0:00.000–0:00.634
shooting out.
0:00.644–0:05.632
I believe that was something I've seen .
0:05.642–0:08.524
The trees are quite reminiscent of a pool.
0:08.534–0:12.562
DeepSeek V4 Pro has just been released in general availability version
0:12.572–0:16.604
, which also happened with the V4 Flash model back about 12 or
0:16.614–0:19.161
13 days ago on 07/31.
0:19.171–0:23.078
Now, that model when tested proved to be a significant
0:23.088–0:26.516
increase in capability over the preview version that was previously
0:26.526–0:27.154
around .
0:27.164–0:29.793
And today we're going to be finding out if the same can be said
0:29.803–0:34.292
for the newly updated checkpoint general availability for V4 Pro
0:34.302–0:35.870
dubbed 0813.
0:35.880–0:38.426
So before we get into it , do feel free to subscribe so we can
0:38.436–0:40.824
hit 70K and then 100K with the plaque .
0:40.834–0:43.861
And with that , basically there's not a lot to say about this right
0:43.871–0:46.693
now because it seems to have been more or less silently released
0:46.693–0:48.780
.
0:48.780–0:49.535
There's no official announcement post yet on X ,
0:49.545–0:52.331
which I do believe is generally where we would see this .
0:52.341–0:55.455
But they do mention here just on the website a couple references
0:55.465–0:59.385
to it where if you use this model identifier right here for V4
0:59.395–1:03.582
Pro, you are being routed to the new V4 Pro 0813 version.
1:03.592–1:06.842
Additionally to that, it is also live on OpenRouter and this
1:06.852–1:10.103
makes me feel comfortable that it is actually genuinely released
1:10.113–1:11.977
and this is the model we're going to be using .
1:11.987–1:14.260
So , from the bit of information we have right here ,
1:14.270–1:17.520
aside from just the model identifier now being today's date ,
1:17.530–1:20.812
depending on what locale one is in, we can see that the model has
1:20.822–1:24.764
a million context window with a maximum output length of 384K ,
1:24.774–1:26.135
which really is a lot .
1:26.145–1:28.947
Now , something that can sometimes happen with DeepSeek ,
1:28.957–1:31.819
at least historically , is there may be a lot of thinking .
1:31.829–1:34.211
Now , this does support selectable thinking modes ,
1:34.221–1:36.045
as you can see if we do click on this .
1:36.055–1:38.120
So , it's possible to kind of tone that down ,
1:38.130–1:40.991
but in the interest of seeing the maximum potential performance
1:41.001–1:44.501
, all of our coding tasks for today are going to be run on this
1:44.511–1:45.697
maximum effort .
1:45.707–1:49.047
Now , I want to mention the following with a grain of salt because
1:49.057–1:51.839
I cannot definitively confirm that this is ground truth ,
1:51.849–1:54.072
but I would imagine it probably is .
1:54.082–1:57.261
There were some references and photos floating around which apparently
1:57.271–1:59.574
came from a group related to DeepSeek .
1:59.584–2:02.821
I believe they said it was a WeChat group about DeepSeek or something
2:02.821–2:08.747
.
2:08.747–2:08.827
And if this is to be believed , we can see that these would hypothetically
2:08.827–2:11.177
be the benchmark results for DeepSeek V4 Pro 0813 or the general
2:11.187–2:12.448
availability model .
2:12.458–2:15.625
And basically my main interest here is just seeing the leap in
2:15.635–2:20.225
capability over the V4 Pro preview to now because I was initially
2:20.235–2:24.272
disappointed with the V4 Pro Preview as well as the V4 Flash preview
2:24.272–2:28.713
.
2:28.713–2:28.793
They felt very undercooked and probably because they were named
2:28.793–2:28.873
preview .
2:28.873–2:29.509
That's actually like fair .
2:29.519–2:30.381
It makes sense .
2:30.391–2:33.273
But again , I want to reiterate the leap in capability between
2:33.283–2:37.518
flash and then the 0731 general availability version of flash was
2:37.528–2:38.753
absolutely massive .
2:38.763–2:42.495
This is like the best local model right now for like a dual 3090
2:42.505–2:42.914
setup .
2:42.924–2:48.182
So I mean in a niche , you know , but still if that leaping capability
2:48.192–2:51.870
increase from 12 .
2:51.880–2:53.058
increase from 12 .
2:53.068–2:54.571
8 to 62 .
2:54.581–2:58.218
7 would then hypothetically put this above things like Opus 3 .
2:58.228–3:02.046
8 8 and make it very very competitive especially that this
3:02.056–3:04.286
is an openweight model or will be .
3:04.296–3:08.276
That's also extremely exciting because this is not like horrifically
3:08.286–3:12.098
gigantic I guess when considering things like Kimi K3 or Qwen
3:12.108–3:12.470
3 .
3:12.480–3:13.291
2 .
3:13.301–3:14.587
It's still big at 1 .
3:14.597–3:15.753
6 trillion parameters .
3:15.763–3:19.089
It is a mixture of experts model with how many active ?
3:19.099–3:22.107
So if we do estimate that this is going to be the same size as
3:22.117–3:24.013
the preview version of V4 Pro .
3:24.023–3:24.650
It's a 1 .
3:24.660–3:28.490
6 trillion parameter mixture of experts model with 49 billion
3:28.500–3:28.970
active .
3:28.980–3:32.810
So it is still very very large but it's not as gigantic as like
3:32.820–3:35.284
closing in on a 3 trillion parameter model.
3:35.294–3:38.623
So additionally to that pricing and this is something that we need
3:38.633–3:41.962
to basically take just as it is right now because they mention
3:41.972–3:44.985
and have been for quite a bit of time that their API pricing is
3:44.995–3:47.289
going to be significantly increased soon .
3:47.299–3:50.469
Now significantly is relative to what it currently is which is
3:50.479–3:52.536
like very very very cheap .
3:52.546–3:56.036
So , right now , pricing per million input tokens for DeepSeek
3:56.046–3:59.250
V4 Pro is 0.43
3:59.260–4:00.080
35 .
4:00.090–4:01.178
So , okay .
4:01.188–4:04.652
No , I know what this is , but like last time I messed up on this
4:04.662–4:08.456
one and I said it was 2 cents and it wasn't and then I got roasted
4:08.456–4:08.966
.
4:08.966–4:14.347
So, this is 43 cents per million input tokens and 87 cents per
4:14.357–4:16.433
million output tokens is the current price .
4:16.443–4:19.162
Now , if this gets like a significant increase ,
4:19.172–4:20.873
say it goes up like five or 10 times ,
4:20.873–4:22.452
it would still hypothetically
4:22.462–4:23.335
be competitive .
4:23.345–4:26.063
So , it's all relative to what it currently costs .
4:26.073–4:28.793
And with that, I am going to be using this from within Cursor
4:28.793–4:30.524
.
4:30.524–4:33.137
When we tested V4 Flash , the general availability model on the
4:33.147–4:36.656
31st of July, I did also use it through Cursor because I believe
4:36.666–4:39.505
they had suggested to do so and the performance was very ,
4:39.515–4:40.850
very good .
4:40.860–4:41.879
I was happy with it .
4:41.889–4:44.381
So, I'm going to be using Cursor as well for this.
4:44.391–4:49.124
Now I should also mention that they did release in beta a desktop
4:49.134–4:51.270
app for Cursor or ChatGPT.
4:51.280–4:54.288
So I also have it routed and working through this desktop app as
4:54.298–4:57.142
well and it is doing a skate game for us in C++.
4:57.152–4:59.099
I would imagine that will take quite a bit of time .
4:59.109–5:01.791
So I did just want to start it and verify this app was actually
5:01.801–5:04.482
working properly on this system and it does seem to .
5:04.492–5:09.430
So keep in mind any mention of GPT or Cursor here is routed directly
5:09.440–5:13.030
to DeepSeek on max reasoning effort for the 0813 model .
5:13.040–5:15.355
And with that , let's take a look at our Okay ,
5:15.365–5:21.369
so apparently the OpenAI beta application set itself to be the
5:21.379–5:23.501
default for all file types .
5:23.511–5:24.525
That's infuriating .
5:24.535–5:25.122
All right .
5:25.132–5:26.060
All right .
5:26.070–5:27.355
Very Gen 3
5:27.365–5:28.772
aesthetic
5:28.782–5:31.642
I'm going to turn sound on because I did see some mentions of it
5:31.652–5:34.833
just talking about like sounds .
5:34.843–5:35.071
All right .
5:35.081–5:37.146
It starts in our notepad .
5:37.156–5:38.900
Can we resize this ?
5:38.910–5:39.618
All right .
5:39.628–5:40.257
Acceptable .
5:40.267–5:41.293
Minimize .
5:41.303–5:42.410
Unminimize .
5:42.420–5:42.968
Good .
5:42.978–5:43.686
The big one .
5:43.696–5:45.281
Is there a right click ?
5:45.291–5:48.720
Yeah , there is .
5:48.730–5:51.898
There is also the correct date and time in my local in the bottom
5:51.908–5:52.137
right .
5:52.147–5:55.011
And there is what I would imagine here to probably be the special
5:55.021–5:55.410
feature .
5:55.420–5:57.326
It is the holographic desktop feature .
5:57.336–5:59.002
So , we will save that for last .
5:59.012–6:02.275
Let's start out by looking at the files app , which I don't think
6:02.285–6:04.909
I've seen a files icon that looks like this before .
6:04.919–6:05.309
All right .
6:05.319–6:08.422
We have some nice transparency to the application backgrounds and
6:08.432–6:11.462
definitely a nice synth wave theme , which I do personally like
6:11.462–6:15.197
.
6:15.197–6:15.277
So , unfortunately , it seems our fake file system is just static
6:15.277–6:15.465
and we can't browse through it .
6:15.475–6:17.866
Nonetheless , can we create new things in it ?
6:17.876–6:18.587
Okay , we can .
6:18.597–6:21.228
Can we remove the new things we created ?
6:21.238–6:22.749
We also can .
6:22.759–6:25.951
Can we remove things that were there by default ?
6:25.961–6:26.591
Yeah , we can .
6:26.601–6:27.231
All right , cool .
6:27.241–6:28.431
That makes me a little happier .
6:28.441–6:30.431
And then up would just go there .
6:30.441–6:30.990
Oh , okay .
6:31.000–6:32.590
Not bad .
6:32.600–6:33.663
Decent bit of immersion ,
6:33.663–6:35.310
at least in terms of what directories
6:35.320–6:37.498
would exist were this a real computer .
6:37.508–6:38.833
Notepad we did kind of see
6:38.833–6:40.284
because it opened with this
6:40.284–6:41.716
just to give us some information .
6:41.726–6:42.991
The only thing I like to see is
6:42.991–6:44.203
will it save it as a text file
6:44.213–6:45.238
on the system .
6:45.248–6:46.988
No such directory .
6:46.998–6:49.138
Didn't I delete the documents directory ?
6:49.148–6:50.252
Hold on a second .
6:50.262–6:51.209
That's okay .
6:51.219–6:53.040
Oh , it remembered that I did that .
6:53.050–6:54.203
I might have messed that up .
6:54.213–6:55.410
So that's going to be
6:55.410–6:56.084
that's on me .
6:56.094–6:57.132
Uh okay .
6:57.142–7:00.112
This right here is just not right .
7:00.122–7:03.495
It should not have a like visible text box .
7:03.505–7:04.945
That happens sometimes .
7:04.955–7:06.637
Help did give us things .
7:06.647–7:09.420
Let's just do neofetch .
7:09.430–7:13.636
That's actually pretty well done right there .
7:13.646–7:16.299
Neofetch or Neofetch .
7:16.309–7:17.630
Neofetch .
7:17.640–7:19.674
Still , that's that's actually all right .
7:19.684–7:21.545
I'm just being kind of mean .
7:21.555–7:22.653
Memory .
7:22.663–7:24.355
This is fine .
7:24.365–7:27.928
Oh , it actually put a bit of cleverness in there .
7:27.938–7:28.552
1 .
7:28.562–7:30.142
44 megs out of infinity .
7:30.152–7:30.822
This is fine .
7:30.832–7:33.854
Isn't that how much an old school floppy disc would hold ?
7:33.864–7:35.914
I don't know cuz those are older than me .
7:35.924–7:37.735
Next up , calculator .
7:37.745–7:44.720
Okay , so 47 47 47 8 8 8 376 .
7:44.730–7:46.636
I haven't done one of these in a while
7:46.636–7:47.424
divided by six .
7:47.434–7:51.519
All right , settings .
7:51.529–7:54.188
Definitely seeing a bit of like Qwen 2.5
7:54.198–7:54.808
8B in here
7:54.808–7:57.171
just in terms of some of the style here.
7:57.181–7:58.848
What is that
7:58.848–8:01.348
and why do I like it ?
8:01.358–8:01.921
Aurora .
8:01.931–8:03.068
Oh , okay .
8:03.078–8:06.262
Sunset City .
8:06.272–8:09.456
This No one would understand that reference .
8:09.466–8:10.734
This looks like something
8:10.734–8:12.405
one of my friends in high school was
8:12.415–8:13.796
making as a game project .
8:13.806–8:15.378
It actually looks like a ton
8:15.378–8:17.564
like the splash screen for his project
8:17.564–8:18.433
.
8:18.433–8:21.413
We have customizable accents as well .
8:21.423–8:24.117
Difficulty easy , normal , and hard .
8:24.127–8:25.100
Oh , for games .
8:25.110–8:26.628
Okay , I probably set it to easy
8:26.628–8:28.146
as you may be familiar with my
8:28.156–8:29.912
quick to anger game playing style .
8:29.922–8:31.862
Next up , let's try our games .
8:31.872–8:33.384
The speaker is on .
8:33.394–8:37.354
So , City Heist .
8:37.364–8:44.554
Oh , you absolute .
8:44.564–8:44.948
Okay .
8:44.958–8:47.711
So , you have to press enter to start
8:47.711–8:49.678
and not click the button
8:49.678–8:50.976
.
8:50.976–8:52.446
Please shift gears .
8:52.456–8:53.827
Ah , okay .
8:53.837–9:00.337
Okay .
9:00.347–9:00.830
All right .
9:00.840–9:11.955
It's Oh , cool .
9:11.965–9:13.847
Oh , okay .
9:13.857–9:16.065
It's troubled,
9:16.065–9:21.486
but the sounds for like the smack
9:21.486–9:23.407
are kind of funny .
9:23.417–9:25.160
Like , all right .
9:25.170–9:25.972
And we wrecked
9:25.972–9:27.508
and we only grabbed a 100 red
9:27.508–9:28.443
out of a thousand,
9:28.453–9:29.466
but it did work .
9:29.476–9:30.168
So , keep in mind
9:30.168–9:32.127
we need to press enter to play these games
9:32.127–9:32.300
and
9:32.310–9:34.740
not hit the button that says launch .
9:34.750–9:36.098
Oh .
9:36.108–9:43.897
Oh , all right .
9:43.907–9:45.950
Do I go through these ?
9:45.960–9:46.272
Yeah .
9:46.282–9:48.364
I I just assume based on the sounds .
9:48.374–9:51.019
I wonder if that's how like dogs are since they don't understand
9:51.029–9:51.744
language .
9:51.754–9:54.640
But like that to me seemed like a positive reinforcement sound
9:54.640–9:56.493
.
9:56.493–9:56.573
So , it makes me want to do it more .
9:56.573–9:59.388
And I think we did just discover something about how dog brains
9:59.398–10:01.803
work .
10:01.813–10:03.654
Okay , the score .
10:03.664–10:07.194
I don't know that we need that many decimal points just to show
10:07.204–10:07.677
a score .
10:07.687–10:12.424
I think maybe just rounding up to 734 or down would have been acceptable
10:12.434–10:16.455
, but I applaud it for its intricacy .
10:16.465–10:17.665
And then , oh , read me .
10:17.675–10:20.891
Okay , so our special feature is something like holographic .
10:20.901–10:23.794
And I would imagine having a couple apps open just when we trigger
10:23.804–10:25.408
that is probably the move .
10:25.418–10:27.747
So , oh yeah .
10:27.757–10:28.149
All right .
10:28.159–10:30.248
So , it just puts like an interesting card effect .
10:30.258–10:31.457
This is what it looks like .
10:31.467–10:34.199
This looks like an Apple Vision Pro like app demo .
10:34.209–10:37.647
This is the way they look from inside the Vision Pro headset .
10:37.657–10:42.548
Um , just if you have used one , you may resonate with what I'm
10:42.558–10:43.655
saying .
10:43.665–10:44.208
Okay .
10:44.218–10:46.025
Overall , it was fine .
10:46.035–10:46.973
Everything worked .
10:46.983–10:48.159
Nothing really blew me away .
10:48.169–10:51.399
But again , that may be a limitation of the test and less of a
10:51.409–10:52.506
limitation of the model .
10:52.516–10:55.131
So alas , huh , it's kind of neat .
10:55.141–10:56.052
We will continue .
10:56.062–10:58.312
So our skateboard game is still going .
10:58.322–11:02.097
And again, I'm using this with the newly released beta version
11:02.107–11:03.985
of the chat GPT desktop app .
11:03.995–11:06.286
I would imagine this will take quite a bit of time .
11:06.296–11:09.486
So let's check to see just like Oh no , cuz we already had looked
11:09.496–11:10.202
at the cost .
11:10.212–11:13.929
And keep in mind this is partially for the skate game as well as
11:13.939–11:14.717
the browser OS .
11:14.727–11:17.655
So 20 cents total across two separate tasks .
11:17.665–11:18.945
One which is still ongoing .
11:18.955–11:19.231
All right .
11:19.241–11:24.748
So , after 68 minutes and 24 seconds , Skate NYC is complete .
11:24.758–11:26.110
So , oh , you idiot .
11:26.120–11:30.886
You spoiled it with these screenshots and you don't even have vision
11:30.886–11:31.396
.
11:31.396–11:32.915
I'm infuriated .
11:32.925–11:36.637
Nonetheless , now that we know what the game looks like ,
11:36.647–11:40.152
I had to stop myself from saying something a little more colorful
11:40.152–11:40.747
.
11:40.747–11:44.147
Let's just take a peek at at the game .
11:44.157–11:47.331
All right , so here's our NYC skate game .
11:47.341–11:48.605
Should be in outputs , right ?
11:48.615–11:51.232
Yes .
11:51.242–11:52.665
All right .
11:52.675–11:56.804
It's not quite as good as the Qwen 2.5
11:56.814–12:04.335
model, but that model does have 800 billion more parameters
12:04.335–12:04.845
.
12:04.845–12:07.101
Does that say NYC skate on the ground ?
12:07.111–12:08.415
It's inverted , but it does .
12:08.425–12:10.798
And it says established 1999 .
12:10.808–12:11.175
All right .
12:11.185–12:12.280
Uhoh .
12:12.290–12:16.396
Sounded like the beginning of a Van Halen song
12:16.406–12:18.840
I don't know what exactly it is
12:18.850–12:19.773
All right .
12:19.783–12:22.766
Uh , let's just maybe not do that same thing .
12:22.776–12:24.020
There are sound effects , though .
12:24.030–12:25.431
I don't normally hear sound effects .
12:25.441–12:27.626
Can we actually move around?
12:27.636–12:28.097
All right .
12:28.107–12:29.273
Not quite .
12:29.283–12:32.176
I will say it's not wonderful .
12:32.186–12:33.533
Oh , that's actually kind of cool , though .
12:33.543–12:37.247
The way it like stylized the game title .
12:37.257–12:37.747
Nice .
12:37.757–12:40.675
Nice .
12:40.685–12:43.890
Well , we don't need to do the drum kit sim .
12:43.900–12:53.313
The trees are quite reminiscent of a poo, if I were to be honest
12:53.323–13:02.090
, but all right , the the legs are like kicking .
13:02.100–13:05.018
This may be the funniest skate game result we've ever seen ,
13:05.028–13:06.283
which is , you know , something .
13:06.293–13:08.419
All right , we have a street sign there .
13:08.429–13:12.690
The text is a little sideways , as are the building faces .
13:12.700–13:14.589
This is definitely a troubled result .
13:14.599–13:17.436
However , I do like the water fountain shooting out .
13:17.446–13:21.988
I believe that was something I've seen .
13:21.998–13:22.894
Okay .
13:22.904–13:23.636
All right .
13:23.646–13:26.150
I've seen that in, I think, the Qwen result.
13:26.160–13:29.357
Now , here's what I would like to just quickly do for some reference
13:29.357–13:30.327
.
13:30.327–13:33.987
This is the result we received for the C++ skateboard test from the
13:33.997–13:36.620
preview version of DeepSeek V4.
13:36.630–13:37.659
We can see right here .
13:37.669–13:39.738
I'm going to turn the speaker off .
13:39.748–13:41.258
This is pretty bad .
13:41.268–13:45.257
It was like so bad that I actually remember to go to this specific
13:45.267–13:46.777
thing just to showcase it .
13:46.787–13:51.256
So, the leap in capability based off of what we see right here
13:51.266–13:54.850
to what we just saw in the skate game that it had created for us
13:54.860–13:58.374
, which we can just look at in these screenshots here is still
13:58.384–13:59.254
pretty significant .
13:59.264–14:01.653
Next up , I'm giving it the Subway FPS test .
14:01.663–14:03.653
I've also added on to the end of the prompt .
14:03.663–14:05.492
Don't worry about checking over the script .
14:05.502–14:08.132
I will manually check it and if there are issues I will report
14:08.142–14:11.491
them back to you, just because this will really extend every
14:11.501–14:16.228
single result by like two times in terms of time-wise length
14:16.238–14:19.808
just in the interest of something I could do visually in like 5
14:19.818–14:21.402
seconds and then be like hey this doesn't work .
14:21.412–14:24.671
So we're going to do that just to maybe like keep things a little
14:24.681–14:25.148
tighter .
14:25.158–14:28.257
All right so it did finish the Subway FPS game .
14:28.267–14:31.127
The problem is it just keeps trying to do like troubleshooting
14:31.137–14:34.635
which it's not good at doing on this system for some reason .
14:34.645–14:37.302
All right , let's check out our subway FPS .
14:37.312–14:39.101
Main Rand Avenue.
14:39.111–14:42.547
All right , going to turn the speaker on .
14:42.557–14:44.063
So , it gave us two options .
14:44.073–14:46.536
We can either explore the station .
14:46.546–14:50.525
Okay, it's... Hey, you know what?
14:50.535–14:54.913
I think that is the best subway car that I've ever seen drawn in
14:54.923–14:56.269
this test .
14:56.279–14:58.623
And you're going to be like , well , it doesn't look that great
14:58.623–15:00.112
.
15:00.112–15:00.192
I know .
15:00.192–15:01.534
It's just that they never really include these .
15:01.544–15:03.769
And when they do , they're not really very good .
15:03.779–15:06.801
So this... actually, what was that?
15:06.811–15:10.870
If that thing starts moving I would have probably become more excited
15:10.880–15:15.177
than would be reasonable for you know responding to such thing
15:15.177–15:15.746
.
15:15.746–15:16.853
This is very clean .
15:16.863–15:17.490
It's simple .
15:17.500–15:20.044
It's nothing that's blowing me away to be completely honest with
15:20.054–15:20.284
you .
15:20.294–15:21.560
Queens Boulevard line .
15:21.570–15:21.880
Okay .
15:21.890–15:22.677
We have stairs .
15:22.687–15:25.150
We have some lighting elements and things of the sort .
15:25.160–15:28.341
We have exits if you want to go that side.
15:28.351–15:34.407
And then tunnels, like indicators for the train, a trash bin.
15:34.417–15:37.121
Floating particles more simple .
15:37.131–15:38.799
And we can't go up the stairs , which is fine .
15:38.809–15:39.677
We don't want to .
15:39.687–15:43.909
Now , of course , the other feature that was included here is the
15:43.919–15:45.106
FPS .
15:45.116–15:48.381
So, let's go back to the menu and then enter the survival FPS.
15:48.391–15:50.377
Let's see if it changes the lighting or anything .
15:50.387–15:51.336
Oh my .
15:51.346–15:53.012
That's a good weapon model .
15:53.022–15:54.370
Does it ?
15:54.380–15:54.769
How ?
15:54.779–15:57.484
Oh , it does .
15:57.494–15:59.480
I think .
15:59.490–16:01.237
Yeah , it does .
16:01.247–16:02.595
It's like smooth now .
16:02.605–16:03.951
I don't see any enemies .
16:03.961–16:05.149
Look at when we run .
16:05.159–16:07.865
It's like a waddle .
16:07.875–16:08.263
All right .
16:08.273–16:12.895
So , uh the big thing I'm noticing here is there are no zombie
16:12.905–16:13.454
enemies .
16:13.464–16:17.047
So , good reload .
16:17.057–16:22.877
The waddle is just kind of funny .
16:22.887–16:24.314
Weird .
16:24.324–16:24.792
Hey , look .
16:24.802–16:26.471
There are shell casings that actually drop .
16:26.481–16:27.672
That's kind of cool .
16:27.682–16:30.235
All right, this is like... yeah, it's kind of mid.
16:30.245–16:33.279
So , I'm going to give it a feedback piece .
16:33.289–16:36.162
Now , I stopped it because it was just like it had finished and
16:36.172–16:38.925
it was like I can't use this in the headless environment to check
16:38.925–16:40.891
.
16:40.891–16:41.608
And at that point , it's easier to just check it and tell it what's
16:41.618–16:42.809
wrong .
16:42.819–16:44.332
So , I'll give it that honest feedback .
16:44.342–16:48.255
So , supposedly our fixed and polished version of the Subway FPS
16:48.265–16:49.376
is completed .
16:49.386–16:51.138
Let's just take a peek at it .
16:51.148–16:52.584
The sound is on .
16:52.594–16:53.460
All right .
16:53.470–16:54.013
Interesting .
16:54.023–16:56.460
I think it swapped the positioning of the buttons .
16:56.470–17:00.130
So , all right .
17:00.140–17:01.015
Seven hostiles .
17:01.025–17:02.626
It's showing me in the top right .
17:02.636–17:07.212
I unfortunately just can't locate them .
17:07.222–17:09.867
I would have expected it to have done better .
17:09.877–17:12.362
And again , this is one where we could go back and reference the
17:12.372–17:14.615
previous model of this .
17:14.625–17:18.639
So , basically the preview version , it was significantly less
17:18.649–17:22.583
detailed than this was .
17:22.593–17:26.043
though unfortunately we don't have enemies and given the amount
17:26.053–17:29.251
of time it took as well as the kind of multiple times I told it
17:29.261–17:32.375
there were no enemies, I'm regrettably going to call this one a next up.
17:32.385–17:36.126
Now I'm giving this a difficult task to create a 3D model of an
17:36.136–17:40.866
RB26 inline 6 engine it must be something that can be 3D printed
17:40.876–17:44.917
with little to no support and fit inside of it a tiny DC motor
17:44.927–17:48.400
called an N20 motor because this is something that I've told it
17:48.410–17:51.430
should like work with my RC car project. Actually, I don't think
17:51.440–17:53.662
that's specifically in this prompt , but that's the whole point
17:53.672–17:57.169
of this, where it should understand and it will scale it
17:57.179–17:58.524
properly to that motor .
17:58.534–18:02.031
So far , I've run this test with a few intelligent models and I've
18:02.041–18:04.263
never really received anything that was actually worth printing
18:04.263–18:07.577
.
18:07.577–18:07.657
So , hopefully sooner than later that will change .
18:07.657–18:10.559
So , our 3D model task is pretty interesting because it's just
18:10.569–18:14.068
doing a lot of research on like the actual proper orientation of
18:14.078–18:16.140
specific items on this engine .
18:16.150–18:18.777
Like where is the downpipe , which is an exhaust component ?
18:18.787–18:19.656
What side is it ?
18:19.666–18:20.775
What is the intake side ?
18:20.785–18:21.975
What is the exhaust side ?
18:21.985–18:23.493
Intake plenum left or right ?
18:23.503–18:28.049
So , I do like to see the amount of like intricate research that's
18:28.059–18:30.607
going into the culmination of this model creation .
18:30.617–18:34.923
And it would be really awesome for like us to get a model for this
18:34.933–18:36.280
task that's actually printable .
18:36.290–18:37.480
It would be the first one .
18:37.490–18:41.816
So , in 1 hour and 10 minutes , we received our 3D model response
18:41.816–18:44.543
.
18:44.543–18:45.914
This is a very, very intricate bit of text about the thing
18:45.924–18:46.717
it made .
18:46.727–18:49.288
Now , of course , the other part of that is actually just taking
18:49.298–18:51.135
a look at what it looks like .
18:51.145–18:53.626
Though , once I remember what I named this folder ,
18:53.636–18:54.991
we will be able to take a peek .
18:55.001–18:56.356
Assembled display STL .
18:56.366–18:58.364
I would imagine that to be the case .
18:58.374–18:59.489
Please be good .
18:59.499–19:03.909
I've really, like, I need a good one for these.
19:03.919–19:06.158
Actually , it's sideways .
19:06.168–19:06.880
Son of a gun .
19:06.890–19:10.335
This did a really good job .
19:10.345–19:15.556
This is actually like so far the best one .
19:15.566–19:17.244
It's not perfect .
19:17.254–19:21.100
We have some sideways turbocharger issues , though .
19:21.110–19:22.383
I mean, this is all right.
19:22.383–19:23.991
Let's peek at this in OpenSCAD
19:23.991–19:24.677
.
19:24.677–19:26.320
And there are a bunch of individual pieces for this .
19:26.330–19:28.907
So , we also want to see if the engine that we told it needs to
19:28.917–19:31.273
fit inside the little remote control car engine .
19:31.283–19:34.634
Let's see if it has any hope of properly fitting.
19:34.634–19:35.340
Research—
19:35.350–19:39.111
It has a research folder right here just showing like a bunch of
19:39.121–19:41.772
images and stats for this motor .
19:41.782–19:42.660
Very cool though .
19:42.670–19:46.392
And here we have some better preview renders .
19:46.402–19:50.501
This is definitely well done .
19:50.511–19:52.041
Turbo side .
19:52.051–19:52.760
Okay .
19:52.770–19:53.966
Underside .
19:53.976–19:54.850
All right .
19:54.860–19:58.466
It's not something that's going to really easily be 3D printable
19:58.466–19:58.976
.
19:58.976–20:01.360
Although , I guess I can't make that determination because this
20:01.370–20:02.968
comes in multiple pieces .
20:02.978–20:06.344
So , it's not designed to be printed as we first saw it just in
20:06.354–20:07.548
one lump like this .
20:07.558–20:09.719
It's designed to be printed separately .
20:09.729–20:11.005
I would like to see .
20:11.015–20:14.381
Let's look now at just the single first piece ,
20:14.391–20:18.720
which is ideally where it would be placing the 3D printed motor
20:18.730–20:21.454
that would go in it .
20:21.464–20:25.473
Unfortunately , printing this without supports is just not an option
20:25.483–20:28.526
in any potential orientation that we see right here ,
20:28.536–20:29.892
which is a bit frustrating .
20:29.902–20:32.626
It almost seems like it has these posts and holes in it to be able
20:32.636–20:35.438
to snap fit on other parts, which is cool.
20:35.448–20:36.243
Okay .
20:36.253–20:36.563
Yep .
20:36.573–20:39.217
And I think those are definitely like attachment posts ,
20:39.227–20:40.262
which is pretty cool .
20:40.272–20:43.717
This one would be able to be printed more or less without any supports
20:43.727–20:46.589
if we were to put it flat .
20:46.599–20:50.712
Then this should have the Nissan logo, or at least the text "Nissan"
20:50.722–20:53.420
written on it .
20:53.430–20:53.883
Yep .
20:53.893–20:55.741
And this one's quite well done .
20:55.751–20:56.668
No issues here .
20:56.678–20:57.781
The oil pan .
20:57.791–20:58.262
Oh , no .
20:58.272–20:58.662
Okay .
20:58.672–21:01.785
This is where we decide we're not going to be
21:01.795–21:02.666
trying to print this .
21:02.676–21:07.391
I'm not disappointed , but I'm not super impressed either .
21:07.401–21:11.315
So , I suppose I am lukewarm when it comes to judging this .
21:11.325–21:14.198
Definitely , if we just look at this as a render ,
21:14.208–21:16.842
even forget about looking at the individual parts there ,
21:16.852–21:20.225
cuz it's not really going to be feasible to print this ,
21:20.235–21:22.963
it did a really decent job just assembling it .
21:22.973–21:25.542
The research that went into this, I think, is probably
21:25.552–21:29.328
like I hate to sound like an AI here , but like the story isn't
21:29.338–21:31.339
the model , it's the research that went into it .
21:31.349–21:33.255
That's like generic AI slop.
21:33.265–21:37.085
It really did a lot of decent research here .
21:37.095–21:41.075
I mean , did it find this and download it just to see ?
21:41.085–21:41.474
No .
21:41.484–21:41.793
Okay .
21:41.803–21:45.464
So , this is something it maybe made or maybe probably found online
21:45.474–21:48.616
just to be able to better assess what the size of this asset is
21:48.626–21:51.768
, but it took a lot of care to ensure that what it tried to generate
21:51.778–21:55.519
would have hypothetically fit the specific specifications that
21:55.529–21:57.674
were denoted in the original prompt .
21:57.684–21:59.669
So , not half bad .
21:59.679–22:02.621
I'm also going to give this a simple front end design test where
22:02.631–22:04.377
it needs to create the watch website .
22:04.387–22:07.643
So , it has to have a pretty cinematic panning effect over a 3D
22:07.653–22:09.222
watch asset that it creates .
22:09.232–22:12.538
I did specifically denote that
22:12.548–22:14.197
the watch needs to be a 3D model .
22:14.207–22:17.039
So , we'll see how competent it is at putting the entirety of this
22:17.049–22:17.513
together .
22:17.523–22:21.817
So , in 1 hour and 30 minutes , we got our watch front end result
22:21.817–22:22.327
.
22:22.327–22:25.803
It again seems to have a pretty substantial like just list of what
22:25.813–22:26.769
was built here .
22:26.779–22:28.460
So , let's take a look and see how it is .
22:28.470–22:32.728
a beautiful watch front-end website.
22:32.738–22:36.096
That's actually, it's so close to excellent.
22:36.106–22:40.745
The big problem is this , I would assume , is the watch strap sticking
22:40.755–22:41.306
out .
22:41.316–22:45.173
This is 3D, so don't be fooled by its 2D-ish look.
22:45.183–22:46.552
It is actually 3D .
22:46.562–22:48.011
It's slowly panning , too .
22:48.021–22:52.088
I really... Okay, so a couple things I do like: the wooden background
22:52.098–22:55.742
and then this like leather card or pad that this watch is placed
22:55.752–22:56.297
on .
22:56.307–22:58.440
In addition to that, the face looks really good.
22:58.450–23:00.347
The materials look very good .
23:00.357–23:03.998
It's just like, overall, it's a high-end aesthetic,
23:04.008–23:07.254
which is exactly what we go for when performing this test .
23:07.264–23:13.563
Now , that's actually that right there looks good .
23:13.573–23:16.975
The... Let's just sit and appreciate it for a minute.
23:16.985–23:17.215
Okay .
23:17.225–23:19.290
Yeah , that maybe I could have done without that ,
23:19.300–23:22.962
but it's definitely quite something .
23:22.972–23:25.435
And then look at the way that goes away as soon as we start to
23:25.445–23:26.633
scroll down .
23:26.643–23:30.225
"Quiet luxury measured in seconds."
23:30.235–23:31.582
"Sunburst dial."
23:31.592–23:33.657
Okay , it has like a material card here .
23:33.667–23:37.249
The sapphire crystal is perhaps something I would have maybe omitted
23:37.259–23:39.564
from this section , but the dial looks good .
23:39.574–23:41.560
"Two watches, one point of view."
23:41.570–23:42.438
Please render them .
23:42.448–23:43.236
Oh , look at that .
23:43.246–23:44.194
That is sick .
23:44.204–23:45.629
That really looks pretty good .
23:45.639–23:47.546
I love the way they're placed on the table .
23:47.556–23:50.420
Really the big thing letting it down is just this leather slab
23:50.430–23:53.533
is inhibiting our view of these beautiful straps .
23:53.543–23:55.688
This one is the stainless steel option .
23:55.698–23:59.679
This one is the gold option and priced accordingly I would imagine
23:59.679–24:01.292
.
24:01.292–24:01.372
Yeah .
24:01.372–24:05.665
Although I would say the jump in price is like a negligible
24:05.675–24:06.702
amount to get the gold one .
24:06.712–24:08.539
I'd probably go with this one personally .
24:08.549–24:08.938
Oh , cool .
24:08.948–24:12.370
And these ones we can freely rotate so that helps a bit .
24:12.380–24:15.643
There is no movement to the watch hands but that's okay .
24:15.653–24:17.240
I'll forgive it for that .
24:17.250–24:19.395
But these are pretty darn cool .
24:19.405–24:22.508
These are well drawn , I would say .
24:22.518–24:25.860
Especially the numeral indicators .
24:25.870–24:28.814
Look at the way that they actually have multi material on them
24:28.814–24:29.770
.
24:29.770–24:31.766
So , they're not just like one static material .
24:31.776–24:35.599
It's gold and then black infill or silver and then black infill
24:35.609–24:35.918
as well .
24:35.928–24:38.473
So , that's some attention to detail I do like seeing .
24:38.483–24:42.383
As well as the kind of brushed material look on the stainless steel
24:42.393–24:42.943
and the gold .
24:42.953–24:46.215
It's probably weird we're like assessing like a 3D watch ,
24:46.225–24:48.690
but it just shows like some capability .
24:48.700–24:50.429
Okay , that looks pretty good right there .
24:50.439–24:52.335
And then we have a clean footer as well .
24:52.345–24:52.997
I like this .
24:53.007–24:54.075
I'm happy with this .
24:54.085–24:58.632
Even though it's like not perfect , it's darn close .
24:58.642–25:00.919
So, next up, just here from within Codex.
25:00.929–25:03.873
And I can't show this full prompt because it has an API key .
25:03.883–25:07.228
I'm giving this the full on cinematic Steve the PC repair game
25:07.238–25:09.624
test where it needs to create a cinematic game .
25:09.634–25:12.900
It needs to create voice assets for these characters just by using
25:12.910–25:16.595
an open AI API key that I've given it and it needs to make a fun
25:16.605–25:18.874
game that is enjoyable and playable .
25:18.884–25:21.940
I will assess this is probably going to take north of an hour .
25:21.950–25:24.694
So, we're just going to basically leave Codex here in the background
25:24.704–25:28.538
the beta desktop version of it to complete this task, and we'll
25:28.548–25:30.255
maybe work on some other things in the meantime .
25:30.265–25:34.365
So , our Steve the PC repairman cinematic game completed in basically
25:34.375–25:35.187
65 minutes .
25:35.197–25:36.755
So , a little over an hour .
25:36.765–25:38.250
All right , here's how to play .
25:38.260–25:41.537
I'll probably just do what it says here , not assume that it's
25:41.547–25:43.777
already running this in some server for us .
25:43.787–25:47.737
So, okay, episode one, the Azure Palm job.
25:47.747–25:49.231
Okay , click to begin .
25:49.241–25:50.502
Headphones recommended .
25:50.512–25:53.084
We have a speaker .
25:53.094–25:57.007
All right .
25:57.017–25:58.867
My name is Steve .
25:58.877–25:59.693
This is good .
25:59.703–26:02.170
I fix computers .
26:02.180–26:05.516
Most of my clients are old enough to remember when a hard drive
26:05.526–26:07.306
was a luxury item .
26:07.316–26:10.768
They pay what they can afford and I make it work .
26:10.778–26:14.143
Mostly honest business .
26:14.153–26:15.855
Oh , thank you , dear .
26:15.865–26:20.261
The screen kept telling me the date was 1980 and I thought, all
26:20.271–26:20.834
fixed.
26:20.844–26:22.554
I have been traveling long enough .
26:22.564–26:23.646
Fresh battery .
26:23.656–26:26.196
Your computer now believes I will bring you oatmeal cookies
26:26.206–26:26.860
on Tuesday .
26:26.870–26:27.720
Same as the rest .
26:27.730–26:28.065
Give Mr.
26:28.075–26:30.110
Buttons a scratch behind his ears for me.
26:30.120–26:33.580
Let me walk you out , ma'am .
26:33.590–26:36.020
I'll be right with you , Mr .
26:36.030–26:38.727
Thomas , the speech overlays itself .
26:38.737–26:40.513
A good boy , Steven .
26:40.523–26:43.953
If that quiet man needs fixing, you be patient.
26:43.963–26:44.560
Punctual .
26:44.570–26:46.061
I like punctual .
26:46.071–26:48.001
Door is unlocked .
26:48.011–26:48.929
Coffee is warm .
26:48.939–26:49.704
A man .
26:49.714–26:50.710
What is the job ?
26:50.720–26:52.723
Came for 2 years.
26:52.733–26:54.813
My life .
26:54.823–26:58.612
See the the head tilt movement back and forth is actually there's
26:58.622–26:59.671
three people talking now .
26:59.681–27:05.838
And he forgets the entire name. Charles buys him a penthouse private
27:05.848–27:08.220
security, security wants his legend. Some people would like to ask
27:08.230–27:12.891
him questions. Him slowly. The client pays 100 upfront. I would like
27:12.901–27:16.850
it's also up in the case. All expenses, first class, two tickets. This
27:16.860–27:22.047
is why we work together. Together, the top tray is a normal file
27:22.057–27:30.980
on a laptop. It's quite my second favorite thing.
27:30.990–27:31.336
Professional .
27:31.346–27:33.116
I will not ask what the first is .
27:33.126–27:36.888
The phone rings when the helicopter is ready .
27:36.898–27:37.873
40 .
27:37.883–27:39.334
Cut scene over .
27:39.344–27:40.279
Thank goodness .
27:40.289–27:41.739
Click to take control .
27:41.749–27:43.373
All right .
27:43.383–27:46.981
Aside from the sounds being just totally messed up ,
27:46.991–27:52.052
this is actually pretty decent .
27:52.062–27:52.654
Okay .
27:52.664–28:00.561
The cat is... There's definitely some geometric issues.
28:00.571–28:02.366
What are we supposed to do ?
28:02.376–28:03.997
Take the briefcase when you're ready .
28:04.007–28:04.256
Okay .
28:04.266–28:06.833
Where is it ?
28:06.843–28:11.414
Well, press buttons are like, "Don't look at me like that."
28:11.424–28:12.146
Buttons .
28:12.156–28:12.700
Easter egg .
28:12.710–28:14.079
A man's got to eat .
28:14.089–28:14.462
Okay .
28:14.472–28:17.600
Where's the briefcase ?
28:17.610–28:19.132
All right .
28:19.142–28:20.127
We have inventory .
28:20.137–28:20.970
Zip ties .
28:20.980–28:22.202
Okay .
28:22.212–28:23.324
E to continue .
28:23.334–28:23.785
All right .
28:23.795–28:25.408
Is it going to put us in Miami?
28:25.418–28:25.608
Nice .
28:25.618–28:30.071
Palm trees , humidity , and one very bad man ,
28:30.081–28:30.308
Mr .
28:30.318–28:31.648
Colton .
28:31.658–28:32.357
Champagne perhaps .
28:32.367–28:36.351
I'm just going to click through cuz it's like like black like my
28:36.361–28:40.530
well, coffee asset at the resort.
28:40.540–28:41.941
Poolside at noon .
28:41.951–28:43.168
Oh , poolside .
28:43.178–28:45.040
17th floor enters.
28:45.050–28:46.830
Camera coverage .
28:46.840–28:47.302
Elevator .
28:47.312–28:48.450
There is no dog .
28:48.460–28:52.831
Behind is the west service stairwell.
28:52.841–28:55.987
Well , I think we just saw a glimpse of the pool .
28:55.997–28:57.456
All right .
28:57.466–28:59.495
Oh , great .
28:59.505–29:00.799
Azure Palm Resort .
29:00.809–29:03.165
Click to take control .
29:03.175–29:07.406
It's even got like the like low rent wallpaper that you'd expect
29:07.416–29:09.201
at one of these joints .
29:09.211–29:10.260
Oh , wow .
29:10.270–29:11.891
This is actually like kind of intricate .
29:11.901–29:12.381
Oh , good .
29:12.391–29:12.626
Okay .
29:12.636–29:15.195
I was worried we wouldn't be able to just walk through the walls
29:15.205–29:16.541
, but okay .
29:16.551–29:18.416
We can toss a coin in the fountain .
29:18.426–29:18.988
Your wish .
29:18.998–29:20.185
No more syntax errors.
29:20.195–29:21.617
50 50 .
29:21.627–29:25.773
It seems like Steve's shop and the airplane scene is like oddly
29:25.783–29:29.299
placed inside of the location we're supposed to be at .
29:29.309–29:31.622
Let's just ignore that .
29:31.632–29:37.710
So , check in at the front desk and then reach the 17th floor .
29:37.720–29:39.711
Welcome to the Azure Palm .
29:39.721–29:41.710
Checking in , Mr .
29:41.720–29:42.750
Me.
29:42.760–29:43.603
Coulter.
29:43.613–29:44.755
Steven Coulter .
29:44.765–29:44.925
Mr .
29:44.935–29:47.198
I believe I have a reservation .
29:47.208–29:51.154
Room 1408 , partial ocean view .
29:51.164–29:52.416
I do apologize .
29:52.426–29:55.363
The ice machine on your floor is out of order .
29:55.373–30:00.141
Story of my life .
30:00.151–30:00.560
All right .
30:00.570–30:02.198
We now have a room key in our inventory .
30:02.208–30:03.882
So , is this the elevator ?
30:03.892–30:12.484
Oh, is that right?
30:12.494–30:12.817
Okay .
30:12.827–30:15.467
West service stairwell .
30:15.477–30:21.017
Let's just assume this is north .
30:21.027–30:45.339
Good. Like, what?
30:45.349–30:51.897
Well , that is frustrating because it gave us this really like
30:51.907–30:53.164
anticipatory music .
30:53.174–30:55.474
I'm sure there's a far better terminology for what I was about
30:55.484–30:56.071
to say .
30:56.081–30:57.188
My name is Steve .
30:57.198–30:57.710
Oh , no .
30:57.720–30:59.830
Don't .
30:59.840–31:03.256
Most of my clients are old enough to remember when a hard drive
31:03.266–31:04.427
was a luxury.
31:04.437–31:04.829
All right .
31:04.839–31:06.970
We we got most of the way .
31:06.980–31:10.849
I think this is a test that more often than not ,
31:10.859–31:12.464
we don't complete it .
31:12.474–31:15.698
I will say the elements of this, judged individually,
31:15.708–31:17.234
Definitely the pieces were all there .
31:17.244–31:19.900
I don't know that the puzzle made the correct picture that was
31:19.910–31:23.052
shown on like the puzzle box , but it was pretty darn close and
31:23.062–31:26.204
I saw some—truthfully, this did a better job than I expected,
31:26.214–31:29.517
especially considering like the subway scene game and things like
31:29.527–31:29.841
that .
31:29.851–31:34.206
So , I'm actually I'm okay with this .
31:34.216–31:37.923
So , the final thing was of course the street yeet game ,
31:37.933–31:41.640
which is weird because this piece of text moving up and down like
31:41.650–31:44.470
that while this one remains completely static makes me uncomfortable
31:44.480–31:46.005
in some level .
31:46.015–31:48.457
So, this is a Bijan issue.
31:48.467–31:49.288
All right , good .
31:49.298–31:53.072
So, I neglected to, um, just run it in a server.
31:53.082–31:57.152
One fist, zero mercy.
31:57.162–31:58.354
I already like this .
31:58.364–32:05.477
Oh , it seems like Good music .
32:05.487–32:05.877
All right .
32:05.887–32:06.276
All right .
32:06.286–32:10.359
All right .
32:10.369–32:16.681
Let's just explore the city real quick .
32:16.691–32:18.603
I don't even know .
32:18.613–32:30.602
Okay, so I saw a car get—like, this one's actually kind of fun
32:30.602–32:34.698
.
32:34.698–32:35.970
Excuse me .
32:35.980–32:37.250
Excuse me .
32:37.260–32:38.131
Pardon me .
32:38.141–32:40.965
It does like a... like a... Okay.
32:40.975–32:42.389
I don't know if we hit...
32:42.399–32:47.562
Is that a Can we hit the bike ?
32:47.572–32:54.399
Bonk .
32:54.409–32:55.114
Uh oh .
32:55.124–33:00.815
Jay walking pedestrian .
33:00.825–33:02.583
Can we punch this parking meter ?
33:02.593–33:05.154
Yeah, we can just hit everything except the buildings.
33:05.164–33:05.556
All right .
33:05.566–33:08.847
There's unfortunately the orientation of some of these things is
33:08.857–33:09.722
not quite right .
33:09.732–33:15.609
We can see that we're moving sideways when we move forward .
33:15.619–33:22.161
And the cars are also Sorry , pal .
33:22.171–33:25.963
Oh , and it's interesting how they uh interact with the scenery
33:25.973–33:26.453
as well .
33:26.463–33:27.763
Can we eat this statue ?
33:27.773–33:30.709
I don't think we can , but Oh , yeah , we can .
33:30.719–33:33.981
All right .
33:33.991–33:37.853
I will say the cartoonishness of this game is actually kind of well
33:37.863–33:38.500
done .
33:38.510–33:41.571
I don't know how easy that is to assess just watching someone play
33:41.581–33:43.917
it , but like the pacing is decent .
33:43.927–33:47.071
I think maybe that's attributed to the music as well ,
33:47.081–33:49.010
but the way that—when you actually... All right,
33:49.020–33:52.412
the loop is getting a little repetitive, but when you actually yeet
33:52.422–33:55.813
the things and they also go and interact with other items in the
33:55.823–33:56.445
environment .
33:56.455–34:00.322
So like how that will bounce off the building and then maybe hit
34:00.332–34:01.745
other things in the path .
34:01.755–34:02.537
That's pretty cool .
34:02.547–34:09.255
And this was actually quite arcade style and playable .
34:09.265–34:10.366
Our score is going up .
34:10.376–34:13.177
We have combo multipliers overall .
34:13.187–34:14.849
38 here is a success .
34:14.859–34:17.596
All right , so that is probably going to conclude our first look
34:17.606–34:21.150
and test of the general availability DeepSeek V4 Pro ,
34:21.160–34:23.388
the 0813 model designation .
34:23.398–34:26.427
Now , this is definitely an improvement over the preview version
34:26.437–34:29.305
that felt pretty undercooked and I was not at all impressed with
34:29.315–34:29.865
it .
34:29.875–34:32.663
But if we look back at some of the tests that were run in that
34:32.673–34:35.062
video , which is on the channel , we ran some of the same ones
34:35.072–34:35.622
today .
34:35.632–34:37.780
This is a significant improvement .
34:37.790–34:41.059
I would say though , I think the flash model impressed me more
34:41.069–34:43.618
than this , especially considering the size of it .
34:43.628–34:47.136
This has a bit more competition , but again , it is a big leap
34:47.146–34:49.854
on its predecessor , which I guess is what we like to see .
34:49.864–34:53.292
I was impressed a few times with some things like basically results
34:53.302–34:54.092
within results .
34:54.102–34:57.931
The cinematic Steve the PC repair game , it actually had some elements
34:57.941–35:00.181
to it that were like , hey , you know , it kind of thought this
35:00.191–35:00.744
through .
35:00.754–35:03.395
Just the way when they were speaking , their heads were like moving
35:03.405–35:04.198
back and forth .
35:04.208–35:05.805
Not the weird chin thing that was happening .
35:05.815–35:07.572
I'm not sure what happened there .
35:07.582–35:10.706
The music it included in both that game and then the Street Yeet
35:10.716–35:14.562
game were actually like they definitely fit the scene that it was
35:14.572–35:17.522
drawing here and it did a good job making the Street Yeet game very
35:17.532–35:20.384
cartoonish and the playability of it felt good ,
35:20.394–35:22.212
I guess , for lack of a better term .
35:22.222–35:26.579
The model that we got of the engine was actually it wasn't printable
35:26.579–35:32.898
.
35:32.898–35:32.978
So , unfortunately , maybe it failed to like check that box in
35:32.978–35:33.058
the specific task .
35:33.058–35:33.934
However , if we just look at one of the previews right here ,
35:33.944–35:35.616
it did a pretty decent job .
35:35.626–35:38.739
Definitely better than the other models I've tested this with .
35:38.749–35:41.941
Though , I have not tested this specific prompt with any big beefy
35:41.951–35:43.143
Frontier models yet .
35:43.153–35:44.744
So , keep that in mind as well .
35:44.754–35:46.186
But, it definitely has some chops here.
35:46.196–35:49.789
And the amount of research it did to actually get to that point
35:49.799–35:50.830
was quite a lot .
35:50.840–35:53.709
We have all the things in here downloading these images when it
35:53.719–35:54.510
doesn't have vision .
35:54.520–35:56.350
So , I'm not quite sure what to make of that .
35:56.360–35:59.870
But nonetheless , I think probably the biggest point of frustration
35:59.880–36:04.750
for me is it just goes all out when trying to actually ensure that
36:04.760–36:08.190
every aspect of every line of code it made is correct .
36:08.200–36:11.790
And with a model like this where it does not have vision ,
36:11.800–36:15.349
sometimes it takes longer and there is a cost thing to it as well
36:15.349–36:22.665
.
36:22.665–36:22.745
Assuming you're going to be paying to use these models ,
36:22.745–36:22.825
it just sometimes is simpler to have it make the result ,
36:22.825–36:24.510
not try to do like overviews and checks and follow ups .
36:24.520–36:26.830
you look at it yourself and then tell it what's wrong and tell
36:26.840–36:27.790
it to fix it .
36:27.800–36:31.469
Having to just go all out and it trying to fix it adds a lot of
36:31.479–36:35.790
time and expense and tokens to the task , which maybe is more of
36:35.800–36:38.876
but I find that's something I've consistently had frustrations
36:38.886–36:41.665
but I find that's something I've consistently had frustrations
36:41.675–36:45.100
with as well as with the flash model where if I'm having them make
36:45.110–36:47.894
like a say street yeet, it's easier if they just spit the script
36:47.904–36:50.371
out , I test it and say , Hey , the buildings are disappeared .
36:50.381–36:53.565
and then it fixes it rather than it trying to go through and like
36:53.575–36:56.280
well let me try to spin up a headless browser and then I'll take
36:56.290–36:59.555
a screenshot and like do a heat map because it can't actually see
36:59.565–37:00.911
it or some other function .
37:00.921–37:04.036
So it's a little token-happy.
37:04.046–37:06.925
I did run it on the highest potential reasoning mode .
37:06.935–37:10.859
So that could be attributed to it as well , but I still would like
37:10.869–37:14.471
to be able to do that without it like going all out to try to make
37:14.481–37:17.602
sure it's polishing every aspect of the script because I want max
37:17.612–37:18.164
performance .
37:18.174–37:19.769
I just don't need that level of depth .
37:19.779–37:22.979
So , that may be a personal complaint , but you know .
37:22.989–37:25.949
So , for the final thing , let's take a look at everything we spent
37:25.959–37:28.359
for the entirety of this test .
37:28.369–37:30.962
We can see right here that all cost 1 .
37:30.972–37:31.810
91 .
37:31.820–37:33.416
So , that was pretty darn cheap .
37:33.426–37:36.466
And you do have to remember this is going to be increased, their
37:36.476–37:37.749
API prices .
37:37.759–37:41.509
They say significant increase expected , but I mean a significant
37:41.519–37:45.071
increase might be like , okay , the entirety of this test was 8
37:45.081–37:48.000
or something , which is still in perspective not horrible .
37:48.010–37:48.881
The Qwen 2
37:48.891–37:52.645
8B max test video that we did a few weeks ago or like a week or
37:52.655–37:54.967
two ago that was about 32 .
37:54.977–37:56.490
So keep that in mind as well .
37:56.500–37:59.373
That may be a bit more performant than this , but there is going
37:59.383–38:02.257
to be a pricing consideration as well .
38:02.267–38:04.499
And this was very very very cheap .
38:04.509–38:08.023
It won't remain that cheap , but even if it gets five times as
38:08.033–38:09.625
expensive , it's still not horrible .
38:09.635–38:12.429
I suppose that's going to conclude our first look and test of the
38:12.439–38:16.834
0813 Deepseek V4 Pro model , the general availability version .
38:16.844–38:18.436
And this is now out of preview .
38:18.446–38:20.438
And that's just pretty exciting .
38:20.448–38:23.411
Basically, now I think I'm going to test Grok 3
38:23.421–38:25.280
because that also came out today.
38:25.290–38:27.767
And then I believe Friday Qwen 2
38:27.777–38:28.820
72B is coming out.
38:28.830–38:30.260
So that'll be exciting as well .
38:30.270–38:33.220
Seems like there's been this has been like a pretty busy week in
38:33.230–38:34.341
terms of model releases .
38:34.351–38:35.700
So that's cool also .
38:35.710–38:37.264
And when MuseTalk 1
38:37.274–38:38.974
2 comes out , that's going to be awesome .
38:38.984–38:41.449
I don't know how big it is , but I'm very excited to try to run
38:41.459–38:42.089
it locally .
38:42.099–38:44.884
So with that , if you have any questions , please feel free to
38:44.894–38:46.002
leave them in the comments .
38:46.012–38:48.000
And thanks for watching .
0:00.000–0:00.634
噴射而出。
0:00.644–0:05.632
我相信那是我曾經看過的東西。
0:05.642–0:08.524
這些樹木讓人聯想到泳池。
0:08.534–0:12.562
DeepSeek V4 Pro 的通用正式版剛剛發布。
0:12.572–0:16.604
這也發生在約12或更早時候的V4 Flash模型上
0:16.614–0:19.161
13 天前的 07/31 發布 V4 Flash 模型的情況相同。
0:19.171–0:23.078
現在，該模型在測試中被證明具有顯著的
0:23.088–0:26.516
在先前預覽版的基礎上，能力有了顯著的提升
0:26.526–0:27.154
左右的預覽版本。
0:27.164–0:29.793
今天我們將要找出是否同樣的情況
0:29.803–0:34.292
適用於 V4 Pro 新更新的通用正式版檢查點
0:34.302–0:35.870
命名為 0813。
0:35.880–0:38.426
所以，在我們深入探討之前，請隨時訂閱，以便我們
0:38.436–0:40.824
能達到 70K，然後是 100K，並獲得獎牌。
0:40.834–0:43.861
就此而言，基本上現在沒什麼好說的，因為它似乎
0:43.871–0:46.693
現在因為它似乎已經在無聲無息中發布
0:46.693–0:48.780
。
0:48.780–0:49.535
目前 X 上還沒有官方公告帖子，
0:49.545–0:52.331
我認為通常我們會在那裡看到這個。
0:52.341–0:55.455
但他們確實只在網站上提到了一些參考
0:55.465–0:59.385
如果在此處使用這個模型識別碼 V4
0:59.395–1:03.582
Pro，您將被導向至新的 V4 Pro 0813 版本。
1:03.592–1:06.842
此外，它也在 OpenRouter 上線，這
1:06.852–1:10.103
讓我感到安心，因為它確實已經正式發布，
1:10.113–1:11.977
而且這也是我們將要使用的模型。
1:11.987–1:14.260
因此，根據我們目前掌握的部分資訊，
1:14.270–1:17.520
除了模型識別碼現在顯示為今天的日期之外，
1:17.530–1:20.812
取決於使用者的地區設定，我們可以得知該模型擁有
1:20.822–1:24.764
一百萬的上下文視窗，最大輸出長度為 384K，
1:24.774–1:26.135
這確實是非常大的數量。
1:26.145–1:28.947
現在，DeepSeek 有時可能會出現的情況是，
1:28.957–1:31.819
至少從歷史紀錄來看，可能會產生大量的思考過程。
1:31.829–1:34.211
現在，這確實支援可選的思考模式，
1:34.221–1:36.045
正如我們點擊此處所見。
1:36.055–1:38.120
因此，可以將這種情況稍微降低，
1:38.130–1:40.991
但為了展現最大潛力的效能
1:41.001–1:44.501
今天所有的程式碼任務都將以
1:44.511–1:45.697
最大努力模式執行。
1:45.707–1:49.047
現在，我想謹慎地提及以下事項，因為
1:49.057–1:51.839
我無法確定地確認這是事實真相，
1:51.849–1:54.072
但我猜測它很可能就是。
1:54.082–1:57.261
網路上流傳著一些參考資料和照片，顯然
1:57.271–1:59.574
來自與 DeepSeek 相關的群組。
1:59.584–2:02.821
據他們表示，那是一個關於 DeepSeek 的微信群組，或是類似的
2:02.821–2:08.747
。
2:08.747–2:08.827
如果這個說法可信，我們可以假設這些是
2:08.827–2:11.177
DeepSeek V4 Pro 0813 或一般
2:11.187–2:12.448
可用性模型的基準測試結果。
2:12.458–2:15.625
基本上，我主要感興趣的是，從 V4 Pro 預覽版到目前版本，
2:15.635–2:20.225
能力上的巨大躍進，因為我最初對 V4 Pro 預覽版以及 V4 Flash 預覽版
2:20.235–2:24.272
。
2:24.272–2:28.713
它們感覺非常不成熟，可能因為它們被命名為
2:28.713–2:28.793
它們感覺非常未熟，可能因為它們被命名為
2:28.793–2:28.873
這其實很合理。
2:28.873–2:29.509
這其實還算公平。
2:29.519–2:30.381
但再次強調，我想重申的是，從
2:30.391–2:33.273
Flash 到 0731 一般可用性版本的 Flash，
2:33.283–2:37.518
Flash，然後0731的Flash一般可用性版本是
2:37.528–2:38.753
這就像是目前最適合雙 3090
2:38.763–2:42.495
這就像是目前雙3090設置下最好的本地模型
2:42.505–2:42.914
設定。
2:42.924–2:48.182
所以我的意思是，在一個小眾領域裡，你知道，但即便如此，如果這種躍升能力
2:48.192–2:51.870
從12增加。
2:51.880–2:53.058
從12增加。
2:53.068–2:54.571
.8 增加到 62
2:54.581–2:58.218
.7，那麼 hypothetically（假設性地）會將它置於像 Opus 3
2:58.228–3:02.046
.8 和 8 之上，並使其非常非常有競爭力，尤其是因為這個
3:02.056–3:04.286
是一個開放權重的模型，或者將成為開放權重模型。
3:04.296–3:08.276
這也極度令人興奮，因為這不像考慮到像 Kimi K3 或 Qwen
3:08.286–3:12.098
考慮到像Kimi K3或Qwen這樣的事物，我猜這是巨大的。
3:12.108–3:12.470
.2
3:12.480–3:13.291
那樣恐怖地巨大，我想。
3:13.301–3:14.587
它仍然很大，有 1
3:14.597–3:15.753
六兆個參數
3:15.763–3:19.089
它是一個混合專家模型，有多少活躍參數？
3:19.099–3:22.107
所以如果我們估算這會和
3:22.117–3:24.013
V4 Pro 的預覽版本大小相同。
3:24.023–3:24.650
它是 1
3:24.660–3:28.490
.6 兆參數的混合專家模型，有 490 億
3:28.500–3:28.970
活躍參數。
3:28.980–3:32.810
所以它仍然非常非常大，但並沒有像接近3兆參數的模型那樣龐大。
3:32.820–3:35.284
關閉接近3兆參數的模型。
3:35.294–3:38.623
因此，在定價方面，這也是我們需要
3:38.633–3:41.962
基本上就照現在這樣看待的事情，因為他們提到
3:41.972–3:44.985
並且在相當長一段時間內表示，他們的API定價
3:44.995–3:47.289
很快就會大幅調漲。
3:47.299–3:50.469
現在「大幅」是相對於它目前的價格，也就是
3:50.479–3:52.536
非常非常非常便宜。
3:52.546–3:56.036
所以，目前DeepSeek每百萬輸入記號的定價
3:56.046–3:59.250
V4 Pro每百萬輸入記號的定價是0.43
3:59.260–4:00.080
[未翻譯]
4:00.090–4:01.178
所以，好吧。
4:01.188–4:04.652
不，我知道這是什麼，但就像上次我搞錯這
4:04.662–4:08.456
點，我說它是2美分，但事實並非如此，然後我被轟炸
4:08.456–4:08.966
。
4:08.966–4:14.347
所以，這是每百萬輸入記號43美分，每百萬輸出記號
4:14.357–4:16.433
87美分是目前的價格。
4:16.443–4:19.162
現在，如果這會像大幅上漲，
4:19.172–4:20.873
假設它上漲了五倍或十倍，
4:20.873–4:22.452
它仍然在假設上
4:22.462–4:23.335
就會具備競爭力。
4:23.345–4:26.063
所以，這一切都取決於它目前的成本。
4:26.073–4:28.793
基於此，我將從 Cursor 內部使用這個功能。
4:28.793–4:30.524
。
4:30.524–4:33.137
在我們測試 V4 Flash 模型於 7 月 31 日正式發布時，
4:33.147–4:36.656
我也透過 Cursor 使用它，因為我認為他們建議這麼做，
4:36.666–4:39.505
他們曾建議這麼做，而且效能非常
4:39.515–4:40.850
非常好。
4:40.860–4:41.879
我對它很滿意。
4:41.889–4:44.381
所以，我也將使用 Cursor。
4:44.391–4:49.124
現在我還應該提到，他們在測試版中發布了 Cursor 或 ChatGPT 的桌面
4:49.134–4:51.270
用於 Cursor 或 ChatGPT 的應用程式。
4:51.280–4:54.288
所以我也將它路由並透過這個桌面應用程式運作，
4:54.298–4:57.142
它正在為我們用 C++ 製作一個滑板遊戲。
4:57.152–4:59.099
我預計這會花費相當多的時間。
4:59.109–5:01.791
所以我只想先啟動它，並驗證這個應用程式在這個系統上
5:01.801–5:04.482
是否正常運作，看起來確實是正常的。
5:04.492–5:09.430
所以請注意，此處任何提及 GPT 或 Cursor 的地方，都是直接路由
5:09.440–5:13.030
到 DeepSeek，使用 0813 模型的極限推理能力。
5:13.040–5:15.355
基於此，讓我們來看看我們的 好的，
5:15.365–5:21.369
所以顯然，OpenAI 的 beta 應用程式被設定為所有檔案類型的預設程式。
5:21.379–5:23.501
這真是令人氣憤。
5:23.511–5:24.525
太令人氣憤了。
5:24.535–5:25.122
好的。
5:25.132–5:26.060
好的。
5:26.070–5:27.355
非常 Gen 3
5:27.365–5:28.772
美學風格
5:28.782–5:31.642
我要開啟聲音，因為我確實看到有人提到它
5:31.652–5:34.833
只是在談論一些聲音。
5:34.843–5:35.071
好的。
5:35.081–5:37.146
它從我們的記事本開始。
5:37.156–5:38.900
我們可以調整這個視窗的大小嗎？
5:38.910–5:39.618
好的。
5:39.628–5:40.257
可以接受。
5:40.267–5:41.293
最小化。
5:41.303–5:42.410
還原最小化。
5:42.420–5:42.968
很好。
5:42.978–5:43.686
重點來了。
5:43.696–5:45.281
有右鍵功能嗎？
5:45.291–5:48.720
有，有的。
5:48.730–5:51.898
我的本地時間和日期也正確顯示在底部
5:51.908–5:52.137
沒錯
5:52.147–5:55.011
而這裡有我認為可能是特別的功能
5:55.021–5:55.410
特色
5:55.420–5:57.326
這是全息桌面功能
5:57.336–5:59.002
所以我們把它留到最後再介紹
5:59.012–6:02.275
我們先來看看檔案應用程式，我從來沒見過這樣的檔案圖示
6:02.285–6:04.909
我好像以前沒看過這樣的檔案圖示
6:04.919–6:05.309
好的
6:05.319–6:08.422
應用程式背景有一些漂亮的透明效果
6:08.432–6:11.462
而且確實有一個很棒的合成波主題，我個人很喜歡
6:11.462–6:15.197
。
6:15.197–6:15.277
所以，不幸的是，我們的假檔案系統似乎是靜態的
6:15.277–6:15.465
我們無法瀏覽它
6:15.475–6:17.866
不過，我們能在裡面建立新東西嗎？
6:17.876–6:18.587
好的，我們可以
6:18.597–6:21.228
我們可以移除我們建立的新東西嗎？
6:21.238–6:22.749
我們也可以
6:22.759–6:25.951
我們可以移除預設就存在的东西嗎？
6:25.961–6:26.591
是的，我們可以
6:26.601–6:27.231
好的，很棒。
6:27.241–6:28.431
這讓我稍微安心一點。
6:28.441–6:30.431
然後「向上」就會直接移到那裡。
6:30.441–6:30.990
哦，好的。
6:31.000–6:32.590
還不錯。
6:32.600–6:33.663
沉浸感還算不錯，
6:33.663–6:35.310
至少在目錄結構方面，
6:35.320–6:37.498
如果這是一台真實的電腦，這些目錄應該會存在。
6:37.508–6:38.833
我們確實看到了一些記事本，
6:38.833–6:40.284
因為它開啟時顯示了這個
6:40.284–6:41.716
只是為了提供一些資訊給我們。
6:41.726–6:42.991
我只希望看到
6:42.991–6:44.203
它是否會將檔案儲存為系統上的文字檔。
6:44.213–6:45.238
沒有這樣的目錄。
6:45.248–6:46.988
我不是已經刪除文件目錄了嗎？
6:46.998–6:49.138
我不是刪除了文件目錄嗎？
6:49.148–6:50.252
等一下
6:50.262–6:51.209
哦，它記得我做了那個動作。
6:51.219–6:53.040
喔，它記得我這麼做了。
6:53.050–6:54.203
我可能搞砸了。
6:54.213–6:55.410
所以這部分
6:55.410–6:56.084
算是我的錯。
6:56.094–6:57.132
呃，好的。
6:57.142–7:00.112
這裡的顯示根本不對。
7:00.122–7:03.495
它不應該出現可見的文字方塊。
7:03.505–7:04.945
這種情況偶爾會發生。
7:04.955–7:06.637
Help 確實給了我們一些東西。
7:06.647–7:09.420
我們就來執行 neofetch 吧。
7:09.430–7:13.636
那裡的 neofetch 其實做得相當不錯。
7:13.646–7:16.299
Neofetch 或 Neofetch。
7:16.309–7:17.630
[未翻譯]
7:17.640–7:19.674
不過，那其實還算可以。
7:19.684–7:21.545
我只是有點嚴苛。
7:21.555–7:22.653
記憶體。
7:22.663–7:24.355
這還行。
7:24.365–7:27.928
哦，它其實在這裡加了一點巧思。
7:27.938–7:28.552
一
7:28.562–7:30.142
從無限中取出 44 MB。
7:30.152–7:30.822
這還行。
7:30.832–7:33.854
這不就是老式軟碟片的容量嗎？
7:33.864–7:35.914
我不知道，因為那些比我還老。
7:35.924–7:37.735
接下來是計算機。
7:37.745–7:44.720
好的，所以 47 47 47 8 8 8 376。
7:44.730–7:46.636
我好久沒做這種事了。
7:46.636–7:47.424
除以六。
7:47.434–7:51.519
好的，設定。
7:51.529–7:54.188
我確實看到了一些類似 Qwen 2.5 8B 的東西。
7:54.198–7:54.808
在風格方面。
7:54.808–7:57.171
就這裡的一些風格而言。
7:57.181–7:58.848
那是什麼？
7:58.848–8:01.348
為什麼我喜歡它？
8:01.358–8:01.921
曙光。
8:01.931–8:03.068
哦，好的。
8:03.078–8:06.262
日落城市。
8:06.272–8:09.456
這沒人會懂這個梗
8:09.466–8:10.734
這看起來像某個
8:10.734–8:12.405
我高中時的一個朋友是
8:12.415–8:13.796
作為遊戲專案製作的東西。
8:13.806–8:15.378
它實際上看起來像很多
8:15.378–8:17.564
就像他專案的啟動畫面
8:17.564–8:18.433
。
8:18.433–8:21.413
我們也支援自訂口音。
8:21.423–8:24.117
難度有簡單、普通和困難。
8:24.127–8:25.100
喔，是遊戲。
8:25.110–8:26.628
好的，我可能把它設為簡單
8:26.628–8:28.146
因為你可能熟悉我
8:28.156–8:29.912
容易生氣的遊戲風格。
8:29.922–8:31.862
接下來，讓我們試試我們的遊戲。
8:31.872–8:33.384
揚聲器已開啟。
8:33.394–8:37.354
所以，城市劫案。
8:37.364–8:44.554
喔，你真是個。
8:44.564–8:44.948
好的。
8:44.958–8:47.711
所以，你必須按 Enter 鍵開始
8:47.711–8:49.678
而不是點擊按鈕
8:49.678–8:50.976
。
8:50.976–8:52.446
請切換檔位。
8:52.456–8:53.827
啊，好的。
8:53.837–9:00.337
好的。
9:00.347–9:00.830
好吧。
9:00.840–9:11.955
它是喔，真酷。
9:11.965–9:13.847
哦，好的。
9:13.857–9:16.065
有點麻煩，
9:16.065–9:21.486
但像是啪嗒聲之類的音效
9:21.486–9:23.407
有點好笑。
9:23.417–9:25.160
像，好吧。
9:25.170–9:25.972
我們搞砸了
9:25.972–9:27.508
而且我們只拿到了100個紅色的
9:27.508–9:28.443
總共一千個，
9:28.453–9:29.466
但它確實起作用了。
9:29.476–9:30.168
所以，請記住
9:30.168–9:32.127
我們需要按 Enter 鍵來玩這些遊戲
9:32.127–9:32.300
並且
9:32.310–9:34.740
不要按寫著「啟動」的按鈕。
9:34.750–9:36.098
哦。
9:36.108–9:43.897
哦，好吧。
9:43.907–9:45.950
我要通過這些嗎？
9:45.960–9:46.272
是的。
9:46.282–9:48.364
我只是根據聲音假設。
9:48.374–9:51.019
我想知道這是否就是狗的方式，因為它們不理解
9:51.029–9:51.744
語言。
9:51.754–9:54.640
但對我來說，那聽起來像是正向強化的聲音
9:54.640–9:56.493
。
9:56.493–9:56.573
所以，這讓我更想去做。
9:56.573–9:59.388
我認為我們剛剛發現了關於狗腦運作方式的某些事情
9:59.398–10:01.803
。
10:01.813–10:03.654
好的，分數。
10:03.664–10:07.194
我不認為我們需要這麼多小數位來顯示
10:07.204–10:07.677
分數。
10:07.687–10:12.424
我認為也許四捨五入到734或向下取整都是可以接受的
10:12.434–10:16.455
，但我讚賞它的複雜性。
10:16.465–10:17.665
然後，哦，讀給我聽。
10:17.675–10:20.891
好的，所以我們的特色功能類似於全息投影。
10:20.901–10:23.794
而且我想象一下，當我們觸發時，只打開幾個應用程式
10:23.804–10:25.408
時，打開幾個應用程式可能是正確的作法。
10:25.418–10:27.747
所以，哦，是的。
10:27.757–10:28.149
好吧。
10:28.159–10:30.248
所以，它只是呈現出一種有趣的卡片效果。
10:30.258–10:31.457
這就是它的樣子。
10:31.467–10:34.199
這看起來像是一個Apple Vision Pro的應用程式演示。
10:34.209–10:37.647
這是從Vision Pro頭戴裝置內部看到的樣子。
10:37.657–10:42.548
嗯，如果你用過的話，可能會對我所說的話產生共鳴。
10:42.558–10:43.655
我說的話。
10:43.665–10:44.208
好的。
10:44.218–10:46.025
整體來說，還算不錯。
10:46.035–10:46.973
一切都運作正常。
10:46.983–10:48.159
沒有什麼真正讓我驚艷的。
10:48.169–10:51.399
但再次強調，這可能是測試的局限性，而非模型本身的局限性。
10:51.409–10:52.506
模型的局限性。
10:52.516–10:55.131
所以唉，嗯，這還挺有趣的。
10:55.141–10:56.052
我們會繼續進行。
10:56.062–10:58.312
所以我們的滑板遊戲仍在進行中。
10:58.322–11:02.097
同樣地，我使用的是新發布的 beta 版 ChatGPT 桌面應用程式。
11:02.107–11:03.985
的 ChatGPT 桌面應用程式。
11:03.995–11:06.286
我預計這會花費相當多的時間。
11:06.296–11:09.486
所以讓我們來看看，就像「哦不」，因為我們已經看過了成本。
11:09.496–11:10.202
成本。
11:10.212–11:13.929
請記住，這部分費用涵蓋了滑板遊戲以及瀏覽器作業系統。
11:13.939–11:14.717
瀏覽器作業系統。
11:14.727–11:17.655
所以兩項獨立任務總共只花了 20 美分。
11:17.665–11:18.945
其中一項任務仍在進行中。
11:18.955–11:19.231
好了。
11:19.241–11:24.748
所以，經過 68 分 24 秒後，Skate NYC 任務完成了。
11:24.758–11:26.110
所以，喔，你這個笨蛋。
11:26.120–11:30.886
你用這些截圖洩露了劇情，你甚至沒有視覺功能。
11:30.886–11:31.396
。
11:31.396–11:32.915
我氣炸了。
11:32.925–11:36.637
不過，既然我們已經知道遊戲長什麼樣子，
11:36.647–11:40.152
我必須克制自己不說些更髒的話。
11:40.152–11:40.747
。
11:40.747–11:44.147
讓我們來看看遊戲吧。
11:44.157–11:47.331
好了，這就是我們的紐約滑板遊戲。
11:47.341–11:48.605
應該在輸出結果裡對吧？
11:48.615–11:51.232
對。
11:51.242–11:52.665
好了。
11:52.675–11:56.804
它沒有 Qwen 2.5
11:56.814–12:04.335
模型那麼好，但那個模型確實多了 8000 億個參數。
12:04.335–12:04.845
。
12:04.845–12:07.101
地面上寫的是「NYC skate」嗎？
12:07.111–12:08.415
它是倒著的，但確實有寫。
12:08.425–12:10.798
而且上面寫著「成立於 1999 年」。
12:10.808–12:11.175
好的。
12:11.185–12:12.280
哎呀。
12:12.290–12:16.396
聽起來像是范·海倫樂團歌曲的開頭。
12:16.406–12:18.840
我不確定這到底是什麼。
12:18.850–12:19.773
好的。
12:19.783–12:22.766
呃，我們也許別再做同樣的事了。
12:22.776–12:24.020
不過還是有音效。
12:24.030–12:25.431
我通常聽不到音效。
12:25.441–12:27.626
我們真的能移動嗎？
12:27.636–12:28.097
好的。
12:28.107–12:29.273
還不太行。
12:29.283–12:32.176
我會說這並不怎麼好。
12:32.186–12:33.533
不過，這其實有點酷。
12:33.543–12:37.247
它將遊戲標題風格化的方式。
12:37.257–12:37.747
不錯。
12:37.757–12:40.675
不錯。
12:40.685–12:43.890
好吧，我們不需要做鼓組模擬。
12:43.900–12:53.313
老實說，這些樹讓人有點聯想到糞便。
12:53.323–13:02.090
不過沒關係，那些腿看起來像是在踢。
13:02.100–13:05.018
這可能是我們看過的最好笑的滑板遊戲結果，
13:05.028–13:06.283
這算是……某種成就吧。
13:06.293–13:08.419
好了，這裡有個路牌。
13:08.429–13:12.690
文字有點側著，建築立面也是。
13:12.700–13:14.589
這絕對是個有問題的結果。
13:14.599–13:17.436
不過，我還挺喜歡噴泉噴射出來的效果。
13:17.446–13:21.988
我相信我之前有看過類似的畫面。
13:21.998–13:22.894
好的。
13:22.904–13:23.636
好了。
13:23.646–13:26.150
我想我在 Qwen 的結果裡有看過這個。
13:26.160–13:29.357
現在，我想快速做一下，作為一些參考
13:29.357–13:30.327
。
13:30.327–13:33.987
這是我們從 C++ 滑板測試中收到的結果，來自
13:33.997–13:36.620
DeepSeek V4 的預覽版收到的 C++ 滑板測試結果。
13:36.630–13:37.659
我們可以直接在這裡看到。
13:37.669–13:39.738
我要把聲音關掉。
13:39.748–13:41.258
這相當糟。
13:41.268–13:45.257
糟糕到讓我甚至記得要特別去拿這個東西來展示
13:45.267–13:46.777
只是為了展示它
13:46.787–13:51.256
所以，根據我們現在看到的，能力的躍進
13:51.266–13:54.850
以及剛才在滑板遊戲中它為我們創造的結果
13:54.860–13:58.374
從這些截圖中可以看出，這仍然
13:58.384–13:59.254
相當顯著
13:59.264–14:01.653
接下來，我將對它進行 Subway FPS 測試。
14:01.663–14:03.653
我也在提示詞的末尾添加了內容。
14:03.663–14:05.492
不用擔心檢查腳本的問題
14:05.502–14:08.132
我會手動檢查，如果有問題，我會回報
14:08.142–14:11.491
給你，因為這確實會將每項結果
14:11.501–14:16.228
在時間長度上，單一結果大約兩倍
14:16.238–14:19.808
只是出於我想在5秒內視覺化呈現的東西
14:19.818–14:21.402
秒內視覺上完成某事，然後說「嘿，這不起作用」。
14:21.412–14:24.671
所以我們這樣做，可能是為了讓事情稍微
14:24.681–14:25.148
緊湊一些。
14:25.158–14:28.257
好的，它完成了 Subway FPS 遊戲。
14:28.267–14:31.127
問題在於它不斷嘗試進行類似故障排除的操作，
14:31.137–14:34.635
不知為何，它在這個系統上並不擅長此類操作。
14:34.645–14:37.302
好的，讓我們來看看我們的 Subway FPS。
14:37.312–14:39.101
主街大道站。
14:39.111–14:42.547
好的，我要打開揚聲器。
14:42.557–14:44.063
所以，它給了我們兩個選項。
14:44.073–14:46.536
我們可以選擇探索車站。
14:46.546–14:50.525
好，它……嘿，你知道嗎？
14:50.535–14:54.913
我認為這是我看過畫得最好的地鐵車廂
14:54.923–14:56.269
這次測試中看過畫得最好的地鐵車廂。
14:56.279–14:58.623
而你可能會說，嗯，它看起來並沒有那麼好。
14:58.623–15:00.112
。
15:00.112–15:00.192
我知道。
15:00.192–15:01.534
只是他們從來沒有真正包含這些。
15:01.544–15:03.769
而當他們包含時，通常也不是很好。
15:03.779–15:06.801
所以這……實際上，那是什麼？
15:06.811–15:10.870
如果那個東西開始移動，我可能會變得比
15:10.880–15:15.177
你認為回應這種情況所合理的興奮程度還要興奮。
15:15.177–15:15.746
。
15:15.746–15:16.853
這非常乾淨。
15:16.863–15:17.490
很簡單。
15:17.500–15:20.044
老實說，這沒什麼讓我驚艷的
15:20.054–15:20.284
你。
15:20.294–15:21.560
皇后大道線。
15:21.570–15:21.880
好的。
15:21.890–15:22.677
我們有樓梯。
15:22.687–15:25.150
我們有一些照明元素之類的東西。
15:25.160–15:28.341
如果想去那邊的話，我們有出口。
15:28.351–15:34.407
然後還有隧道，像是列車的指示燈、垃圾桶。
15:34.417–15:37.121
浮動粒子更簡單。
15:37.131–15:38.799
我們不能走上樓梯，這沒關係。
15:38.809–15:39.677
我們也不想上去。
15:39.687–15:43.909
現在，當然，這裡包含的另一個功能是
15:43.919–15:45.106
幀率（FPS）。
15:45.116–15:48.381
所以，讓我們回到選單，然後進入生存模式的幀率。
15:48.391–15:50.377
讓我們看看它是否改變了燈光或其他東西。
15:50.387–15:51.336
天啊。
15:51.346–15:53.012
那個武器模型不錯。
15:53.022–15:54.370
有嗎？
15:54.380–15:54.769
怎麼說？
15:54.779–15:57.484
哦，確實有。
15:57.494–15:59.480
我想是。
15:59.490–16:01.237
是的，確實有。
16:01.247–16:02.595
現在看起來很順暢。
16:02.605–16:03.951
我看不到任何敵人。
16:03.961–16:05.149
看看我們跑起來的時候。
16:05.159–16:07.865
就像在搖擺。
16:07.875–16:08.263
好的。
16:08.273–16:12.895
所以，嗯，我現在注意到的一大重點是這裡沒有殭屍
16:12.905–16:13.454
敵人。
16:13.464–16:17.047
所以，重新裝填得不錯。
16:17.057–16:22.877
那搖擺的動作有點好笑。
16:22.887–16:24.314
很奇怪。
16:24.324–16:24.792
嘿，看。
16:24.802–16:26.471
實際上會有彈殼掉落。
16:26.481–16:27.672
這有點酷。
16:27.682–16:30.235
好的，這就像……是啊，有點平庸。
16:30.245–16:33.279
所以，我會給它一些反饋。
16:33.289–16:36.162
現在，我把它停下來，因為它看起來像是已經完成了，
16:36.172–16:38.925
而且感覺我無法在無頭環境中使用它來檢查
16:38.925–16:40.891
。
16:40.891–16:41.608
在這種情況下，直接檢查並告訴它哪裡出錯會更容易。
16:41.618–16:42.809
錯誤。
16:42.819–16:44.332
所以，我會給它誠實的反饋。
16:44.342–16:48.255
所以，據說我們修復並優化過的《地鐵射擊遊戲》
16:48.265–16:49.376
版本已經完成了。
16:49.386–16:51.138
讓我們來看看它吧。
16:51.148–16:52.584
聲音已開啟。
16:52.594–16:53.460
好的。
16:53.470–16:54.013
很有趣。
16:54.023–16:56.460
我認為它交換了按鈕的位置。
16:56.470–17:00.130
所以，好的。
17:00.140–17:01.015
七個敵方目標。
17:01.025–17:02.626
它顯示我在右上角。
17:02.636–17:07.212
不幸的是，我無法找到它們。
17:07.222–17:09.867
我原本預期它會表現得更好。
17:09.877–17:12.362
同樣地，這是我們可以回頭參考
17:12.372–17:14.615
該模型先前版本的情況。
17:14.625–17:18.639
所以，基本上，預覽版本的細節程度
17:18.649–17:22.583
遠低於這個版本。
17:22.593–17:26.043
雖然不幸的是我們沒有敵人，而且考慮到
17:26.053–17:29.251
所花費的時間，以及我多次告訴它
17:29.261–17:32.375
沒有敵人的情況，我很遺憾地將這個版本列為下一個。
17:32.385–17:36.126
現在我給它一個困難的任務，要創建一個
17:36.136–17:40.866
RB26 直列六缸引擎的 3D 模型，它必須是可以
17:40.876–17:44.917
以極少或無需支撐結構進行 3D 列印，
17:44.927–17:48.400
並且能容納一個稱為 N20 馬達的小型直流馬達，因為這是我曾告訴它
17:48.410–17:51.430
應該能跟我的遙控車專案配合。其實，我不認為
17:51.440–17:53.662
這在提示詞裡有特別提到，但這就是重點
17:53.672–17:57.169
所在，它應該會理解，並會將它
17:57.179–17:58.524
正確地縮放至該馬達的尺寸。
17:58.534–18:02.031
到目前為止，我用過幾個智慧模型來執行這項測試，我從來
18:02.041–18:04.263
沒有真正收到任何值得列印的東西
18:04.263–18:07.577
。
18:07.577–18:07.657
所以，希望這能盡快改變。
18:07.657–18:10.559
所以，我們的 3D 模型任務相當有趣，因為它只是
18:10.569–18:14.068
針對該引擎上特定物品的實際正確方向
18:14.078–18:16.140
這引擎上的特定物品。
18:16.150–18:18.777
例如，排氣管（downpipe）在哪裡？那是排氣組件嗎？
18:18.787–18:19.656
它在哪一側？
18:19.666–18:20.775
進氣側在哪裡？
18:20.785–18:21.975
排氣側在哪裡？
18:21.985–18:23.493
進氣歧管（Intake plenum）是在左側還是右側？
18:23.503–18:28.049
所以，我很喜歡看到這種細緻研究的程度，
18:28.059–18:30.607
這些研究都匯聚在這個模型創作的成果中。
18:30.617–18:34.923
如果我們能為這項任務獲得一個
18:34.933–18:36.280
真正可以列印的模型，那將非常棒。
18:36.290–18:37.480
這將是第一個。
18:37.490–18:41.816
因此，在1小時10分鐘後，我們收到了3D模型的回覆。
18:41.816–18:44.543
。
18:44.543–18:45.914
這是一段非常、非常複雜的文字，描述它所生成的東西。
18:45.924–18:46.717
它生成的東西。
18:46.727–18:49.288
現在，當然，另一部分就是實際上去看看它長什麼樣子。
18:49.298–18:51.135
看看它長什麼樣子。
18:51.145–18:53.626
不過，一旦我記起我給這個資料夾取的名字，
18:53.636–18:54.991
我們就能夠偷偷看一眼。
18:55.001–18:56.356
組裝顯示器STL。
18:56.366–18:58.364
我猜應該是這樣沒錯。
18:58.374–18:59.489
請爭點氣。
18:59.499–19:03.909
我真的，呃，我需要一個好的來做這些。
19:03.919–19:06.158
實際上，它是側著的。
19:06.168–19:06.880
見鬼。
19:06.890–19:10.335
這做得相當好。
19:10.345–19:15.556
這實際上是目前為止最好的。
19:15.566–19:17.244
它並不完美。
19:17.254–19:21.100
不過，我們有一些渦輪增壓器側向的問題。
19:21.110–19:22.383
我的意思是，這還行。
19:22.383–19:23.991
讓我們在 OpenSCAD 裡來看看這個
19:23.991–19:24.677
。
19:24.677–19:26.320
這裡有好幾個單獨的零件。
19:26.330–19:28.907
所以，我們也想確認一下，我們告訴它需要
19:28.917–19:31.273
塞進那個小型遙控車引擎裡。
19:31.283–19:34.634
讓我們看看它有沒有機會正確地裝進去。
19:34.634–19:35.340
研究——
19:35.350–19:39.111
這裡有個研究資料夾，裡面顯示了
19:39.121–19:41.772
這個馬達的各種圖片和數據。
19:41.782–19:42.660
真的很酷。
19:42.670–19:46.392
這裡我們有一些更好的預覽渲染圖。
19:46.402–19:50.501
這絕對做得很好。
19:50.511–19:52.041
渦輪側面。
19:52.051–19:52.760
好的。
19:52.770–19:53.966
底部。
19:53.976–19:54.850
好吧。
19:54.860–19:58.466
這不太可能輕易地用 3D 列印出來
19:58.466–19:58.976
。
19:58.976–20:01.360
不過，我想我無法做出這樣的判斷，因為這個
20:01.370–20:02.968
是由多個零件組成的。
20:02.978–20:06.344
所以，它並非設計成我們最初看到的那樣，作為一個整體來列印。
20:06.354–20:07.548
像這樣的一整塊。
20:07.558–20:09.719
它是設計成分開列印的。
20:09.729–20:11.005
我想看看。
20:11.015–20:14.381
現在讓我們只看單獨的第一個零件，
20:14.391–20:18.720
這正是理想情況下放置 3D 列印馬達的位置，
20:18.730–20:21.454
馬達將安裝在裡面。
20:21.464–20:25.473
不幸的是，在我們目前看到的任何潛在方向中，
20:25.483–20:28.526
不使用支撐結構來列印這部分根本不可行，
20:28.536–20:29.892
這有點令人沮喪。
20:29.902–20:32.626
這看起來似乎有這些柱子和孔洞，
20:32.636–20:35.438
以便卡扣安裝到其他零件上，這很酷。
20:35.448–20:36.243
好的。
20:36.253–20:36.563
是的。
20:36.573–20:39.217
我認為那些絕對像是附件安裝點，
20:39.227–20:40.262
這真的很酷
20:40.272–20:43.717
這個幾乎不需要任何支撐結構就能列印。
20:43.727–20:46.589
這個零件基本上可以不用任何支撐結構來列印。
20:46.599–20:50.712
然後這上面應該會有 Nissan 的標誌，或者至少寫上「Nissan」的文字，
20:50.722–20:53.420
印在上面。
20:53.430–20:53.883
沒錯。
20:53.893–20:55.741
這一個做得相當不錯。
20:55.751–20:56.668
這裡沒有問題。
20:56.678–20:57.781
油底殼。
20:57.791–20:58.262
喔，不。
20:58.272–20:58.662
好吧。
20:58.672–21:01.785
這裡我們決定不再嘗試列印這個。
21:01.795–21:02.666
試著列印它。
21:02.676–21:07.391
我不會失望，但也沒有特別驚艷。
21:07.401–21:11.315
所以，我想在評價這個時，我算是態度中立。
21:11.325–21:14.198
絕對地，如果我們只把它當作一張渲染圖來看，
21:14.208–21:16.842
甚至別去管那裡面的個別零件，
21:16.852–21:20.225
因為實際上列印這個並不可行，
21:20.235–21:22.963
它只是組裝起來就做得相當不錯了。
21:22.973–21:25.542
我認為這項研究投入，大概就像
21:25.552–21:29.328
我討厭在這裡聽起來像個AI，但故事不是
21:29.338–21:31.339
模型，而是背後的研究。
21:31.349–21:33.255
那就像是一般AI生成的垃圾內容。
21:33.265–21:37.085
它確實做了很多不錯的研究。
21:37.095–21:41.075
我的意思是，它找到並下載這個只是為了看看嗎？
21:41.085–21:41.474
不。
21:41.484–21:41.793
好的。
21:41.803–21:45.464
所以，這可能是它自己生成的，也可能是從網路上找到的
21:45.474–21:48.616
以便更好地評估這個資產的大小
21:48.626–21:51.768
，但它非常謹慎，以確保它嘗試生成的內容
21:51.778–21:55.519
在假設上符合原始提示中指定的具體規格。
21:55.529–21:57.674
原始提示中指定的具體規格。
21:57.684–21:59.669
所以，還不錯。
21:59.679–22:02.621
我還要對這個進行簡單的前端設計測試，
22:02.631–22:04.377
它需要創建手錶網站。
22:04.387–22:07.643
所以，它必須具備相當電影感的3D橫掃效果。
22:07.653–22:09.222
手錶資產進行相當有電影感的平移效果。
22:09.232–22:12.538
我特別標註了那點
22:12.548–22:14.197
手錶必須是 3D 模型。
22:14.207–22:17.039
所以，我們來看看它整合這一切的
22:17.049–22:17.513
能力如何。
22:17.523–22:21.817
所以，在 1 小時 30 分鐘後，我們得到了手錶前端結果
22:21.817–22:22.327
。
22:22.327–22:25.803
它再次似乎有一個相當實質的，就像只是列出了這裡
22:25.813–22:26.769
所構建的內容清單。
22:26.779–22:28.460
那麼，讓我們來看看它的情況如何。
22:28.470–22:32.728
一個精美的手錶前端網站。
22:32.738–22:36.096
實際上，它已經非常接近優秀了。
22:36.106–22:40.745
我認為最大的問題在於，手錶錶帶突出來了。
22:40.755–22:41.306
突出來了。
22:41.316–22:45.173
這是3D效果，不要被它看似2D的外觀所欺騙。
22:45.183–22:46.552
它確實是3D的。
22:46.562–22:48.011
它也在緩慢地平移。
22:48.021–22:52.088
我真的……好吧，我有幾點喜歡的地方：木質背景。
22:52.098–22:55.742
以及這個類似皮革卡片或墊子，手錶就放在上面。
22:55.752–22:56.297
放在上面。
22:56.307–22:58.440
除此之外，錶盤看起來真的很不錯。
22:58.450–23:00.347
材質看起來非常好。
23:00.357–23:03.998
總體來說，它呈現出高端的美感，
23:04.008–23:07.254
這正是我們進行這項測試時所追求的。
23:07.264–23:13.563
現在，那裡的確看起來不錯。
23:13.573–23:16.975
讓我們靜下來欣賞它一分鐘。
23:16.985–23:17.215
好的。
23:17.225–23:19.290
是的，也許我可以不要那個，
23:19.300–23:22.962
但它確實相當引人注目。
23:22.972–23:25.435
接著看看我們開始往下捲動時，它如何消失
23:25.445–23:26.633
往下捲動
23:26.643–23:30.225
以秒計量的靜奢風
23:30.235–23:31.582
放射紋錶盤
23:31.592–23:33.657
好的，這裡有一張像是材質說明卡的東西
23:33.667–23:37.249
藍寶石水晶鏡面或許是我本來會省略的部分
23:37.259–23:39.564
雖然這部分我可能不會保留，但錶盤看起來不錯
23:39.574–23:41.560
兩款手錶，同一個觀點
23:41.570–23:42.438
請幫它們渲染出來
23:42.448–23:43.236
哇，你看那個
23:43.246–23:44.194
這太酷了
23:44.204–23:45.629
這看起來真的很不錯
23:45.639–23:47.546
我喜歡它們在桌上的擺放方式
23:47.556–23:50.420
其實真正拖後腿的就是這塊皮革板
23:50.430–23:53.533
這塊皮革板擋住了我們欣賞這些精美錶帶的視線
23:53.543–23:55.688
這是不鏽鋼的選項
23:55.698–23:59.679
這款是金色選項，價格自然也會相應提高
23:59.679–24:01.292
這是鈦金屬選項，這是黃金選項，價格也相應較高，我想
24:01.292–24:01.372
沒錯
24:01.372–24:05.665
不過我會說，價格的跳動幅度微乎其微
24:05.675–24:06.702
為了拿到金色款所增加的金額其實微乎其微
24:06.712–24:08.539
我個人大概會選這一款
24:08.549–24:08.938
喔，不錯
24:08.948–24:12.370
而且這些可以隨意旋轉，所以還是有點幫助
24:12.380–24:15.643
雖然錶針沒有動，不過沒關係
24:15.653–24:17.240
我願意原諒它這一點
24:17.250–24:19.395
但這些真的非常酷
24:19.405–24:22.508
我覺得這些畫得很精緻
24:22.518–24:25.860
特別是那些數字指示器
24:25.870–24:28.814
看看他們實際上的多材質設計
24:28.814–24:29.770
所以它們並非只有一種靜態材質，而是有金色搭配黑色填充，或是銀色搭配黑色填充。
24:29.770–24:31.766
所以它們並非只有一種靜態材質
24:31.776–24:35.599
是金色搭配黑色填充，或是銀色搭配黑色填充
24:35.609–24:35.918
也是
24:35.928–24:38.473
所以我很欣賞這種注重細節的設計
24:38.483–24:42.383
以及不鏽鋼與金色部分那種刷紋材質的質感
24:42.393–24:42.943
以及金色部分
24:42.953–24:46.215
這可能有點奇怪，我們居然在評估一款3D手錶
24:46.225–24:48.690
但這確實展現了某種能力
24:48.700–24:50.429
好的，這部分看起來相當不錯。
24:50.439–24:52.335
接著我們還有一個乾淨的頁尾。
24:52.345–24:52.997
我喜歡這個。
24:53.007–24:54.075
我對這個很滿意。
24:54.085–24:58.632
雖然它不是完美的，但也已經非常接近了。
24:58.642–25:00.919
所以，接下來，我們直接在 Codex 裡面操作。
25:00.929–25:03.873
我無法顯示完整的提示詞，因為裡面包含 API 金鑰。
25:03.883–25:07.228
我將對它進行完整的「史蒂夫，電腦維修遊戲」電影風格測試，
25:07.238–25:09.624
測試它是否能創建一個電影風格的遊戲。
25:09.634–25:12.900
它需要僅憑我提供的 OpenAI API 金鑰，為這些角色建立語音素材
25:12.910–25:16.595
OpenAI API 金鑰，為這些角色創建語音資產，並製作一個有趣
25:16.605–25:18.874
且令人愉悅、可遊玩的遊戲。
25:18.884–25:21.940
我評估這大概需要一個小時以上的時間。
25:21.950–25:24.694
所以，我們基本上就把 Codex 留在背景中，
25:24.704–25:28.538
使用它的 Beta 桌面版來完成這個任務，同時我們
25:28.548–25:30.255
也許可以順便處理其他一些事情。
25:30.265–25:34.365
所以，我們的「史蒂夫，電腦維修員」電影風格遊戲，基本上在
25:34.375–25:35.187
65 分鐘內完成了。
25:35.197–25:36.755
也就是說，稍微超過一個小時。
25:36.765–25:38.250
好的，這是遊戲玩法說明。
25:38.260–25:41.537
我大概會照這裡寫的做，不會假設它
25:41.547–25:43.777
已經在我們的某個伺服器上運行這個了。
25:43.787–25:47.737
好的，第一集，Azure Palm 的工作。
25:47.747–25:49.231
好的，點擊開始。
25:49.241–25:50.502
建議使用耳機。
25:50.512–25:53.084
我們有喇叭。
25:53.094–25:57.007
好的。
25:57.017–25:58.867
我是 Steve。
25:58.877–25:59.693
這很好。
25:59.703–26:02.170
我修理電腦。
26:02.180–26:05.516
我的大多數客戶都夠老，記得硬碟
26:05.526–26:07.306
曾經是奢侈品。
26:07.316–26:10.768
他們付得起多少我就收多少，然後讓它運作。
26:10.778–26:14.143
主要是誠實的交易。
26:14.153–26:15.855
哦，謝謝你，親愛的。
26:15.865–26:20.261
螢幕一直告訴我日期是 1980 年，我想，全
26:20.271–26:20.834
修好了。
26:20.844–26:22.554
我已經旅行夠久了。
26:22.564–26:23.646
全新的電池。
26:23.656–26:26.196
你的電腦現在相信我會帶燕麥餅乾
26:26.206–26:26.860
在星期二來。
26:26.870–26:27.720
跟其他人一樣。
26:27.730–26:28.065
請替我幫 Mr.
26:28.075–26:30.110
Buttons 撓撓他耳朵後面。
26:30.120–26:33.580
讓我送您出去，夫人。
26:33.590–26:36.020
我馬上就來，Thomas 先生，這段演講會互相覆蓋。
26:36.030–26:38.727
好男孩，Steven。
26:38.737–26:40.513
如果那個安靜的男人需要修理，你要耐心點。
26:40.523–26:43.953
如果那個安靜的男人需要修理，你要有耐心。
26:43.963–26:44.560
我喜歡守時。
26:44.570–26:46.061
我喜歡守時
26:46.071–26:48.001
咖啡是溫的。
26:48.011–26:48.929
一個男人。
26:48.939–26:49.704
這份工作是什麼？
26:49.714–26:50.710
來了兩年。
26:50.720–26:52.723
我的生活。
26:52.733–26:54.813
你看那個頭部前後晃動的動作，其實現在有三個人正在說話。
26:54.823–26:58.612
而且他忘記了整個名字。Charles 給他買了一間私人頂樓公寓，
26:58.622–26:59.671
安保人員想要他的傳奇。有些人想問
26:59.681–27:05.838
然後他連整個名字都忘了。查爾斯買給他一間私人頂層公寓
27:05.848–27:08.220
安全部門，安全部門想要他的傳奇。有些人想問他問題。他慢慢地。客戶先付100。我也希望這也適用於這個案例。所有費用，頭等艙，兩張機票。這就是我們合作的原因。一起，頂層托盤是一份普通檔案
27:08.230–27:12.891
向他提問題。慢慢地。客戶先付100。我希望
27:12.901–27:16.850
這在案件中也同樣適用。所有費用，頭等艙，兩張機票。這
27:16.860–27:22.047
就是我們合作的原因。一起，頂層托盤是一個普通檔案
27:22.057–27:30.980
在筆記型電腦上。這是我第二喜歡的東西。
27:30.990–27:31.336
專業人士
27:31.346–27:33.116
我不會問第一個是什麼。
27:33.126–27:36.888
當直升機準備好時電話響了。
27:36.898–27:37.873
四十
27:37.883–27:39.334
場景切換結束。
27:39.344–27:40.279
謝天謝地。
27:40.289–27:41.739
點擊以控制。
27:41.749–27:43.373
好吧。
27:43.383–27:46.981
除了聲音完全亂七八糟之外，
27:46.991–27:52.052
這其實相當不錯。
27:52.062–27:52.654
好的。
27:52.664–28:00.561
貓是……確實有一些幾何問題。
28:00.571–28:02.366
我們應該做什麼？
28:02.376–28:03.997
當你準備好時，拿走手提箱。
28:04.007–28:04.256
好的。
28:04.266–28:06.833
它在哪裡？
28:06.843–28:11.414
嗯，按按鈕就像在說「別那樣看我」。
28:11.424–28:12.146
按鈕。
28:12.156–28:12.700
彩蛋。
28:12.710–28:14.079
人總得吃飯。
28:14.089–28:14.462
好的。
28:14.472–28:17.600
手提箱在哪裡？
28:17.610–28:19.132
好吧。
28:19.142–28:20.127
我們有庫存。
28:20.137–28:20.970
束線帶。
28:20.980–28:22.202
好的。
28:22.212–28:23.324
按 E 繼續。
28:23.334–28:23.785
好吧。
28:23.795–28:25.408
它會把我們帶到邁阿密嗎？
28:25.418–28:25.608
不錯。
28:25.618–28:30.071
棕櫚樹、濕氣，還有一個非常壞的男人，
28:30.081–28:30.308
科爾頓先生。
28:30.318–28:31.648
科爾頓。
28:31.658–28:32.357
也許來點香檳。
28:32.367–28:36.351
我直接點過去，因為它就像我的
28:36.361–28:40.530
度假村裡的咖啡資產一樣黑。
28:40.540–28:41.941
中午在泳池邊。
28:41.951–28:43.168
哦，泳池邊。
28:43.178–28:45.040
十七樓進入。
28:45.050–28:46.830
攝影機覆蓋範圍。
28:46.840–28:47.302
電梯。
28:47.312–28:48.450
沒有狗。
28:48.460–28:52.831
後面是西側服務樓梯間。
28:52.841–28:55.987
嗯，我想我們剛才瞥見了泳池。
28:55.997–28:57.456
好吧。
28:57.466–28:59.495
哦，太棒了。
28:59.505–29:00.799
Azure Palm 度假村
29:00.809–29:03.165
點擊以取得控制權。
29:03.175–29:07.406
它甚至還有那種你預期會在這種地方看到的低檔壁紙。
29:07.416–29:09.201
在這種地方。
29:09.211–29:10.260
哦，哇。
29:10.270–29:11.891
這其實有點複雜
29:11.901–29:12.381
哦，很好。
29:12.391–29:12.626
好的。
29:12.636–29:15.195
我擔心我們無法直接穿牆而過，
29:15.205–29:16.541
但沒關係。
29:16.551–29:18.416
我們可以在噴泉裡擲硬幣。
29:18.426–29:18.988
你的願望。
29:18.998–29:20.185
不再有語法錯誤。
29:20.195–29:21.617
五五開。
29:21.627–29:25.773
史蒂夫的商店和飛機場景似乎被怪異地
29:25.783–29:29.299
放置在了我們應該所在的地點內部。
29:29.309–29:31.622
我們就假裝沒看到吧。
29:31.632–29:37.710
所以，先去櫃檯報到，然後前往17樓。
29:37.720–29:39.711
歡迎來到藍棕櫚酒店。
29:39.721–29:41.710
報到，庫爾特先生。
29:41.720–29:42.750
我。
29:42.760–29:43.603
庫爾特。
29:43.613–29:44.755
史蒂文·庫爾特。
29:44.765–29:44.925
先生。
29:44.935–29:47.198
我有訂房紀錄
29:47.208–29:51.154
1408號房，部分海景。
29:51.164–29:52.416
我很抱歉。
29:52.426–29:55.363
您這層的製冰機故障了。
29:55.373–30:00.141
這真是我人生的寫照。
30:00.151–30:00.560
好吧。
30:00.570–30:02.198
我們現在庫存裡有一張房門鑰匙。
30:02.208–30:03.882
所以，這是電梯嗎？
30:03.892–30:12.484
哦，是這樣嗎？
30:12.494–30:12.817
好的。
30:12.827–30:15.467
西側服務樓梯間。
30:15.477–30:21.017
我們就假設這方向是北。
30:21.027–30:45.339
很好。像什麼？
30:45.349–30:51.897
嗯，這很令人沮喪，因為它給了我們一種非常像
30:51.907–30:53.164
預期的音樂。
30:53.174–30:55.474
我確定對於我接下來要說的話，有更專業的術語。
30:55.484–30:56.071
要說的話。
30:56.081–30:57.188
我叫史蒂夫。
30:57.198–30:57.710
哦，不。
30:57.720–30:59.830
別這樣。
30:59.840–31:03.256
我的大多數客戶都夠老了，記得硬碟
31:03.266–31:04.427
曾經是奢侈品。
31:04.437–31:04.829
好的。
31:04.839–31:06.970
我們已經完成了大部分。
31:06.980–31:10.849
我認為這是一個測試，而且往往
31:10.859–31:12.464
我們無法完成它。
31:12.474–31:15.698
我會說，這些元素若單獨來看，
31:15.708–31:17.234
絕對來說，所有零件都在那裡。
31:17.244–31:19.900
我不確定這個拼圖拼出的正確畫面，
31:19.910–31:23.052
是否與拼圖盒上展示的相符，但確實非常接近，而且
31:23.062–31:26.204
我看到了一些——誠實說，這比我預期做得更好，
31:26.214–31:29.517
特別是考慮到像地鐵場景遊戲這類東西。
31:29.527–31:29.841
那樣。
31:29.851–31:34.206
所以，我其實對這個沒意見。
31:34.216–31:37.923
所以，最後一項當然是街頭投擲遊戲，
31:37.933–31:41.640
這很奇怪，因為這段文字上下移動，
31:41.650–31:44.470
而這個完全靜止不動，這種對比讓我不舒服
31:44.480–31:46.005
在某種程度上感到不適。
31:46.015–31:48.457
所以，這是個 Bijan 的問題。
31:48.467–31:49.288
好的，不錯。
31:49.298–31:53.072
所以，我疏忽了，呃，只是把它在伺服器上運行。
31:53.082–31:57.152
一拳，零仁慈。
31:57.162–31:58.354
我已經喜歡這個了。
31:58.364–32:05.477
哦，這聽起來像好音樂。
32:05.487–32:05.877
好的。
32:05.887–32:06.276
好的。
32:06.286–32:10.359
好了。
32:10.369–32:16.681
我們就快速探索一下這座城市吧。
32:16.691–32:18.603
我甚至不知道。
32:18.613–32:30.602
好，我看到一輛車撞上了——像這輛其實有點好玩
32:30.602–32:34.698
。
32:34.698–32:35.970
不好意思。
32:35.980–32:37.250
不好意思。
32:37.260–32:38.131
抱歉。
32:38.141–32:40.965
它會做一種……像是一種……好了。
32:40.975–32:42.389
我不知道我們有沒有撞到……
32:42.399–32:47.562
我們可以撞到那輛自行車嗎？
32:47.572–32:54.399
砰。
32:54.409–32:55.114
哎呀。
32:55.124–33:00.815
亂穿馬路的行人。
33:00.825–33:02.583
我們可以揍這台停車計費器嗎？
33:02.593–33:05.154
是啊，我們可以撞所有東西，除了建築物。
33:05.164–33:05.556
好了。
33:05.566–33:08.847
不幸的是，這些東西中的一些方向
33:08.857–33:09.722
方向不太對
33:09.732–33:15.609
我們可以看到，當我們向前移動時，我們是在側向移動。
33:15.619–33:22.161
而且這些車子也……抱歉，老兄
33:22.171–33:25.963
對了，他們與景物的互動方式很有趣
33:25.973–33:26.453
還有
33:26.463–33:27.763
這座雕像可以吃嗎
33:27.773–33:30.709
我本來以為不行，但喔，對，我們可以吃。
33:30.719–33:33.981
好吧
33:33.991–33:37.853
我會說這款遊戲的卡通風格其實做得還不錯
33:37.863–33:38.500
完成得不錯
33:38.510–33:41.571
我不知道光看別人玩能不能輕易評估這一點
33:41.581–33:43.917
不過節奏還算不錯
33:43.927–33:47.071
我認為這可能也與音樂有關
33:47.081–33:49.010
但當你真的……好吧，那種方式
33:49.020–33:52.412
循環有點重複，但當你實際把東西丟出去時
33:52.422–33:55.813
把東西丟出去，它們還會和其他環境中的物品產生互動
33:55.823–33:56.445
環境
33:56.455–34:00.322
所以就像它會彈到建築物上然後可能撞到
34:00.332–34:01.745
路徑上的其他物體
34:01.755–34:02.537
這真的很酷
34:02.547–34:09.255
而且這其實很有街機風格，也可以玩
34:09.265–34:10.366
我們的分數正在上升
34:10.376–34:13.177
我們有整體的連擊倍率。
34:13.187–34:14.849
這裡的 38 分算是成功了。
34:14.859–34:17.596
好的，這大概會結束我們對 DeepSeek V4 Pro 首次亮相
34:17.606–34:21.150
以及對一般可用性 DeepSeek V4 Pro 的測試，
34:21.160–34:23.388
也就是 0813 型號的介紹。
34:23.398–34:26.427
現在這版本絕對比預覽版來得好
34:26.437–34:29.305
那個預覽版感覺相當不成熟，我完全沒有印象
34:29.315–34:29.865
它。
34:29.875–34:32.663
但如果我們回顧之前影片中執行的一些測試，
34:32.673–34:35.062
該影片在頻道上，我們今天也執行了一些相同的測試。
34:35.072–34:35.622
今天。
34:35.632–34:37.780
這是顯著的進步
34:37.790–34:41.059
不過我會說，我認為 flash 模型讓我印象更深刻，
34:41.069–34:43.618
比這個更甚，特別是考慮到它的體積。
34:43.628–34:47.136
這個模型面臨的競爭稍多，但再次強調，這是對其前代產品的大幅躍進，
34:47.146–34:49.854
我想這也是我們樂見其成的。
34:49.864–34:53.292
我幾次對某些事情印象深刻，比如結果中的結果。
34:53.302–34:54.092
結果中的結果。
34:54.102–34:57.931
電影版 Steve 的電腦維修遊戲，它實際上包含了一些元素，
34:57.941–35:00.181
那些元素讓人覺得，嘿，你知道，它似乎有點思考了
35:00.191–35:00.744
透過。
35:00.754–35:03.395
只是當他們說話時，他們的頭部像是
35:03.405–35:04.198
前後晃動。
35:04.208–35:05.805
不是那種奇怪的下巴動作。
35:05.815–35:07.572
我不確定那裡發生了什麼事。
35:07.582–35:10.706
那款遊戲以及之後的 Street Yeet
35:10.716–35:14.562
遊戲中收錄的音樂，實際上非常符合
35:14.572–35:17.522
這裡所呈現的場景，而且它很好地讓 Street Yeet 遊戲顯得
35:17.532–35:20.384
卡通化，其可玩性感覺不錯，
35:20.394–35:22.212
我想，如果沒有更好的詞彙來形容的話。
35:22.222–35:26.579
我們獲得的引擎模型實際上無法列印。
35:26.579–35:32.898
。
35:32.898–35:32.978
所以，不幸的是，它可能沒能完成這項任務的檢查
35:32.978–35:33.058
在特定任務中勾選該選項。
35:33.058–35:33.934
然而，如果我們只看這裡的其中一個預覽，
35:33.944–35:35.616
它做得相當不錯。
35:35.626–35:38.739
絕對比我用過的其他模型測試結果更好。
35:38.749–35:41.941
雖然，我還沒有用任何大型強大的
35:41.951–35:43.143
Frontier 模型測試過這個特定的提示詞。
35:43.153–35:44.744
所以，也請記住這一點。
35:44.754–35:46.186
不過，它確實具備一些實力。
35:46.196–35:49.789
而且它為了達到那個目標所進行的研究量
35:49.799–35:50.830
相當龐大。
35:50.840–35:53.709
這裡包含了所有下載這些影像的步驟，即使它
35:53.719–35:54.510
沒有視覺功能。
35:54.520–35:56.350
所以，我還不太確定該如何看待這一點。
35:56.360–35:59.870
不過話雖如此，我認為讓我感到最挫折的
35:59.880–36:04.750
重點是，它會全力以赴，試圖確保它編寫的
36:04.760–36:08.190
每一行程式碼的每個面向都是正確的。
36:08.200–36:11.790
對於像這樣沒有視覺功能的模型，
36:11.800–36:15.349
有時會花費更長的時間，而且也會產生成本。
36:15.349–36:22.665
。
36:22.665–36:22.745
假設你將為使用這些模型付費，
36:22.745–36:22.825
有時讓它直接產生結果會更簡單，
36:22.825–36:24.510
而不是嘗試進行概覽、檢查和後續追蹤。
36:24.520–36:26.830
你自己查看結果，然後告訴它哪裡有錯，並指示
36:26.840–36:27.790
它進行修正。
36:27.800–36:31.469
必須全力以赴並讓它嘗試修正，這會增加大量的
36:31.479–36:35.790
時間、費用和 token 成本，這可能更多是
36:35.800–36:38.876
但我發現這是我一直感到挫折的地方
36:38.886–36:41.665
但我也發現，這是我一直感到挫折的地方
36:41.675–36:45.100
以及在使用閃電模型時，如果我要它們製作
36:45.110–36:47.894
比如說街景，讓它們直接輸出腳本會比較容易
36:47.904–36:50.371
我測試後說：「嘿，建築物消失了。」
36:50.381–36:53.565
然後它進行修復，而不是嘗試去
36:53.575–36:56.280
好吧，讓我嘗試啟動無頭瀏覽器，然後我會截圖
36:56.290–36:59.555
並像做熱力圖那樣，因為它實際上無法看見
36:59.565–37:00.911
它或其他功能。
37:00.921–37:04.036
所以它有點過度消耗 token。
37:04.046–37:06.925
我確實是在最高潛力的推理模式下運行的。
37:06.935–37:10.859
這也可能歸因於此，但我仍然希望
37:10.869–37:14.471
希望能做到這點，而不需要它全力嘗試去確保腳本的每個面向都經過打磨，因為我只求最高效能，不需要那麼深的程度
37:14.481–37:17.602
它打磨腳本的每個方面的情況下做到這一點，因為我想要最大
37:17.612–37:18.164
效能。
37:18.174–37:19.769
我只是不需要那麼深的程度。
37:19.779–37:22.979
所以，這可能是個人的抱怨，但你知道。
37:22.989–37:25.949
所以，最後一點，讓我們看看這次測試中我們花費的所有
37:25.959–37:28.359
涵蓋整個測試期間
37:28.369–37:30.962
我們可以看到這裡的總成本為 1。
37:30.972–37:31.810
[未翻譯]
37:31.820–37:33.416
所以，這真的非常便宜。
37:33.426–37:36.466
而且你必須記住，這將會增加，他們的
37:36.476–37:37.749
API 價格將會上漲。
37:37.759–37:41.509
他們表示預期會有顯著增加，但我的意思是，顯著增加可能就像，好吧，整個測試總共才 8
37:41.519–37:45.071
元或類似金額，這在整體來看還不算太糟。
37:45.081–37:48.000
或者類似的情況，這在整體來看還不算太糟。
37:48.010–37:48.881
8B 最大測試影片是我們幾週前，或大約一兩週前做的，那次的費用大約是 32。
37:48.891–37:52.645
我們幾週前或大約一週前做的最大 8B 測試影片
37:52.655–37:54.967
這可能比這個效能更好，但也會有價格方面的考量。
37:54.977–37:56.490
而且這真的非常非常便宜。
37:56.500–37:59.373
它不會一直這麼便宜，但即使價格漲到五倍，也還不算太糟。
37:59.383–38:02.257
我想這將結束我們對
38:02.267–38:04.499
0813 Deepseek V4 Pro 模型一般可用性版本的首次檢視和測試。
38:04.509–38:08.023
它不會一直這麼便宜，但就算漲到五倍，也不算太糟
38:08.033–38:09.625
這真的很令人興奮。
38:09.635–38:12.429
基本上，現在我想我會去測試 Grok 3
38:12.439–38:16.834
0813 Deepseek V4 Pro 模型，一般可用性版本
38:16.844–38:18.436
現在已經正式推出，不再處於預覽階段。
38:18.446–38:20.438
這真的令人相當興奮
38:20.448–38:23.411
基本上，我想我現在要來測試 Grok 3
38:23.421–38:25.280
因為它也是今天發布的。
38:25.290–38:27.767
然後我相信週五 Qwen 2
38:27.777–38:28.820
72B 也會發布。
38:28.830–38:30.260
所以這也會令人興奮。
38:30.270–38:33.220
看來這週在模型發布方面相當忙碌
38:33.230–38:34.341
模型發布方面相當忙碌。
38:34.351–38:35.700
所以這也很酷。
38:35.710–38:37.264
當 MuseTalk 1
38:37.274–38:38.974
2 發布時，那將非常棒。
38:38.984–38:41.449
我不知道它有多大，但我非常興奮想嘗試在
38:41.459–38:42.089
本地運行它。
38:42.099–38:44.884
因此，如果你有任何問題，請隨時
38:44.894–38:46.002
在留言區留言。
38:46.012–38:48.000
感謝觀看
0:00.000–0:00.634
shooting out.
噴射而出。
0:00.644–0:05.632
I believe that was something I've seen .
我相信那是我曾經看過的東西。
0:05.642–0:08.524
The trees are quite reminiscent of a pool.
這些樹木讓人聯想到泳池。
0:08.534–0:12.562
DeepSeek V4 Pro has just been released in general availability version
DeepSeek V4 Pro 的通用正式版剛剛發布。
0:12.572–0:16.604
, which also happened with the V4 Flash model back about 12 or
這也發生在約12或更早時候的V4 Flash模型上
0:16.614–0:19.161
13 days ago on 07/31.
13 天前的 07/31 發布 V4 Flash 模型的情況相同。
0:19.171–0:23.078
Now, that model when tested proved to be a significant
現在，該模型在測試中被證明具有顯著的
0:23.088–0:26.516
increase in capability over the preview version that was previously
在先前預覽版的基礎上，能力有了顯著的提升
0:26.526–0:27.154
around .
左右的預覽版本。
0:27.164–0:29.793
And today we're going to be finding out if the same can be said
今天我們將要找出是否同樣的情況
0:29.803–0:34.292
for the newly updated checkpoint general availability for V4 Pro
適用於 V4 Pro 新更新的通用正式版檢查點
0:34.302–0:35.870
dubbed 0813.
命名為 0813。
0:35.880–0:38.426
So before we get into it , do feel free to subscribe so we can
所以，在我們深入探討之前，請隨時訂閱，以便我們
0:38.436–0:40.824
hit 70K and then 100K with the plaque .
能達到 70K，然後是 100K，並獲得獎牌。
0:40.834–0:43.861
And with that , basically there's not a lot to say about this right
就此而言，基本上現在沒什麼好說的，因為它似乎
0:43.871–0:46.693
now because it seems to have been more or less silently released
現在因為它似乎已經在無聲無息中發布
0:46.693–0:48.780
.
。
0:48.780–0:49.535
There's no official announcement post yet on X ,
目前 X 上還沒有官方公告帖子，
0:49.545–0:52.331
which I do believe is generally where we would see this .
我認為通常我們會在那裡看到這個。
0:52.341–0:55.455
But they do mention here just on the website a couple references
但他們確實只在網站上提到了一些參考
0:55.465–0:59.385
to it where if you use this model identifier right here for V4
如果在此處使用這個模型識別碼 V4
0:59.395–1:03.582
Pro, you are being routed to the new V4 Pro 0813 version.
Pro，您將被導向至新的 V4 Pro 0813 版本。
1:03.592–1:06.842
Additionally to that, it is also live on OpenRouter and this
此外，它也在 OpenRouter 上線，這
1:06.852–1:10.103
makes me feel comfortable that it is actually genuinely released
讓我感到安心，因為它確實已經正式發布，
1:10.113–1:11.977
and this is the model we're going to be using .
而且這也是我們將要使用的模型。
1:11.987–1:14.260
So , from the bit of information we have right here ,
因此，根據我們目前掌握的部分資訊，
1:14.270–1:17.520
aside from just the model identifier now being today's date ,
除了模型識別碼現在顯示為今天的日期之外，
1:17.530–1:20.812
depending on what locale one is in, we can see that the model has
取決於使用者的地區設定，我們可以得知該模型擁有
1:20.822–1:24.764
a million context window with a maximum output length of 384K ,
一百萬的上下文視窗，最大輸出長度為 384K，
1:24.774–1:26.135
which really is a lot .
這確實是非常大的數量。
1:26.145–1:28.947
Now , something that can sometimes happen with DeepSeek ,
現在，DeepSeek 有時可能會出現的情況是，
1:28.957–1:31.819
at least historically , is there may be a lot of thinking .
至少從歷史紀錄來看，可能會產生大量的思考過程。
1:31.829–1:34.211
Now , this does support selectable thinking modes ,
現在，這確實支援可選的思考模式，
1:34.221–1:36.045
as you can see if we do click on this .
正如我們點擊此處所見。
1:36.055–1:38.120
So , it's possible to kind of tone that down ,
因此，可以將這種情況稍微降低，
1:38.130–1:40.991
but in the interest of seeing the maximum potential performance
但為了展現最大潛力的效能
1:41.001–1:44.501
, all of our coding tasks for today are going to be run on this
今天所有的程式碼任務都將以
1:44.511–1:45.697
maximum effort .
最大努力模式執行。
1:45.707–1:49.047
Now , I want to mention the following with a grain of salt because
現在，我想謹慎地提及以下事項，因為
1:49.057–1:51.839
I cannot definitively confirm that this is ground truth ,
我無法確定地確認這是事實真相，
1:51.849–1:54.072
but I would imagine it probably is .
但我猜測它很可能就是。
1:54.082–1:57.261
There were some references and photos floating around which apparently
網路上流傳著一些參考資料和照片，顯然
1:57.271–1:59.574
came from a group related to DeepSeek .
來自與 DeepSeek 相關的群組。
1:59.584–2:02.821
I believe they said it was a WeChat group about DeepSeek or something
據他們表示，那是一個關於 DeepSeek 的微信群組，或是類似的
2:02.821–2:08.747
.
。
2:08.747–2:08.827
And if this is to be believed , we can see that these would hypothetically
如果這個說法可信，我們可以假設這些是
2:08.827–2:11.177
be the benchmark results for DeepSeek V4 Pro 0813 or the general
DeepSeek V4 Pro 0813 或一般
2:11.187–2:12.448
availability model .
可用性模型的基準測試結果。
2:12.458–2:15.625
And basically my main interest here is just seeing the leap in
基本上，我主要感興趣的是，從 V4 Pro 預覽版到目前版本，
2:15.635–2:20.225
capability over the V4 Pro preview to now because I was initially
能力上的巨大躍進，因為我最初對 V4 Pro 預覽版以及 V4 Flash 預覽版
2:20.235–2:24.272
disappointed with the V4 Pro Preview as well as the V4 Flash preview
。
2:24.272–2:28.713
.
它們感覺非常不成熟，可能因為它們被命名為
2:28.713–2:28.793
They felt very undercooked and probably because they were named
它們感覺非常未熟，可能因為它們被命名為
2:28.793–2:28.873
preview .
這其實很合理。
2:28.873–2:29.509
That's actually like fair .
這其實還算公平。
2:29.519–2:30.381
It makes sense .
但再次強調，我想重申的是，從
2:30.391–2:33.273
But again , I want to reiterate the leap in capability between
Flash 到 0731 一般可用性版本的 Flash，
2:33.283–2:37.518
flash and then the 0731 general availability version of flash was
Flash，然後0731的Flash一般可用性版本是
2:37.528–2:38.753
absolutely massive .
這就像是目前最適合雙 3090
2:38.763–2:42.495
This is like the best local model right now for like a dual 3090
這就像是目前雙3090設置下最好的本地模型
2:42.505–2:42.914
setup .
設定。
2:42.924–2:48.182
So I mean in a niche , you know , but still if that leaping capability
所以我的意思是，在一個小眾領域裡，你知道，但即便如此，如果這種躍升能力
2:48.192–2:51.870
increase from 12 .
從12增加。
2:51.880–2:53.058
increase from 12 .
從12增加。
2:53.068–2:54.571
8 to 62 .
.8 增加到 62
2:54.581–2:58.218
7 would then hypothetically put this above things like Opus 3 .
.7，那麼 hypothetically（假設性地）會將它置於像 Opus 3
2:58.228–3:02.046
8 8 and make it very very competitive especially that this
.8 和 8 之上，並使其非常非常有競爭力，尤其是因為這個
3:02.056–3:04.286
is an openweight model or will be .
是一個開放權重的模型，或者將成為開放權重模型。
3:04.296–3:08.276
That's also extremely exciting because this is not like horrifically
這也極度令人興奮，因為這不像考慮到像 Kimi K3 或 Qwen
3:08.286–3:12.098
gigantic I guess when considering things like Kimi K3 or Qwen
考慮到像Kimi K3或Qwen這樣的事物，我猜這是巨大的。
3:12.108–3:12.470
3 .
.2
3:12.480–3:13.291
2 .
那樣恐怖地巨大，我想。
3:13.301–3:14.587
It's still big at 1 .
它仍然很大，有 1
3:14.597–3:15.753
6 trillion parameters .
六兆個參數
3:15.763–3:19.089
It is a mixture of experts model with how many active ?
它是一個混合專家模型，有多少活躍參數？
3:19.099–3:22.107
So if we do estimate that this is going to be the same size as
所以如果我們估算這會和
3:22.117–3:24.013
the preview version of V4 Pro .
V4 Pro 的預覽版本大小相同。
3:24.023–3:24.650
It's a 1 .
它是 1
3:24.660–3:28.490
6 trillion parameter mixture of experts model with 49 billion
.6 兆參數的混合專家模型，有 490 億
3:28.500–3:28.970
active .
活躍參數。
3:28.980–3:32.810
So it is still very very large but it's not as gigantic as like
所以它仍然非常非常大，但並沒有像接近3兆參數的模型那樣龐大。
3:32.820–3:35.284
closing in on a 3 trillion parameter model.
關閉接近3兆參數的模型。
3:35.294–3:38.623
So additionally to that pricing and this is something that we need
因此，在定價方面，這也是我們需要
3:38.633–3:41.962
to basically take just as it is right now because they mention
基本上就照現在這樣看待的事情，因為他們提到
3:41.972–3:44.985
and have been for quite a bit of time that their API pricing is
並且在相當長一段時間內表示，他們的API定價
3:44.995–3:47.289
going to be significantly increased soon .
很快就會大幅調漲。
3:47.299–3:50.469
Now significantly is relative to what it currently is which is
現在「大幅」是相對於它目前的價格，也就是
3:50.479–3:52.536
like very very very cheap .
非常非常非常便宜。
3:52.546–3:56.036
So , right now , pricing per million input tokens for DeepSeek
所以，目前DeepSeek每百萬輸入記號的定價
3:56.046–3:59.250
V4 Pro is 0.43
V4 Pro每百萬輸入記號的定價是0.43
3:59.260–4:00.080
35 .
[未翻譯]
4:00.090–4:01.178
So , okay .
所以，好吧。
4:01.188–4:04.652
No , I know what this is , but like last time I messed up on this
不，我知道這是什麼，但就像上次我搞錯這
4:04.662–4:08.456
one and I said it was 2 cents and it wasn't and then I got roasted
點，我說它是2美分，但事實並非如此，然後我被轟炸
4:08.456–4:08.966
.
。
4:08.966–4:14.347
So, this is 43 cents per million input tokens and 87 cents per
所以，這是每百萬輸入記號43美分，每百萬輸出記號
4:14.357–4:16.433
million output tokens is the current price .
87美分是目前的價格。
4:16.443–4:19.162
Now , if this gets like a significant increase ,
現在，如果這會像大幅上漲，
4:19.172–4:20.873
say it goes up like five or 10 times ,
假設它上漲了五倍或十倍，
4:20.873–4:22.452
it would still hypothetically
它仍然在假設上
4:22.462–4:23.335
be competitive .
就會具備競爭力。
4:23.345–4:26.063
So , it's all relative to what it currently costs .
所以，這一切都取決於它目前的成本。
4:26.073–4:28.793
And with that, I am going to be using this from within Cursor
基於此，我將從 Cursor 內部使用這個功能。
4:28.793–4:30.524
.
。
4:30.524–4:33.137
When we tested V4 Flash , the general availability model on the
在我們測試 V4 Flash 模型於 7 月 31 日正式發布時，
4:33.147–4:36.656
31st of July, I did also use it through Cursor because I believe
我也透過 Cursor 使用它，因為我認為他們建議這麼做，
4:36.666–4:39.505
they had suggested to do so and the performance was very ,
他們曾建議這麼做，而且效能非常
4:39.515–4:40.850
very good .
非常好。
4:40.860–4:41.879
I was happy with it .
我對它很滿意。
4:41.889–4:44.381
So, I'm going to be using Cursor as well for this.
所以，我也將使用 Cursor。
4:44.391–4:49.124
Now I should also mention that they did release in beta a desktop
現在我還應該提到，他們在測試版中發布了 Cursor 或 ChatGPT 的桌面
4:49.134–4:51.270
app for Cursor or ChatGPT.
用於 Cursor 或 ChatGPT 的應用程式。
4:51.280–4:54.288
So I also have it routed and working through this desktop app as
所以我也將它路由並透過這個桌面應用程式運作，
4:54.298–4:57.142
well and it is doing a skate game for us in C++.
它正在為我們用 C++ 製作一個滑板遊戲。
4:57.152–4:59.099
I would imagine that will take quite a bit of time .
我預計這會花費相當多的時間。
4:59.109–5:01.791
So I did just want to start it and verify this app was actually
所以我只想先啟動它，並驗證這個應用程式在這個系統上
5:01.801–5:04.482
working properly on this system and it does seem to .
是否正常運作，看起來確實是正常的。
5:04.492–5:09.430
So keep in mind any mention of GPT or Cursor here is routed directly
所以請注意，此處任何提及 GPT 或 Cursor 的地方，都是直接路由
5:09.440–5:13.030
to DeepSeek on max reasoning effort for the 0813 model .
到 DeepSeek，使用 0813 模型的極限推理能力。
5:13.040–5:15.355
And with that , let's take a look at our Okay ,
基於此，讓我們來看看我們的 好的，
5:15.365–5:21.369
so apparently the OpenAI beta application set itself to be the
所以顯然，OpenAI 的 beta 應用程式被設定為所有檔案類型的預設程式。
5:21.379–5:23.501
default for all file types .
這真是令人氣憤。
5:23.511–5:24.525
That's infuriating .
太令人氣憤了。
5:24.535–5:25.122
All right .
好的。
5:25.132–5:26.060
All right .
好的。
5:26.070–5:27.355
Very Gen 3
非常 Gen 3
5:27.365–5:28.772
aesthetic
美學風格
5:28.782–5:31.642
I'm going to turn sound on because I did see some mentions of it
我要開啟聲音，因為我確實看到有人提到它
5:31.652–5:34.833
just talking about like sounds .
只是在談論一些聲音。
5:34.843–5:35.071
All right .
好的。
5:35.081–5:37.146
It starts in our notepad .
它從我們的記事本開始。
5:37.156–5:38.900
Can we resize this ?
我們可以調整這個視窗的大小嗎？
5:38.910–5:39.618
All right .
好的。
5:39.628–5:40.257
Acceptable .
可以接受。
5:40.267–5:41.293
Minimize .
最小化。
5:41.303–5:42.410
Unminimize .
還原最小化。
5:42.420–5:42.968
Good .
很好。
5:42.978–5:43.686
The big one .
重點來了。
5:43.696–5:45.281
Is there a right click ?
有右鍵功能嗎？
5:45.291–5:48.720
Yeah , there is .
有，有的。
5:48.730–5:51.898
There is also the correct date and time in my local in the bottom
我的本地時間和日期也正確顯示在底部
5:51.908–5:52.137
right .
沒錯
5:52.147–5:55.011
And there is what I would imagine here to probably be the special
而這裡有我認為可能是特別的功能
5:55.021–5:55.410
feature .
特色
5:55.420–5:57.326
It is the holographic desktop feature .
這是全息桌面功能
5:57.336–5:59.002
So , we will save that for last .
所以我們把它留到最後再介紹
5:59.012–6:02.275
Let's start out by looking at the files app , which I don't think
我們先來看看檔案應用程式，我從來沒見過這樣的檔案圖示
6:02.285–6:04.909
I've seen a files icon that looks like this before .
我好像以前沒看過這樣的檔案圖示
6:04.919–6:05.309
All right .
好的
6:05.319–6:08.422
We have some nice transparency to the application backgrounds and
應用程式背景有一些漂亮的透明效果
6:08.432–6:11.462
definitely a nice synth wave theme , which I do personally like
而且確實有一個很棒的合成波主題，我個人很喜歡
6:11.462–6:15.197
.
。
6:15.197–6:15.277
So , unfortunately , it seems our fake file system is just static
所以，不幸的是，我們的假檔案系統似乎是靜態的
6:15.277–6:15.465
and we can't browse through it .
我們無法瀏覽它
6:15.475–6:17.866
Nonetheless , can we create new things in it ?
不過，我們能在裡面建立新東西嗎？
6:17.876–6:18.587
Okay , we can .
好的，我們可以
6:18.597–6:21.228
Can we remove the new things we created ?
我們可以移除我們建立的新東西嗎？
6:21.238–6:22.749
We also can .
我們也可以
6:22.759–6:25.951
Can we remove things that were there by default ?
我們可以移除預設就存在的东西嗎？
6:25.961–6:26.591
Yeah , we can .
是的，我們可以
6:26.601–6:27.231
All right , cool .
好的，很棒。
6:27.241–6:28.431
That makes me a little happier .
這讓我稍微安心一點。
6:28.441–6:30.431
And then up would just go there .
然後「向上」就會直接移到那裡。
6:30.441–6:30.990
Oh , okay .
哦，好的。
6:31.000–6:32.590
Not bad .
還不錯。
6:32.600–6:33.663
Decent bit of immersion ,
沉浸感還算不錯，
6:33.663–6:35.310
at least in terms of what directories
至少在目錄結構方面，
6:35.320–6:37.498
would exist were this a real computer .
如果這是一台真實的電腦，這些目錄應該會存在。
6:37.508–6:38.833
Notepad we did kind of see
我們確實看到了一些記事本，
6:38.833–6:40.284
because it opened with this
因為它開啟時顯示了這個
6:40.284–6:41.716
just to give us some information .
只是為了提供一些資訊給我們。
6:41.726–6:42.991
The only thing I like to see is
我只希望看到
6:42.991–6:44.203
will it save it as a text file
它是否會將檔案儲存為系統上的文字檔。
6:44.213–6:45.238
on the system .
沒有這樣的目錄。
6:45.248–6:46.988
No such directory .
我不是已經刪除文件目錄了嗎？
6:46.998–6:49.138
Didn't I delete the documents directory ?
我不是刪除了文件目錄嗎？
6:49.148–6:50.252
Hold on a second .
等一下
6:50.262–6:51.209
That's okay .
哦，它記得我做了那個動作。
6:51.219–6:53.040
Oh , it remembered that I did that .
喔，它記得我這麼做了。
6:53.050–6:54.203
I might have messed that up .
我可能搞砸了。
6:54.213–6:55.410
So that's going to be
所以這部分
6:55.410–6:56.084
that's on me .
算是我的錯。
6:56.094–6:57.132
Uh okay .
呃，好的。
6:57.142–7:00.112
This right here is just not right .
這裡的顯示根本不對。
7:00.122–7:03.495
It should not have a like visible text box .
它不應該出現可見的文字方塊。
7:03.505–7:04.945
That happens sometimes .
這種情況偶爾會發生。
7:04.955–7:06.637
Help did give us things .
Help 確實給了我們一些東西。
7:06.647–7:09.420
Let's just do neofetch .
我們就來執行 neofetch 吧。
7:09.430–7:13.636
That's actually pretty well done right there .
那裡的 neofetch 其實做得相當不錯。
7:13.646–7:16.299
Neofetch or Neofetch .
Neofetch 或 Neofetch。
7:16.309–7:17.630
Neofetch .
[未翻譯]
7:17.640–7:19.674
Still , that's that's actually all right .
不過，那其實還算可以。
7:19.684–7:21.545
I'm just being kind of mean .
我只是有點嚴苛。
7:21.555–7:22.653
Memory .
記憶體。
7:22.663–7:24.355
This is fine .
這還行。
7:24.365–7:27.928
Oh , it actually put a bit of cleverness in there .
哦，它其實在這裡加了一點巧思。
7:27.938–7:28.552
1 .
一
7:28.562–7:30.142
44 megs out of infinity .
從無限中取出 44 MB。
7:30.152–7:30.822
This is fine .
這還行。
7:30.832–7:33.854
Isn't that how much an old school floppy disc would hold ?
這不就是老式軟碟片的容量嗎？
7:33.864–7:35.914
I don't know cuz those are older than me .
我不知道，因為那些比我還老。
7:35.924–7:37.735
Next up , calculator .
接下來是計算機。
7:37.745–7:44.720
Okay , so 47 47 47 8 8 8 376 .
好的，所以 47 47 47 8 8 8 376。
7:44.730–7:46.636
I haven't done one of these in a while
我好久沒做這種事了。
7:46.636–7:47.424
divided by six .
除以六。
7:47.434–7:51.519
All right , settings .
好的，設定。
7:51.529–7:54.188
Definitely seeing a bit of like Qwen 2.5
我確實看到了一些類似 Qwen 2.5 8B 的東西。
7:54.198–7:54.808
8B in here
在風格方面。
7:54.808–7:57.171
just in terms of some of the style here.
就這裡的一些風格而言。
7:57.181–7:58.848
What is that
那是什麼？
7:58.848–8:01.348
and why do I like it ?
為什麼我喜歡它？
8:01.358–8:01.921
Aurora .
曙光。
8:01.931–8:03.068
Oh , okay .
哦，好的。
8:03.078–8:06.262
Sunset City .
日落城市。
8:06.272–8:09.456
This No one would understand that reference .
這沒人會懂這個梗
8:09.466–8:10.734
This looks like something
這看起來像某個
8:10.734–8:12.405
one of my friends in high school was
我高中時的一個朋友是
8:12.415–8:13.796
making as a game project .
作為遊戲專案製作的東西。
8:13.806–8:15.378
It actually looks like a ton
它實際上看起來像很多
8:15.378–8:17.564
like the splash screen for his project
就像他專案的啟動畫面
8:17.564–8:18.433
.
。
8:18.433–8:21.413
We have customizable accents as well .
我們也支援自訂口音。
8:21.423–8:24.117
Difficulty easy , normal , and hard .
難度有簡單、普通和困難。
8:24.127–8:25.100
Oh , for games .
喔，是遊戲。
8:25.110–8:26.628
Okay , I probably set it to easy
好的，我可能把它設為簡單
8:26.628–8:28.146
as you may be familiar with my
因為你可能熟悉我
8:28.156–8:29.912
quick to anger game playing style .
容易生氣的遊戲風格。
8:29.922–8:31.862
Next up , let's try our games .
接下來，讓我們試試我們的遊戲。
8:31.872–8:33.384
The speaker is on .
揚聲器已開啟。
8:33.394–8:37.354
So , City Heist .
所以，城市劫案。
8:37.364–8:44.554
Oh , you absolute .
喔，你真是個。
8:44.564–8:44.948
Okay .
好的。
8:44.958–8:47.711
So , you have to press enter to start
所以，你必須按 Enter 鍵開始
8:47.711–8:49.678
and not click the button
而不是點擊按鈕
8:49.678–8:50.976
.
。
8:50.976–8:52.446
Please shift gears .
請切換檔位。
8:52.456–8:53.827
Ah , okay .
啊，好的。
8:53.837–9:00.337
Okay .
好的。
9:00.347–9:00.830
All right .
好吧。
9:00.840–9:11.955
It's Oh , cool .
它是喔，真酷。
9:11.965–9:13.847
Oh , okay .
哦，好的。
9:13.857–9:16.065
It's troubled,
有點麻煩，
9:16.065–9:21.486
but the sounds for like the smack
但像是啪嗒聲之類的音效
9:21.486–9:23.407
are kind of funny .
有點好笑。
9:23.417–9:25.160
Like , all right .
像，好吧。
9:25.170–9:25.972
And we wrecked
我們搞砸了
9:25.972–9:27.508
and we only grabbed a 100 red
而且我們只拿到了100個紅色的
9:27.508–9:28.443
out of a thousand,
總共一千個，
9:28.453–9:29.466
but it did work .
但它確實起作用了。
9:29.476–9:30.168
So , keep in mind
所以，請記住
9:30.168–9:32.127
we need to press enter to play these games
我們需要按 Enter 鍵來玩這些遊戲
9:32.127–9:32.300
and
並且
9:32.310–9:34.740
not hit the button that says launch .
不要按寫著「啟動」的按鈕。
9:34.750–9:36.098
Oh .
哦。
9:36.108–9:43.897
Oh , all right .
哦，好吧。
9:43.907–9:45.950
Do I go through these ?
我要通過這些嗎？
9:45.960–9:46.272
Yeah .
是的。
9:46.282–9:48.364
I I just assume based on the sounds .
我只是根據聲音假設。
9:48.374–9:51.019
I wonder if that's how like dogs are since they don't understand
我想知道這是否就是狗的方式，因為它們不理解
9:51.029–9:51.744
language .
語言。
9:51.754–9:54.640
But like that to me seemed like a positive reinforcement sound
但對我來說，那聽起來像是正向強化的聲音
9:54.640–9:56.493
.
。
9:56.493–9:56.573
So , it makes me want to do it more .
所以，這讓我更想去做。
9:56.573–9:59.388
And I think we did just discover something about how dog brains
我認為我們剛剛發現了關於狗腦運作方式的某些事情
9:59.398–10:01.803
work .
。
10:01.813–10:03.654
Okay , the score .
好的，分數。
10:03.664–10:07.194
I don't know that we need that many decimal points just to show
我不認為我們需要這麼多小數位來顯示
10:07.204–10:07.677
a score .
分數。
10:07.687–10:12.424
I think maybe just rounding up to 734 or down would have been acceptable
我認為也許四捨五入到734或向下取整都是可以接受的
10:12.434–10:16.455
, but I applaud it for its intricacy .
，但我讚賞它的複雜性。
10:16.465–10:17.665
And then , oh , read me .
然後，哦，讀給我聽。
10:17.675–10:20.891
Okay , so our special feature is something like holographic .
好的，所以我們的特色功能類似於全息投影。
10:20.901–10:23.794
And I would imagine having a couple apps open just when we trigger
而且我想象一下，當我們觸發時，只打開幾個應用程式
10:23.804–10:25.408
that is probably the move .
時，打開幾個應用程式可能是正確的作法。
10:25.418–10:27.747
So , oh yeah .
所以，哦，是的。
10:27.757–10:28.149
All right .
好吧。
10:28.159–10:30.248
So , it just puts like an interesting card effect .
所以，它只是呈現出一種有趣的卡片效果。
10:30.258–10:31.457
This is what it looks like .
這就是它的樣子。
10:31.467–10:34.199
This looks like an Apple Vision Pro like app demo .
這看起來像是一個Apple Vision Pro的應用程式演示。
10:34.209–10:37.647
This is the way they look from inside the Vision Pro headset .
這是從Vision Pro頭戴裝置內部看到的樣子。
10:37.657–10:42.548
Um , just if you have used one , you may resonate with what I'm
嗯，如果你用過的話，可能會對我所說的話產生共鳴。
10:42.558–10:43.655
saying .
我說的話。
10:43.665–10:44.208
Okay .
好的。
10:44.218–10:46.025
Overall , it was fine .
整體來說，還算不錯。
10:46.035–10:46.973
Everything worked .
一切都運作正常。
10:46.983–10:48.159
Nothing really blew me away .
沒有什麼真正讓我驚艷的。
10:48.169–10:51.399
But again , that may be a limitation of the test and less of a
但再次強調，這可能是測試的局限性，而非模型本身的局限性。
10:51.409–10:52.506
limitation of the model .
模型的局限性。
10:52.516–10:55.131
So alas , huh , it's kind of neat .
所以唉，嗯，這還挺有趣的。
10:55.141–10:56.052
We will continue .
我們會繼續進行。
10:56.062–10:58.312
So our skateboard game is still going .
所以我們的滑板遊戲仍在進行中。
10:58.322–11:02.097
And again, I'm using this with the newly released beta version
同樣地，我使用的是新發布的 beta 版 ChatGPT 桌面應用程式。
11:02.107–11:03.985
of the chat GPT desktop app .
的 ChatGPT 桌面應用程式。
11:03.995–11:06.286
I would imagine this will take quite a bit of time .
我預計這會花費相當多的時間。
11:06.296–11:09.486
So let's check to see just like Oh no , cuz we already had looked
所以讓我們來看看，就像「哦不」，因為我們已經看過了成本。
11:09.496–11:10.202
at the cost .
成本。
11:10.212–11:13.929
And keep in mind this is partially for the skate game as well as
請記住，這部分費用涵蓋了滑板遊戲以及瀏覽器作業系統。
11:13.939–11:14.717
the browser OS .
瀏覽器作業系統。
11:14.727–11:17.655
So 20 cents total across two separate tasks .
所以兩項獨立任務總共只花了 20 美分。
11:17.665–11:18.945
One which is still ongoing .
其中一項任務仍在進行中。
11:18.955–11:19.231
All right .
好了。
11:19.241–11:24.748
So , after 68 minutes and 24 seconds , Skate NYC is complete .
所以，經過 68 分 24 秒後，Skate NYC 任務完成了。
11:24.758–11:26.110
So , oh , you idiot .
所以，喔，你這個笨蛋。
11:26.120–11:30.886
You spoiled it with these screenshots and you don't even have vision
你用這些截圖洩露了劇情，你甚至沒有視覺功能。
11:30.886–11:31.396
.
。
11:31.396–11:32.915
I'm infuriated .
我氣炸了。
11:32.925–11:36.637
Nonetheless , now that we know what the game looks like ,
不過，既然我們已經知道遊戲長什麼樣子，
11:36.647–11:40.152
I had to stop myself from saying something a little more colorful
我必須克制自己不說些更髒的話。
11:40.152–11:40.747
.
。
11:40.747–11:44.147
Let's just take a peek at at the game .
讓我們來看看遊戲吧。
11:44.157–11:47.331
All right , so here's our NYC skate game .
好了，這就是我們的紐約滑板遊戲。
11:47.341–11:48.605
Should be in outputs , right ?
應該在輸出結果裡對吧？
11:48.615–11:51.232
Yes .
對。
11:51.242–11:52.665
All right .
好了。
11:52.675–11:56.804
It's not quite as good as the Qwen 2.5
它沒有 Qwen 2.5
11:56.814–12:04.335
model, but that model does have 800 billion more parameters
模型那麼好，但那個模型確實多了 8000 億個參數。
12:04.335–12:04.845
.
。
12:04.845–12:07.101
Does that say NYC skate on the ground ?
地面上寫的是「NYC skate」嗎？
12:07.111–12:08.415
It's inverted , but it does .
它是倒著的，但確實有寫。
12:08.425–12:10.798
And it says established 1999 .
而且上面寫著「成立於 1999 年」。
12:10.808–12:11.175
All right .
好的。
12:11.185–12:12.280
Uhoh .
哎呀。
12:12.290–12:16.396
Sounded like the beginning of a Van Halen song
聽起來像是范·海倫樂團歌曲的開頭。
12:16.406–12:18.840
I don't know what exactly it is
我不確定這到底是什麼。
12:18.850–12:19.773
All right .
好的。
12:19.783–12:22.766
Uh , let's just maybe not do that same thing .
呃，我們也許別再做同樣的事了。
12:22.776–12:24.020
There are sound effects , though .
不過還是有音效。
12:24.030–12:25.431
I don't normally hear sound effects .
我通常聽不到音效。
12:25.441–12:27.626
Can we actually move around?
我們真的能移動嗎？
12:27.636–12:28.097
All right .
好的。
12:28.107–12:29.273
Not quite .
還不太行。
12:29.283–12:32.176
I will say it's not wonderful .
我會說這並不怎麼好。
12:32.186–12:33.533
Oh , that's actually kind of cool , though .
不過，這其實有點酷。
12:33.543–12:37.247
The way it like stylized the game title .
它將遊戲標題風格化的方式。
12:37.257–12:37.747
Nice .
不錯。
12:37.757–12:40.675
Nice .
不錯。
12:40.685–12:43.890
Well , we don't need to do the drum kit sim .
好吧，我們不需要做鼓組模擬。
12:43.900–12:53.313
The trees are quite reminiscent of a poo, if I were to be honest
老實說，這些樹讓人有點聯想到糞便。
12:53.323–13:02.090
, but all right , the the legs are like kicking .
不過沒關係，那些腿看起來像是在踢。
13:02.100–13:05.018
This may be the funniest skate game result we've ever seen ,
這可能是我們看過的最好笑的滑板遊戲結果，
13:05.028–13:06.283
which is , you know , something .
這算是……某種成就吧。
13:06.293–13:08.419
All right , we have a street sign there .
好了，這裡有個路牌。
13:08.429–13:12.690
The text is a little sideways , as are the building faces .
文字有點側著，建築立面也是。
13:12.700–13:14.589
This is definitely a troubled result .
這絕對是個有問題的結果。
13:14.599–13:17.436
However , I do like the water fountain shooting out .
不過，我還挺喜歡噴泉噴射出來的效果。
13:17.446–13:21.988
I believe that was something I've seen .
我相信我之前有看過類似的畫面。
13:21.998–13:22.894
Okay .
好的。
13:22.904–13:23.636
All right .
好了。
13:23.646–13:26.150
I've seen that in, I think, the Qwen result.
我想我在 Qwen 的結果裡有看過這個。
13:26.160–13:29.357
Now , here's what I would like to just quickly do for some reference
現在，我想快速做一下，作為一些參考
13:29.357–13:30.327
.
。
13:30.327–13:33.987
This is the result we received for the C++ skateboard test from the
這是我們從 C++ 滑板測試中收到的結果，來自
13:33.997–13:36.620
preview version of DeepSeek V4.
DeepSeek V4 的預覽版收到的 C++ 滑板測試結果。
13:36.630–13:37.659
We can see right here .
我們可以直接在這裡看到。
13:37.669–13:39.738
I'm going to turn the speaker off .
我要把聲音關掉。
13:39.748–13:41.258
This is pretty bad .
這相當糟。
13:41.268–13:45.257
It was like so bad that I actually remember to go to this specific
糟糕到讓我甚至記得要特別去拿這個東西來展示
13:45.267–13:46.777
thing just to showcase it .
只是為了展示它
13:46.787–13:51.256
So, the leap in capability based off of what we see right here
所以，根據我們現在看到的，能力的躍進
13:51.266–13:54.850
to what we just saw in the skate game that it had created for us
以及剛才在滑板遊戲中它為我們創造的結果
13:54.860–13:58.374
, which we can just look at in these screenshots here is still
從這些截圖中可以看出，這仍然
13:58.384–13:59.254
pretty significant .
相當顯著
13:59.264–14:01.653
Next up , I'm giving it the Subway FPS test .
接下來，我將對它進行 Subway FPS 測試。
14:01.663–14:03.653
I've also added on to the end of the prompt .
我也在提示詞的末尾添加了內容。
14:03.663–14:05.492
Don't worry about checking over the script .
不用擔心檢查腳本的問題
14:05.502–14:08.132
I will manually check it and if there are issues I will report
我會手動檢查，如果有問題，我會回報
14:08.142–14:11.491
them back to you, just because this will really extend every
給你，因為這確實會將每項結果
14:11.501–14:16.228
single result by like two times in terms of time-wise length
在時間長度上，單一結果大約兩倍
14:16.238–14:19.808
just in the interest of something I could do visually in like 5
只是出於我想在5秒內視覺化呈現的東西
14:19.818–14:21.402
seconds and then be like hey this doesn't work .
秒內視覺上完成某事，然後說「嘿，這不起作用」。
14:21.412–14:24.671
So we're going to do that just to maybe like keep things a little
所以我們這樣做，可能是為了讓事情稍微
14:24.681–14:25.148
tighter .
緊湊一些。
14:25.158–14:28.257
All right so it did finish the Subway FPS game .
好的，它完成了 Subway FPS 遊戲。
14:28.267–14:31.127
The problem is it just keeps trying to do like troubleshooting
問題在於它不斷嘗試進行類似故障排除的操作，
14:31.137–14:34.635
which it's not good at doing on this system for some reason .
不知為何，它在這個系統上並不擅長此類操作。
14:34.645–14:37.302
All right , let's check out our subway FPS .
好的，讓我們來看看我們的 Subway FPS。
14:37.312–14:39.101
Main Rand Avenue.
主街大道站。
14:39.111–14:42.547
All right , going to turn the speaker on .
好的，我要打開揚聲器。
14:42.557–14:44.063
So , it gave us two options .
所以，它給了我們兩個選項。
14:44.073–14:46.536
We can either explore the station .
我們可以選擇探索車站。
14:46.546–14:50.525
Okay, it's... Hey, you know what?
好，它……嘿，你知道嗎？
14:50.535–14:54.913
I think that is the best subway car that I've ever seen drawn in
我認為這是我看過畫得最好的地鐵車廂
14:54.923–14:56.269
this test .
這次測試中看過畫得最好的地鐵車廂。
14:56.279–14:58.623
And you're going to be like , well , it doesn't look that great
而你可能會說，嗯，它看起來並沒有那麼好。
14:58.623–15:00.112
.
。
15:00.112–15:00.192
I know .
我知道。
15:00.192–15:01.534
It's just that they never really include these .
只是他們從來沒有真正包含這些。
15:01.544–15:03.769
And when they do , they're not really very good .
而當他們包含時，通常也不是很好。
15:03.779–15:06.801
So this... actually, what was that?
所以這……實際上，那是什麼？
15:06.811–15:10.870
If that thing starts moving I would have probably become more excited
如果那個東西開始移動，我可能會變得比
15:10.880–15:15.177
than would be reasonable for you know responding to such thing
你認為回應這種情況所合理的興奮程度還要興奮。
15:15.177–15:15.746
.
。
15:15.746–15:16.853
This is very clean .
這非常乾淨。
15:16.863–15:17.490
It's simple .
很簡單。
15:17.500–15:20.044
It's nothing that's blowing me away to be completely honest with
老實說，這沒什麼讓我驚艷的
15:20.054–15:20.284
you .
你。
15:20.294–15:21.560
Queens Boulevard line .
皇后大道線。
15:21.570–15:21.880
Okay .
好的。
15:21.890–15:22.677
We have stairs .
我們有樓梯。
15:22.687–15:25.150
We have some lighting elements and things of the sort .
我們有一些照明元素之類的東西。
15:25.160–15:28.341
We have exits if you want to go that side.
如果想去那邊的話，我們有出口。
15:28.351–15:34.407
And then tunnels, like indicators for the train, a trash bin.
然後還有隧道，像是列車的指示燈、垃圾桶。
15:34.417–15:37.121
Floating particles more simple .
浮動粒子更簡單。
15:37.131–15:38.799
And we can't go up the stairs , which is fine .
我們不能走上樓梯，這沒關係。
15:38.809–15:39.677
We don't want to .
我們也不想上去。
15:39.687–15:43.909
Now , of course , the other feature that was included here is the
現在，當然，這裡包含的另一個功能是
15:43.919–15:45.106
FPS .
幀率（FPS）。
15:45.116–15:48.381
So, let's go back to the menu and then enter the survival FPS.
所以，讓我們回到選單，然後進入生存模式的幀率。
15:48.391–15:50.377
Let's see if it changes the lighting or anything .
讓我們看看它是否改變了燈光或其他東西。
15:50.387–15:51.336
Oh my .
天啊。
15:51.346–15:53.012
That's a good weapon model .
那個武器模型不錯。
15:53.022–15:54.370
Does it ?
有嗎？
15:54.380–15:54.769
How ?
怎麼說？
15:54.779–15:57.484
Oh , it does .
哦，確實有。
15:57.494–15:59.480
I think .
我想是。
15:59.490–16:01.237
Yeah , it does .
是的，確實有。
16:01.247–16:02.595
It's like smooth now .
現在看起來很順暢。
16:02.605–16:03.951
I don't see any enemies .
我看不到任何敵人。
16:03.961–16:05.149
Look at when we run .
看看我們跑起來的時候。
16:05.159–16:07.865
It's like a waddle .
就像在搖擺。
16:07.875–16:08.263
All right .
好的。
16:08.273–16:12.895
So , uh the big thing I'm noticing here is there are no zombie
所以，嗯，我現在注意到的一大重點是這裡沒有殭屍
16:12.905–16:13.454
enemies .
敵人。
16:13.464–16:17.047
So , good reload .
所以，重新裝填得不錯。
16:17.057–16:22.877
The waddle is just kind of funny .
那搖擺的動作有點好笑。
16:22.887–16:24.314
Weird .
很奇怪。
16:24.324–16:24.792
Hey , look .
嘿，看。
16:24.802–16:26.471
There are shell casings that actually drop .
實際上會有彈殼掉落。
16:26.481–16:27.672
That's kind of cool .
這有點酷。
16:27.682–16:30.235
All right, this is like... yeah, it's kind of mid.
好的，這就像……是啊，有點平庸。
16:30.245–16:33.279
So , I'm going to give it a feedback piece .
所以，我會給它一些反饋。
16:33.289–16:36.162
Now , I stopped it because it was just like it had finished and
現在，我把它停下來，因為它看起來像是已經完成了，
16:36.172–16:38.925
it was like I can't use this in the headless environment to check
而且感覺我無法在無頭環境中使用它來檢查
16:38.925–16:40.891
.
。
16:40.891–16:41.608
And at that point , it's easier to just check it and tell it what's
在這種情況下，直接檢查並告訴它哪裡出錯會更容易。
16:41.618–16:42.809
wrong .
錯誤。
16:42.819–16:44.332
So , I'll give it that honest feedback .
所以，我會給它誠實的反饋。
16:44.342–16:48.255
So , supposedly our fixed and polished version of the Subway FPS
所以，據說我們修復並優化過的《地鐵射擊遊戲》
16:48.265–16:49.376
is completed .
版本已經完成了。
16:49.386–16:51.138
Let's just take a peek at it .
讓我們來看看它吧。
16:51.148–16:52.584
The sound is on .
聲音已開啟。
16:52.594–16:53.460
All right .
好的。
16:53.470–16:54.013
Interesting .
很有趣。
16:54.023–16:56.460
I think it swapped the positioning of the buttons .
我認為它交換了按鈕的位置。
16:56.470–17:00.130
So , all right .
所以，好的。
17:00.140–17:01.015
Seven hostiles .
七個敵方目標。
17:01.025–17:02.626
It's showing me in the top right .
它顯示我在右上角。
17:02.636–17:07.212
I unfortunately just can't locate them .
不幸的是，我無法找到它們。
17:07.222–17:09.867
I would have expected it to have done better .
我原本預期它會表現得更好。
17:09.877–17:12.362
And again , this is one where we could go back and reference the
同樣地，這是我們可以回頭參考
17:12.372–17:14.615
previous model of this .
該模型先前版本的情況。
17:14.625–17:18.639
So , basically the preview version , it was significantly less
所以，基本上，預覽版本的細節程度
17:18.649–17:22.583
detailed than this was .
遠低於這個版本。
17:22.593–17:26.043
though unfortunately we don't have enemies and given the amount
雖然不幸的是我們沒有敵人，而且考慮到
17:26.053–17:29.251
of time it took as well as the kind of multiple times I told it
所花費的時間，以及我多次告訴它
17:29.261–17:32.375
there were no enemies, I'm regrettably going to call this one a next up.
沒有敵人的情況，我很遺憾地將這個版本列為下一個。
17:32.385–17:36.126
Now I'm giving this a difficult task to create a 3D model of an
現在我給它一個困難的任務，要創建一個
17:36.136–17:40.866
RB26 inline 6 engine it must be something that can be 3D printed
RB26 直列六缸引擎的 3D 模型，它必須是可以
17:40.876–17:44.917
with little to no support and fit inside of it a tiny DC motor
以極少或無需支撐結構進行 3D 列印，
17:44.927–17:48.400
called an N20 motor because this is something that I've told it
並且能容納一個稱為 N20 馬達的小型直流馬達，因為這是我曾告訴它
17:48.410–17:51.430
should like work with my RC car project. Actually, I don't think
應該能跟我的遙控車專案配合。其實，我不認為
17:51.440–17:53.662
that's specifically in this prompt , but that's the whole point
這在提示詞裡有特別提到，但這就是重點
17:53.672–17:57.169
of this, where it should understand and it will scale it
所在，它應該會理解，並會將它
17:57.179–17:58.524
properly to that motor .
正確地縮放至該馬達的尺寸。
17:58.534–18:02.031
So far , I've run this test with a few intelligent models and I've
到目前為止，我用過幾個智慧模型來執行這項測試，我從來
18:02.041–18:04.263
never really received anything that was actually worth printing
沒有真正收到任何值得列印的東西
18:04.263–18:07.577
.
。
18:07.577–18:07.657
So , hopefully sooner than later that will change .
所以，希望這能盡快改變。
18:07.657–18:10.559
So , our 3D model task is pretty interesting because it's just
所以，我們的 3D 模型任務相當有趣，因為它只是
18:10.569–18:14.068
doing a lot of research on like the actual proper orientation of
針對該引擎上特定物品的實際正確方向
18:14.078–18:16.140
specific items on this engine .
這引擎上的特定物品。
18:16.150–18:18.777
Like where is the downpipe , which is an exhaust component ?
例如，排氣管（downpipe）在哪裡？那是排氣組件嗎？
18:18.787–18:19.656
What side is it ?
它在哪一側？
18:19.666–18:20.775
What is the intake side ?
進氣側在哪裡？
18:20.785–18:21.975
What is the exhaust side ?
排氣側在哪裡？
18:21.985–18:23.493
Intake plenum left or right ?
進氣歧管（Intake plenum）是在左側還是右側？
18:23.503–18:28.049
So , I do like to see the amount of like intricate research that's
所以，我很喜歡看到這種細緻研究的程度，
18:28.059–18:30.607
going into the culmination of this model creation .
這些研究都匯聚在這個模型創作的成果中。
18:30.617–18:34.923
And it would be really awesome for like us to get a model for this
如果我們能為這項任務獲得一個
18:34.933–18:36.280
task that's actually printable .
真正可以列印的模型，那將非常棒。
18:36.290–18:37.480
It would be the first one .
這將是第一個。
18:37.490–18:41.816
So , in 1 hour and 10 minutes , we received our 3D model response
因此，在1小時10分鐘後，我們收到了3D模型的回覆。
18:41.816–18:44.543
.
。
18:44.543–18:45.914
This is a very, very intricate bit of text about the thing
這是一段非常、非常複雜的文字，描述它所生成的東西。
18:45.924–18:46.717
it made .
它生成的東西。
18:46.727–18:49.288
Now , of course , the other part of that is actually just taking
現在，當然，另一部分就是實際上去看看它長什麼樣子。
18:49.298–18:51.135
a look at what it looks like .
看看它長什麼樣子。
18:51.145–18:53.626
Though , once I remember what I named this folder ,
不過，一旦我記起我給這個資料夾取的名字，
18:53.636–18:54.991
we will be able to take a peek .
我們就能夠偷偷看一眼。
18:55.001–18:56.356
Assembled display STL .
組裝顯示器STL。
18:56.366–18:58.364
I would imagine that to be the case .
我猜應該是這樣沒錯。
18:58.374–18:59.489
Please be good .
請爭點氣。
18:59.499–19:03.909
I've really, like, I need a good one for these.
我真的，呃，我需要一個好的來做這些。
19:03.919–19:06.158
Actually , it's sideways .
實際上，它是側著的。
19:06.168–19:06.880
Son of a gun .
見鬼。
19:06.890–19:10.335
This did a really good job .
這做得相當好。
19:10.345–19:15.556
This is actually like so far the best one .
這實際上是目前為止最好的。
19:15.566–19:17.244
It's not perfect .
它並不完美。
19:17.254–19:21.100
We have some sideways turbocharger issues , though .
不過，我們有一些渦輪增壓器側向的問題。
19:21.110–19:22.383
I mean, this is all right.
我的意思是，這還行。
19:22.383–19:23.991
Let's peek at this in OpenSCAD
讓我們在 OpenSCAD 裡來看看這個
19:23.991–19:24.677
.
。
19:24.677–19:26.320
And there are a bunch of individual pieces for this .
這裡有好幾個單獨的零件。
19:26.330–19:28.907
So , we also want to see if the engine that we told it needs to
所以，我們也想確認一下，我們告訴它需要
19:28.917–19:31.273
fit inside the little remote control car engine .
塞進那個小型遙控車引擎裡。
19:31.283–19:34.634
Let's see if it has any hope of properly fitting.
讓我們看看它有沒有機會正確地裝進去。
19:34.634–19:35.340
Research—
研究——
19:35.350–19:39.111
It has a research folder right here just showing like a bunch of
這裡有個研究資料夾，裡面顯示了
19:39.121–19:41.772
images and stats for this motor .
這個馬達的各種圖片和數據。
19:41.782–19:42.660
Very cool though .
真的很酷。
19:42.670–19:46.392
And here we have some better preview renders .
這裡我們有一些更好的預覽渲染圖。
19:46.402–19:50.501
This is definitely well done .
這絕對做得很好。
19:50.511–19:52.041
Turbo side .
渦輪側面。
19:52.051–19:52.760
Okay .
好的。
19:52.770–19:53.966
Underside .
底部。
19:53.976–19:54.850
All right .
好吧。
19:54.860–19:58.466
It's not something that's going to really easily be 3D printable
這不太可能輕易地用 3D 列印出來
19:58.466–19:58.976
.
。
19:58.976–20:01.360
Although , I guess I can't make that determination because this
不過，我想我無法做出這樣的判斷，因為這個
20:01.370–20:02.968
comes in multiple pieces .
是由多個零件組成的。
20:02.978–20:06.344
So , it's not designed to be printed as we first saw it just in
所以，它並非設計成我們最初看到的那樣，作為一個整體來列印。
20:06.354–20:07.548
one lump like this .
像這樣的一整塊。
20:07.558–20:09.719
It's designed to be printed separately .
它是設計成分開列印的。
20:09.729–20:11.005
I would like to see .
我想看看。
20:11.015–20:14.381
Let's look now at just the single first piece ,
現在讓我們只看單獨的第一個零件，
20:14.391–20:18.720
which is ideally where it would be placing the 3D printed motor
這正是理想情況下放置 3D 列印馬達的位置，
20:18.730–20:21.454
that would go in it .
馬達將安裝在裡面。
20:21.464–20:25.473
Unfortunately , printing this without supports is just not an option
不幸的是，在我們目前看到的任何潛在方向中，
20:25.483–20:28.526
in any potential orientation that we see right here ,
不使用支撐結構來列印這部分根本不可行，
20:28.536–20:29.892
which is a bit frustrating .
這有點令人沮喪。
20:29.902–20:32.626
It almost seems like it has these posts and holes in it to be able
這看起來似乎有這些柱子和孔洞，
20:32.636–20:35.438
to snap fit on other parts, which is cool.
以便卡扣安裝到其他零件上，這很酷。
20:35.448–20:36.243
Okay .
好的。
20:36.253–20:36.563
Yep .
是的。
20:36.573–20:39.217
And I think those are definitely like attachment posts ,
我認為那些絕對像是附件安裝點，
20:39.227–20:40.262
which is pretty cool .
這真的很酷
20:40.272–20:43.717
This one would be able to be printed more or less without any supports
這個幾乎不需要任何支撐結構就能列印。
20:43.727–20:46.589
if we were to put it flat .
這個零件基本上可以不用任何支撐結構來列印。
20:46.599–20:50.712
Then this should have the Nissan logo, or at least the text "Nissan"
然後這上面應該會有 Nissan 的標誌，或者至少寫上「Nissan」的文字，
20:50.722–20:53.420
written on it .
印在上面。
20:53.430–20:53.883
Yep .
沒錯。
20:53.893–20:55.741
And this one's quite well done .
這一個做得相當不錯。
20:55.751–20:56.668
No issues here .
這裡沒有問題。
20:56.678–20:57.781
The oil pan .
油底殼。
20:57.791–20:58.262
Oh , no .
喔，不。
20:58.272–20:58.662
Okay .
好吧。
20:58.672–21:01.785
This is where we decide we're not going to be
這裡我們決定不再嘗試列印這個。
21:01.795–21:02.666
trying to print this .
試著列印它。
21:02.676–21:07.391
I'm not disappointed , but I'm not super impressed either .
我不會失望，但也沒有特別驚艷。
21:07.401–21:11.315
So , I suppose I am lukewarm when it comes to judging this .
所以，我想在評價這個時，我算是態度中立。
21:11.325–21:14.198
Definitely , if we just look at this as a render ,
絕對地，如果我們只把它當作一張渲染圖來看，
21:14.208–21:16.842
even forget about looking at the individual parts there ,
甚至別去管那裡面的個別零件，
21:16.852–21:20.225
cuz it's not really going to be feasible to print this ,
因為實際上列印這個並不可行，
21:20.235–21:22.963
it did a really decent job just assembling it .
它只是組裝起來就做得相當不錯了。
21:22.973–21:25.542
The research that went into this, I think, is probably
我認為這項研究投入，大概就像
21:25.552–21:29.328
like I hate to sound like an AI here , but like the story isn't
我討厭在這裡聽起來像個AI，但故事不是
21:29.338–21:31.339
the model , it's the research that went into it .
模型，而是背後的研究。
21:31.349–21:33.255
That's like generic AI slop.
那就像是一般AI生成的垃圾內容。
21:33.265–21:37.085
It really did a lot of decent research here .
它確實做了很多不錯的研究。
21:37.095–21:41.075
I mean , did it find this and download it just to see ?
我的意思是，它找到並下載這個只是為了看看嗎？
21:41.085–21:41.474
No .
不。
21:41.484–21:41.793
Okay .
好的。
21:41.803–21:45.464
So , this is something it maybe made or maybe probably found online
所以，這可能是它自己生成的，也可能是從網路上找到的
21:45.474–21:48.616
just to be able to better assess what the size of this asset is
以便更好地評估這個資產的大小
21:48.626–21:51.768
, but it took a lot of care to ensure that what it tried to generate
，但它非常謹慎，以確保它嘗試生成的內容
21:51.778–21:55.519
would have hypothetically fit the specific specifications that
在假設上符合原始提示中指定的具體規格。
21:55.529–21:57.674
were denoted in the original prompt .
原始提示中指定的具體規格。
21:57.684–21:59.669
So , not half bad .
所以，還不錯。
21:59.679–22:02.621
I'm also going to give this a simple front end design test where
我還要對這個進行簡單的前端設計測試，
22:02.631–22:04.377
it needs to create the watch website .
它需要創建手錶網站。
22:04.387–22:07.643
So , it has to have a pretty cinematic panning effect over a 3D
所以，它必須具備相當電影感的3D橫掃效果。
22:07.653–22:09.222
watch asset that it creates .
手錶資產進行相當有電影感的平移效果。
22:09.232–22:12.538
I did specifically denote that
我特別標註了那點
22:12.548–22:14.197
the watch needs to be a 3D model .
手錶必須是 3D 模型。
22:14.207–22:17.039
So , we'll see how competent it is at putting the entirety of this
所以，我們來看看它整合這一切的
22:17.049–22:17.513
together .
能力如何。
22:17.523–22:21.817
So , in 1 hour and 30 minutes , we got our watch front end result
所以，在 1 小時 30 分鐘後，我們得到了手錶前端結果
22:21.817–22:22.327
.
。
22:22.327–22:25.803
It again seems to have a pretty substantial like just list of what
它再次似乎有一個相當實質的，就像只是列出了這裡
22:25.813–22:26.769
was built here .
所構建的內容清單。
22:26.779–22:28.460
So , let's take a look and see how it is .
那麼，讓我們來看看它的情況如何。
22:28.470–22:32.728
a beautiful watch front-end website.
一個精美的手錶前端網站。
22:32.738–22:36.096
That's actually, it's so close to excellent.
實際上，它已經非常接近優秀了。
22:36.106–22:40.745
The big problem is this , I would assume , is the watch strap sticking
我認為最大的問題在於，手錶錶帶突出來了。
22:40.755–22:41.306
out .
突出來了。
22:41.316–22:45.173
This is 3D, so don't be fooled by its 2D-ish look.
這是3D效果，不要被它看似2D的外觀所欺騙。
22:45.183–22:46.552
It is actually 3D .
它確實是3D的。
22:46.562–22:48.011
It's slowly panning , too .
它也在緩慢地平移。
22:48.021–22:52.088
I really... Okay, so a couple things I do like: the wooden background
我真的……好吧，我有幾點喜歡的地方：木質背景。
22:52.098–22:55.742
and then this like leather card or pad that this watch is placed
以及這個類似皮革卡片或墊子，手錶就放在上面。
22:55.752–22:56.297
on .
放在上面。
22:56.307–22:58.440
In addition to that, the face looks really good.
除此之外，錶盤看起來真的很不錯。
22:58.450–23:00.347
The materials look very good .
材質看起來非常好。
23:00.357–23:03.998
It's just like, overall, it's a high-end aesthetic,
總體來說，它呈現出高端的美感，
23:04.008–23:07.254
which is exactly what we go for when performing this test .
這正是我們進行這項測試時所追求的。
23:07.264–23:13.563
Now , that's actually that right there looks good .
現在，那裡的確看起來不錯。
23:13.573–23:16.975
The... Let's just sit and appreciate it for a minute.
讓我們靜下來欣賞它一分鐘。
23:16.985–23:17.215
Okay .
好的。
23:17.225–23:19.290
Yeah , that maybe I could have done without that ,
是的，也許我可以不要那個，
23:19.300–23:22.962
but it's definitely quite something .
但它確實相當引人注目。
23:22.972–23:25.435
And then look at the way that goes away as soon as we start to
接著看看我們開始往下捲動時，它如何消失
23:25.445–23:26.633
scroll down .
往下捲動
23:26.643–23:30.225
"Quiet luxury measured in seconds."
以秒計量的靜奢風
23:30.235–23:31.582
"Sunburst dial."
放射紋錶盤
23:31.592–23:33.657
Okay , it has like a material card here .
好的，這裡有一張像是材質說明卡的東西
23:33.667–23:37.249
The sapphire crystal is perhaps something I would have maybe omitted
藍寶石水晶鏡面或許是我本來會省略的部分
23:37.259–23:39.564
from this section , but the dial looks good .
雖然這部分我可能不會保留，但錶盤看起來不錯
23:39.574–23:41.560
"Two watches, one point of view."
兩款手錶，同一個觀點
23:41.570–23:42.438
Please render them .
請幫它們渲染出來
23:42.448–23:43.236
Oh , look at that .
哇，你看那個
23:43.246–23:44.194
That is sick .
這太酷了
23:44.204–23:45.629
That really looks pretty good .
這看起來真的很不錯
23:45.639–23:47.546
I love the way they're placed on the table .
我喜歡它們在桌上的擺放方式
23:47.556–23:50.420
Really the big thing letting it down is just this leather slab
其實真正拖後腿的就是這塊皮革板
23:50.430–23:53.533
is inhibiting our view of these beautiful straps .
這塊皮革板擋住了我們欣賞這些精美錶帶的視線
23:53.543–23:55.688
This one is the stainless steel option .
這是不鏽鋼的選項
23:55.698–23:59.679
This one is the gold option and priced accordingly I would imagine
這款是金色選項，價格自然也會相應提高
23:59.679–24:01.292
.
這是鈦金屬選項，這是黃金選項，價格也相應較高，我想
24:01.292–24:01.372
Yeah .
沒錯
24:01.372–24:05.665
Although I would say the jump in price is like a negligible
不過我會說，價格的跳動幅度微乎其微
24:05.675–24:06.702
amount to get the gold one .
為了拿到金色款所增加的金額其實微乎其微
24:06.712–24:08.539
I'd probably go with this one personally .
我個人大概會選這一款
24:08.549–24:08.938
Oh , cool .
喔，不錯
24:08.948–24:12.370
And these ones we can freely rotate so that helps a bit .
而且這些可以隨意旋轉，所以還是有點幫助
24:12.380–24:15.643
There is no movement to the watch hands but that's okay .
雖然錶針沒有動，不過沒關係
24:15.653–24:17.240
I'll forgive it for that .
我願意原諒它這一點
24:17.250–24:19.395
But these are pretty darn cool .
但這些真的非常酷
24:19.405–24:22.508
These are well drawn , I would say .
我覺得這些畫得很精緻
24:22.518–24:25.860
Especially the numeral indicators .
特別是那些數字指示器
24:25.870–24:28.814
Look at the way that they actually have multi material on them
看看他們實際上的多材質設計
24:28.814–24:29.770
.
所以它們並非只有一種靜態材質，而是有金色搭配黑色填充，或是銀色搭配黑色填充。
24:29.770–24:31.766
So , they're not just like one static material .
所以它們並非只有一種靜態材質
24:31.776–24:35.599
It's gold and then black infill or silver and then black infill
是金色搭配黑色填充，或是銀色搭配黑色填充
24:35.609–24:35.918
as well .
也是
24:35.928–24:38.473
So , that's some attention to detail I do like seeing .
所以我很欣賞這種注重細節的設計
24:38.483–24:42.383
As well as the kind of brushed material look on the stainless steel
以及不鏽鋼與金色部分那種刷紋材質的質感
24:42.393–24:42.943
and the gold .
以及金色部分
24:42.953–24:46.215
It's probably weird we're like assessing like a 3D watch ,
這可能有點奇怪，我們居然在評估一款3D手錶
24:46.225–24:48.690
but it just shows like some capability .
但這確實展現了某種能力
24:48.700–24:50.429
Okay , that looks pretty good right there .
好的，這部分看起來相當不錯。
24:50.439–24:52.335
And then we have a clean footer as well .
接著我們還有一個乾淨的頁尾。
24:52.345–24:52.997
I like this .
我喜歡這個。
24:53.007–24:54.075
I'm happy with this .
我對這個很滿意。
24:54.085–24:58.632
Even though it's like not perfect , it's darn close .
雖然它不是完美的，但也已經非常接近了。
24:58.642–25:00.919
So, next up, just here from within Codex.
所以，接下來，我們直接在 Codex 裡面操作。
25:00.929–25:03.873
And I can't show this full prompt because it has an API key .
我無法顯示完整的提示詞，因為裡面包含 API 金鑰。
25:03.883–25:07.228
I'm giving this the full on cinematic Steve the PC repair game
我將對它進行完整的「史蒂夫，電腦維修遊戲」電影風格測試，
25:07.238–25:09.624
test where it needs to create a cinematic game .
測試它是否能創建一個電影風格的遊戲。
25:09.634–25:12.900
It needs to create voice assets for these characters just by using
它需要僅憑我提供的 OpenAI API 金鑰，為這些角色建立語音素材
25:12.910–25:16.595
an open AI API key that I've given it and it needs to make a fun
OpenAI API 金鑰，為這些角色創建語音資產，並製作一個有趣
25:16.605–25:18.874
game that is enjoyable and playable .
且令人愉悅、可遊玩的遊戲。
25:18.884–25:21.940
I will assess this is probably going to take north of an hour .
我評估這大概需要一個小時以上的時間。
25:21.950–25:24.694
So, we're just going to basically leave Codex here in the background
所以，我們基本上就把 Codex 留在背景中，
25:24.704–25:28.538
the beta desktop version of it to complete this task, and we'll
使用它的 Beta 桌面版來完成這個任務，同時我們
25:28.548–25:30.255
maybe work on some other things in the meantime .
也許可以順便處理其他一些事情。
25:30.265–25:34.365
So , our Steve the PC repairman cinematic game completed in basically
所以，我們的「史蒂夫，電腦維修員」電影風格遊戲，基本上在
25:34.375–25:35.187
65 minutes .
65 分鐘內完成了。
25:35.197–25:36.755
So , a little over an hour .
也就是說，稍微超過一個小時。
25:36.765–25:38.250
All right , here's how to play .
好的，這是遊戲玩法說明。
25:38.260–25:41.537
I'll probably just do what it says here , not assume that it's
我大概會照這裡寫的做，不會假設它
25:41.547–25:43.777
already running this in some server for us .
已經在我們的某個伺服器上運行這個了。
25:43.787–25:47.737
So, okay, episode one, the Azure Palm job.
好的，第一集，Azure Palm 的工作。
25:47.747–25:49.231
Okay , click to begin .
好的，點擊開始。
25:49.241–25:50.502
Headphones recommended .
建議使用耳機。
25:50.512–25:53.084
We have a speaker .
我們有喇叭。
25:53.094–25:57.007
All right .
好的。
25:57.017–25:58.867
My name is Steve .
我是 Steve。
25:58.877–25:59.693
This is good .
這很好。
25:59.703–26:02.170
I fix computers .
我修理電腦。
26:02.180–26:05.516
Most of my clients are old enough to remember when a hard drive
我的大多數客戶都夠老，記得硬碟
26:05.526–26:07.306
was a luxury item .
曾經是奢侈品。
26:07.316–26:10.768
They pay what they can afford and I make it work .
他們付得起多少我就收多少，然後讓它運作。
26:10.778–26:14.143
Mostly honest business .
主要是誠實的交易。
26:14.153–26:15.855
Oh , thank you , dear .
哦，謝謝你，親愛的。
26:15.865–26:20.261
The screen kept telling me the date was 1980 and I thought, all
螢幕一直告訴我日期是 1980 年，我想，全
26:20.271–26:20.834
fixed.
修好了。
26:20.844–26:22.554
I have been traveling long enough .
我已經旅行夠久了。
26:22.564–26:23.646
Fresh battery .
全新的電池。
26:23.656–26:26.196
Your computer now believes I will bring you oatmeal cookies
你的電腦現在相信我會帶燕麥餅乾
26:26.206–26:26.860
on Tuesday .
在星期二來。
26:26.870–26:27.720
Same as the rest .
跟其他人一樣。
26:27.730–26:28.065
Give Mr.
請替我幫 Mr.
26:28.075–26:30.110
Buttons a scratch behind his ears for me.
Buttons 撓撓他耳朵後面。
26:30.120–26:33.580
Let me walk you out , ma'am .
讓我送您出去，夫人。
26:33.590–26:36.020
I'll be right with you , Mr .
我馬上就來，Thomas 先生，這段演講會互相覆蓋。
26:36.030–26:38.727
Thomas , the speech overlays itself .
好男孩，Steven。
26:38.737–26:40.513
A good boy , Steven .
如果那個安靜的男人需要修理，你要耐心點。
26:40.523–26:43.953
If that quiet man needs fixing, you be patient.
如果那個安靜的男人需要修理，你要有耐心。
26:43.963–26:44.560
Punctual .
我喜歡守時。
26:44.570–26:46.061
I like punctual .
我喜歡守時
26:46.071–26:48.001
Door is unlocked .
咖啡是溫的。
26:48.011–26:48.929
Coffee is warm .
一個男人。
26:48.939–26:49.704
A man .
這份工作是什麼？
26:49.714–26:50.710
What is the job ?
來了兩年。
26:50.720–26:52.723
Came for 2 years.
我的生活。
26:52.733–26:54.813
My life .
你看那個頭部前後晃動的動作，其實現在有三個人正在說話。
26:54.823–26:58.612
See the the head tilt movement back and forth is actually there's
而且他忘記了整個名字。Charles 給他買了一間私人頂樓公寓，
26:58.622–26:59.671
three people talking now .
安保人員想要他的傳奇。有些人想問
26:59.681–27:05.838
And he forgets the entire name. Charles buys him a penthouse private
然後他連整個名字都忘了。查爾斯買給他一間私人頂層公寓
27:05.848–27:08.220
security, security wants his legend. Some people would like to ask
安全部門，安全部門想要他的傳奇。有些人想問他問題。他慢慢地。客戶先付100。我也希望這也適用於這個案例。所有費用，頭等艙，兩張機票。這就是我們合作的原因。一起，頂層托盤是一份普通檔案
27:08.230–27:12.891
him questions. Him slowly. The client pays 100 upfront. I would like
向他提問題。慢慢地。客戶先付100。我希望
27:12.901–27:16.850
it's also up in the case. All expenses, first class, two tickets. This
這在案件中也同樣適用。所有費用，頭等艙，兩張機票。這
27:16.860–27:22.047
is why we work together. Together, the top tray is a normal file
就是我們合作的原因。一起，頂層托盤是一個普通檔案
27:22.057–27:30.980
on a laptop. It's quite my second favorite thing.
在筆記型電腦上。這是我第二喜歡的東西。
27:30.990–27:31.336
Professional .
專業人士
27:31.346–27:33.116
I will not ask what the first is .
我不會問第一個是什麼。
27:33.126–27:36.888
The phone rings when the helicopter is ready .
當直升機準備好時電話響了。
27:36.898–27:37.873
40 .
四十
27:37.883–27:39.334
Cut scene over .
場景切換結束。
27:39.344–27:40.279
Thank goodness .
謝天謝地。
27:40.289–27:41.739
Click to take control .
點擊以控制。
27:41.749–27:43.373
All right .
好吧。
27:43.383–27:46.981
Aside from the sounds being just totally messed up ,
除了聲音完全亂七八糟之外，
27:46.991–27:52.052
this is actually pretty decent .
這其實相當不錯。
27:52.062–27:52.654
Okay .
好的。
27:52.664–28:00.561
The cat is... There's definitely some geometric issues.
貓是……確實有一些幾何問題。
28:00.571–28:02.366
What are we supposed to do ?
我們應該做什麼？
28:02.376–28:03.997
Take the briefcase when you're ready .
當你準備好時，拿走手提箱。
28:04.007–28:04.256
Okay .
好的。
28:04.266–28:06.833
Where is it ?
它在哪裡？
28:06.843–28:11.414
Well, press buttons are like, "Don't look at me like that."
嗯，按按鈕就像在說「別那樣看我」。
28:11.424–28:12.146
Buttons .
按鈕。
28:12.156–28:12.700
Easter egg .
彩蛋。
28:12.710–28:14.079
A man's got to eat .
人總得吃飯。
28:14.089–28:14.462
Okay .
好的。
28:14.472–28:17.600
Where's the briefcase ?
手提箱在哪裡？
28:17.610–28:19.132
All right .
好吧。
28:19.142–28:20.127
We have inventory .
我們有庫存。
28:20.137–28:20.970
Zip ties .
束線帶。
28:20.980–28:22.202
Okay .
好的。
28:22.212–28:23.324
E to continue .
按 E 繼續。
28:23.334–28:23.785
All right .
好吧。
28:23.795–28:25.408
Is it going to put us in Miami?
它會把我們帶到邁阿密嗎？
28:25.418–28:25.608
Nice .
不錯。
28:25.618–28:30.071
Palm trees , humidity , and one very bad man ,
棕櫚樹、濕氣，還有一個非常壞的男人，
28:30.081–28:30.308
Mr .
科爾頓先生。
28:30.318–28:31.648
Colton .
科爾頓。
28:31.658–28:32.357
Champagne perhaps .
也許來點香檳。
28:32.367–28:36.351
I'm just going to click through cuz it's like like black like my
我直接點過去，因為它就像我的
28:36.361–28:40.530
well, coffee asset at the resort.
度假村裡的咖啡資產一樣黑。
28:40.540–28:41.941
Poolside at noon .
中午在泳池邊。
28:41.951–28:43.168
Oh , poolside .
哦，泳池邊。
28:43.178–28:45.040
17th floor enters.
十七樓進入。
28:45.050–28:46.830
Camera coverage .
攝影機覆蓋範圍。
28:46.840–28:47.302
Elevator .
電梯。
28:47.312–28:48.450
There is no dog .
沒有狗。
28:48.460–28:52.831
Behind is the west service stairwell.
後面是西側服務樓梯間。
28:52.841–28:55.987
Well , I think we just saw a glimpse of the pool .
嗯，我想我們剛才瞥見了泳池。
28:55.997–28:57.456
All right .
好吧。
28:57.466–28:59.495
Oh , great .
哦，太棒了。
28:59.505–29:00.799
Azure Palm Resort .
Azure Palm 度假村
29:00.809–29:03.165
Click to take control .
點擊以取得控制權。
29:03.175–29:07.406
It's even got like the like low rent wallpaper that you'd expect
它甚至還有那種你預期會在這種地方看到的低檔壁紙。
29:07.416–29:09.201
at one of these joints .
在這種地方。
29:09.211–29:10.260
Oh , wow .
哦，哇。
29:10.270–29:11.891
This is actually like kind of intricate .
這其實有點複雜
29:11.901–29:12.381
Oh , good .
哦，很好。
29:12.391–29:12.626
Okay .
好的。
29:12.636–29:15.195
I was worried we wouldn't be able to just walk through the walls
我擔心我們無法直接穿牆而過，
29:15.205–29:16.541
, but okay .
但沒關係。
29:16.551–29:18.416
We can toss a coin in the fountain .
我們可以在噴泉裡擲硬幣。
29:18.426–29:18.988
Your wish .
你的願望。
29:18.998–29:20.185
No more syntax errors.
不再有語法錯誤。
29:20.195–29:21.617
50 50 .
五五開。
29:21.627–29:25.773
It seems like Steve's shop and the airplane scene is like oddly
史蒂夫的商店和飛機場景似乎被怪異地
29:25.783–29:29.299
placed inside of the location we're supposed to be at .
放置在了我們應該所在的地點內部。
29:29.309–29:31.622
Let's just ignore that .
我們就假裝沒看到吧。
29:31.632–29:37.710
So , check in at the front desk and then reach the 17th floor .
所以，先去櫃檯報到，然後前往17樓。
29:37.720–29:39.711
Welcome to the Azure Palm .
歡迎來到藍棕櫚酒店。
29:39.721–29:41.710
Checking in , Mr .
報到，庫爾特先生。
29:41.720–29:42.750
Me.
我。
29:42.760–29:43.603
Coulter.
庫爾特。
29:43.613–29:44.755
Steven Coulter .
史蒂文·庫爾特。
29:44.765–29:44.925
Mr .
先生。
29:44.935–29:47.198
I believe I have a reservation .
我有訂房紀錄
29:47.208–29:51.154
Room 1408 , partial ocean view .
1408號房，部分海景。
29:51.164–29:52.416
I do apologize .
我很抱歉。
29:52.426–29:55.363
The ice machine on your floor is out of order .
您這層的製冰機故障了。
29:55.373–30:00.141
Story of my life .
這真是我人生的寫照。
30:00.151–30:00.560
All right .
好吧。
30:00.570–30:02.198
We now have a room key in our inventory .
我們現在庫存裡有一張房門鑰匙。
30:02.208–30:03.882
So , is this the elevator ?
所以，這是電梯嗎？
30:03.892–30:12.484
Oh, is that right?
哦，是這樣嗎？
30:12.494–30:12.817
Okay .
好的。
30:12.827–30:15.467
West service stairwell .
西側服務樓梯間。
30:15.477–30:21.017
Let's just assume this is north .
我們就假設這方向是北。
30:21.027–30:45.339
Good. Like, what?
很好。像什麼？
30:45.349–30:51.897
Well , that is frustrating because it gave us this really like
嗯，這很令人沮喪，因為它給了我們一種非常像
30:51.907–30:53.164
anticipatory music .
預期的音樂。
30:53.174–30:55.474
I'm sure there's a far better terminology for what I was about
我確定對於我接下來要說的話，有更專業的術語。
30:55.484–30:56.071
to say .
要說的話。
30:56.081–30:57.188
My name is Steve .
我叫史蒂夫。
30:57.198–30:57.710
Oh , no .
哦，不。
30:57.720–30:59.830
Don't .
別這樣。
30:59.840–31:03.256
Most of my clients are old enough to remember when a hard drive
我的大多數客戶都夠老了，記得硬碟
31:03.266–31:04.427
was a luxury.
曾經是奢侈品。
31:04.437–31:04.829
All right .
好的。
31:04.839–31:06.970
We we got most of the way .
我們已經完成了大部分。
31:06.980–31:10.849
I think this is a test that more often than not ,
我認為這是一個測試，而且往往
31:10.859–31:12.464
we don't complete it .
我們無法完成它。
31:12.474–31:15.698
I will say the elements of this, judged individually,
我會說，這些元素若單獨來看，
31:15.708–31:17.234
Definitely the pieces were all there .
絕對來說，所有零件都在那裡。
31:17.244–31:19.900
I don't know that the puzzle made the correct picture that was
我不確定這個拼圖拼出的正確畫面，
31:19.910–31:23.052
shown on like the puzzle box , but it was pretty darn close and
是否與拼圖盒上展示的相符，但確實非常接近，而且
31:23.062–31:26.204
I saw some—truthfully, this did a better job than I expected,
我看到了一些——誠實說，這比我預期做得更好，
31:26.214–31:29.517
especially considering like the subway scene game and things like
特別是考慮到像地鐵場景遊戲這類東西。
31:29.527–31:29.841
that .
那樣。
31:29.851–31:34.206
So , I'm actually I'm okay with this .
所以，我其實對這個沒意見。
31:34.216–31:37.923
So , the final thing was of course the street yeet game ,
所以，最後一項當然是街頭投擲遊戲，
31:37.933–31:41.640
which is weird because this piece of text moving up and down like
這很奇怪，因為這段文字上下移動，
31:41.650–31:44.470
that while this one remains completely static makes me uncomfortable
而這個完全靜止不動，這種對比讓我不舒服
31:44.480–31:46.005
in some level .
在某種程度上感到不適。
31:46.015–31:48.457
So, this is a Bijan issue.
所以，這是個 Bijan 的問題。
31:48.467–31:49.288
All right , good .
好的，不錯。
31:49.298–31:53.072
So, I neglected to, um, just run it in a server.
所以，我疏忽了，呃，只是把它在伺服器上運行。
31:53.082–31:57.152
One fist, zero mercy.
一拳，零仁慈。
31:57.162–31:58.354
I already like this .
我已經喜歡這個了。
31:58.364–32:05.477
Oh , it seems like Good music .
哦，這聽起來像好音樂。
32:05.487–32:05.877
All right .
好的。
32:05.887–32:06.276
All right .
好的。
32:06.286–32:10.359
All right .
好了。
32:10.369–32:16.681
Let's just explore the city real quick .
我們就快速探索一下這座城市吧。
32:16.691–32:18.603
I don't even know .
我甚至不知道。
32:18.613–32:30.602
Okay, so I saw a car get—like, this one's actually kind of fun
好，我看到一輛車撞上了——像這輛其實有點好玩
32:30.602–32:34.698
.
。
32:34.698–32:35.970
Excuse me .
不好意思。
32:35.980–32:37.250
Excuse me .
不好意思。
32:37.260–32:38.131
Pardon me .
抱歉。
32:38.141–32:40.965
It does like a... like a... Okay.
它會做一種……像是一種……好了。
32:40.975–32:42.389
I don't know if we hit...
我不知道我們有沒有撞到……
32:42.399–32:47.562
Is that a Can we hit the bike ?
我們可以撞到那輛自行車嗎？
32:47.572–32:54.399
Bonk .
砰。
32:54.409–32:55.114
Uh oh .
哎呀。
32:55.124–33:00.815
Jay walking pedestrian .
亂穿馬路的行人。
33:00.825–33:02.583
Can we punch this parking meter ?
我們可以揍這台停車計費器嗎？
33:02.593–33:05.154
Yeah, we can just hit everything except the buildings.
是啊，我們可以撞所有東西，除了建築物。
33:05.164–33:05.556
All right .
好了。
33:05.566–33:08.847
There's unfortunately the orientation of some of these things is
不幸的是，這些東西中的一些方向
33:08.857–33:09.722
not quite right .
方向不太對
33:09.732–33:15.609
We can see that we're moving sideways when we move forward .
我們可以看到，當我們向前移動時，我們是在側向移動。
33:15.619–33:22.161
And the cars are also Sorry , pal .
而且這些車子也……抱歉，老兄
33:22.171–33:25.963
Oh , and it's interesting how they uh interact with the scenery
對了，他們與景物的互動方式很有趣
33:25.973–33:26.453
as well .
還有
33:26.463–33:27.763
Can we eat this statue ?
這座雕像可以吃嗎
33:27.773–33:30.709
I don't think we can , but Oh , yeah , we can .
我本來以為不行，但喔，對，我們可以吃。
33:30.719–33:33.981
All right .
好吧
33:33.991–33:37.853
I will say the cartoonishness of this game is actually kind of well
我會說這款遊戲的卡通風格其實做得還不錯
33:37.863–33:38.500
done .
完成得不錯
33:38.510–33:41.571
I don't know how easy that is to assess just watching someone play
我不知道光看別人玩能不能輕易評估這一點
33:41.581–33:43.917
it , but like the pacing is decent .
不過節奏還算不錯
33:43.927–33:47.071
I think maybe that's attributed to the music as well ,
我認為這可能也與音樂有關
33:47.081–33:49.010
but the way that—when you actually... All right,
但當你真的……好吧，那種方式
33:49.020–33:52.412
the loop is getting a little repetitive, but when you actually yeet
循環有點重複，但當你實際把東西丟出去時
33:52.422–33:55.813
the things and they also go and interact with other items in the
把東西丟出去，它們還會和其他環境中的物品產生互動
33:55.823–33:56.445
environment .
環境
33:56.455–34:00.322
So like how that will bounce off the building and then maybe hit
所以就像它會彈到建築物上然後可能撞到
34:00.332–34:01.745
other things in the path .
路徑上的其他物體
34:01.755–34:02.537
That's pretty cool .
這真的很酷
34:02.547–34:09.255
And this was actually quite arcade style and playable .
而且這其實很有街機風格，也可以玩
34:09.265–34:10.366
Our score is going up .
我們的分數正在上升
34:10.376–34:13.177
We have combo multipliers overall .
我們有整體的連擊倍率。
34:13.187–34:14.849
38 here is a success .
這裡的 38 分算是成功了。
34:14.859–34:17.596
All right , so that is probably going to conclude our first look
好的，這大概會結束我們對 DeepSeek V4 Pro 首次亮相
34:17.606–34:21.150
and test of the general availability DeepSeek V4 Pro ,
以及對一般可用性 DeepSeek V4 Pro 的測試，
34:21.160–34:23.388
the 0813 model designation .
也就是 0813 型號的介紹。
34:23.398–34:26.427
Now , this is definitely an improvement over the preview version
現在這版本絕對比預覽版來得好
34:26.437–34:29.305
that felt pretty undercooked and I was not at all impressed with
那個預覽版感覺相當不成熟，我完全沒有印象
34:29.315–34:29.865
it .
它。
34:29.875–34:32.663
But if we look back at some of the tests that were run in that
但如果我們回顧之前影片中執行的一些測試，
34:32.673–34:35.062
video , which is on the channel , we ran some of the same ones
該影片在頻道上，我們今天也執行了一些相同的測試。
34:35.072–34:35.622
today .
今天。
34:35.632–34:37.780
This is a significant improvement .
這是顯著的進步
34:37.790–34:41.059
I would say though , I think the flash model impressed me more
不過我會說，我認為 flash 模型讓我印象更深刻，
34:41.069–34:43.618
than this , especially considering the size of it .
比這個更甚，特別是考慮到它的體積。
34:43.628–34:47.136
This has a bit more competition , but again , it is a big leap
這個模型面臨的競爭稍多，但再次強調，這是對其前代產品的大幅躍進，
34:47.146–34:49.854
on its predecessor , which I guess is what we like to see .
我想這也是我們樂見其成的。
34:49.864–34:53.292
I was impressed a few times with some things like basically results
我幾次對某些事情印象深刻，比如結果中的結果。
34:53.302–34:54.092
within results .
結果中的結果。
34:54.102–34:57.931
The cinematic Steve the PC repair game , it actually had some elements
電影版 Steve 的電腦維修遊戲，它實際上包含了一些元素，
34:57.941–35:00.181
to it that were like , hey , you know , it kind of thought this
那些元素讓人覺得，嘿，你知道，它似乎有點思考了
35:00.191–35:00.744
through .
透過。
35:00.754–35:03.395
Just the way when they were speaking , their heads were like moving
只是當他們說話時，他們的頭部像是
35:03.405–35:04.198
back and forth .
前後晃動。
35:04.208–35:05.805
Not the weird chin thing that was happening .
不是那種奇怪的下巴動作。
35:05.815–35:07.572
I'm not sure what happened there .
我不確定那裡發生了什麼事。
35:07.582–35:10.706
The music it included in both that game and then the Street Yeet
那款遊戲以及之後的 Street Yeet
35:10.716–35:14.562
game were actually like they definitely fit the scene that it was
遊戲中收錄的音樂，實際上非常符合
35:14.572–35:17.522
drawing here and it did a good job making the Street Yeet game very
這裡所呈現的場景，而且它很好地讓 Street Yeet 遊戲顯得
35:17.532–35:20.384
cartoonish and the playability of it felt good ,
卡通化，其可玩性感覺不錯，
35:20.394–35:22.212
I guess , for lack of a better term .
我想，如果沒有更好的詞彙來形容的話。
35:22.222–35:26.579
The model that we got of the engine was actually it wasn't printable
我們獲得的引擎模型實際上無法列印。
35:26.579–35:32.898
.
。
35:32.898–35:32.978
So , unfortunately , maybe it failed to like check that box in
所以，不幸的是，它可能沒能完成這項任務的檢查
35:32.978–35:33.058
the specific task .
在特定任務中勾選該選項。
35:33.058–35:33.934
However , if we just look at one of the previews right here ,
然而，如果我們只看這裡的其中一個預覽，
35:33.944–35:35.616
it did a pretty decent job .
它做得相當不錯。
35:35.626–35:38.739
Definitely better than the other models I've tested this with .
絕對比我用過的其他模型測試結果更好。
35:38.749–35:41.941
Though , I have not tested this specific prompt with any big beefy
雖然，我還沒有用任何大型強大的
35:41.951–35:43.143
Frontier models yet .
Frontier 模型測試過這個特定的提示詞。
35:43.153–35:44.744
So , keep that in mind as well .
所以，也請記住這一點。
35:44.754–35:46.186
But, it definitely has some chops here.
不過，它確實具備一些實力。
35:46.196–35:49.789
And the amount of research it did to actually get to that point
而且它為了達到那個目標所進行的研究量
35:49.799–35:50.830
was quite a lot .
相當龐大。
35:50.840–35:53.709
We have all the things in here downloading these images when it
這裡包含了所有下載這些影像的步驟，即使它
35:53.719–35:54.510
doesn't have vision .
沒有視覺功能。
35:54.520–35:56.350
So , I'm not quite sure what to make of that .
所以，我還不太確定該如何看待這一點。
35:56.360–35:59.870
But nonetheless , I think probably the biggest point of frustration
不過話雖如此，我認為讓我感到最挫折的
35:59.880–36:04.750
for me is it just goes all out when trying to actually ensure that
重點是，它會全力以赴，試圖確保它編寫的
36:04.760–36:08.190
every aspect of every line of code it made is correct .
每一行程式碼的每個面向都是正確的。
36:08.200–36:11.790
And with a model like this where it does not have vision ,
對於像這樣沒有視覺功能的模型，
36:11.800–36:15.349
sometimes it takes longer and there is a cost thing to it as well
有時會花費更長的時間，而且也會產生成本。
36:15.349–36:22.665
.
。
36:22.665–36:22.745
Assuming you're going to be paying to use these models ,
假設你將為使用這些模型付費，
36:22.745–36:22.825
it just sometimes is simpler to have it make the result ,
有時讓它直接產生結果會更簡單，
36:22.825–36:24.510
not try to do like overviews and checks and follow ups .
而不是嘗試進行概覽、檢查和後續追蹤。
36:24.520–36:26.830
you look at it yourself and then tell it what's wrong and tell
你自己查看結果，然後告訴它哪裡有錯，並指示
36:26.840–36:27.790
it to fix it .
它進行修正。
36:27.800–36:31.469
Having to just go all out and it trying to fix it adds a lot of
必須全力以赴並讓它嘗試修正，這會增加大量的
36:31.479–36:35.790
time and expense and tokens to the task , which maybe is more of
時間、費用和 token 成本，這可能更多是
36:35.800–36:38.876
but I find that's something I've consistently had frustrations
但我發現這是我一直感到挫折的地方
36:38.886–36:41.665
but I find that's something I've consistently had frustrations
但我也發現，這是我一直感到挫折的地方
36:41.675–36:45.100
with as well as with the flash model where if I'm having them make
以及在使用閃電模型時，如果我要它們製作
36:45.110–36:47.894
like a say street yeet, it's easier if they just spit the script
比如說街景，讓它們直接輸出腳本會比較容易
36:47.904–36:50.371
out , I test it and say , Hey , the buildings are disappeared .
我測試後說：「嘿，建築物消失了。」
36:50.381–36:53.565
and then it fixes it rather than it trying to go through and like
然後它進行修復，而不是嘗試去
36:53.575–36:56.280
well let me try to spin up a headless browser and then I'll take
好吧，讓我嘗試啟動無頭瀏覽器，然後我會截圖
36:56.290–36:59.555
a screenshot and like do a heat map because it can't actually see
並像做熱力圖那樣，因為它實際上無法看見
36:59.565–37:00.911
it or some other function .
它或其他功能。
37:00.921–37:04.036
So it's a little token-happy.
所以它有點過度消耗 token。
37:04.046–37:06.925
I did run it on the highest potential reasoning mode .
我確實是在最高潛力的推理模式下運行的。
37:06.935–37:10.859
So that could be attributed to it as well , but I still would like
這也可能歸因於此，但我仍然希望
37:10.869–37:14.471
to be able to do that without it like going all out to try to make
希望能做到這點，而不需要它全力嘗試去確保腳本的每個面向都經過打磨，因為我只求最高效能，不需要那麼深的程度
37:14.481–37:17.602
sure it's polishing every aspect of the script because I want max
它打磨腳本的每個方面的情況下做到這一點，因為我想要最大
37:17.612–37:18.164
performance .
效能。
37:18.174–37:19.769
I just don't need that level of depth .
我只是不需要那麼深的程度。
37:19.779–37:22.979
So , that may be a personal complaint , but you know .
所以，這可能是個人的抱怨，但你知道。
37:22.989–37:25.949
So , for the final thing , let's take a look at everything we spent
所以，最後一點，讓我們看看這次測試中我們花費的所有
37:25.959–37:28.359
for the entirety of this test .
涵蓋整個測試期間
37:28.369–37:30.962
We can see right here that all cost 1 .
我們可以看到這裡的總成本為 1。
37:30.972–37:31.810
91 .
[未翻譯]
37:31.820–37:33.416
So , that was pretty darn cheap .
所以，這真的非常便宜。
37:33.426–37:36.466
And you do have to remember this is going to be increased, their
而且你必須記住，這將會增加，他們的
37:36.476–37:37.749
API prices .
API 價格將會上漲。
37:37.759–37:41.509
They say significant increase expected , but I mean a significant
他們表示預期會有顯著增加，但我的意思是，顯著增加可能就像，好吧，整個測試總共才 8
37:41.519–37:45.071
increase might be like , okay , the entirety of this test was 8
元或類似金額，這在整體來看還不算太糟。
37:45.081–37:48.000
or something , which is still in perspective not horrible .
或者類似的情況，這在整體來看還不算太糟。
37:48.010–37:48.881
The Qwen 2
8B 最大測試影片是我們幾週前，或大約一兩週前做的，那次的費用大約是 32。
37:48.891–37:52.645
8B max test video that we did a few weeks ago or like a week or
我們幾週前或大約一週前做的最大 8B 測試影片
37:52.655–37:54.967
two ago that was about 32 .
這可能比這個效能更好，但也會有價格方面的考量。
37:54.977–37:56.490
So keep that in mind as well .
而且這真的非常非常便宜。
37:56.500–37:59.373
That may be a bit more performant than this , but there is going
它不會一直這麼便宜，但即使價格漲到五倍，也還不算太糟。
37:59.383–38:02.257
to be a pricing consideration as well .
我想這將結束我們對
38:02.267–38:04.499
And this was very very very cheap .
0813 Deepseek V4 Pro 模型一般可用性版本的首次檢視和測試。
38:04.509–38:08.023
It won't remain that cheap , but even if it gets five times as
它不會一直這麼便宜，但就算漲到五倍，也不算太糟
38:08.033–38:09.625
expensive , it's still not horrible .
這真的很令人興奮。
38:09.635–38:12.429
I suppose that's going to conclude our first look and test of the
基本上，現在我想我會去測試 Grok 3
38:12.439–38:16.834
0813 Deepseek V4 Pro model , the general availability version .
0813 Deepseek V4 Pro 模型，一般可用性版本
38:16.844–38:18.436
And this is now out of preview .
現在已經正式推出，不再處於預覽階段。
38:18.446–38:20.438
And that's just pretty exciting .
這真的令人相當興奮
38:20.448–38:23.411
Basically, now I think I'm going to test Grok 3
基本上，我想我現在要來測試 Grok 3
38:23.421–38:25.280
because that also came out today.
因為它也是今天發布的。
38:25.290–38:27.767
And then I believe Friday Qwen 2
然後我相信週五 Qwen 2
38:27.777–38:28.820
72B is coming out.
72B 也會發布。
38:28.830–38:30.260
So that'll be exciting as well .
所以這也會令人興奮。
38:30.270–38:33.220
Seems like there's been this has been like a pretty busy week in
看來這週在模型發布方面相當忙碌
38:33.230–38:34.341
terms of model releases .
模型發布方面相當忙碌。
38:34.351–38:35.700
So that's cool also .
所以這也很酷。
38:35.710–38:37.264
And when MuseTalk 1
當 MuseTalk 1
38:37.274–38:38.974
2 comes out , that's going to be awesome .
2 發布時，那將非常棒。
38:38.984–38:41.449
I don't know how big it is , but I'm very excited to try to run
我不知道它有多大，但我非常興奮想嘗試在
38:41.459–38:42.089
it locally .
本地運行它。
38:42.099–38:44.884
So with that , if you have any questions , please feel free to
因此，如果你有任何問題，請隨時
38:44.894–38:46.002
leave them in the comments .
在留言區留言。
38:46.012–38:48.000
And thanks for watching .
感謝觀看
影片筆記：DeepSeek V4 Pro Is HERE – Is THIS the BEST Open Model Yet?
一句話總結
影片測試 DeepSeek 最新發布的 V4 Pro (0813 checkpoint) 通用版，透過 Cursor、ChatGPT Desktop Beta 及 Codex 進行多項複雜任務（2D/3D 遊戲生成、3D 模型設計、前端網頁開發），評估其技術規格、定價競爭力、生成質量與效率，並與預覽版及其他模型進行對比。
核心重點
- 模型發布與規格：DeepSeek V4 Pro 通用版（0813 checkpoint）已發布，擁有 100 萬上下文窗口、最大輸出長度 384K、1.6 兆總參數（混合專家架構，490 億活躍參數）。
- 定價優勢：目前 API 定價極低（輸入 $0.4335/百萬 token，輸出 $0.87/百萬 token），預計未來漲價後仍具競爭力。
- 測試環境：使用 Cursor 編輯器、ChatGPT Desktop App (Beta) 及 OpenAI API Key 驅動 Codex (Beta)，路由至 DeepSeek V4 Pro 0813，並開啟最大推理努力（Maximum Effort/Thinking Mode）。
- 生成能力評估：
- 2D 遊戲：能生成具有合成波風格的模擬操作系統界面及街機風格遊戲（如 Street Yeet），但存在音頻混亂、幾何錯誤及操作邏輯問題。
- 3D 模型：能生成詳細的 RB26 引擎模型，但結構複雜度導致 3D 列印可行性低；手錶前端設計評價極高，接近優秀。
- 進步顯著：相比 V4 Pro 預覽版，通用版在能力上有顯著提升，雖不及 Flash 模型令人印象深刻，但作為預覽版的升級是顯著的。
- 效率與痛點：模型在無視覺反饋下，過度嘗試自我檢查（如啟動無頭瀏覽器截圖、做熱圖），導致耗時長且消耗大量 Token。測試者建議簡化流程，直接輸出腳本後手動修正。
- 成本效益：整個測試過程總成本約 $1.91，即使未來漲價五倍，總成本仍不算糟糕。
詳細大綱
1. DeepSeek V4 Pro 0813 發布概況
- 發布資訊：DeepSeek V4 Pro 通用版（0813 checkpoint）已發布，此前 V4 Flash 通用版於 7/31 發布。無官方 X 平台公告，僅在網站和 OpenRouter 上可見。
- 技術規格：
- 上下文窗口：100 萬。
- 最大輸出長度：384K。
- 參數量：1.6 兆總參數，490 億活躍參數（混合專家模型）。
- 可選思考模式（Thinking Modes），測試時選擇最大努力。
- 性能預期：與 V4 Flash 預覽版及通用版相比有性能躍升，引用疑似來自 DeepSeek 相關群組的基準測試結果，提及超越 Opus 3.8 的潛力。
2. 定價策略
- 當前定價：輸入每百萬 token $0.4335，輸出每百萬 token $0.87。
- 未來預測：預計未來定價將大幅上漲，但目前仍具競爭力。
3. 測試工具與環境設定
- 工具：Cursor 編輯器、ChatGPT Desktop App (Beta 版本)、OpenAI Beta Application。
- 路由設定：直接指向 DeepSeek V4 Pro 0813，開啟最大推理努力。
4. 實際測試過程與反饋
A. 初步測試：Gen 3 美學風格模擬操作系統
- 界面風格：合成波（Synth Wave）主題，透明背景，Gen 3 美學。
- 功能測試：
- 文件系統：靜態但可互動（創建/刪除）。
- 記事本：能記住目錄刪除操作。
- 終端（Neofetch）：記憶體顯示為 1.44 MB（講者幽默稱其為「infinity」的一部分）。
- 計算機：運算正確。
- 設定：風格類似 Qwen 2.5 8B，包含遊戲難度選擇。
- 遊戲測試：
- City Heist：聲音效果有趣，操作需按 Enter。
- 全息桌面：觸發後呈現類似 Apple Vision Pro 的卡片疊加效果。
- 滑板遊戲：使用 ChatGPT 桌面版生成 C++ 滑板遊戲，耗時 68 分 24 秒，總費用約 20 美分。畫面評價：樹木像糞便（poo），但可能是最有趣的滑板遊戲結果之一。
B. DeepSeek V4 (Preview Version) 回顧與複雜任務測試
- Subway FPS Test：
- 生成地鐵站場景與生存射擊遊戲。
- 評價：中等（mid）。優點包括地鐵車廂繪製精美、彈殼掉落動畫真實、武器模型平滑；缺點包括缺乏敵人、角色移動動作滑稽（waddle）、按鈕位置錯位。
- 修復版檢查：按鈕位置錯位、敵人缺失問題未解決，判定為失敗。
- RB26 Inline 6 Engine 3D 模型任務：
- 需求：3D 列印（少支援）、容納 N20 DC 馬達、RC 車專案。
- 過程：耗時 1 小時 10 分鐘，模型展現詳細研究過程（排氣管、進氣側等）。
- 結果：渲染效果良好，但整體結構難以無支援列印，且非單一整體列印設計。評語為「溫吞（lukewarm）」。
- Watch Front-end Design Test：
- 需求：電影級平移效果、3D 手錶資產、網站前端。
- 過程：耗時 1 小時 30 分鐘。
- 結果：評價極高，「非常接近優秀」。優點包括高級美學、木質背景、皮革墊、錶盤與材質（藍寶石水晶、不鏽鋼、黃金）渲染精美。缺點包括錶帶突出遮擋視線、皮革板遮擋錶帶、錶針無動作。
C. DeepSeek V4 Pro (0813) 最終評估與遊戲生成
- 工具使用：使用 OpenAI API Key 驅動 Codex（Beta 桌面版）在後台運行任務。
- 遊戲一：Steve the PC repairman (電影風格遊戲)
- 耗時約 65 分鐘。
- 內容：角色自我介紹為 Steve，電腦維修員，場景涉及 Azure Palm Resort。
- 反饋：音頻混亂（語音重疊）、幾何問題、場景放置奇怪，但整體相當不錯，比預期的好。
- 遊戲二：Street Yeet (街機風格遊戲)
- 機制：玩家控制角色在城市中移動，撞擊自行車、行人等，具有連擊倍數和分數系統。
- 反饋：卡通風格做得相當好，節奏不錯，互動效果令人驚訝地酷。缺點包括文本上下移動而另一部分靜止，部分物體方向不正確。
- 模型評價：
- 相比於感覺「未煮熟」的預覽版，這是顯著改進。
- 雖然進步巨大，但 Flash 模型仍給測試者留下更深刻印象。
- 痛點：模型沒有視覺能力，卻試圖通過啟動無頭瀏覽器截圖、做熱圖來自我檢查，導致過度消耗 Token 和時間。
- 成本分析：
- 整個測試總成本約 $1.91。
- 即使 DeepSeek 預計將顯著提高 API 價格，但目前價格仍具競爭力。
5. 結尾
- 邀請觀眾在評論區留言。
- 感謝觀眾觀看影片。
- 後續計劃：測試 Grok 3、Qwen 2 72B，期待 MuseTalk 1.2 的發布。
工具 / 模型 / 名詞整理
- DeepSeek V4 Pro (0813 checkpoint / 通用版)
- DeepSeek V4 Flash (0731 通用版)
- DeepSeek V4 Pro Preview
- DeepSeek V4 Flash Preview
- OpenRouter
- Cursor (編輯器)
- ChatGPT Desktop App (Beta 版本)
- OpenAI Beta Application
- Codex (工具，Beta 桌面版)
- Open AI API key (工具/憑證)
- X (平台)
- WeChat (群組)
- Opus 3.8 (提及的競爭模型)
- Kimi K3 (提及的模型)
- Qwen 3.2 (提及的模型)
- Qwen 2.5 8B (風格對比)
- Qwen 2.5 (提及的模型)
- Qwen 2 8B max (用於成本對比)
- Qwen 2 72B (提及的模型)
- Grok 3 (提及的模型)
- MuseTalk 1.2 (模型/工具名稱)
- Apple Vision Pro (參考界面風格)
- Neofetch (命令/工具)
- C++ (程式語言)
- City Heist (遊戲名稱)
- Skate NYC (遊戲名稱)
- Van Halen (音樂團體)
- Subway FPS (遊戲名稱)
- Steve the PC repair game (遊戲名稱)
- Street Yeet (遊戲名稱)
- RB26 inline 6 engine (引擎)
- N20 motor (DC 馬達)
- Downpipe (排氣管)
- Intake plenum (進氣歧管)
- Turbocharger (渦輪增壓器)
- Oil pan (油底殼)
- STL (檔案格式)
- Nissan (標誌/文字)
- OpenSCAD (軟體)
- Azure Palm Resort (遊戲內場景名稱)
- Charles (遊戲內客戶角色)
- Mr. Buttons (遊戲內角色/寵物)
- Mr. Thomas (遊戲內角色)
- Mr. Colton / Coulter (遊戲內反派角色)
- Headless browser (技術術語)
- Heat map (技術術語)
- General availability (產品發布狀態術語)
- Flash model (模型名稱，指代 DeepSeek 的 Flash 版本)
操作流程整理
- 環境準備：
- 安裝並開啟 Cursor 編輯器及 ChatGPT Desktop App (Beta)。
- 配置路由至 DeepSeek V4 Pro 0813 模型。
- 開啟最大推理努力（Maximum Effort/Thinking Mode）。
- 準備 OpenAI API Key 以驅動 Codex (Beta 桌面版)。
- 任務執行：
- 2D 遊戲生成：在 ChatGPT Desktop App 中輸入提示詞，生成具有合成波風格的模擬操作系統界面及街機風格遊戲（如 Street Yeet）。
- 3D 模型設計：在 Codex 中輸入詳細提示詞（如 RB26 引擎規格、手錶前端設計需求），生成 3D 模型或前端代碼。
- 電影風格遊戲生成：使用 Codex 生成「Steve the PC repairman」遊戲，耗時約 65 分鐘。
- 評估與反饋：
- 檢查生成內容的視覺效果、音頻、幾何結構及互動性。
- 識別問題（如敵人缺失、幾何錯誤、音頻混亂）。
- 評估模型在無視覺反饋下的自我檢查行為（如啟動無頭瀏覽器截圖）。
- 成本與效率分析：
- 記錄生成過程的耗時與 Token 消耗。
- 計算總成本（約 $1.91）。
- 對比不同模型（Flash vs. Pro）的性能與成本效益。
值得注意的限制或風險
- 無視覺反饋的缺陷：模型沒有視覺能力，卻試圖通過啟動無頭瀏覽器截圖、做熱圖來自我檢查，導致過度消耗 Token 和時間。
- 效率問題：測試者認為更簡單的方法是讓模型直接輸出腳本，測試者手動檢查並指出錯誤，讓模型修復，而非模型試圖全方位自我完善。
- 生成質量不穩定：
- 2D 遊戲存在音頻混亂、幾何錯誤及操作邏輯問題。
- 3D 模型結構複雜度導致 3D 列印可行性低。
- 部分遊戲場景放置奇怪，角色動作滑稽。
- 定價上漲風險：DeepSeek 預計將顯著提高 API 價格，雖然目前價格具競爭力，但未來漲價可能影響成本效益。
逐字稿辨識疑點
- poo：逐字稿中形容樹木「reminiscent of a poo」，疑為口誤或聽寫錯誤，可能意指「pool」（泳池）或其他詞彙，但依規則保留原樣。
- 1.44 megs out of infinity：Neofetch 顯示記憶體為 1.44 MB，講者稱其為「infinity」（無限）的一部分，疑為幽默或口誤，實際記憶體單位應為 MB 或 GB，但依規則保留原樣。
- 1.6 trillion parameter mixture of experts model with 49 billion active：參數數量描述，疑點在於「how many active ?」後的問號，可能為語法不完整或口語停頓，但數值 1.6 兆與 490 億依規則保留。
- 0.43 35：定價描述，疑點在於小數點後的空格或斷句，實際應為 0.4335，但依規則保留原樣。
- 12 . 8 to 62 . 7：基準測試數值，疑點在於小數點後的空格，實際應為 12.8 到 62.7，但依規則保留原樣。
- Gen 3 aesthetic：形容界面風格，疑點在於是否為特定術語或口誤，但依規則保留原樣。
- Synth wave theme：形容界面風格，依規則保留原樣。
- Holographic desktop feature：全息桌面功能，依規則保留原樣。
- City Heist：遊戲名稱，依規則保留原樣。
- Skate NYC：遊戲名稱，依規則保留原樣。
- Van Halen：音樂團體，依規則保留原樣。
- Opus 3 . 8 8：提及的模型版本，疑點在於重複的 8，實際應為 Opus 3.8，但依規則保留原樣。
- Kimi K3：提及的模型，依規則保留原樣。
- Qwen 3 . 2：提及的模型，疑點在於小數點後的空格，實際應為 Qwen 3.2，但依規則保留原樣。
- Qwen 2.5 8B：提及的模型，依規則保留原樣。
- Qwen 2.5：提及的模型，依規則保留原樣。
- Apple Vision Pro：參考界面風格，依規則保留原樣。
- Neofetch：命令/工具，依規則保留原樣。
- C++：程式語言，依規則保留原樣。
- Cursor：編輯器，依規則保留原樣。
- ChatGPT Desktop App：產品名稱，依規則保留原樣。
- OpenAI Beta Application：產品名稱，依規則保留原樣。
- X：平台，依規則保留原樣。
- WeChat：平台，依規則保留原樣。
- OpenRouter：平台，依規則保留原樣。
- DeepSeek V4 Pro：模型名稱，依規則保留原樣。
- DeepSeek V4 Flash：模型名稱，依規則保留原樣。
- DeepSeek V4 Pro Preview：模型名稱，依規則保留原樣。
- DeepSeek V4 Flash Preview：模型名稱，依規則保留原樣。
- Main Rand Avenue：逐字稿中出現此詞，疑為「Main Street」或特定地鐵站名稱之聽寫錯誤，需查證。
- Seven hostiles：在 Subway FPS 測試中，UI 顯示「Seven hostiles」但使用者找不到敵人，此為逐字稿原文記錄。
- Steve the PC repair game：此為
生字列表
| 生字 | 讀音 | 類型 | 中文 |
|---|---|---|---|
| reminiscent | /rɪˈmɪnɪsənt/ | adj. | 讓人聯想到...的 |
| significant | /sɪɡˈnɪfɪkənt/ | adj. | 顯著的 |
| dubbed | /dʌbd/ | v. | 被命名為 |
| genuinely | /ˈdʒenjuɪnli/ | adv. | 真正地 |
| hypothetically | /ˌhaɪpəˈθetɪkli/ | adv. | 假設性地 |
| undercooked | /ˌʌndərˈkʊkt/ | adj. | 未熟的 / 不成熟的 |
| reiterate | /riˈɪtəreɪt/ | v. | 重申 |
| niche | /niːʃ/ | n. | 小眾領域 |
| negligible | /ˈneɡlɪdʒəbl/ | adj. | 微不足道的 |
| infuriating | /ɪnˈfjʊərieɪtɪŋ/ | adj. | 令人氣憤的 |
| intricacy | /ˈɪntrɪkəsi/ | n. | 複雜性 / 精細之處 |
| resonate | /ˈrezəneɪt/ | v. | 產生共鳴 |
| competent | /ˈkɒmpɪtənt/ | adj. | 有能力的 / 勝任的 |
| inhibiting | /ɪnˈhɪbɪtɪŋ/ | v. | 阻礙 / 妨礙 |
| punctual | /ˈpʌŋktʃuəl/ | adj. | 準時的 |
生字解說
reminiscent /rɪˈmɪnɪsənt/
意思：讓人聯想到...的
解說：用於描述某物在視覺、氣味或感覺上讓人想起另一件事物。
影片原句
The trees are quite reminiscent of a pool.
這些樹木讓人聯想到泳池。
The trees are quite reminiscent of a pool.
這些樹木讓人聯想到泳池。
延伸例句
The smell of the ocean is reminiscent of my childhood summers.
大海的味道讓我聯想到童年的夏天。
The smell of the ocean is reminiscent of my childhood summers.
大海的味道讓我聯想到童年的夏天。
significant /sɪɡˈnɪfɪkənt/
意思：顯著的
解說：指數量、程度或影響力大到足以被注意到。
影片原句
Now, that model when tested proved to be a significant increase in capability over the preview version
現在，該模型在測試中被證明具有顯著的能力提升（相較於預覽版）。
Now, that model when tested proved to be a significant increase in capability over the preview version
現在，該模型在測試中被證明具有顯著的能力提升（相較於預覽版）。
延伸例句
There has been a significant increase in the number of electric cars on the road.
路上電動車的數量有了顯著的增加。
There has been a significant increase in the number of electric cars on the road.
路上電動車的數量有了顯著的增加。
dubbed /dʌbd/
意思：被命名為
解說：給予某人或某物一個特定的名稱或綽號。
影片原句
for the newly updated checkpoint general availability for V4 Pro dubbed 0813.
適用於 V4 Pro 新更新的通用正式版檢查點，命名為 0813。
for the newly updated checkpoint general availability for V4 Pro dubbed 0813.
適用於 V4 Pro 新更新的通用正式版檢查點，命名為 0813。
延伸例句
The new stadium was dubbed 「The Cauldron」 by the local fans.
這座新體育場被當地球迷命名為「大鍋」。
The new stadium was dubbed 「The Cauldron」 by the local fans.
這座新體育場被當地球迷命名為「大鍋」。
genuinely /ˈdʒenjuɪnli/
意思：真正地
解說：強調某事是真實的，而非偽造或表面上的。
影片原句
makes me feel comfortable that it is actually genuinely released
讓我感到安心，因為它確實已經正式發布。
makes me feel comfortable that it is actually genuinely released
讓我感到安心，因為它確實已經正式發布。
延伸例句
I was genuinely surprised by the unexpected news.
我對這個意外的消息感到真正驚訝。
I was genuinely surprised by the unexpected news.
我對這個意外的消息感到真正驚訝。
hypothetically /ˌhaɪpəˈθetɪkli/
意思：假設性地
解說：基於假設而非事實地討論某種可能性。
影片原句
we can see that these would hypothetically be the benchmark results
我們可以假設這些是基準測試結果。
we can see that these would hypothetically be the benchmark results
我們可以假設這些是基準測試結果。
延伸例句
Hypothetically, if you won the lottery, what would you do first?
假設你中了樂透，你會先做什麼？
Hypothetically, if you won the lottery, what would you do first?
假設你中了樂透，你會先做什麼？
undercooked /ˌʌndərˈkʊkt/
意思：未熟的 / 不成熟的
解說：原指食物沒煮熟，在此處比喻產品開發不足、尚未完善。
影片原句
They felt very undercooked and probably because they were named preview.
它們感覺非常未熟，可能因為它們被命名為預覽版。
They felt very undercooked and probably because they were named preview.
它們感覺非常未熟，可能因為它們被命名為預覽版。
延伸例句
The first version of the app felt undercooked and crashed frequently.
該應用程式的第一個版本感覺很不成熟，且經常崩潰。
The first version of the app felt undercooked and crashed frequently.
該應用程式的第一個版本感覺很不成熟，且經常崩潰。
reiterate /riˈɪtəreɪt/
意思：重申
解說：為了強調或確保對方理解而再次說明。
影片原句
But again , I want to reiterate the leap in capability
但再次強調，我想重申能力上的巨大躍進。
But again , I want to reiterate the leap in capability
但再次強調，我想重申能力上的巨大躍進。
延伸例句
Let me reiterate that safety is our top priority.
讓我重申，安全是我們的首要任務。
Let me reiterate that safety is our top priority.
讓我重申，安全是我們的首要任務。
niche /niːʃ/
意思：小眾領域
解說：指針對特定小群體或特定需求的市場或領域。
影片原句
So I mean in a niche , you know , but still
所以我的意思是，在一個小眾領域裡，你知道，但即便如此。
So I mean in a niche , you know , but still
所以我的意思是，在一個小眾領域裡，你知道，但即便如此。
延伸例句
He found a niche market for handmade organic soaps.
他為手工有機肥皂找到了一個小眾市場。
He found a niche market for handmade organic soaps.
他為手工有機肥皂找到了一個小眾市場。
negligible /ˈneɡlɪdʒəbl/
意思：微不足道的
解說：數量或影響小到可以忽略不計。
影片原句
Although I would say the jump in price is like a negligible amount
不過我會說，價格的跳動幅度微乎其微。
Although I would say the jump in price is like a negligible amount
不過我會說，價格的跳動幅度微乎其微。
延伸例句
The difference in cost between the two options is negligible.
這兩個選項之間的成本差異微不足道。
The difference in cost between the two options is negligible.
這兩個選項之間的成本差異微不足道。
infuriating /ɪnˈfjʊərieɪtɪŋ/
意思：令人氣憤的
解說：使人感到極度憤怒或挫折的。
影片原句
That's infuriating .
太令人氣憤了。
That's infuriating .
太令人氣憤了。
延伸例句
It is infuriating when the computer freezes right before you save your work.
在儲存工作前電腦突然當機真的令人氣憤。
It is infuriating when the computer freezes right before you save your work.
在儲存工作前電腦突然當機真的令人氣憤。
intricacy /ˈɪntrɪkəsi/
意思：複雜性 / 精細之處
解說：指事物由許多細小且複雜的部分組成。
影片原句
but I applaud it for its intricacy .
，但我讚賞它的複雜性。
but I applaud it for its intricacy .
，但我讚賞它的複雜性。
延伸例句
The intricacy of the watch's internal mechanism is amazing.
這隻手錶內部機制的複雜程度令人驚嘆。
The intricacy of the watch's internal mechanism is amazing.
這隻手錶內部機制的複雜程度令人驚嘆。
resonate /ˈrezəneɪt/
意思：產生共鳴
解說：指某種想法或感覺讓他人感到認同或有相同體驗。
影片原句
you may resonate with what I'm saying .
可能會對我所說的話產生共鳴。
you may resonate with what I'm saying .
可能會對我所說的話產生共鳴。
延伸例句
The themes of the movie resonate with people of all ages.
這部電影的主題讓所有年齡層的人都能產生共鳴。
The themes of the movie resonate with people of all ages.
這部電影的主題讓所有年齡層的人都能產生共鳴。
competent /ˈkɒmpɪtənt/
意思：有能力的 / 勝任的
解說：具備完成某項工作所需的技能或知識。
影片原句
we'll see how competent it is at putting the entirety of this together .
我們來看看它整合這一切的能力如何。
we'll see how competent it is at putting the entirety of this together .
我們來看看它整合這一切的能力如何。
延伸例句
She is a highly competent manager who handles stress well.
她是一位非常有能力的經理，能很好地處理壓力。
She is a highly competent manager who handles stress well.
她是一位非常有能力的經理，能很好地處理壓力。
inhibiting /ɪnˈhɪbɪtɪŋ/
意思：阻礙 / 妨礙
解說：阻止某事發生或限制某人的行動/視線。
影片原句
is inhibiting our view of these beautiful straps .
這塊皮革板擋住了我們欣賞這些精美錶帶的視線。
is inhibiting our view of these beautiful straps .
這塊皮革板擋住了我們欣賞這些精美錶帶的視線。
延伸例句
Fear of failure can be inhibiting to creative growth.
對失敗的恐懼可能會阻礙創造力的成長。
Fear of failure can be inhibiting to creative growth.
對失敗的恐懼可能會阻礙創造力的成長。
punctual /ˈpʌŋktʃuəl/
意思：準時的
解說：在約定的時間準時到達或完成。
影片原句
I like punctual .
我喜歡守時。
I like punctual .
我喜歡守時。
延伸例句
Please be punctual for the meeting tomorrow morning.
請在明天早上的會議中準時出席。
Please be punctual for the meeting tomorrow morning.
請在明天早上的會議中準時出席。
句型解說（含實例）
feel free to [verb]
意思：請隨時... / 儘管...
接續：feel free to + 原形動詞
解說：一種禮貌的邀請，告訴對方不需要猶豫，可以隨意採取某個行動。
影片原句
do feel free to subscribe so we can hit 70K
請隨時訂閱，以便我們能達到 70K。
do feel free to subscribe so we can hit 70K
請隨時訂閱，以便我們能達到 70K。
實例
- Feel free to ask any questions after the presentation.
簡報結束後，請隨時提出任何問題。 - Feel free to contact me if you need further assistance.
如果您需要進一步的協助，請儘管聯繫我。
with a grain of salt
意思：持保留態度 / 不要完全相信
接續：take [something] with a grain of salt
解說：慣用語，指對某個資訊保持懷疑，認為其可能不完全正確或被誇大了。
影片原句
I want to mention the following with a grain of salt
我想謹慎地提及以下事項（請持保留態度看待）。
I want to mention the following with a grain of salt
我想謹慎地提及以下事項（請持保留態度看待）。
實例
- You should take everything he says with a grain of salt.
你對他說的所有話都應該持保留態度。 - I read the rumors online, but I took them with a grain of salt.
我讀了網路上的傳聞，但我並沒有完全相信。
depending on [something]
意思：取決於...
接續：depending on + 名詞/名詞短語
解說：用來表示結果會根據不同的條件而有所改變。
影片原句
depending on what locale one is in, we can see that the model has
取決於使用者的地區設定，我們可以得知該模型擁有...
depending on what locale one is in, we can see that the model has
取決於使用者的地區設定，我們可以得知該模型擁有...
實例
- The price of the hotel varies depending on the season.
飯店的價格取決於季節而有所不同。 - Depending on the weather, we might cancel the picnic.
取決於天氣情況，我們可能會取消野餐。
in the interest of [doing something]
意思：為了... / 出於...的考量
接續：in the interest of + 名詞/動名詞
解說：表示採取某個行動是為了達到某個目的或為了某種利益。
影片原句
but in the interest of seeing the maximum potential performance
但為了展現最大潛力的效能。
but in the interest of seeing the maximum potential performance
但為了展現最大潛力的效能。
實例
- In the interest of time, let's move to the next topic.
為了節省時間，讓我們進入下一個主題。 - The company updated its policy in the interest of employee safety.
公司更新了政策，以確保員工的安全。
more often than not
意思：通常 / 往往
接續：more often than not (副詞短語)
解說：表示某事發生的機率很高，相當於 「usually」 或 「most of the time」。
影片原句
I think this is a test that more often than not , we don't complete it .
我認為這是一個測試，而且往往我們無法完成它。
I think this is a test that more often than not , we don't complete it .
我認為這是一個測試，而且往往我們無法完成它。
實例
- More often than not, he forgets his keys at home.
他通常會把鑰匙忘在家裡。 - The trains are, more often than not, delayed during winter.
冬天時，火車往往會延誤。