution, and review.
    我們可以從三個階段來思考它：規劃、執行和審查。
  2. Think of it as a hierarchy where the top level manages the bottom.
    我們可以將其視為一個層次結構，其中頂層管理底層。

What you might not know is that...

意思：你可能不知道的是，...

接續：What + [Subject] + might not know + is that + [Clause]

解說：用於引入聽眾可能缺乏的額外資訊。原文指出觀眾可能不知道團隊還與企業主合作實施 AI。

影片原句
What you might not know is that alongside covering AI news , we work with business owners...
你可能不知道的是，除了報導 AI 新聞外，我們還與企業主合作，
實例
  1. What you might not know is that the restaurant is closed on Mondays.
    你可能不知道的是，餐廳週一休息。
  2. What you might not know is that this feature is available for free.
    你可能不知道的是，這個功能是免費提供的。

It's trying to...

意思：它正試圖...

接續：It's + trying to + [Verb Phrase]

解說：用於描述主體（此處為 Google）的意圖或戰略目標。原文指出 Google 試圖確保用戶與 Gemini 之間只隔一個產品。

影片原句
It's trying to make sure you're never more than one product away from Gemini ,
它正試圖確保無論您在 Google 裝置上或 Google 應用程式中進行什麼操作，您與 Gemini 之間永遠只隔一個產品。
實例
  1. The company is trying to reduce its carbon footprint.
    該公司正試圖減少其碳足跡。
  2. We are trying to find a balance between speed and quality.
    我們正試圖在速度和質量之間找到平衡。

If there's one thing worth...

意思：如果有一件事值得...

接續：If there's one thing + [Noun] + worth + [Gerund/Adjective]

解說：用於強調某件事的極高價值或重要性。原文建議觀眾這週最值得嘗試的是 Deep Research。

影片原句
If there's one thing worth trying this week, it's Deep Research
如果這週只推薦嘗試一件事，那就是 Deep Research
實例
  1. If there's one thing worth remembering, it's to always back up your data.
    如果有一件事值得記住，那就是始終備份你的數據。
  2. If there's one thing worth noting, it's the change in policy.
    如果有一件事值得注意，那就是政策的變化。