Would You Still Pay for Claude After Seeing This? (GLM 5.2 vs DeepSeek V4)

Edward Donner 首頁 每日列表 原始影片
實際影片長度：15:01.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:03.320
It's one of the most common questions I get about coding agents.
0:03.600–0:08.960
Are we ready yet to use open source models to build real world projects?
0:09.160–0:11.960
I'm here to tell you that yes, we are ready.
0:12.280–0:13.240
And I'm going to show you.
0:13.500–0:14.988
In just about 10 minutes time,
0:14.988–0:18.260
we're going to be looking at real world products like this and this
0:18.260–0:22.340
built by coding agents using open source models.
0:22.340–0:25.940
And we'll be comparing it to what Claude can do with the same brief.
0:26.040–0:28.660
And I'm cramming so many interesting things into this one video.
0:28.660–0:32.158
Instead of using Claude Code, we're going to be using Pi,
0:32.158–0:34.540
which is a super popular coding agent.
0:34.800–0:36.220
Just recently become really popular.
0:36.520–0:38.499
We're going to be using GLM 5.2,
0:38.499–0:41.240
the latest groundbreaking model from ZAI.
0:41.580–0:43.720
And we're going to be using DeepSeq v4.
0:43.880–0:46.760
And we're going to be using dev containers for extra safety.
0:47.060–0:48.080
A lot going on.
0:48.340–0:52.540
Now, this website is one of the most important resources for an AI engineer.
0:52.740–0:54.080
And I'm going to link to it in the description.
0:54.080–0:58.021
It's artificialanalysis.ai with tons of information,
0:58.021–1:00.620
a lot of data about different models.
1:01.000–1:02.783
In particular, if you scroll down,
1:02.783–1:05.180
there is this chart intelligence versus cost.
1:05.340–1:08.000
It's a bit buried, but it's a really valuable chart.
1:08.300–1:12.060
And people on my AI Engineer CoreTrack course know this chart well.
1:12.380–1:14.420
Each dot represents a different model.
1:14.780–1:19.160
The x-axis is how much does it cost to run the model for a task.
1:19.560–1:21.360
More to the right means it's more expensive.
1:21.360–1:24.708
the y-axis is how intelligent is the model.
1:24.708–1:28.520
A smarter model goes higher up. So you're ideally
1:28.520–1:32.120
looking to fit in that green quadrant on the top left.
1:32.120–1:34.820
That is where models go if they are cheap
1:34.820–1:36.225
and they're also smart. And
1:36.225–1:39.280
today we're going to try out two models that are kind of on the
1:39.280–1:44.554
rim of this line here. They're kind of at this frontier of being smart but
1:44.554–1:45.960
cheap. We're going
1:45.960–1:47.679
to look at DeepSeq v4 Pro.
1:47.679–1:52.100
This is one that has a great reputation for being really low cost,
1:52.220–1:55.746
and we'll see. And we're also going to look at GLM 5.
1:55.746–1:58.600
2 from z.ai, and that has a reputation for
1:58.600–2:01.933
being even smarter than this diagram might lead you to believe,
2:01.933–2:03.820
and we'll judge that for ourselves.
2:03.820–2:08.409
So GLM 5.2, which the community is super excited about,
2:08.409–2:11.900
and DeepSeq v4, let's put them to the test.
2:11.900–2:17.655
So Pi is an example of a coding agent like a Claude code or more like an open code because
2:17.655–2:23.409
it's one of these coding agents that are designed to not be tied to any LLM but to allow y
2:23.409–2:25.760
ou to plug in any open source model.
2:25.760–2:28.895
And in particular, Pi has a great reputation because
2:28.895–2:30.320
it's very lightweight.
2:30.580–2:31.340
It is minimal.
2:31.680–2:34.240
It's easy to extend it in different directions.
2:34.380–2:37.080
But what you get out of the box is very simple.
2:37.300–2:39.184
And it just has a simple prompt and
2:39.184–2:41.540
a simple approach to context management.
2:41.540–2:44.460
And it lets you take it in whatever direction you want.
2:44.720–2:46.260
And that's exactly what we're going to do.
2:46.500–2:48.480
Okay, the moment has arrived to put this to the test.
2:48.540–2:50.356
And I would say whether or not you're technical,
2:50.356–2:51.900
you're going to want to watch this just to
2:51.900–2:54.923
get a sense for how it all works, how Pi works, and
2:54.923–2:57.720
how these models perform. We are going to put
2:57.720–3:00.908
GLM and DeepSeek to the test. We're going to do it right now.
3:00.908–3:02.340
We're going to build out an
3:02.340–3:05.840
interesting project. And here it is. I've got a folder here,
3:05.840–3:07.800
which is almost empty. It just has
3:07.800–3:10.545
a few things, including an agents.md file. And
3:10.545–3:13.360
that describes what we'd like to build. We would
3:13.360–3:16.474
like to build a personal CRM system,
3:16.474–3:21.560
a personal Salesforce kind of platform, just for you, just
3:21.560–3:25.155
for me, our very own version of it, like a Salesforce,
3:25.155–3:27.280
a pipe drive that I cover in some
3:27.280–3:29.146
of my other courses, that kind of thing. But
3:29.146–3:31.220
we're going to want to build the whole thing from
3:31.220–3:33.523
scratch. We want a coding agent to do that.
3:33.523–3:36.220
Just a simple one, mind. And I've got the business
3:36.220–3:40.112
requirements here written from a business person's point of view.
3:40.112–3:41.780
There is one small technical
3:41.780–3:45.248
section in here, but otherwise it's about the look and
3:45.248–3:47.560
feel. It's about what's in scope and
3:47.560–3:48.452
not in scope and
3:48.452–3:49.344
what the CRM can
3:49.344–3:52.500
do from a business person's point of view, just a little
3:52.500–3:55.727
bit of technical guidance to keep the model on rails.
3:55.727–3:58.220
But the most important thing we do that is
3:58.220–4:02.671
absolutely mandatory when working with open source models is this.
4:02.671–4:04.780
We lay out the project in terms of
4:04.780–4:09.679
phases, bite-sized phases that can each be implemented a step at a time.
4:09.679–4:10.740
For each phase,
4:10.740–4:14.372
we explain the features in business terms, and then
4:14.372–4:17.160
we give success criteria, ways that the
4:17.160–4:21.060
coding agent can verify that it has passed this phase,
4:21.060–4:23.660
it's done, it can move on to the next.
4:23.920–4:27.113
And then at the end, we give overall success criteria.
4:27.113–4:29.100
I've got six phases and then final
4:29.100–4:31.069
success criteria at the end. And
4:31.069–4:35.080
it cannot declare victory on this project until all these success
4:35.080–4:37.788
criteria are met. And I know I go on about it a lot,
4:37.788–4:40.360
but success criteria are so crucial to making
4:40.360–4:43.250
sure that you keep the coding agent on track. And
4:43.250–4:46.140
importantly, you also have to give it the skills
4:46.140–4:48.436
it needs to be able to test thoroughly,
4:48.436–4:51.880
like bring up the screens in a browser window and check it.
4:51.980–4:54.196
That way it's got the feedback loop and
4:54.196–4:57.520
it can keep iterating until the success criteria are met.
4:57.920–5:03.103
So the agents.md file has all of the instructions for our coding agent,
5:03.103–5:04.860
for Py to follow, but the
5:04.860–5:07.263
readme has instructions for you and me.
5:07.263–5:10.540
This tells us what we need to do to get Py up and running.
5:10.540–5:12.267
And it's super simple because
5:12.267–5:15.860
I'm using something called dev containers built into VS Code.
5:15.860–5:18.840
so that we can easily run everything in a Docker container.
5:18.980–5:22.220
That means that we're sandboxing everything, we're nice and secure,
5:22.440–5:24.540
and also makes it really easy to set up.
5:24.760–5:25.940
So follow these instructions.
5:25.940–5:30.120
It includes having an open router key in a .emv file.
5:30.460–5:31.632
And then when you're ready,
5:31.632–5:33.620
we're going to bring this up in a dev container.
5:33.980–5:38.880
So I'm going to do the Ctrl-Shift-P on a PC or Command-Shift-P on a Mac.
5:39.140–5:42.660
And I'm going to say dev containers reopen in container.
5:42.660–5:46.300
and this is going to now launch a dev container.
5:46.600–5:48.440
I'm going to close this chat window here.
5:48.520–5:50.620
This is launching a Linux container,
5:50.880–5:54.060
which is going to be running installing Pi right now.
5:54.180–5:56.320
And we now have Pi installed
5:56.320–5:58.100
and we're ready to try this out.
5:58.360–6:00.160
And now as explained in the readme,
6:00.360–6:02.720
I'm running a command that will add a skill
6:02.720–6:03.940
to our coding agent.
6:03.940–6:05.980
It's a skill called agent browser
6:05.980–6:07.600
from the great people of Vercel.
6:07.780–6:09.500
And it's a skill which is designed
6:09.500–6:11.840
to give any coding agent the ability
6:11.840–6:15.520
to run a Chrome-based browser behind the scenes.
6:15.640–6:18.280
It's asking me which agent I would like to install it for.
6:18.500–6:21.240
I'm going to come all the way down and select Pi.
6:21.480–6:22.800
Find Pi in this list.
6:22.920–6:23.500
There it is.
6:23.600–6:25.660
And press the space bar and then enter.
6:26.060–6:27.420
Install it for the project.
6:27.980–6:28.980
Yes, we're done.
6:29.060–6:33.460
We've just installed the agent browser skill for Pi.
6:33.700–6:35.720
And now I'm going to run the command to start Pi.
6:35.720–6:39.580
As described in the readme, I just call Pi dash dash models.
6:39.580–6:44.073
and I pass in the name of the two models through OpenRouter that I want to use.
6:44.073–6:45.000
And this is your
6:45.000–6:48.121
first sighting of Pi. This is Pi running.
6:48.121–6:51.700
It's very simple and we're about to use it to try
6:51.700–6:54.381
building a sales CRM platform of our very own.
6:54.381–6:56.920
The first thing I show you is that if you hold
6:56.920–7:00.269
down control P, you can toggle the different models and
7:00.269–7:02.380
the models show on the bottom right
7:02.380–7:05.411
here. You can see it says DeepSeek V4 Pro.
7:05.411–7:08.620
I can do control P and it will flip to GLM 5.2.
7:08.620–7:11.512
These are the two models that I started it with, so
7:11.512–7:13.840
we can switch between them at any point.
7:14.180–7:16.900
And we're going to use both. Of course, we're going to use both.
7:17.100–7:22.020
The other thing to show you is that Shift-Tab chooses the different reasoning levels.
7:22.140–7:27.140
We can toggle High to Extra High to Thinking Off back to High and Extra High.
7:27.200–7:29.080
And we're going to go on Extra High.
7:29.260–7:32.860
We're going to have Deep Seek V4 Pro and also Control-P.
7:32.980–7:36.500
We're going to have GLM 5.2 on Extra High mode.
7:36.500–7:39.629
And now all I have to do wonderfully is say, look,
7:39.629–7:41.820
please build the entire project.
7:41.820–7:44.381
As described in agents.md, don't stop until
7:44.381–7:46.873
the success criteria are met and the server
7:46.873–7:48.880
is running and ready for me to test.
7:49.060–7:52.000
I can kick that off and off it goes.
7:52.180–7:53.510
You can see down here,
7:53.510–7:56.540
it tells you how much of the context is being used.
7:56.800–7:59.980
It's got 0.7% of the 1 million context.
8:00.000–8:02.544
percent of the 1 million context window for glm 5.
8:02.544–8:04.840
2 and it will also tell you how much it's spent
8:04.840–8:08.375
and we've already spent 0.9 of a cent uh already and
8:08.375–8:11.320
we've just spent our first cent on this and
8:11.320–8:14.960
We'll be keeping a very close watch on the spend, but it is off.
8:15.080–8:16.480
It's going to keep going.
8:16.720–8:19.560
It's going to go through each of the different stages of build,
8:19.700–8:22.560
and it's going to check the success criteria at every point.
8:22.940–8:24.660
Well, it's been going away for about five minutes.
8:24.800–8:25.640
Lots has been going on.
8:25.680–8:27.380
We've now spent 10 cents.
8:27.660–8:29.980
But meanwhile, I'm going to do something cool.
8:30.300–8:34.280
I've got another VS Code window running over here.
8:34.600–8:35.660
But this is, in fact,
8:35.660–8:38.340
another dev container on exactly the same project.
8:38.460–8:41.300
I've cloned it separately to a different directory.
8:41.320–8:43.751
It's called now CRM DeepSeek,
8:43.751–8:46.960
which is a clue as to what I'm about to do.
8:46.960–8:47.520
I am, of course,
8:47.520–8:49.500
I'm going to make the screen a little bit bigger for you.
8:49.900–8:54.960
I am, of course, now set to use DeepSeek instead of GLM 5.2.
8:55.120–8:57.560
And I'm going to use the same exact command.
8:57.840–8:59.620
Please build the entire project.
9:00.020–9:05.380
So I'm now going to set off in this other screen the same project in parallel.
9:05.920–9:09.269
Off it goes with DeepSeek V4 Pro running, also
9:09.269–9:11.120
on extra high reasoning.
9:11.320–9:14.414
Look at it go. And if I just shrink this window,
9:14.414–9:18.160
you'll see that the other one, GLM 5.2, is also on the go.
9:18.280–9:20.931
I have both coding agents, both using Pi,
9:20.931–9:23.427
both models on the go at the same time, b
9:23.427–9:26.000
uilding out the same project so we can se
9:26.000–9:28.417
e how they do, both in terms of cost and
9:28.417–9:32.160
speed, and also, of course, in terms of the final product.
9:32.160–9:34.229
And this is now one second later for you, but
9:34.229–9:36.241
it's like an hour and a half later for me, an
9:36.241–9:37.620
d both of them have finished.
9:37.620–9:41.012
And wonderfully, you can go into OpenRouter and
9:41.012–9:44.150
look at the logs and see the messages going to
9:44.150–9:46.100
GLM 5.2 and DeepSeek V4 Pro.
9:46.340–9:49.511
And you can also look at things like the activity charts and
9:49.511–9:51.840
see the usage and explore what was going on.
9:52.180–9:54.389
We look at the past three hours and
9:54.389–9:56.676
we can see everything that was happ
9:56.676–10:00.620
ening and use that to figure out how much did we spend on each.
10:01.580–10:05.700
And this page in the OpenRouter activity screen shows you everything that went on.
10:05.700–10:08.071
You can see the activity there. In the end,
10:08.071–10:11.120
if I back out things when I ran something a second time,
10:11.520–10:16.362
for GLM 5.2, it cost $4.15 to run this. $4.15, and
10:16.362–10:20.720
it took an hour and a quarter. For DeepSeek,
10:20.760–10:24.091
it was cheaper. It was $2.56 in total, and
10:24.091–10:28.500
it was a little bit faster. It took just about one hour,
10:28.800–10:31.719
a bit faster as well. So DeepSeek was faster and
10:31.719–10:34.040
cheaper. What were the results like?
10:34.040–10:37.557
And here is DeepSeek's implementation of a personal CRM.
10:37.557–10:40.928
Your personal CRM, like a sales force. You've got a dash
10:40.928–10:44.299
board with deals one and revenue. It's got some dummy da
10:44.299–10:47.596
ta in here to show it off. Something is overdue here. Yo
10:47.596–10:51.113
u've got organizations, the companies that you have with
10:51.113–10:54.484
an edit and delete, a Dell button, that's all right. Con
10:54.484–10:57.928
tacts and deals listed out here. Contacts, the people yo
10:57.928–10:59.540
u know at these companies.
10:59.540–11:05.440
deals, which are the deals that are at various stages of the pipeline shown in the stage here.
11:05.440–11:13.300
And the pipeline itself is a Kanban style board of deals moving from new all the way through to
11:13.300–11:17.402
one or lost. And you can drag them across to see what happens when
11:17.402–11:19.260
you do that. And you should be
11:19.260–11:21.887
able to add and add and search throughout.
11:21.887–11:24.740
Let's search for someone in here. There we go.
11:24.740–11:28.377
The search works too. It is a functioning platform.
11:28.377–11:30.660
It's backed by SQLite database.
11:31.340–11:34.255
This is a working but simple sales CRM. And
11:34.255–11:37.420
you've got to admit, it's pretty impressive.
11:37.780–11:40.328
But you know what's more impressive? GLM 5.
11:40.328–11:43.300
2's version is more impressive. Let's switch over
11:43.300–11:45.845
to it. This is the dashboard. And come on,
11:45.845–11:48.540
immediately you see this is better. This is
11:48.540–11:52.291
better. We've got these kind of hot metrics being surfaced at the top.
11:52.291–11:53.520
We've got deals one and
11:53.520–11:58.144
revenue, which just is like a better form chart with more rich data.
11:58.144–11:59.960
We've got these tasks here
11:59.960–12:02.912
that could be checked off. We've got recent activity and
12:02.912–12:04.420
then pipeline by stage at the
12:04.420–12:07.211
bottom and an open board link there. It's really nice.
12:07.211–12:09.320
We've got these icons here. We can switch
12:09.320–12:11.535
to organizations. This looks similar, but
12:11.535–12:14.180
of course we've got proper icons, not a Dell button.
12:14.180–12:19.545
We've got contacts as before, deals here with this really nice badges.
12:19.545–12:21.000
And then this is the
12:21.000–12:22.688
pipeline. It's just sharper.
12:22.688–12:26.400
Let's go back over here again to the pipeline. Compare this is a bit
12:26.400–12:30.186
more bland. This is really gorgeous, a really nice,
12:30.186–12:33.620
sharp pipeline. And of course, we will still no
12:33.620–12:35.675
doubt have the ability to drag and drop.
12:35.675–12:38.540
Everything works with the same stages and won and lost.
12:38.920–12:41.608
This is a pretty impressive screen,
12:41.608–12:45.820
a really nice product from GLM 5.2. But I know what you're
12:45.820–12:49.475
thinking. I know you so well. You're thinking, okay,
12:49.475–12:52.300
Ed, we get it. This is impressive. GLM 5.2
12:52.300–12:57.175
did a decent job, but how would Opus have done on the same thing?
12:57.175–12:59.800
How would Opus 4.8, the strongest
12:59.800–13:02.516
model on the planet, at least that's available to us all,
13:02.516–13:04.480
how would that do with this? And how would
13:04.480–13:06.859
it do if it was on ultra code mode and
13:06.859–13:10.960
allowed to do its absolute best? Well, of course, I've given
13:10.960–13:14.993
that a try too. And it's running on this tab right here.
13:14.993–13:18.220
It took an hour and it would have cost $21.27
13:18.220–13:23.321
if I'd used API directly rather than just using a usage of my Claude AI account.
13:23.321–13:24.420
And let's go over
13:24.420–13:26.985
to it. Here we go. This is the dashboard.
13:26.985–13:30.560
You can see it's quite similar. It has a chart. It has the
13:30.560–13:33.891
same called out kinds of numbers and a checklist here,
13:33.891–13:36.260
but it is a bit nicer, right? Look at the
13:36.260–13:40.496
dark background here it looks a little bit more professional but
13:40.496–13:42.300
only slightly only by a hair
13:42.300–13:47.800
and clicking through the tables you'll see a very similar construct clicking through them here it's
13:47.800–13:50.118
just slightly better perhaps and
13:50.118–13:54.920
the pipeline here we have another nice pipeline here is a kanban board
13:54.920–13:56.793
but check this out when i drag and
13:56.793–14:00.400
drop again you get this dotted line thing it's just a little bit
14:00.400–14:02.884
better. And so in summary, GLM 5.
14:02.884–14:07.760
2 was neither the cheapest nor the best, but it was quite close on
14:07.760–14:11.098
both fronts. It was low cost and it was really good.
14:11.098–14:13.800
And so I would say, yes, this is a credible
14:13.800–14:17.894
alternative to frontier models that can be used for your own projects.
14:17.894–14:19.560
It's not super cheap. You're
14:19.560–14:22.111
still spending north of $4 for a product like this,
14:22.111–14:24.420
but there's a lot of capability here. I didn't
14:24.420–14:26.788
even show you. You can click into organizations and
14:26.788–14:28.660
edit them and then click here to go over to
14:28.660–14:32.268
the contact screen, edit a log activity here in the pipeline.
14:32.268–14:34.320
You can click on cards to bring them
14:34.320–14:37.153
up. There's tons of functionality and
14:37.153–14:40.960
it built all of this first time one shot and it met the
14:40.960–14:43.219
success criteria and worked right away. And
14:43.219–14:45.600
if you'd like to give this a try yourself and use
14:45.600–14:48.089
Pi, there are details in the description below and
14:48.089–14:50.340
I'd love to see what you come up with. And if you
14:50.340–14:52.523
enjoyed this video, then please do like and subscribe.
14:52.523–14:53.900
That's the way I know you're really
14:53.900–14:57.685
there. And it's the way the YouTube algorithm knows you're really there as well.
14:57.685–14:58.420
And either way,
14:58.420–15:00.340
I hope to see you very soon for another video.
0:00.000–0:03.320
這是關於程式碼代理程式我最常收到的問題之一。
0:03.600–0:08.960
我們是否已經準備好使用開源模型來建構真實世界的專案？
0:09.160–0:11.960
我來告訴你們，是的，我們已經準備好了。
0:12.280–0:13.240
我將向你們展示。
0:13.500–0:14.988
只需大約 10 分鐘，
0:14.988–0:18.260
我們將檢視真實世界的產品，像是這個和這個
0:18.260–0:22.340
由程式碼代理程式使用開源模型所建構。
0:22.340–0:25.940
我們將把它與 Claude 在相同簡報下能做的事進行比較。
0:26.040–0:28.660
我將許多有趣的事物塞進這支影片裡。
0:28.660–0:32.158
我們不使用 Claude Code，而是使用 Pi，
0:32.158–0:34.540
這是一個超級受歡迎的程式碼代理程式。
0:34.800–0:36.220
最近才真正變得非常受歡迎。
0:36.520–0:38.499
我們將使用 GLM 5.2，
0:38.499–0:41.240
來自 ZAI 的最新突破性模型。
0:41.580–0:43.720
我們將使用 DeepSeq v4。
0:43.880–0:46.760
我們將使用開發容器以獲得額外的安全性。
0:47.060–0:48.080
有很多事情要做。
0:48.340–0:52.540
現在，這個網站是 AI 工程師最重要的資源之一。
0:52.740–0:54.080
我將在描述中連結到它。
0:54.080–0:58.021
它是 artificialanalysis.ai，擁有大量資訊，
0:58.021–1:00.620
關於不同模型的大量數據。
1:01.000–1:02.783
特別是，如果你往下捲動，
1:02.783–1:05.180
你會看到這張「智慧與成本」的圖表。
1:05.340–1:08.000
它有點被埋沒，但這是一張非常有價值的圖表。
1:08.300–1:12.060
我的 AI 工程師 CoreTrack 課程的學員對這張圖表很熟悉。
1:12.380–1:14.420
每個點代表不同的模型。
1:14.780–1:19.160
橫軸是執行模型任務的成本。
1:19.560–1:21.360
越往右代表成本越高。
1:21.360–1:24.708
縱軸是模型的智慧程度。
1:24.708–1:28.520
更聰明的模型會位置更高。所以理想上
1:28.520–1:32.120
你希望進入左上角的綠色區域。
1:32.120–1:34.820
如果模型既便宜
1:34.820–1:36.225
又聰明，它們就會出現在那裡。而
1:36.225–1:39.280
今天我們將嘗試兩種位於
1:39.280–1:44.554
這條線邊緣的模型。它們處於智慧但
1:44.554–1:45.960
便宜的邊界。我們將
1:45.960–1:47.679
檢視 DeepSeq v4 Pro。
1:47.679–1:52.100
這是一款以極低成本著稱的模型，
1:52.220–1:55.746
我們將看看結果。我們也將檢視來自 z.ai 的 GLM 5.2，
1:55.746–1:58.600
它享有
1:58.600–2:01.933
甚至比這張圖表可能讓你認為的還要聰明，
2:01.933–2:03.820
我們將自行評判。
2:03.820–2:08.409
因此，對於 GLM 5.2，社群相當興奮，
2:08.409–2:11.900
以及 DeepSeq v4，讓我們來測試它們。
2:11.900–2:17.655
因此，Pi 是一個編碼代理程式的範例，像是 Claude code 或更像 open code，因為
2:17.655–2:23.409
它是這些設計為不綁定任何大型語言模型，但允許你
2:23.409–2:25.760
插入任何開源模型的編碼代理程式之一。
2:25.760–2:28.895
特別是，Pi 有很好的聲譽，因為
2:28.895–2:30.320
它非常輕量。
2:30.580–2:31.340
它是最小化的。
2:31.680–2:34.240
很容易向不同方向擴展它。
2:34.380–2:37.080
但是，你开箱即用的東西非常簡單。
2:37.300–2:39.184
它只有一個簡單的提示和
2:39.184–2:41.540
簡單的上下文管理方法。
2:41.540–2:44.460
它讓你按照你想要的任何方向進行。
2:44.720–2:46.260
而這正是我們將要做的。
2:46.500–2:48.480
好的，測試的時刻已經到來。
2:48.540–2:50.356
我認為，無論你是否具備技術背景，
2:50.356–2:51.900
你都想要觀看這個，只是為了
2:51.900–2:54.923
了解它是如何運作的，Pi 是如何運作的，以及
2:54.923–2:57.720
這些模型的表現如何。我們將對
2:57.720–3:00.908
GLM 和 DeepSeek 進行測試。我們將現在就進行。
3:00.908–3:02.340
我們將建立一個
3:02.340–3:05.840
有趣的專案。就在這裡。我這裡有一個資料夾，
3:05.840–3:07.800
幾乎是空的。它只有
3:07.800–3:10.545
一些東西，包括一個 agents.md 檔案。而
3:10.545–3:13.360
那描述了我們想要建立的。我們想要
3:13.360–3:16.474
建立一個個人 CRM 系統，
3:16.474–3:21.560
一個類似 Salesforce 的平台，只為你，只
3:21.560–3:25.155
為我，我們自己的版本，就像 Salesforce，
3:25.155–3:27.280
一個 pipe drive，我在我的其他
3:27.280–3:29.146
課程中涵蓋過，類似這樣的东西。但是
3:29.146–3:31.220
我們想要從頭開始建立整個東西。我們想要一個編碼代理程式來做這件事。
3:31.220–3:33.523
只是簡單的一個，請注意。而我這裡有從商業人士的角度撰寫的業務
3:33.523–3:36.220
需求。這裡有一個小技術
3:36.220–3:40.112
部分，但否則它是關於外觀和
3:40.112–3:41.780
感覺。它是關於範圍內和
3:41.780–3:45.248
範圍外以及
3:45.248–3:47.560
CRM 從商業人士的角度來看可以
3:47.560–3:48.452
做什麼，只有一點
3:48.452–3:49.344
從業務人員的角度來看，CRM 能
3:49.344–3:52.500
做什麼，只需一點點
3:52.500–3:55.727
技術指導，讓模型保持在正軌上。
3:55.727–3:58.220
但我們所做且
3:58.220–4:02.671
在使用開源模型時絕對必要的事情是這個。
4:02.671–4:04.780
我們將專案以
4:04.780–4:09.679
階段的方式規劃，這些階段被切割成小塊，可以一步步實施。
4:09.679–4:10.740
對於每個階段，
4:10.740–4:14.372
我們用業務術語解釋功能，然後
4:14.372–4:17.160
我們提供成功標準，讓
4:17.160–4:21.060
程式碼生成代理（coding agent）可以驗證它是否已通過此階段，
4:21.060–4:23.660
即已完成，可以進入下一個階段。
4:23.920–4:27.113
然後在最後，我們提供整體成功標準。
4:27.113–4:29.100
我有六個階段，然後在最後
4:29.100–4:31.069
的成功標準。並且
4:31.069–4:35.080
在滿足所有這些成功標準之前，它不能宣告專案勝利。我知道我經常強調這一點，
4:35.080–4:37.788
但成功標準對於確保
4:37.788–4:40.360
程式碼生成代理保持在正軌上至關重要。並且
4:40.360–4:43.250
重要的是，你還必須賦予它
4:43.250–4:46.140
它進行徹底測試所需的技能，
4:46.140–4:48.436
例如在瀏覽器視窗中開啟畫面並進行檢查。
4:48.436–4:51.880
這樣它就有了回饋迴圈，
4:51.980–4:54.196
並且可以持續迭代，直到滿足成功標準。
4:54.196–4:57.520
因此，agents.md 檔案包含了我們程式碼生成代理
4:57.920–5:03.103
Py 需要遵循的所有指示，但
5:03.103–5:04.860
readme 則包含給你和我的指示。
5:04.860–5:07.263
這告訴我們需要做什麼才能讓 Py 啟動並運行。
5:07.263–5:10.540
而且這非常簡單，因為
5:10.540–5:12.267
我使用了 VS Code 內建的
5:12.267–5:15.860
稱為開發容器（dev containers）的功能。
5:15.860–5:18.840
這樣我們就可以輕鬆地在 Docker 容器中運行一切。
5:18.980–5:22.220
這意味著我們將一切隔離在沙盒中，安全性很高，
5:22.440–5:24.540
並且也極大地簡化了設定過程。
5:24.760–5:25.940
請遵循這些指示。
5:25.940–5:30.120
其中包括在 .env 檔案中擁有 Open Router 金鑰。
5:30.460–5:31.632
然後當你準備好時，
5:31.632–5:33.620
我們將以開發容器啟動它。
5:33.980–5:38.880
所以我將在 PC 上按 Ctrl-Shift-P，或在 Mac 上按 Command-Shift-P。
5:39.140–5:42.660
然後我會輸入 dev containers reopen in container。
5:42.660–5:46.300
這將現在啟動一個開發容器。
5:46.600–5:48.440
我現在要關閉這個聊天視窗。
5:48.520–5:50.620
這正在啟動一個 Linux 容器，
5:50.880–5:54.060
它現在正在安裝 Pi。
5:54.180–5:56.320
現在 Pi 已經安裝好了，
5:56.320–5:58.100
我們準備好要來試試看了。
5:58.360–6:00.160
根據 README 中的說明，
6:00.360–6:02.720
我正在執行一個指令，用來新增一個技能
6:02.720–6:03.940
給我們的程式碼代理程式（coding agent）。
6:03.940–6:05.980
這個技能叫做 agent browser
6:05.980–6:07.600
來自 Vercel 的優秀團隊。
6:07.780–6:09.500
這是一個設計用來
6:09.500–6:11.840
賦予任何程式碼代理程式能力
6:11.840–6:15.520
在背景執行基於 Chrome 的瀏覽器的技能。
6:15.640–6:18.280
它詢問我要為哪個代理程式安裝此技能。
6:18.500–6:21.240
我會一直往下捲動並選擇 Pi。
6:21.480–6:22.800
在這個清單中找到 Pi。
6:22.920–6:23.500
就是它。
6:23.600–6:25.660
然後按下空白鍵並輸入。
6:26.060–6:27.420
為專案安裝它。
6:27.980–6:28.980
是的，我們完成了。
6:29.060–6:33.460
我們剛剛為 Pi 安裝了 agent browser 技能。
6:33.700–6:35.720
現在我要執行指令來啟動 Pi。
6:35.720–6:39.580
根據 README 的描述，我只需呼叫 Pi 並加上 --models 參數。
6:39.580–6:44.073
並透過 OpenRouter 傳入我想使用的兩個模型名稱。
6:44.073–6:45.000
這是你
6:45.000–6:48.121
第一次見到 Pi。這是正在執行的 Pi。
6:48.121–6:51.700
它非常簡單，我們即將使用它來嘗試
6:51.700–6:54.381
建構我們自己的銷售 CRM 平台。
6:54.381–6:56.920
我要先展示的是，如果你按住
6:56.920–7:00.269
Control P，你可以切換不同的模型，
7:00.269–7:02.380
模型會顯示在右下角
7:02.380–7:05.411
這裡。你可以看到它顯示 DeepSeek V4 Pro。
7:05.411–7:08.620
我可以按 Control P，它會切換到 GLM 5.2。
7:08.620–7:11.512
這兩個是我啟動時使用的模型，所以
7:11.512–7:13.840
我們可以隨時在它們之間切換。
7:14.180–7:16.900
我們將同時使用這兩個。當然，我們會同時使用這兩個。
7:17.100–7:22.020
另一個要展示的是，Shift-Tab 可以選擇不同的推理層級。
7:22.140–7:27.140
我們可以在 High、Extra High、Thinking Off、High 和 Extra High 之間切換。
7:27.200–7:29.080
我們將選擇 Extra High。
7:29.260–7:32.860
我們將使用 Deep Seek V4 Pro 並同時按 Control-P。
7:32.980–7:36.500
我們將以 Extra High 模式使用 GLM 5.2。
7:36.500–7:39.629
現在我只需要很完美地說：「看，
7:39.629–7:41.820
請建立整個專案。
7:41.820–7:44.381
如 agents.md 所述，直到
7:44.381–7:46.873
成功標準達成且伺服器
7:46.873–7:48.880
正在運行並準備好讓我測試之前，不要停止。
7:49.060–7:52.000
我可以啟動它，然後它就開始運作了。
7:52.180–7:53.510
你可以在這裡看到，
7:53.510–7:56.540
它會告訴你使用了多少上下文。
7:56.800–7:59.980
它使用了 100 萬上下文中的 0.7%。
8:00.000–8:02.544
GLM 5.2 的 100 萬上下文視窗的
8:02.544–8:04.840
百分比，它也會告訴你花費了多少
8:04.840–8:08.375
我們已經花費了 0.9 美分，呃，已經
8:08.375–8:11.320
我們剛剛花費了第一美分在這個專案上，並且
8:11.320–8:14.960
我們會密切關注花費，但它已經關閉了。
8:15.080–8:16.480
它會繼續進行。
8:16.720–8:19.560
它會經過建構的不同階段，
8:19.700–8:22.560
並且它會在每個點檢查成功標準。
8:22.940–8:24.660
嗯，它已經運行了大約五分鐘。
8:24.800–8:25.640
發生了很多事情。
8:25.680–8:27.380
我們現在已經花費了 10 美分。
8:27.660–8:29.980
但與此同時，我要做一些很酷的事情。
8:30.300–8:34.280
我在這裡有另一個 VS Code 視窗在運行。
8:34.600–8:35.660
但這實際上，
8:35.660–8:38.340
是完全相同專案的另一個開發容器。
8:38.460–8:41.300
我已將其單獨複製到不同的目錄。
8:41.320–8:43.751
它現在被稱為 CRM DeepSeek，
8:43.751–8:46.960
這是我即將要做的事情的線索。
8:46.960–8:47.520
我當然，
8:47.520–8:49.500
我會把畫面稍微放大一點給你看。
8:49.900–8:54.960
我當然，現在設定為使用 DeepSeek 而不是 GLM 5.2。
8:55.120–8:57.560
我將使用完全相同的指令。
8:57.840–8:59.620
請建立整個專案。
9:00.020–9:05.380
所以現在我將在另一個螢幕中平行啟動相同的專案。
9:05.920–9:09.269
它開始運行 DeepSeek V4 Pro，也
9:09.269–9:11.120
處於 Extra High 推理模式。
9:11.320–9:14.414
看它運行。如果我只是縮小這個視窗，
9:14.414–9:18.160
你會看到另一個，GLM 5.2，也在運行。
9:18.280–9:20.931
我有兩個編碼代理，都使用 Pi，
9:20.931–9:23.427
兩個模型同時運行，構
9:23.427–9:26.000
建立相同的專案，以便我們觀察
9:26.000–9:28.417
它們在成本和
9:28.417–9:32.160
速度方面的表現，當然還有最終產品的品質。
9:32.160–9:34.229
這現在對你來說是一秒後的畫面，但
9:34.229–9:36.241
對我來說已經過了一個半小時，兩
9:36.241–9:37.620
個專案都已經完成了。
9:37.620–9:41.012
而且很棒的是，你可以進入 OpenRouter
9:41.012–9:44.150
查看日誌，並看到發送到
9:44.150–9:46.100
GLM 5.2 和 DeepSeek V4 Pro 的訊息。
9:46.340–9:49.511
你也可以查看活動圖表等內容，
9:49.511–9:51.840
了解使用情況並探索當時發生了什麼。
9:52.180–9:54.389
我們查看過去三個小時的記錄，
9:54.389–9:56.676
可以看到所有正在發生的
9:56.676–10:00.620
事情，並用這些資訊來計算我們每個專案花費了多少。
10:01.580–10:05.700
OpenRouter 活動畫面中的這個頁面顯示了所有發生的事情。
10:05.700–10:08.071
你可以在那裡看到活動記錄。最後，
10:08.071–10:11.120
如果我第二次執行時排除某些內容，
10:11.520–10:16.362
對於 GLM 5.2，這次執行的成本是 4.15 美元。4.15 美元，
10:16.362–10:20.720
並且耗時一個小時一刻。對於 DeepSeek，
10:20.760–10:24.091
它更便宜。總共是 2.56 美元，
10:24.091–10:28.500
而且速度稍快。它只花了一個小時，
10:28.800–10:31.719
速度也稍快一些。所以 DeepSeek 更快且
10:31.719–10:34.040
更便宜。結果如何呢？
10:34.040–10:37.557
這裡是 DeepSeek 實現的個人 CRM。
10:37.557–10:40.928
你的個人 CRM，類似於銷售團隊。你有一個儀
10:40.928–10:44.299
表板，顯示交易一和收入。這裡有一些虛構數
10:44.299–10:47.596
據來展示功能。這裡有一項已逾期。你
10:47.596–10:51.113
可以看到組織，即你與之合作的
10:51.113–10:54.484
公司，有編輯和刪除功能，刪除按鈕沒問題。
10:54.484–10:57.928
這裡列出了聯絡人和交易。聯絡人是你
10:57.928–10:59.540
在這些公司認識的人。
10:59.540–11:05.440
交易則是處於管道各階段的交易，如這裡所示。
11:05.440–11:13.300
管道本身是一個看板風格的交易板，交易從新客戶一直推進到
11:13.300–11:17.402
成交或丟失。你可以拖曳它們來查看發生什麼情況。
11:17.402–11:19.260
你應該能夠新增、新增並進行搜尋。
11:19.260–11:21.887
讓我們在這裡搜尋某人。找到了。
11:21.887–11:24.740
搜尋功能也正常運作。這是一個功能完整的平台。
11:24.740–11:28.377
它由 SQLite 資料庫支援。
11:28.377–11:30.660
這是一個運作中但簡單的銷售 CRM。而
11:31.340–11:34.255
你必須承認，這相當令人印象深刻。
11:34.255–11:37.420
你不得不承認，它相當令人印象深刻。
11:37.780–11:40.328
但你知道什麼更令人印象深刻嗎？GLM 5.2 的版本。
11:40.328–11:43.300
讓我們切換過去看看。
11:43.300–11:45.845
這是儀表板。說真的，
11:45.845–11:48.540
你馬上就會發現這個更好。這個
11:48.540–11:52.291
更好。我們在頂部看到了這些類型的熱門指標。
11:52.291–11:53.520
我們有交易一和
11:53.520–11:58.144
營收，這就像是一個擁有更豐富數據的更好的圖表。
11:58.144–11:59.960
我們這裡有這些
11:59.960–12:02.912
可以勾選的任務。我們有最近的活動和
12:02.912–12:04.420
底部的管道階段以及
12:04.420–12:07.211
這裡的開放看板連結。真的很不錯。
12:07.211–12:09.320
我們這裡有這些圖標。我們可以切換
12:09.320–12:11.535
到組織。這看起來相似，但是
12:11.535–12:14.180
當然我們有正確的圖標，而不是戴爾按鈕。
12:14.180–12:19.545
我們有聯絡人，這裡有交易，還有這些非常漂亮的徽章。
12:19.545–12:21.000
然後這是
12:21.000–12:22.688
管道。它更銳利。
12:22.688–12:26.400
讓我們回到這裡的管道。比較一下，這個有點
12:26.400–12:30.186
平淡。這個真的很漂亮，一個非常漂亮、
12:30.186–12:33.620
銳利的管道。當然，我們仍然無疑擁有拖放功能。
12:33.620–12:35.675
所有功能都使用相同的階段，以及贏和輸。
12:35.675–12:38.540
這是一個相當令人印象深刻的畫面，
12:38.920–12:41.608
來自 GLM 5.2 的非常棒的产品。但我知道你在想什麼。
12:41.608–12:45.820
我這麼了解你。你在想，好吧，
12:45.820–12:49.475
Ed，我們明白了。這很令人印象深刻。GLM 5.2
12:49.475–12:52.300
做得還不錯，但是 Opus 在同樣的事情上會表現如何呢？
12:52.300–12:57.175
Opus 4.8，地球上最強大的
12:57.175–12:59.800
模型，至少是我們所有人都能使用的，
12:59.800–13:02.516
它會如何處理這個？如果它處於超代碼模式，
13:02.516–13:04.480
並被允許發揮其絕對最佳水平，它會表現如何？
13:04.480–13:06.859
當然，我也試過了。它正在這裡的標籤上運行。
13:06.859–13:10.960
它花了一個小時，如果我直接使用 API 而不是使用我的 Claude AI 帳戶，
13:10.960–13:14.993
這將花費 21.27 美元。
13:14.993–13:18.220
讓我們過去看看。
13:18.220–13:23.321
我們開始吧。這是儀表板。
13:23.321–13:24.420
你可以看到它非常相似。它有一個圖表。它有
13:24.420–13:26.985
同樣列出的數字類型和這裡的檢查清單，
13:26.985–13:30.560
你可以看到它非常相似。它有一個圖表。它有
13:30.560–13:33.891
相同列出的數值，這裡還有一個檢查清單，
13:33.891–13:36.260
但看起來好一點，對吧？看看
13:36.260–13:40.496
這裡的深色背景，看起來更專業一點，但
13:40.496–13:42.300
只有一點點，僅微乎其微
13:42.300–13:47.800
點擊瀏覽表格時，你會看到非常相似的結構，在這裡點擊瀏覽它們時，它
13:47.800–13:50.118
只是稍微好一點，也許
13:50.118–13:54.920
這裡的管道，我們有另一個很棒的管道，這裡是一個看板
13:54.920–13:56.793
但看看這個，當我拖放
13:56.793–14:00.400
再次拖放時，你會看到這個虛線的東西，只是稍微
14:00.400–14:02.884
好一點。所以總結來說，GLM 5.
14:02.884–14:07.760
2 既不是最便宜的，也不是最好的，但在
14:07.760–14:11.098
兩個方面都非常接近。它成本低，而且真的很好。
14:11.098–14:13.800
所以我想說，是的，這是一個可信的
14:13.800–14:17.894
替代前沿模型的方案，可以用於您自己的專案。
14:17.894–14:19.560
它不是超級便宜。你
14:19.560–14:22.111
仍然要為這樣的產品支付超過 4 美元的費用，
14:22.111–14:24.420
但這裡有很多功能。我甚至
14:24.420–14:26.788
甚至展示給你。你可以點擊進入組織，並
14:26.788–14:28.660
編輯它們，然後點擊這裡前往
14:28.660–14:32.268
聯絡畫面，編輯管道中的日誌活動。
14:32.268–14:34.320
你可以點擊卡片來
14:34.320–14:37.153
顯示它們。這裡有大量的功能，並且
14:37.153–14:40.960
它第一次就構建了所有這些，一次完成，並且符合
14:40.960–14:43.219
成功標準，立即生效。並且
14:43.219–14:45.600
如果你想自己嘗試這個並使用
14:45.600–14:48.089
Pi，下方的描述中有詳細資訊，並且
14:48.089–14:50.340
我很想看看你會想出什麼。如果你
14:50.340–14:52.523
喜歡這個影片，請按讚並訂閱。
14:52.523–14:53.900
那是我知道你真正
14:53.900–14:57.685
在那裡的方式。這也是 YouTube 演算法知道你真的在那裡的方式。
14:57.685–14:58.420
無論哪種方式，
14:58.420–15:00.340
我希望很快再見，為你帶來另一個影片。
0:00.000–0:03.320
It's one of the most common questions I get about coding agents.
這是關於程式碼代理程式我最常收到的問題之一。
0:03.600–0:08.960
Are we ready yet to use open source models to build real world projects?
我們是否已經準備好使用開源模型來建構真實世界的專案？
0:09.160–0:11.960
I'm here to tell you that yes, we are ready.
我來告訴你們，是的，我們已經準備好了。
0:12.280–0:13.240
And I'm going to show you.
我將向你們展示。
0:13.500–0:14.988
In just about 10 minutes time,
只需大約 10 分鐘，
0:14.988–0:18.260
we're going to be looking at real world products like this and this
我們將檢視真實世界的產品，像是這個和這個
0:18.260–0:22.340
built by coding agents using open source models.
由程式碼代理程式使用開源模型所建構。
0:22.340–0:25.940
And we'll be comparing it to what Claude can do with the same brief.
我們將把它與 Claude 在相同簡報下能做的事進行比較。
0:26.040–0:28.660
And I'm cramming so many interesting things into this one video.
我將許多有趣的事物塞進這支影片裡。
0:28.660–0:32.158
Instead of using Claude Code, we're going to be using Pi,
我們不使用 Claude Code，而是使用 Pi，
0:32.158–0:34.540
which is a super popular coding agent.
這是一個超級受歡迎的程式碼代理程式。
0:34.800–0:36.220
Just recently become really popular.
最近才真正變得非常受歡迎。
0:36.520–0:38.499
We're going to be using GLM 5.2,
我們將使用 GLM 5.2，
0:38.499–0:41.240
the latest groundbreaking model from ZAI.
來自 ZAI 的最新突破性模型。
0:41.580–0:43.720
And we're going to be using DeepSeq v4.
我們將使用 DeepSeq v4。
0:43.880–0:46.760
And we're going to be using dev containers for extra safety.
我們將使用開發容器以獲得額外的安全性。
0:47.060–0:48.080
A lot going on.
有很多事情要做。
0:48.340–0:52.540
Now, this website is one of the most important resources for an AI engineer.
現在，這個網站是 AI 工程師最重要的資源之一。
0:52.740–0:54.080
And I'm going to link to it in the description.
我將在描述中連結到它。
0:54.080–0:58.021
It's artificialanalysis.ai with tons of information,
它是 artificialanalysis.ai，擁有大量資訊，
0:58.021–1:00.620
a lot of data about different models.
關於不同模型的大量數據。
1:01.000–1:02.783
In particular, if you scroll down,
特別是，如果你往下捲動，
1:02.783–1:05.180
there is this chart intelligence versus cost.
你會看到這張「智慧與成本」的圖表。
1:05.340–1:08.000
It's a bit buried, but it's a really valuable chart.
它有點被埋沒，但這是一張非常有價值的圖表。
1:08.300–1:12.060
And people on my AI Engineer CoreTrack course know this chart well.
我的 AI 工程師 CoreTrack 課程的學員對這張圖表很熟悉。
1:12.380–1:14.420
Each dot represents a different model.
每個點代表不同的模型。
1:14.780–1:19.160
The x-axis is how much does it cost to run the model for a task.
橫軸是執行模型任務的成本。
1:19.560–1:21.360
More to the right means it's more expensive.
越往右代表成本越高。
1:21.360–1:24.708
the y-axis is how intelligent is the model.
縱軸是模型的智慧程度。
1:24.708–1:28.520
A smarter model goes higher up. So you're ideally
更聰明的模型會位置更高。所以理想上
1:28.520–1:32.120
looking to fit in that green quadrant on the top left.
你希望進入左上角的綠色區域。
1:32.120–1:34.820
That is where models go if they are cheap
如果模型既便宜
1:34.820–1:36.225
and they're also smart. And
又聰明，它們就會出現在那裡。而
1:36.225–1:39.280
today we're going to try out two models that are kind of on the
今天我們將嘗試兩種位於
1:39.280–1:44.554
rim of this line here. They're kind of at this frontier of being smart but
這條線邊緣的模型。它們處於智慧但
1:44.554–1:45.960
cheap. We're going
便宜的邊界。我們將
1:45.960–1:47.679
to look at DeepSeq v4 Pro.
檢視 DeepSeq v4 Pro。
1:47.679–1:52.100
This is one that has a great reputation for being really low cost,
這是一款以極低成本著稱的模型，
1:52.220–1:55.746
and we'll see. And we're also going to look at GLM 5.
我們將看看結果。我們也將檢視來自 z.ai 的 GLM 5.2，
1:55.746–1:58.600
2 from z.ai, and that has a reputation for
它享有
1:58.600–2:01.933
being even smarter than this diagram might lead you to believe,
甚至比這張圖表可能讓你認為的還要聰明，
2:01.933–2:03.820
and we'll judge that for ourselves.
我們將自行評判。
2:03.820–2:08.409
So GLM 5.2, which the community is super excited about,
因此，對於 GLM 5.2，社群相當興奮，
2:08.409–2:11.900
and DeepSeq v4, let's put them to the test.
以及 DeepSeq v4，讓我們來測試它們。
2:11.900–2:17.655
So Pi is an example of a coding agent like a Claude code or more like an open code because
因此，Pi 是一個編碼代理程式的範例，像是 Claude code 或更像 open code，因為
2:17.655–2:23.409
it's one of these coding agents that are designed to not be tied to any LLM but to allow y
它是這些設計為不綁定任何大型語言模型，但允許你
2:23.409–2:25.760
ou to plug in any open source model.
插入任何開源模型的編碼代理程式之一。
2:25.760–2:28.895
And in particular, Pi has a great reputation because
特別是，Pi 有很好的聲譽，因為
2:28.895–2:30.320
it's very lightweight.
它非常輕量。
2:30.580–2:31.340
It is minimal.
它是最小化的。
2:31.680–2:34.240
It's easy to extend it in different directions.
很容易向不同方向擴展它。
2:34.380–2:37.080
But what you get out of the box is very simple.
但是，你开箱即用的東西非常簡單。
2:37.300–2:39.184
And it just has a simple prompt and
它只有一個簡單的提示和
2:39.184–2:41.540
a simple approach to context management.
簡單的上下文管理方法。
2:41.540–2:44.460
And it lets you take it in whatever direction you want.
它讓你按照你想要的任何方向進行。
2:44.720–2:46.260
And that's exactly what we're going to do.
而這正是我們將要做的。
2:46.500–2:48.480
Okay, the moment has arrived to put this to the test.
好的，測試的時刻已經到來。
2:48.540–2:50.356
And I would say whether or not you're technical,
我認為，無論你是否具備技術背景，
2:50.356–2:51.900
you're going to want to watch this just to
你都想要觀看這個，只是為了
2:51.900–2:54.923
get a sense for how it all works, how Pi works, and
了解它是如何運作的，Pi 是如何運作的，以及
2:54.923–2:57.720
how these models perform. We are going to put
這些模型的表現如何。我們將對
2:57.720–3:00.908
GLM and DeepSeek to the test. We're going to do it right now.
GLM 和 DeepSeek 進行測試。我們將現在就進行。
3:00.908–3:02.340
We're going to build out an
我們將建立一個
3:02.340–3:05.840
interesting project. And here it is. I've got a folder here,
有趣的專案。就在這裡。我這裡有一個資料夾，
3:05.840–3:07.800
which is almost empty. It just has
幾乎是空的。它只有
3:07.800–3:10.545
a few things, including an agents.md file. And
一些東西，包括一個 agents.md 檔案。而
3:10.545–3:13.360
that describes what we'd like to build. We would
那描述了我們想要建立的。我們想要
3:13.360–3:16.474
like to build a personal CRM system,
建立一個個人 CRM 系統，
3:16.474–3:21.560
a personal Salesforce kind of platform, just for you, just
一個類似 Salesforce 的平台，只為你，只
3:21.560–3:25.155
for me, our very own version of it, like a Salesforce,
為我，我們自己的版本，就像 Salesforce，
3:25.155–3:27.280
a pipe drive that I cover in some
一個 pipe drive，我在我的其他
3:27.280–3:29.146
of my other courses, that kind of thing. But
課程中涵蓋過，類似這樣的东西。但是
3:29.146–3:31.220
we're going to want to build the whole thing from
我們想要從頭開始建立整個東西。我們想要一個編碼代理程式來做這件事。
3:31.220–3:33.523
scratch. We want a coding agent to do that.
只是簡單的一個，請注意。而我這裡有從商業人士的角度撰寫的業務
3:33.523–3:36.220
Just a simple one, mind. And I've got the business
需求。這裡有一個小技術
3:36.220–3:40.112
requirements here written from a business person's point of view.
部分，但否則它是關於外觀和
3:40.112–3:41.780
There is one small technical
感覺。它是關於範圍內和
3:41.780–3:45.248
section in here, but otherwise it's about the look and
範圍外以及
3:45.248–3:47.560
feel. It's about what's in scope and
CRM 從商業人士的角度來看可以
3:47.560–3:48.452
not in scope and
做什麼，只有一點
3:48.452–3:49.344
what the CRM can
從業務人員的角度來看，CRM 能
3:49.344–3:52.500
do from a business person's point of view, just a little
做什麼，只需一點點
3:52.500–3:55.727
bit of technical guidance to keep the model on rails.
技術指導，讓模型保持在正軌上。
3:55.727–3:58.220
But the most important thing we do that is
但我們所做且
3:58.220–4:02.671
absolutely mandatory when working with open source models is this.
在使用開源模型時絕對必要的事情是這個。
4:02.671–4:04.780
We lay out the project in terms of
我們將專案以
4:04.780–4:09.679
phases, bite-sized phases that can each be implemented a step at a time.
階段的方式規劃，這些階段被切割成小塊，可以一步步實施。
4:09.679–4:10.740
For each phase,
對於每個階段，
4:10.740–4:14.372
we explain the features in business terms, and then
我們用業務術語解釋功能，然後
4:14.372–4:17.160
we give success criteria, ways that the
我們提供成功標準，讓
4:17.160–4:21.060
coding agent can verify that it has passed this phase,
程式碼生成代理（coding agent）可以驗證它是否已通過此階段，
4:21.060–4:23.660
it's done, it can move on to the next.
即已完成，可以進入下一個階段。
4:23.920–4:27.113
And then at the end, we give overall success criteria.
然後在最後，我們提供整體成功標準。
4:27.113–4:29.100
I've got six phases and then final
我有六個階段，然後在最後
4:29.100–4:31.069
success criteria at the end. And
的成功標準。並且
4:31.069–4:35.080
it cannot declare victory on this project until all these success
在滿足所有這些成功標準之前，它不能宣告專案勝利。我知道我經常強調這一點，
4:35.080–4:37.788
criteria are met. And I know I go on about it a lot,
但成功標準對於確保
4:37.788–4:40.360
but success criteria are so crucial to making
程式碼生成代理保持在正軌上至關重要。並且
4:40.360–4:43.250
sure that you keep the coding agent on track. And
重要的是，你還必須賦予它
4:43.250–4:46.140
importantly, you also have to give it the skills
它進行徹底測試所需的技能，
4:46.140–4:48.436
it needs to be able to test thoroughly,
例如在瀏覽器視窗中開啟畫面並進行檢查。
4:48.436–4:51.880
like bring up the screens in a browser window and check it.
這樣它就有了回饋迴圈，
4:51.980–4:54.196
That way it's got the feedback loop and
並且可以持續迭代，直到滿足成功標準。
4:54.196–4:57.520
it can keep iterating until the success criteria are met.
因此，agents.md 檔案包含了我們程式碼生成代理
4:57.920–5:03.103
So the agents.md file has all of the instructions for our coding agent,
Py 需要遵循的所有指示，但
5:03.103–5:04.860
for Py to follow, but the
readme 則包含給你和我的指示。
5:04.860–5:07.263
readme has instructions for you and me.
這告訴我們需要做什麼才能讓 Py 啟動並運行。
5:07.263–5:10.540
This tells us what we need to do to get Py up and running.
而且這非常簡單，因為
5:10.540–5:12.267
And it's super simple because
我使用了 VS Code 內建的
5:12.267–5:15.860
I'm using something called dev containers built into VS Code.
稱為開發容器（dev containers）的功能。
5:15.860–5:18.840
so that we can easily run everything in a Docker container.
這樣我們就可以輕鬆地在 Docker 容器中運行一切。
5:18.980–5:22.220
That means that we're sandboxing everything, we're nice and secure,
這意味著我們將一切隔離在沙盒中，安全性很高，
5:22.440–5:24.540
and also makes it really easy to set up.
並且也極大地簡化了設定過程。
5:24.760–5:25.940
So follow these instructions.
請遵循這些指示。
5:25.940–5:30.120
It includes having an open router key in a .emv file.
其中包括在 .env 檔案中擁有 Open Router 金鑰。
5:30.460–5:31.632
And then when you're ready,
然後當你準備好時，
5:31.632–5:33.620
we're going to bring this up in a dev container.
我們將以開發容器啟動它。
5:33.980–5:38.880
So I'm going to do the Ctrl-Shift-P on a PC or Command-Shift-P on a Mac.
所以我將在 PC 上按 Ctrl-Shift-P，或在 Mac 上按 Command-Shift-P。
5:39.140–5:42.660
And I'm going to say dev containers reopen in container.
然後我會輸入 dev containers reopen in container。
5:42.660–5:46.300
and this is going to now launch a dev container.
這將現在啟動一個開發容器。
5:46.600–5:48.440
I'm going to close this chat window here.
我現在要關閉這個聊天視窗。
5:48.520–5:50.620
This is launching a Linux container,
這正在啟動一個 Linux 容器，
5:50.880–5:54.060
which is going to be running installing Pi right now.
它現在正在安裝 Pi。
5:54.180–5:56.320
And we now have Pi installed
現在 Pi 已經安裝好了，
5:56.320–5:58.100
and we're ready to try this out.
我們準備好要來試試看了。
5:58.360–6:00.160
And now as explained in the readme,
根據 README 中的說明，
6:00.360–6:02.720
I'm running a command that will add a skill
我正在執行一個指令，用來新增一個技能
6:02.720–6:03.940
to our coding agent.
給我們的程式碼代理程式（coding agent）。
6:03.940–6:05.980
It's a skill called agent browser
這個技能叫做 agent browser
6:05.980–6:07.600
from the great people of Vercel.
來自 Vercel 的優秀團隊。
6:07.780–6:09.500
And it's a skill which is designed
這是一個設計用來
6:09.500–6:11.840
to give any coding agent the ability
賦予任何程式碼代理程式能力
6:11.840–6:15.520
to run a Chrome-based browser behind the scenes.
在背景執行基於 Chrome 的瀏覽器的技能。
6:15.640–6:18.280
It's asking me which agent I would like to install it for.
它詢問我要為哪個代理程式安裝此技能。
6:18.500–6:21.240
I'm going to come all the way down and select Pi.
我會一直往下捲動並選擇 Pi。
6:21.480–6:22.800
Find Pi in this list.
在這個清單中找到 Pi。
6:22.920–6:23.500
There it is.
就是它。
6:23.600–6:25.660
And press the space bar and then enter.
然後按下空白鍵並輸入。
6:26.060–6:27.420
Install it for the project.
為專案安裝它。
6:27.980–6:28.980
Yes, we're done.
是的，我們完成了。
6:29.060–6:33.460
We've just installed the agent browser skill for Pi.
我們剛剛為 Pi 安裝了 agent browser 技能。
6:33.700–6:35.720
And now I'm going to run the command to start Pi.
現在我要執行指令來啟動 Pi。
6:35.720–6:39.580
As described in the readme, I just call Pi dash dash models.
根據 README 的描述，我只需呼叫 Pi 並加上 --models 參數。
6:39.580–6:44.073
and I pass in the name of the two models through OpenRouter that I want to use.
並透過 OpenRouter 傳入我想使用的兩個模型名稱。
6:44.073–6:45.000
And this is your
這是你
6:45.000–6:48.121
first sighting of Pi. This is Pi running.
第一次見到 Pi。這是正在執行的 Pi。
6:48.121–6:51.700
It's very simple and we're about to use it to try
它非常簡單，我們即將使用它來嘗試
6:51.700–6:54.381
building a sales CRM platform of our very own.
建構我們自己的銷售 CRM 平台。
6:54.381–6:56.920
The first thing I show you is that if you hold
我要先展示的是，如果你按住
6:56.920–7:00.269
down control P, you can toggle the different models and
Control P，你可以切換不同的模型，
7:00.269–7:02.380
the models show on the bottom right
模型會顯示在右下角
7:02.380–7:05.411
here. You can see it says DeepSeek V4 Pro.
這裡。你可以看到它顯示 DeepSeek V4 Pro。
7:05.411–7:08.620
I can do control P and it will flip to GLM 5.2.
我可以按 Control P，它會切換到 GLM 5.2。
7:08.620–7:11.512
These are the two models that I started it with, so
這兩個是我啟動時使用的模型，所以
7:11.512–7:13.840
we can switch between them at any point.
我們可以隨時在它們之間切換。
7:14.180–7:16.900
And we're going to use both. Of course, we're going to use both.
我們將同時使用這兩個。當然，我們會同時使用這兩個。
7:17.100–7:22.020
The other thing to show you is that Shift-Tab chooses the different reasoning levels.
另一個要展示的是，Shift-Tab 可以選擇不同的推理層級。
7:22.140–7:27.140
We can toggle High to Extra High to Thinking Off back to High and Extra High.
我們可以在 High、Extra High、Thinking Off、High 和 Extra High 之間切換。
7:27.200–7:29.080
And we're going to go on Extra High.
我們將選擇 Extra High。
7:29.260–7:32.860
We're going to have Deep Seek V4 Pro and also Control-P.
我們將使用 Deep Seek V4 Pro 並同時按 Control-P。
7:32.980–7:36.500
We're going to have GLM 5.2 on Extra High mode.
我們將以 Extra High 模式使用 GLM 5.2。
7:36.500–7:39.629
And now all I have to do wonderfully is say, look,
現在我只需要很完美地說：「看，
7:39.629–7:41.820
please build the entire project.
請建立整個專案。
7:41.820–7:44.381
As described in agents.md, don't stop until
如 agents.md 所述，直到
7:44.381–7:46.873
the success criteria are met and the server
成功標準達成且伺服器
7:46.873–7:48.880
is running and ready for me to test.
正在運行並準備好讓我測試之前，不要停止。
7:49.060–7:52.000
I can kick that off and off it goes.
我可以啟動它，然後它就開始運作了。
7:52.180–7:53.510
You can see down here,
你可以在這裡看到，
7:53.510–7:56.540
it tells you how much of the context is being used.
它會告訴你使用了多少上下文。
7:56.800–7:59.980
It's got 0.7% of the 1 million context.
它使用了 100 萬上下文中的 0.7%。
8:00.000–8:02.544
percent of the 1 million context window for glm 5.
GLM 5.2 的 100 萬上下文視窗的
8:02.544–8:04.840
2 and it will also tell you how much it's spent
百分比，它也會告訴你花費了多少
8:04.840–8:08.375
and we've already spent 0.9 of a cent uh already and
我們已經花費了 0.9 美分，呃，已經
8:08.375–8:11.320
we've just spent our first cent on this and
我們剛剛花費了第一美分在這個專案上，並且
8:11.320–8:14.960
We'll be keeping a very close watch on the spend, but it is off.
我們會密切關注花費，但它已經關閉了。
8:15.080–8:16.480
It's going to keep going.
它會繼續進行。
8:16.720–8:19.560
It's going to go through each of the different stages of build,
它會經過建構的不同階段，
8:19.700–8:22.560
and it's going to check the success criteria at every point.
並且它會在每個點檢查成功標準。
8:22.940–8:24.660
Well, it's been going away for about five minutes.
嗯，它已經運行了大約五分鐘。
8:24.800–8:25.640
Lots has been going on.
發生了很多事情。
8:25.680–8:27.380
We've now spent 10 cents.
我們現在已經花費了 10 美分。
8:27.660–8:29.980
But meanwhile, I'm going to do something cool.
但與此同時，我要做一些很酷的事情。
8:30.300–8:34.280
I've got another VS Code window running over here.
我在這裡有另一個 VS Code 視窗在運行。
8:34.600–8:35.660
But this is, in fact,
但這實際上，
8:35.660–8:38.340
another dev container on exactly the same project.
是完全相同專案的另一個開發容器。
8:38.460–8:41.300
I've cloned it separately to a different directory.
我已將其單獨複製到不同的目錄。
8:41.320–8:43.751
It's called now CRM DeepSeek,
它現在被稱為 CRM DeepSeek，
8:43.751–8:46.960
which is a clue as to what I'm about to do.
這是我即將要做的事情的線索。
8:46.960–8:47.520
I am, of course,
我當然，
8:47.520–8:49.500
I'm going to make the screen a little bit bigger for you.
我會把畫面稍微放大一點給你看。
8:49.900–8:54.960
I am, of course, now set to use DeepSeek instead of GLM 5.2.
我當然，現在設定為使用 DeepSeek 而不是 GLM 5.2。
8:55.120–8:57.560
And I'm going to use the same exact command.
我將使用完全相同的指令。
8:57.840–8:59.620
Please build the entire project.
請建立整個專案。
9:00.020–9:05.380
So I'm now going to set off in this other screen the same project in parallel.
所以現在我將在另一個螢幕中平行啟動相同的專案。
9:05.920–9:09.269
Off it goes with DeepSeek V4 Pro running, also
它開始運行 DeepSeek V4 Pro，也
9:09.269–9:11.120
on extra high reasoning.
處於 Extra High 推理模式。
9:11.320–9:14.414
Look at it go. And if I just shrink this window,
看它運行。如果我只是縮小這個視窗，
9:14.414–9:18.160
you'll see that the other one, GLM 5.2, is also on the go.
你會看到另一個，GLM 5.2，也在運行。
9:18.280–9:20.931
I have both coding agents, both using Pi,
我有兩個編碼代理，都使用 Pi，
9:20.931–9:23.427
both models on the go at the same time, b
兩個模型同時運行，構
9:23.427–9:26.000
uilding out the same project so we can se
建立相同的專案，以便我們觀察
9:26.000–9:28.417
e how they do, both in terms of cost and
它們在成本和
9:28.417–9:32.160
speed, and also, of course, in terms of the final product.
速度方面的表現，當然還有最終產品的品質。
9:32.160–9:34.229
And this is now one second later for you, but
這現在對你來說是一秒後的畫面，但
9:34.229–9:36.241
it's like an hour and a half later for me, an
對我來說已經過了一個半小時，兩
9:36.241–9:37.620
d both of them have finished.
個專案都已經完成了。
9:37.620–9:41.012
And wonderfully, you can go into OpenRouter and
而且很棒的是，你可以進入 OpenRouter
9:41.012–9:44.150
look at the logs and see the messages going to
查看日誌，並看到發送到
9:44.150–9:46.100
GLM 5.2 and DeepSeek V4 Pro.
GLM 5.2 和 DeepSeek V4 Pro 的訊息。
9:46.340–9:49.511
And you can also look at things like the activity charts and
你也可以查看活動圖表等內容，
9:49.511–9:51.840
see the usage and explore what was going on.
了解使用情況並探索當時發生了什麼。
9:52.180–9:54.389
We look at the past three hours and
我們查看過去三個小時的記錄，
9:54.389–9:56.676
we can see everything that was happ
可以看到所有正在發生的
9:56.676–10:00.620
ening and use that to figure out how much did we spend on each.
事情，並用這些資訊來計算我們每個專案花費了多少。
10:01.580–10:05.700
And this page in the OpenRouter activity screen shows you everything that went on.
OpenRouter 活動畫面中的這個頁面顯示了所有發生的事情。
10:05.700–10:08.071
You can see the activity there. In the end,
你可以在那裡看到活動記錄。最後，
10:08.071–10:11.120
if I back out things when I ran something a second time,
如果我第二次執行時排除某些內容，
10:11.520–10:16.362
for GLM 5.2, it cost $4.15 to run this. $4.15, and
對於 GLM 5.2，這次執行的成本是 4.15 美元。4.15 美元，
10:16.362–10:20.720
it took an hour and a quarter. For DeepSeek,
並且耗時一個小時一刻。對於 DeepSeek，
10:20.760–10:24.091
it was cheaper. It was $2.56 in total, and
它更便宜。總共是 2.56 美元，
10:24.091–10:28.500
it was a little bit faster. It took just about one hour,
而且速度稍快。它只花了一個小時，
10:28.800–10:31.719
a bit faster as well. So DeepSeek was faster and
速度也稍快一些。所以 DeepSeek 更快且
10:31.719–10:34.040
cheaper. What were the results like?
更便宜。結果如何呢？
10:34.040–10:37.557
And here is DeepSeek's implementation of a personal CRM.
這裡是 DeepSeek 實現的個人 CRM。
10:37.557–10:40.928
Your personal CRM, like a sales force. You've got a dash
你的個人 CRM，類似於銷售團隊。你有一個儀
10:40.928–10:44.299
board with deals one and revenue. It's got some dummy da
表板，顯示交易一和收入。這裡有一些虛構數
10:44.299–10:47.596
ta in here to show it off. Something is overdue here. Yo
據來展示功能。這裡有一項已逾期。你
10:47.596–10:51.113
u've got organizations, the companies that you have with
可以看到組織，即你與之合作的
10:51.113–10:54.484
an edit and delete, a Dell button, that's all right. Con
公司，有編輯和刪除功能，刪除按鈕沒問題。
10:54.484–10:57.928
tacts and deals listed out here. Contacts, the people yo
這裡列出了聯絡人和交易。聯絡人是你
10:57.928–10:59.540
u know at these companies.
在這些公司認識的人。
10:59.540–11:05.440
deals, which are the deals that are at various stages of the pipeline shown in the stage here.
交易則是處於管道各階段的交易，如這裡所示。
11:05.440–11:13.300
And the pipeline itself is a Kanban style board of deals moving from new all the way through to
管道本身是一個看板風格的交易板，交易從新客戶一直推進到
11:13.300–11:17.402
one or lost. And you can drag them across to see what happens when
成交或丟失。你可以拖曳它們來查看發生什麼情況。
11:17.402–11:19.260
you do that. And you should be
你應該能夠新增、新增並進行搜尋。
11:19.260–11:21.887
able to add and add and search throughout.
讓我們在這裡搜尋某人。找到了。
11:21.887–11:24.740
Let's search for someone in here. There we go.
搜尋功能也正常運作。這是一個功能完整的平台。
11:24.740–11:28.377
The search works too. It is a functioning platform.
它由 SQLite 資料庫支援。
11:28.377–11:30.660
It's backed by SQLite database.
這是一個運作中但簡單的銷售 CRM。而
11:31.340–11:34.255
This is a working but simple sales CRM. And
你必須承認，這相當令人印象深刻。
11:34.255–11:37.420
you've got to admit, it's pretty impressive.
你不得不承認，它相當令人印象深刻。
11:37.780–11:40.328
But you know what's more impressive? GLM 5.
但你知道什麼更令人印象深刻嗎？GLM 5.2 的版本。
11:40.328–11:43.300
2's version is more impressive. Let's switch over
讓我們切換過去看看。
11:43.300–11:45.845
to it. This is the dashboard. And come on,
這是儀表板。說真的，
11:45.845–11:48.540
immediately you see this is better. This is
你馬上就會發現這個更好。這個
11:48.540–11:52.291
better. We've got these kind of hot metrics being surfaced at the top.
更好。我們在頂部看到了這些類型的熱門指標。
11:52.291–11:53.520
We've got deals one and
我們有交易一和
11:53.520–11:58.144
revenue, which just is like a better form chart with more rich data.
營收，這就像是一個擁有更豐富數據的更好的圖表。
11:58.144–11:59.960
We've got these tasks here
我們這裡有這些
11:59.960–12:02.912
that could be checked off. We've got recent activity and
可以勾選的任務。我們有最近的活動和
12:02.912–12:04.420
then pipeline by stage at the
底部的管道階段以及
12:04.420–12:07.211
bottom and an open board link there. It's really nice.
這裡的開放看板連結。真的很不錯。
12:07.211–12:09.320
We've got these icons here. We can switch
我們這裡有這些圖標。我們可以切換
12:09.320–12:11.535
to organizations. This looks similar, but
到組織。這看起來相似，但是
12:11.535–12:14.180
of course we've got proper icons, not a Dell button.
當然我們有正確的圖標，而不是戴爾按鈕。
12:14.180–12:19.545
We've got contacts as before, deals here with this really nice badges.
我們有聯絡人，這裡有交易，還有這些非常漂亮的徽章。
12:19.545–12:21.000
And then this is the
然後這是
12:21.000–12:22.688
pipeline. It's just sharper.
管道。它更銳利。
12:22.688–12:26.400
Let's go back over here again to the pipeline. Compare this is a bit
讓我們回到這裡的管道。比較一下，這個有點
12:26.400–12:30.186
more bland. This is really gorgeous, a really nice,
平淡。這個真的很漂亮，一個非常漂亮、
12:30.186–12:33.620
sharp pipeline. And of course, we will still no
銳利的管道。當然，我們仍然無疑擁有拖放功能。
12:33.620–12:35.675
doubt have the ability to drag and drop.
所有功能都使用相同的階段，以及贏和輸。
12:35.675–12:38.540
Everything works with the same stages and won and lost.
這是一個相當令人印象深刻的畫面，
12:38.920–12:41.608
This is a pretty impressive screen,
來自 GLM 5.2 的非常棒的产品。但我知道你在想什麼。
12:41.608–12:45.820
a really nice product from GLM 5.2. But I know what you're
我這麼了解你。你在想，好吧，
12:45.820–12:49.475
thinking. I know you so well. You're thinking, okay,
Ed，我們明白了。這很令人印象深刻。GLM 5.2
12:49.475–12:52.300
Ed, we get it. This is impressive. GLM 5.2
做得還不錯，但是 Opus 在同樣的事情上會表現如何呢？
12:52.300–12:57.175
did a decent job, but how would Opus have done on the same thing?
Opus 4.8，地球上最強大的
12:57.175–12:59.800
How would Opus 4.8, the strongest
模型，至少是我們所有人都能使用的，
12:59.800–13:02.516
model on the planet, at least that's available to us all,
它會如何處理這個？如果它處於超代碼模式，
13:02.516–13:04.480
how would that do with this? And how would
並被允許發揮其絕對最佳水平，它會表現如何？
13:04.480–13:06.859
it do if it was on ultra code mode and
當然，我也試過了。它正在這裡的標籤上運行。
13:06.859–13:10.960
allowed to do its absolute best? Well, of course, I've given
它花了一個小時，如果我直接使用 API 而不是使用我的 Claude AI 帳戶，
13:10.960–13:14.993
that a try too. And it's running on this tab right here.
這將花費 21.27 美元。
13:14.993–13:18.220
It took an hour and it would have cost $21.27
讓我們過去看看。
13:18.220–13:23.321
if I'd used API directly rather than just using a usage of my Claude AI account.
我們開始吧。這是儀表板。
13:23.321–13:24.420
And let's go over
你可以看到它非常相似。它有一個圖表。它有
13:24.420–13:26.985
to it. Here we go. This is the dashboard.
同樣列出的數字類型和這裡的檢查清單，
13:26.985–13:30.560
You can see it's quite similar. It has a chart. It has the
你可以看到它非常相似。它有一個圖表。它有
13:30.560–13:33.891
same called out kinds of numbers and a checklist here,
相同列出的數值，這裡還有一個檢查清單，
13:33.891–13:36.260
but it is a bit nicer, right? Look at the
但看起來好一點，對吧？看看
13:36.260–13:40.496
dark background here it looks a little bit more professional but
這裡的深色背景，看起來更專業一點，但
13:40.496–13:42.300
only slightly only by a hair
只有一點點，僅微乎其微
13:42.300–13:47.800
and clicking through the tables you'll see a very similar construct clicking through them here it's
點擊瀏覽表格時，你會看到非常相似的結構，在這裡點擊瀏覽它們時，它
13:47.800–13:50.118
just slightly better perhaps and
只是稍微好一點，也許
13:50.118–13:54.920
the pipeline here we have another nice pipeline here is a kanban board
這裡的管道，我們有另一個很棒的管道，這裡是一個看板
13:54.920–13:56.793
but check this out when i drag and
但看看這個，當我拖放
13:56.793–14:00.400
drop again you get this dotted line thing it's just a little bit
再次拖放時，你會看到這個虛線的東西，只是稍微
14:00.400–14:02.884
better. And so in summary, GLM 5.
好一點。所以總結來說，GLM 5.
14:02.884–14:07.760
2 was neither the cheapest nor the best, but it was quite close on
2 既不是最便宜的，也不是最好的，但在
14:07.760–14:11.098
both fronts. It was low cost and it was really good.
兩個方面都非常接近。它成本低，而且真的很好。
14:11.098–14:13.800
And so I would say, yes, this is a credible
所以我想說，是的，這是一個可信的
14:13.800–14:17.894
alternative to frontier models that can be used for your own projects.
替代前沿模型的方案，可以用於您自己的專案。
14:17.894–14:19.560
It's not super cheap. You're
它不是超級便宜。你
14:19.560–14:22.111
still spending north of $4 for a product like this,
仍然要為這樣的產品支付超過 4 美元的費用，
14:22.111–14:24.420
but there's a lot of capability here. I didn't
但這裡有很多功能。我甚至
14:24.420–14:26.788
even show you. You can click into organizations and
甚至展示給你。你可以點擊進入組織，並
14:26.788–14:28.660
edit them and then click here to go over to
編輯它們，然後點擊這裡前往
14:28.660–14:32.268
the contact screen, edit a log activity here in the pipeline.
聯絡畫面，編輯管道中的日誌活動。
14:32.268–14:34.320
You can click on cards to bring them
你可以點擊卡片來
14:34.320–14:37.153
up. There's tons of functionality and
顯示它們。這裡有大量的功能，並且
14:37.153–14:40.960
it built all of this first time one shot and it met the
它第一次就構建了所有這些，一次完成，並且符合
14:40.960–14:43.219
success criteria and worked right away. And
成功標準，立即生效。並且
14:43.219–14:45.600
if you'd like to give this a try yourself and use
如果你想自己嘗試這個並使用
14:45.600–14:48.089
Pi, there are details in the description below and
Pi，下方的描述中有詳細資訊，並且
14:48.089–14:50.340
I'd love to see what you come up with. And if you
我很想看看你會想出什麼。如果你
14:50.340–14:52.523
enjoyed this video, then please do like and subscribe.
喜歡這個影片，請按讚並訂閱。
14:52.523–14:53.900
That's the way I know you're really
那是我知道你真正
14:53.900–14:57.685
there. And it's the way the YouTube algorithm knows you're really there as well.
在那裡的方式。這也是 YouTube 演算法知道你真的在那裡的方式。
14:57.685–14:58.420
And either way,
無論哪種方式，
14:58.420–15:00.340
I hope to see you very soon for another video.
我希望很快再見，為你帶來另一個影片。

影片筆記：Would You Still Pay for Claude After Seeing This? (GLM 5.2 vs DeepSeek V4)

一句話總結

影片透過開源編碼代理程式 Pi 實測建構個人 CRM 系統，對比 GLM 5.2 與 DeepSeq v4 Pro（講者口語稱 DeepSeek）在成本、速度與產出品質上的差異，並進一步評估 GLM 5.2 與 Opus 4.8 的視覺與功能表現，結論認為 GLM 5.2 是具備高功能性且成本可控的前線模型可信替代品。

核心重點

  • 開源模型已具備實戰能力：講者確認開源模型配合編碼代理程式（Coding Agents）已能建構真實世界專案。
  • 工具架構：使用開源代理程式 Pi，搭配 GLM 5.2 (來自 ZAI/z.ai) 與 DeepSeq v4 Pro，透過 OpenRouter 調用。
  • 專案管理方法：採用 agents.md 檔案將專案拆解為多個階段（Phases），並設定明確的「成功標準」（Success Criteria），代理程式必須滿足所有標準才能宣告完成。
  • GLM 5.2 vs DeepSeq v4 Pro 對比：
  • DeepSeq v4 Pro：成本較低 ($2.56)、速度較快 (約 1 小時)，產出基礎功能完整的 CRM (Dashboard, Deals, Contacts, Kanban Board, SQLite 後端)。
  • GLM 5.2：成本較高 ($4.15)、速度較慢 (1 小時 15 分)，但講者認為其生成的介面與功能細節（數據視覺化、任務檢查清單、活動紀錄）更為出色。
  • GLM 5.2 vs Opus 4.8 對比：
  • Opus 4.8 (Ultra Code Mode)：耗時一小時，成本極高 ($21.27)，視覺效果僅比 GLM 5.2 略好一點點（如深色背景、拖放虛線提示）。
  • GLM 5.2：被定位為前線模型（frontier models）的可信替代品，具備高能力，雖非最便宜也非最佳，但在兩者之間非常接近，且能一次生成成功（first time one shot）。

詳細大綱

  1. 開源模型與編碼代理程式的現狀
  • 確認開源模型已可用於建構真實世界專案。
  • 介紹工具組合：Pi (代理程式)、GLM 5.2、DeepSeq v4 Pro。
  1. 模型評估指標：智能與成本
  • 引用 artificialanalysis.ai 網站的「智能 vs 成本」圖表。
  • 目標是尋找位於左上角綠色區域（高智能、低成本）的模型。
  • 選擇 GLM 5.2 與 DeepSeq v4 Pro 作為邊界測試對象。
  1. 編碼代理程式 Pi 的特性
  • 設計理念：輕量、最小化、不綁定特定 LLM、允許插入任何開源模型。
  • 預設功能簡單，但具備高度可擴展性。
  1. 專案建構策略：階段性與成功標準
  • 使用 agents.md 定義專案需求。
  • 將專案拆解為 bite-sized phases (小階段)。
  • 每個階段需包含業務描述與「成功標準」（Success Criteria）。
  • 代理程式必須通過所有成功標準才能繼續或宣告完成。
  • 需提供測試技能（如瀏覽器操作）以形成反饋迴圈。
  1. 環境設置與執行準備
  • 使用 VS Code 的 Dev Containers 建立沙盒環境。
  • 安裝 agent browser 技能（來自 Vercel），賦予代理程式運行 Chrome 瀏覽器的能力。
  • 透過 pi --models 命令啟動 Pi，並指定 GLM 5.2 與 DeepSeq v4 Pro。
  • 設定推理層級（Reasoning Levels）為 Extra High。
  1. 平行實測：GLM 5.2 vs DeepSeq v4 Pro
  • 同時在兩個 Dev Containers 中執行相同的建構指令。
  • 監控 OpenRouter 的日誌、活動圖表與成本消耗。
  1. 實測結果初步分析
  • 成本與速度：DeepSeq v4 Pro 較便宜 ($2.56) 且較快 (約 1 小時)；GLM 5.2 較貴 ($4.15) 且較慢 (1 小時 15 分)。
  • 產出品質：
  • DeepSeq v4 Pro 產出基礎功能完整的 CRM。
  • GLM 5.2 產出被講者認為「更令人印象深刻」的版本，具備更豐富的數據視覺化、任務檢查清單及活動紀錄。
  1. GLM 5.2 生成結果評估
  • 介面特徵：擁有 proper icons、contacts、deals 及帶有 badges 的 pipeline。
  • 視覺評價：被形容為「gorgeous」、「sharp」，比另一個較單調的版本更好。
  • 功能完整性：支援拖放（drag and drop），包含 stages、won 和 lost 的狀態。
  • 成本與效率：成本相對較低，但並非最低。
  1. Opus 4.8 生成結果評估
  • 測試條件：使用 Opus 4.8（被稱為目前地球上最強模型），並啟用 Ultra Code Mode。
  • 成本與耗時：耗時一小時，若直接透過 API 使用需花費 21.27 美元（目前使用 Claude AI 帳戶）。
  • 視覺與功能對比：
  • Dashboard：有圖表、相同數值、清單，深色背景看起來稍顯專業，但僅好「一點點」（by a hair）。
  • Pipeline：同樣是 Kanban board，拖放時出現虛線（dotted line thing），被認為「稍微好一點」。
  • 結論：Opus 4.8 的結果僅在細微處優於 GLM 5.2。
  1. 最終總結與建議
  • GLM 5.2 定位：既非最便宜也非最佳，但在兩者之間非常接近。
  • 價值主張：具備高能力（capability），是前線模型的可信替代品（credible alternative），可用於個人專案。
  • 成本效益：單次產品生成成本仍超過 4 美元，但功能豐富。
  • 功能細節：支援編輯組織、聯絡人、日誌活動、點擊卡片等，所有功能首次生成即成功（first time one shot）。

工具 / 模型 / 名詞整理

  • 編碼代理程式 (Coding Agents):
  • Pi (講者強調其為開源、輕量、可擴展的代理程式)
  • Claude Code (作為對比參考)
  • Open Code (講者提及 Pi 類似於此類代理程式)
  • 大型語言模型 (LLMs):
  • GLM 5.2 (來自 ZAI / z.ai)
  • DeepSeq v4 Pro (講者口語提及 DeepSeek，但筆稿為 DeepSeq)
  • Opus 4.8 (被稱為目前地球上最強模型)
  • Claude (用於後續比較)
  • 平台與服務:
  • OpenRouter (用於調用模型、查看日誌與成本)
  • artificialanalysis.ai (提供模型智能與成本數據的網站)
  • Claude AI (帳戶)
  • YouTube
  • 開發工具與環境:
  • VS Code (Visual Studio Code)
  • Dev Containers (開發容器，用於沙盒化與安全執行)
  • Docker (容器技術)
  • Agent Browser (由 Vercel 提供的技能，用於讓代理程式運行瀏覽器)
  • 專案相關:
  • agents.md (專案指令與成功標準檔案)
  • readme (環境設置說明)
  • .env (存放 OpenRouter Key)
  • SQLite (DeepSeq 版本使用的資料庫)
  • Salesforce, Pipe Drive (講者提及的商業 CRM 產品作為參考)
  • Kanban board
  • Ultra Code Mode
  • API

操作流程整理

  1. 準備階段：
  • 參考 artificialanalysis.ai 選擇高智能低成本的模型（GLM 5.2 與 DeepSeq v4 Pro）。
  • 在 VS Code 中使用 Dev Containers 建立沙盒環境。
  • 安裝 agent browser 技能（來自 Vercel）。
  • 設定 .env 存放 OpenRouter Key。
  1. 專案定義：
  • 撰寫 agents.md 檔案，將專案拆解為 bite-sized phases。
  • 為每個階段設定業務描述與「成功標準」（Success Criteria）。
  1. 執行代理程式：
  • 透過 pi --models 命令啟動 Pi。
  • 指定模型為 GLM 5.2 與 DeepSeq v4 Pro。
  • 設定推理層級（Reasoning Levels）為 Extra High。
  • 在兩個 Dev Containers 中平行執行相同的建構指令。
  1. 監控與評估：
  • 監控 OpenRouter 的日誌、活動圖表與成本消耗。
  • 比較產出結果：
  • DeepSeq v4 Pro：基礎功能完整，成本 $2.56，耗時約 1 小時。
  • GLM 5.2：介面與細節更出色，成本 $4.15，耗時 1 小時 15 分。
  • 後續測試 Opus 4.8 (Ultra Code Mode)：耗時一小時，成本 $21.27，視覺效果僅略優於 GLM 5.2。

值得注意的限制或風險

  • 成本問題：即使使用開源模型，單次產品生成成本仍超過 4 美元（GLM 5.2）甚至高達 21.27 美元（Opus 4.8），對於個人專案雖可接受，但並非最低成本方案。
  • 速度限制：生成完整專案耗時約 1 至 1.25 小時，速度相對較慢。
  • 模型名稱不確定性：影片中出現的模型名稱（如 GLM 5.2, DeepSeq v4 Pro, Opus 4.8）需查證其準確性與對應關係。
  • 代理程式 Pi 的語境不明：在部分段落中提及 "use Pi"，但未明確解釋 Pi 為何物，可能存在特定工具或口誤的情況。

逐字稿辨識疑點

  • DeepSeq v4 / DeepSeek：逐字稿中多次出現「DeepSeq v4」及「DeepSeq v4 Pro」，但在講者口語部分提到「DeepSeek」。由於要求不得自行更正，故保留「DeepSeq」作為模型名稱，並標註講者口語提及「DeepSeek」為疑點。
  • ZAI / z.ai：逐字稿中模型 GLM 5.2 的來源分別被稱為「ZAI」與「z.ai」。
  • agents.md：講者口語中曾出現「Py」，但檔案名稱明確為 agents.md，且代理程式名稱為 Pi，此處「Py」可能為口誤或聽寫錯誤。
  • Ctrl-Shift-P / Command-Shift-P：講者提及在 PC 上按 Ctrl-Shift-P，Mac 上按 Command-Shift-P 來開啟 Dev Containers，此為 VS Code 標準快捷鍵，但講者口語描述較為冗長。
  • 0.7% of the 1 million context：講者提及 GLM 5.2 使用了 100 萬上下文視窗的 0.7%，此數據具體但需查證該模型是否確實支援 1M 上下文。
  • Dell button：講者在描述 DeepSeq 產出的 CRM 介面時，提到「edit and delete, a Dell button」，「Dell」疑似為「Delete」的聽寫錯誤或口誤。
  • Hot metrics：講者描述 GLM 5.2 產出的介面時，提到「hot metrics」，此詞彙在 UI/UX 描述中較不常見，可能為「Key metrics」或「Hotspot metrics」的口誤，但保留原樣。
  • GLM 5.2：逐字稿中多次出現此名稱，需查證是否為特定模型版本或口誤（常見模型為 GPT 或 Gemini 等，但嚴格遵循逐字稿保留）。
  • Opus 4.8：需查證此模型名稱及版本號是否準確。
  • Pi：在 "use Pi" 一詞中出現，語境不明，需查證是否為特定工具、平台或口誤。
  • called out kinds of numbers：語意稍顯模糊，需查證是否為特定術語或口誤。
  • dotted line thing：描述拖放時的視覺效果，用詞較為口語化，需查證是否為特定 UI 元件名稱。

可延伸追問

  • GLM 5.2、DeepSeq v4 Pro 與 Opus 4.8 的準確模型對應關係為何？
  • 代理程式 "Pi" 的具體技術架構與開源連結為何？
  • agents.md 的成功標準（Success Criteria）具體範例為何？
  • 如何進一步降低使用開源模型建構專案的成本與時間？
  • Agent Browser 技能在反饋迴圈中的具體運作機制為何？

生字列表

生字讀音類型中文
coding agentこーでぃんぐえいじぇんと程式碼代理程式
open sourceおーぷんそーす形容動詞開源的
groundbreakingぐらうんどぶれいく形容動詞突破性的
dev containersでぶこんてな名詞開發容器
artificialanalysis.aiあーてぃふぃしゃるあなりしす.えーあい名詞Artificial Analysis AI（網站名稱）
x-axisえっくすあきす名詞橫軸
y-axisわいあきす名詞縱軸
quadrantくおらんと名詞象限
frontierふろんてぃあ名詞邊界；前沿
out of the boxあうとおぶざぼっくす副詞開箱即用；預設
promptぷろむと名詞提示；提示詞
context managementこんてくすとまねじめんと名詞上下文管理
sandboxingさんどぼっくしんぐ名詞沙盒隔離
reasoning levelsりーずにんぐれべるず名詞推理層級
success criteriaさっくすくたりあ名詞成功標準
feedback loopふぃーどばっくるーぷ名詞回饋迴圈

生字解說

coding agent こーでぃんぐえいじぇんと

· N1

意思：程式碼代理程式

解說：指能自動編寫、修改或除錯程式碼的 AI 工具或系統。

影片原句
Are we ready yet to use open source models to build real world projects?
我們是否已經準備好使用開源模型來建構真實世界的專案？
延伸例句
このコーディングエージェントは、バグを自動で見つけて修正してくれます。
這個程式碼代理程式能自動發現並修正錯誤。

open source おーぷんそーす

形容動詞 · N1

意思：開源的

解說：指軟體的原始碼公開，允許任何人檢視、修改和分發。

影片原句
In just about 10 minutes time, we're going to be looking at real world products like this and this built by coding agents using open source models.
只需大約 10 分鐘，我們將檢視真實世界的產品，像是這個和這個由程式碼代理程式使用開源模型所建構。
延伸例句
オープンソースのソフトウェアは、コミュニティによって継続的に改善されます。
開源軟體會由社群持續改善。

groundbreaking ぐらうんどぶれいく

形容動詞 · N1

意思：突破性的

解說：指具有創新性、能開創新領域或帶來重大改變的。

影片原句
We're going to be using GLM 5.2, the latest groundbreaking model from ZAI.
我們將使用 GLM 5.2，來自 ZAI 的最新突破性模型。
延伸例句
彼女は画期的な研究論文を発表しました。
她發表了一篇突破性的研究論文。

dev containers でぶこんてな

名詞 · N1

意思：開發容器

解說：一種將開發環境（如依賴庫、設定）打包在容器中的技術，確保環境一致性。

影片原句
And we're going to be using dev containers for extra safety.
我們將使用開發容器以獲得額外的安全性。
延伸例句
Dev containersを使うと、環境構築の手間が省けます。
使用開發容器可以節省環境設定的時間。

artificialanalysis.ai あーてぃふぃしゃるあなりしす.えーあい

名詞 · N1

意思：Artificial Analysis AI（網站名稱）

解說：一個提供 AI 模型性能與成本數據分析的網站資源。

影片原句
It's artificialanalysis.ai with tons of information, a lot of data about different models.
它是 artificialanalysis.ai，擁有大量資訊，關於不同模型的大量數據。
延伸例句
Artificialanalysis.aiは、AIモデルの比較に役立ちます。
Artificialanalysis.ai 有助於比較 AI 模型。

x-axis えっくすあきす

名詞 · N1

意思：橫軸

解說：座標系中水平的軸，通常表示自變量。

影片原句
The x-axis is how much does it cost to run the model for a task.
橫軸是執行模型任務的成本。
延伸例句
グラフのx軸は時間を表しています。
圖表的橫軸代表時間。

y-axis わいあきす

名詞 · N1

意思：縱軸

解說：座標系中垂直的軸，通常表示因變量。

影片原句
the y-axis is how intelligent is the model.
縱軸是模型的智慧程度。
延伸例句
y軸は売上高を示しています。
縱軸顯示銷售額。

quadrant くおらんと

名詞 · N1

意思：象限

解說：座標系中被軸線劃分的四個區域之一。

影片原句
looking to fit in that green quadrant on the top left.
你希望進入左上角的綠色區域。
延伸例句
ビジネスの戦略は、右上の象限に位置しています。
商業策略位於右上角的象限。

frontier ふろんてぃあ

名詞 · N1

意思：邊界；前沿

解說：指知識或技術發展的最前線或極限。

影片原句
They're kind of at this frontier of being smart but cheap.
它們處於智慧但便宜的邊界。
延伸例句
科学のフロンティアを探求しています。
探索科學的前沿。

out of the box あうとおぶざぼっくす

副詞 · N1

意思：開箱即用；預設

解說：指產品或軟體在未經額外配置或修改的情況下即可使用。

影片原句
But what you get out of the box is very simple.
但是，你开箱即用的東西非常簡單。
延伸例句
このソフトウェアは、箱から出してすぐに使えます。
這套軟體開箱即用。

prompt ぷろむと

名詞 · N1

意思：提示；提示詞

解說：指輸入給 AI 模型的指令或問題，以引導其生成回應。

影片原句
And it just has a simple prompt and a simple approach to context management.
它只有一個簡單的提示和簡單的上下文管理方法。
延伸例句
効果的なプロンプトを書くことが重要です。
撰寫有效的提示詞很重要。

context management こんてくすとまねじめんと

名詞 · N1

意思：上下文管理

解說：指在對話或程式執行過程中，維護和管理相關資訊（如歷史記錄、變量）的機制。

影片原句
And it just has a simple prompt and a simple approach to context management.
它只有一個簡單的提示和簡單的上下文管理方法。
延伸例句
コンテキスト管理は、長期記憶を持つAIにとって重要です。
上下文管理對於具有長期記憶的 AI 來說很重要。

sandboxing さんどぼっくしんぐ

名詞 · N1

意思：沙盒隔離

解說：指將程式或程式碼限制在一個隔離的環境中運行，以防止對系統其他部分造成影響。

影片原句
That means that we're sandboxing everything, we're nice and secure,
這意味著我們將一切隔離在沙盒中，安全性很高，
延伸例句
サンドボックス環境でテストを行うことで、セキュリティリスクを減らせます。
在沙盒環境中進行測試可以降低安全風險。

reasoning levels りーずにんぐれべるず

名詞 · N1

意思：推理層級

解說：指 AI 模型在生成回應時進行邏輯思考和推演的深度或強度設定。

影片原句
The other thing to show you is that Shift-Tab chooses the different reasoning levels.
另一個要展示的是，Shift-Tab 可以選擇不同的推理層級。
延伸例句
高度な推論レベルを選択すると、より正確な回答が得られます。
選擇高級推理層級會獲得更準確的回答。

success criteria さっくすくたりあ

名詞 · N1

意思：成功標準

解說：指判斷專案或任務是否完成並達到預期目標的具體條件。

影片原句
we give success criteria, ways that the coding agent can verify that it has passed this phase,
我們提供成功標準，讓程式碼生成代理（coding agent）可以驗證它是否已通過此階段，
延伸例句
プロジェクトの成功基準を明確に定義する必要があります。
必須明確定義專案的成功標準。

feedback loop ふぃーどばっくるーぷ

名詞 · N1

意思：回饋迴圈

解說：指系統將輸出結果反饋回輸入端，以便進行調整或改進的循環過程。

影片原句
That way it's got the feedback loop and it can keep iterating until the success criteria are met.
這樣它就有了回饋迴圈，並且可以持續迭代，直到滿足成功標準。
延伸例句
フィードバックループを通じて、製品を改善していきます。
透過回饋迴圈來改善產品。

句型解說（含實例）

～之一

意思：……之一

接續：名詞 + の + 一つ / 一つ + の + 名詞

解說：表示所提及的事物是某個群體或類別中的一員。

影片原句
It's one of the most common questions I get about coding agents.
這是關於程式碼代理程式我最常收到的問題之一。
實例
  1. これは私が持っている本の一つです。
    這是我擁有的書之一。
  2. 彼はクラスで最も優秀な学生の一人です。
    他是班上最優秀的學生之一。

～ていく / ～てくる

意思：（動作）逐漸進行 / （動作）從過去持續到現在或向未來發展

接續：動詞て形 + いく / くる

解說：表示動作或狀態隨時間推移而持續、發展或變化。

影片原句
And we'll be comparing it to what Claude can do with the same brief.
我們將把它與 Claude 在相同簡報下能做的事進行比較。
實例
  1. このプロジェクトは順調に進んでいきます。
    這個專案正順利進行著。
  2. 最近、健康意識が高まってきています。
    最近，健康意識逐漸提高。

～ようにする

意思：努力使……；確保……

接續：動詞辞書形 / ない形 + ようにする

解說：表示為了達到某種狀態或結果而採取行動或保持某種習慣。

影片原句
But the most important thing we do that is absolutely mandatory when working with open source models is this.
但我們所做且在使用開源模型時絕對必要的事情是這個。
實例
  1. 毎日運動するようにしています。
    我努力每天運動。
  2. 遅刻しないように気をつけてください。
    請注意不要遲到。

～ために

意思：為了……

接續：動詞辞書形 / 名詞 + の + ために

解說：表示目的或原因。

影片原句
And we're going to be using dev containers for extra safety.
我們將使用開發容器以獲得額外的安全性。
實例
  1. 留学のために日本語を勉強しています。
    為了留學而學習日語。
  2. 健康のために野菜を食べるようにしています。
    為了健康而吃蔬菜。

～ておく

意思：事先……；提前準備……

接續：動詞て形 + おく

解說：表示為了未來的某個目的而提前完成某個動作。

影片原句
And I'm going to link to it in the description.
我將在描述中連結到它。
實例
  1. 事前に準備しておいてください。
    請事先準備好。
  2. チケットを買っておきました。
    我已經提前買了票。

～てしまう

意思：完全……；……完了（帶有遺憾或強調語氣）

接續：動詞て形 + しまう

解說：表示動作完成，有時帶有遺憾、驚訝或強調的語氣。

影片原句
And we've already spent 0.9 of a cent uh already and we've just spent our first cent on this and
我們已經花費了 0.9 美分，呃，已經我們剛剛花費了第一美分在這個專案上，並且
實例
  1. 時間を無駄にしてしまった。
    浪費了時間。
  2. 全部食べてしまった。
    全部吃光了。

～てみる

意思：試著……

接續：動詞て形 + みる

解說：表示嘗試做某事以確認結果或效果。

影片原句
today we're going to try out two models that are kind of on the rim of this line here.
今天我們將嘗試兩種位於這條線邊緣的模型。
實例
  1. この料理を食べてみてください。
    請試著吃這個料理。
  2. 一度日本に行ってみてください。
    請試著去一次日本。

～てくれる

意思：（他人）為（我）做……

接續：動詞て形 + くれる

解說：表示他人為說話者或說話者一方的人做某事，帶有感激或受益的語氣。

影片原句
And it's a skill which is designed to give any coding agent the ability to run a Chrome-based browser behind the scenes.
這是一個設計用來賦予任何程式碼代理程式能力在背景執行基於 Chrome 的瀏覽器的技能。
實例
  1. 友達が手伝ってくれました。
    朋友幫了我。
  2. 先生が教えてくれました。
    老師教了我。