Muse Spark 1.1: Meta's First Paid Model (No Longer Open ??)

AI with Surya 首頁 每日列表 原始影片
實際影片長度：11:34.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:05.160
Meta just released its most capable AI model, which is MuseSpark 1.1.
0:05.540–0:13.300
And according to Zugg, this is a model that can act,
0:08.127–0:10.713
it can watch a video, use tools and get it as done f
0:10.713–0:13.300
or you, not just answer questions.
0:13.820–0:15.660
And later in the video, I'm going to talk about all of that.
0:15.820–0:21.760
Mark Zuckerberg, actually, the way he announced it on the X platform,
0:18.790–0:21.760
I think he came back on it after a few years.
0:21.760–0:31.000
And look at what he leads with, a strong agentic and
0:26.380–0:31.000
coding model at a very low price through our new Meta model API.
0:31.780–0:34.000
A low price, a paid API.
0:34.680–0:40.700
Now think about this, from a company that for years made free and
0:37.690–0:40.700
open source AI its entire identity.
0:41.360–0:49.400
So before I get into what Spark can actually do,
0:44.040–0:46.720
I think it'll be interesting to understand what
0:46.720–0:49.400
pushed Meta to change the strategy like this.
0:49.400–0:52.420
So let's go back to history and rewind, right?
0:52.480–0:54.500
Because the backstory explains everything.
0:55.020–0:57.740
For years, Meta was the champion of open source AI.
0:57.980–1:05.520
Back in July 2024, Zuck published an essay titled,
1:01.750–1:05.520
literally, Open Source AI is the Path Forward.
1:06.160–1:14.300
Free downloadable models were Meta's whole identity,
1:08.873–1:11.587
and he positioned the company as the open alternativ
1:11.587–1:14.300
e to closed labs like OpenAI.
1:15.000–1:16.400
Then it changed fast.
1:16.400–1:24.040
Meta brought in Scale AI's Alexander Wang as their chief AI officer,
1:20.220–1:24.040
a $14 billion deal.
1:24.660–1:34.200
They stood up a brand new super intelligence lab,
1:27.840–1:31.020
and back in April, shipped the first Mewk Spark m
1:31.020–1:34.200
odel, which was closed, no waits, no download available.
1:34.940–1:36.940
That was a real break from open source.
1:37.440–1:38.820
For the first time, right?
1:38.900–1:45.260
And Meta's frontier model were locked behind its own doors,
1:42.080–1:45.260
running inside its apps instead of out in the open.
1:45.260–1:51.000
And this week, Meta took the next step, a big upgrade, Mews Spark 1.1.
1:51.320–1:58.260
And for the first time ever, a public paid API with pricing,
1:54.790–1:58.260
Zuckerberg called very aggressive.
1:58.880–2:05.600
So Meta didn't just close its best model,
2:02.240–2:05.600
it put a price tag on it and open it up for anyone to build it.
2:05.960–2:08.180
Meta didn't just undercut everyone on price.
2:08.180–2:11.800
It's about a quarter of what GPT and Claude cost.
2:11.800–2:16.880
They also made Mews Spark drop straight into the tools developers already use.
2:17.380–2:21.440
OpenAI setup, sure, but also Claude code, Anthropic's own coding tool.
2:21.900–2:28.960
You pointed at Meta's model change basically one line,
2:24.253–2:26.607
and now you're suddenly running on Mews Spark at a qua
2:26.607–2:28.960
rter of the price.
2:29.120–2:35.340
Meta built a side door out of OpenAI and Anthropic and
2:32.230–2:35.340
hung a discount sign on it.
2:35.340–2:39.520
This isn't just a new model,
2:37.430–2:39.520
it's a direct play for their rivals' customers.
2:40.160–2:41.740
Now, two things to keep honest.
2:41.900–2:45.020
One, Lama isn't dead, it's still open, still downloadable.
2:45.480–2:47.880
What changes, Meta's best models are closed now.
2:48.020–2:51.440
And two, as wild as the flip feels, it makes sense, right?
2:51.520–2:55.200
Meta's spending north of $100 billion a year on AI.
2:55.660–2:58.000
You don't spend that and give your best model away.
2:58.360–3:04.160
And if you want a personal agent in the hands of 3 billion people,
3:01.260–3:04.160
you want to own the model doing the work.
3:04.160–3:07.740
So what did they actually build, and is it worth your time?
3:08.260–3:13.580
I started using it for the last couple of days, and
3:10.920–3:13.580
I also built a little app on Meta's model.
3:14.180–3:22.400
I handed it a video of anything that you probably want to sell,
3:16.920–3:19.660
and asked it to give me a full price, ready-to-post listing on
3:19.660–3:22.400
Facebook's marketplace.
3:22.900–3:24.320
So that's one demo that we will see.
3:24.480–3:29.000
Second, I also ran it on Claude code's Anthropic,
3:26.740–3:29.000
so that'll be an interesting one.
3:29.000–3:36.760
And third, I wanted to test the real multimodality,
3:31.587–3:34.173
so I gave it a picture of the ingredients in the fr
3:34.173–3:36.760
idge, and it was able to pinpoint the price tag on it, right?
3:37.080–3:47.280
So I'll show you all of the fun parts as we go through,
3:40.480–3:43.880
but then, interesting part, as I mentioned earlier, is
3:43.880–3:47.280
Meta's shift to the paid strategy versus just keeping it open source.
3:47.560–3:47.780
All right?
3:47.980–3:53.560
Before we dive deep into it, a quick disclaimer,
3:50.770–3:53.560
all opinions are my own and do not belong to my employer.
3:53.820–3:54.080
All right?
3:54.080–3:55.820
With that, let's get into it.
3:56.720–4:04.860
All right, so one of the fastest ways to experience Spark is to actually come to Meta.
4:00.790–4:04.860
ai and change the mode from Instant to Thinking mode.
4:05.200–4:09.840
So if you ask any question, like, for example,
4:07.520–4:09.840
here I'm asking which particular Meta model are you using?
4:10.200–4:14.200
You would see that it is going to confirm that it is using Spark 1.1.
4:14.640–4:20.380
So you just need to change the mode to Thinking,
4:17.510–4:20.380
and you will get to leverage the Spark 1.1 model, right?
4:20.380–4:23.380
So it was launched literally a couple of days back, okay?
4:24.080–4:28.820
But then, from a developer perspective,
4:26.450–4:28.820
if you want to use it, you need to go into dev.meta.ai.
4:29.220–4:40.060
So here I am in dev.meta.ai, and I was able to create an API key,
4:32.833–4:36.447
and you can already see detailed docs and how you're able to use
4:36.447–4:40.060
this in plot code.
4:40.060–4:43.040
And these are all the details that they have provided, codecs.
4:43.320–4:46.180
Then if you want to just call it via Python as well as curl, right?
4:46.220–4:48.460
These are some of the details that have been given.
4:48.920–4:51.460
So what you could do is you can create an API key.
4:51.580–4:55.060
So what I did was, as I mentioned earlier,
4:53.320–4:55.060
I'm going to be doing three different demos.
4:55.060–5:00.500
So the first demo here is,
4:57.780–5:00.500
think of it like a Facebook Marketplace listing generator.
5:00.820–5:02.720
So here I would be using MuseSpark.
5:03.280–5:05.740
So what I'm going to do here is upload a video here.
5:05.860–5:07.200
So you can see the video very clearly.
5:07.780–5:13.400
This is a video of a bike of a kid where I've just taken like a very short nine seconds.
5:13.520–5:18.440
I just wanted to make it a little bit tough, and
5:15.980–5:18.440
I'm going to ask it to generate listing, right?
5:18.480–5:20.840
So there are multiple things which are happening here, right?
5:20.840–5:24.040
So obviously it is calling the MuseSpark 1.1 model.
5:24.500–5:28.200
So the first thing what's happening here is it is going to be identifying the item.
5:28.560–5:31.540
So this is the response back from the first step.
5:31.940–5:36.220
So it has correctly identified that this is a Global Primo Ping three-wheel kit scooter.
5:36.500–5:39.080
And obviously it is used, so it has understood that as well.
5:39.420–5:40.580
So very good job done.
5:40.700–5:44.000
Very quickly, you can see the amount of time it took was very short.
5:44.520–5:47.900
Then the agent is going to now do a pricing research.
5:47.900–5:53.280
So it's going to look into some other prices so
5:50.590–5:53.280
that it can give us some comparison pricing, right?
5:53.660–5:59.960
Once you have that idea of the comparison pricing,
5:56.810–5:59.960
then it will consolidate all of that and create the listing.
6:00.080–6:00.860
So you can see that.
6:01.240–6:04.640
I'll search current resale listing for your global this.
6:04.760–6:06.660
In the US market, it is around this.
6:06.760–6:08.380
And you can see all the pricing over here.
6:08.540–6:10.580
In UK, it is something of this sort, right?
6:10.600–6:11.820
In Canada, it is this, right?
6:11.820–6:16.940
And because this is coming from Meta and Facebook,
6:14.380–6:16.940
so this is Marketplace API.
6:17.240–6:21.100
So you will be able to actually get the ride because
6:19.170–6:21.100
they already have the data.
6:21.200–6:23.340
So now it has already created the listing.
6:23.500–6:28.460
So you can clearly see that Global Primo Ping three-wheel kit scooter,
6:25.980–6:28.460
hot pink and black.
6:28.540–6:31.700
It also identified this, the best for scooter for toddlers and kids.
6:32.100–6:33.540
This condition is very good.
6:33.800–6:34.780
The retail is for this.
6:35.140–6:38.000
But then here we are charging it at this particular price, right?
6:38.000–6:39.920
So that's what it was able to do.
6:40.060–6:42.320
And you can see that it did a pretty good job.
6:42.420–6:47.080
And you were able to then copy this and just take it and
6:44.750–6:47.080
paste it in Marketplace, right?
6:47.500–6:54.900
So the reason I wanted to do this was to show you how you're able to create something like this very quickly using new Spark.
6:55.080–7:00.420
And it really demonstrates an agentic behavioral understanding of a multimodal input in this case video.
7:00.780–7:02.240
And also doing a quick search.
7:02.240–7:08.780
And then not only just limited to the US, but also
7:04.420–7:06.600
search across the board and then providing you lik
7:06.600–7:08.780
e a competitive pricing, right?
7:08.820–7:11.040
So I was very impressed with what I saw.
7:11.260–7:12.960
Okay, so that's the first demo.
7:13.080–7:13.960
I hope you enjoyed it.
7:14.280–7:20.600
Now, what I want to do here is I actually want to show you how you are also
7:17.440–7:20.600
able to use it directly in CloudCode.
7:20.720–7:23.560
So for that, let's just open Cloud, right?
7:23.780–7:28.140
So in this case, I've already configured Cloud's connector.
7:28.140–7:35.020
So if I show you this, you can see that here I have got the new Spark 1.
7:31.580–7:35.020
1 agent here, right?
7:35.100–7:38.440
So the way I was able to do this was I basically followed this specific instruction.
7:38.800–7:42.240
Here I provided the API key, which I already showed you.
7:42.380–7:45.360
And I ran this in PowerShell first, right?
7:45.420–7:48.680
So before actually running Cloud, I basically ran this, right?
7:48.720–7:55.400
So once I have this, now Cloud is being forced to use this particular model,
7:52.060–7:55.400
which is new Spark 1.1, right?
7:55.400–7:59.400
So you are welcome to try this and
7:57.400–7:59.400
see what kind of results you're getting.
8:00.280–8:05.340
All right, so for the third demo,
8:01.967–8:03.653
I decided to actually use one of
8:03.653–8:05.340
Meta's cookbooks, which they have given.
8:05.600–8:10.160
And they have really done a great job and provided 10,
8:07.880–8:10.160
actually 13 different use cases.
8:10.720–8:14.560
So here in this one,
8:12.640–8:14.560
I actually decided to use this perception grounding, right?
8:14.560–8:28.420
So the use case here is you have your fridge filled with some food objects and
8:19.180–8:23.800
can Meta's multimodal model be able to get and identify each one of the food o
8:23.800–8:28.420
bjects and provide some sort of a score, right?
8:28.800–8:34.840
So if I go in detail, like this is the original image,
8:31.820–8:34.840
you can see all the different food items over here.
8:34.840–8:45.120
And once you basically run this particular program and
8:38.267–8:41.693
provide your API key, it should be able to identify ea
8:41.693–8:45.120
ch one of these items and provide the score, right?
8:45.180–8:47.700
So what I did was I ran this particular prompt, right?
8:47.760–8:51.700
I'm a pescetarian with high cholesterol,
8:49.730–8:51.700
put green dots on recommended food.
8:51.700–8:59.400
And now that it has run this particular output,
8:54.267–8:56.833
once I give this particular command, I will be
8:56.833–8:59.400
able to see the output, right?
8:59.700–9:05.220
You can see it is now generating the HTML overlay so
9:02.460–9:05.220
that it will be able to identify that, right?
9:05.220–9:08.840
So I want to also show you the output of what it produces.
9:09.240–9:14.100
So again, just for the reference,
9:11.670–9:14.100
this is how the original picture looks like, right?
9:14.200–9:16.500
So if you look at this is how the picture looks like.
9:16.500–9:23.560
Now, what we will be able to generate is something like this,
9:20.030–9:23.560
which is after it has generated the output, right?
9:23.620–9:27.520
So pescetarian plus high cholesterol,
9:25.570–9:27.520
fridge guide recommended and non-recommended.
9:27.600–9:30.060
And you can see the scores pretty much well done.
9:30.180–9:31.260
So this is the output.
9:31.760–9:39.480
And I again thought like this was amazing because
9:34.333–9:36.907
you can see orange juice, we all think that it is
9:36.907–9:39.480
great, but for some reason it is giving not a great score.
9:39.720–9:43.420
So let's see if I eat something which is not recommended.
9:43.540–9:45.040
So cheese pack is not recommended.
9:45.040–9:46.840
Yellow cheese is not recommended.
9:47.080–9:48.940
Whereas this butter is definitely not recommended.
9:49.100–9:50.200
So pretty cool, right?
9:50.240–9:55.820
So you are able to use this out of the box and
9:53.030–9:55.820
able to get this label and it's fast, it's cheaper as well.
9:56.360–9:57.880
So this is what I wanted to cover.
9:58.120–10:04.020
The main idea is a make you aware of this is a brand new agentic model out there.
10:04.120–10:05.420
So definitely give it a try.
10:05.540–10:09.180
You can see there like this is a good playground and a dashboard.
10:09.660–10:10.660
So I've been playing with it.
10:10.740–10:14.120
It provides like the usage and stuff like that very well, etc.
10:14.120–10:15.300
All of those things.
10:15.380–10:17.640
And then they have also given like $20 free.
10:18.140–10:21.580
So, so far I've ran it a few times and
10:19.860–10:21.580
I only spent less than a dollar.
10:21.740–10:23.840
And there's another way for you to try, which is playground.
10:24.020–10:26.720
So you can upload an image or ask certain questions.
10:26.720–10:28.680
You can also change some of the settings over here.
10:28.880–10:30.720
And then there are some advanced settings as well.
10:30.760–10:30.900
Right.
10:30.900–10:33.000
You can add some JSON schema and stuff like that.
10:33.000–10:35.120
So what I would recommend is give it a shot.
10:35.240–10:38.280
Also try it in combination with Cloud Code.
10:38.460–10:44.020
The app that I built, I asked Antigravity to actually build the app,
10:41.240–10:44.020
but leveraging the Spark API.
10:44.220–10:51.440
You could do like all of these types of combination where you use your ID and
10:47.830–10:51.440
build an app like this and see for yourself what you feel as the performance.
10:51.620–10:51.740
Right.
10:51.740–10:55.600
And obviously on the benchmarks,
10:53.670–10:55.600
they have talked about the benchmarks over here.
10:55.680–10:58.960
They've compared themselves against Gemini 3.1, 4.8.
10:59.280–11:00.780
These are here for your reading.
11:00.940–11:02.540
And I will share this as well.
11:02.900–11:06.820
Of course, it does all different types of use cases from an agentic perspective.
11:07.340–11:08.700
It also does computer use.
11:08.840–11:10.440
It definitely writes code.
11:10.800–11:13.780
All of these things are something which they have actually explained.
11:13.960–11:14.000
Right.
11:14.020–11:15.740
So multimodal is something which we saw live.
11:16.620–11:18.220
Again, hopefully this was helpful.
11:18.220–11:21.160
It added some extra knowledge to your existing knowledge base.
11:21.500–11:24.920
Let me know if you guys have any questions and
11:23.210–11:24.920
what you feel after trying this.
11:25.060–11:25.920
Thank you very much for your time.
11:26.000–11:27.840
If you like the video, please hit that like button.
11:27.940–11:30.560
And if you're new here, please hit that subscribe button as well.
11:30.960–11:33.180
Thank you for watching and I will see you in the next one.
0:00.000–0:05.160
Meta 剛剛發布了它最強大的 AI 模型，即 MuseSpark 1.1。
0:05.540–0:13.300
根據 Zugg 的說法，這是一個能夠執行動作的模型，
0:08.127–0:10.713
它可以觀看影片、使用工具並完成任務，
0:10.713–0:13.300
而不只是回答問題。
0:13.820–0:15.660
在影片後面，我會詳細討論這些內容。
0:15.820–0:21.760
實際上，馬克·祖克柏在 X 平台上宣布的方式，
0:18.790–0:21.760
我認為他在幾年後又重新關注了這一點。
0:21.760–0:31.000
看看他強調的重點：一個強大的代理（agentic）和
0:26.380–0:31.000
透過我們新的 Meta 模型 API，以非常低的價格提供編碼模型。
0:31.780–0:34.000
低廉的價格，付費 API。
0:34.680–0:40.700
現在想想這一點，來自一家多年來以免費和
0:37.690–0:40.700
開源 AI 為其整個身份的公司的做法。
0:41.360–0:49.400
所以在深入探討 Spark 實際能做什麼之前，
0:44.040–0:46.720
我認為了解什麼
0:46.720–0:49.400
促使 Meta 改變策略會很有趣。
0:49.400–0:52.420
所以讓我們回到歷史，倒帶回去，對吧？
0:52.480–0:54.500
因為背景故事解釋了一切。
0:55.020–0:57.740
多年來，Meta 一直是開源 AI 的冠軍。
0:57.980–1:05.520
早在 2024 年 7 月，祖克發布了一篇標題為
1:01.750–1:05.520
字面意義上「開源 AI 是未來的道路」的文章。
1:06.160–1:14.300
免費可下載的模型是 Meta 的全部身份，
1:08.873–1:11.587
他將公司定位為像 OpenAI 這樣的閉源實驗室
1:11.587–1:14.300
的開源替代方案。
1:15.000–1:16.400
然後變化很快。
1:16.400–1:24.040
Meta 聘請了 Scale AI 的 Alexander Wang 擔任首席 AI 官，
1:20.220–1:24.040
這是一筆 140 億美元的交易。
1:24.660–1:34.200
他們建立了一個全新的超級智能實驗室，
1:27.840–1:31.020
並在四月推出了第一個 Mewk Spark m
1:31.020–1:34.200
模型，這是閉源的，沒有等待，也不提供下載。
1:34.940–1:36.940
這確實是與開源的分離。
1:37.440–1:38.820
這是第一次，對吧？
1:38.900–1:45.260
而 Meta 的前沿模型被鎖在它們自己的門後，
1:42.080–1:45.260
運行在它們的應用程式內部，而不是公開開放。
1:45.260–1:51.000
本週，Meta 邁出了下一步，一次重大升級，Mews Spark 1.1。
1:51.320–1:58.260
並且有史以來第一次，公開付費 API 並提供定價，
1:54.790–1:58.260
祖克柏稱其為「非常具侵略性」。
1:58.880–2:05.600
所以 Meta 不僅封閉了它最好的模型，
2:02.240–2:05.600
還給它貼上了價格標籤，並開放給任何人來構建它。
2:05.960–2:08.180
Meta 不僅在價格上壓倒所有人。
2:08.180–2:11.800
它只有 GPT 和 Claude 成本的四分之一。
2:11.800–2:16.880
他們還讓 Mews Spark 直接嵌入開發人員已經使用的工具中。
2:17.380–2:21.440
無論是 OpenAI 的設定，還是 Claude Code（Anthropic 自家的編碼工具），都可以。
2:21.900–2:28.960
你基本上只需針對 Meta 的模型更改一行程式碼，
2:24.253–2:26.607
現在你突然就能以四分之一的價格運行 Mews Spark。
2:26.607–2:28.960
四分之一的價格。
2:29.120–2:35.340
Meta 在 OpenAI 和 Anthropic 之間開了一扇側門，
2:32.230–2:35.340
並在門上掛了一個折扣標籤。
2:35.340–2:39.520
這不僅僅是一個新模型，
2:37.430–2:39.520
這是直接針對其競爭對手客戶的搶攻。
2:40.160–2:41.740
現在，有兩件事需要誠實說明。
2:41.900–2:45.020
第一，Llama 並沒有死，它仍然是開放的，仍可下載。
2:45.480–2:47.880
改變的是，Meta 的最佳模型現在是封閉的。
2:48.020–2:51.440
第二，雖然這個轉變感覺很瘋狂，但這是合理的，對吧？
2:51.520–2:55.200
Meta 每年在 AI 上的支出超過 1000 億美元。
2:55.660–2:58.000
你不會花了這麼多錢，卻把最好的模型免費送人。
2:58.360–3:04.160
如果你希望 30 億人手裡都有個人助理，
3:01.260–3:04.160
你就需要擁有執行這項工作的模型。
3:04.160–3:07.740
那麼他們到底建立了什麼？這值得你花時間嗎？
3:08.260–3:13.580
過去幾天我開始使用它，
3:10.920–3:13.580
我還基於 Meta 的模型建構了一個小應用程式。
3:14.180–3:22.400
我給它一段你可能想出售的商品影片，
3:16.920–3:19.660
並要求它提供完整價格、準備好發布的
3:19.660–3:22.400
Facebook 市集上架清單。
3:22.900–3:24.320
所以這將是我們看到的第一個示範。
3:24.480–3:29.000
其次，我也在 Anthropic 的 Claude Code 上運行它，
3:26.740–3:29.000
所以這將是一個有趣的測試。
3:29.000–3:36.760
第三，我想測試真正的多模態能力，
3:31.587–3:34.173
所以我給它一張冰箱裡食材的照片，
3:34.173–3:36.760
它能夠標註出上面的價格標籤，對吧？
3:37.080–3:47.280
所以我們會逐步展示所有有趣的部分，
3:40.480–3:43.880
但接著，正如我之前提到的，有趣的部分在於
3:43.880–3:47.280
Meta 轉向付費策略，而非僅僅保持開源。
3:47.560–3:47.780
好嗎？
3:47.980–3:53.560
在深入探討之前，先做個快速聲明，
3:50.770–3:53.560
所有意見均為我個人觀點，不代表我的僱主。
3:53.820–3:54.080
好嗎？
3:54.080–3:55.820
有了這個聲明，讓我們開始吧。
3:56.720–4:04.860
體驗 Spark 最快的方法之一，其實是來到 Meta。
4:00.790–4:04.860
ai 並將模式從即時（Instant）切換為思考（Thinking）模式。
4:05.200–4:09.840
所以如果你問任何問題，例如，
4:07.520–4:09.840
這裡我問的是，你正在使用哪個特定的 Meta 模型？
4:10.200–4:14.200
你會看到它確認正在使用 Spark 1.1。
4:14.640–4:20.380
所以，你只需要將模式切換為「思考模式」，
4:17.510–4:20.380
你就能夠利用 Spark 1.1 模型，對吧？
4:20.380–4:23.380
所以它就在幾天前剛剛發布，好吧？
4:24.080–4:28.820
但是，從開發者的角度來看，
4:26.450–4:28.820
如果你想使用它，你需要進入 dev.meta.ai。
4:29.220–4:40.060
所以這裡我在 dev.meta.ai，並且我能夠建立一個 API 金鑰，
4:32.833–4:36.447
你可以看到詳細的文件，以及如何將
4:36.447–4:40.060
這用於程式碼中。
4:40.060–4:43.040
這些都是他們提供的詳細資訊，包括編解碼器。
4:43.320–4:46.180
然後，如果你想透過 Python 以及 curl 來呼叫它，對吧？
4:46.220–4:48.460
這些是一些已提供的詳細資訊。
4:48.920–4:51.460
所以你可以做的是建立一個 API 金鑰。
4:51.580–4:55.060
所以我所做的，正如我之前提到的，
4:53.320–4:55.060
我將進行三個不同的示範。
4:55.060–5:00.500
所以這裡的第一個示範是，
4:57.780–5:00.500
把它想像成一個 Facebook Marketplace 的清單生成器。
5:00.820–5:02.720
所以這裡我會使用 MuseSpark。
5:03.280–5:05.740
所以我要在這裡上傳一段影片。
5:05.860–5:07.200
所以你可以很清楚地看到這段影片。
5:07.780–5:13.400
這是一段關於小孩腳踏車的影片，我只截取了一小段九秒鐘。
5:13.520–5:18.440
我只是想讓它稍微困難一點，並且
5:15.980–5:18.440
我會要求它生成清單，對吧？
5:18.480–5:20.840
所以這裡發生了多件事情，對吧？
5:20.840–5:24.040
所以顯然它正在呼叫 MuseSpark 1.1 模型。
5:24.500–5:28.200
所以這裡發生的第一件事是它將識別該物品。
5:28.560–5:31.540
所以這是來自第一步的回應。
5:31.940–5:36.220
所以它正確地識別出這是一輛 Global Primo Ping 三輪滑板車。
5:36.500–5:39.080
而且顯然它是二手的，所以它也理解了這一點。
5:39.420–5:40.580
所以做得非常好。
5:40.700–5:44.000
非常快，你可以看到所花費的時間非常短。
5:44.520–5:47.900
然後，代理程式現在將進行定價研究。
5:47.900–5:53.280
所以它將查看其他一些價格，以便
5:50.590–5:53.280
它可以給我們一些比較價格，對吧？
5:53.660–5:59.960
一旦你有了比較價格的概念，
5:56.810–5:59.960
它將整合所有這些資訊並生成清單。
6:00.080–6:00.860
所以你可以看到。
6:01.240–6:04.640
我將搜尋你這輛 Global 的目前二手清單。
6:04.760–6:06.660
在美國市場，大約是這個價格。
6:06.760–6:08.380
你可以在這裡看到所有的定價。
6:08.540–6:10.580
在英國，價格大約是這樣，對吧？
6:10.600–6:11.820
在加拿大，價格是這樣，對吧？
6:11.820–6:16.940
因為這是來自 Meta 和 Facebook 的資料，
6:14.380–6:16.940
所以這是 Marketplace API。
6:17.240–6:21.100
因此，你實際上能夠獲得這個騎乘服務，因為
6:19.170–6:21.100
他們已經擁有這些資料。
6:21.200–6:23.340
所以現在它已經建立了這個上架項目。
6:23.500–6:28.460
所以你可以清楚地看到 Global Primo Ping 三輪車套件滑板車，
6:25.980–6:28.460
熱粉紅色和黑色。
6:28.540–6:31.700
它也識別出這是適合幼兒和兒童的滑板車。
6:32.100–6:33.540
這個狀況非常好。
6:33.800–6:34.780
零售價是這樣。
6:35.140–6:38.000
但我們在這裡以這個特定價格收費，對吧？
6:38.000–6:39.920
所以這就是它能夠做到的事情。
6:40.060–6:42.320
你可以看到它做得相當不錯。
6:42.420–6:47.080
然後你可以複製這個，並直接將其
6:44.750–6:47.080
貼上到 Marketplace，對吧？
6:47.500–6:54.900
所以我想要這樣做的理由，是要向你展示如何能夠使用新的 Spark 非常快速地建立類似這樣的內容。
6:55.080–7:00.420
這確實展示了對多模態輸入（在此情況下為影片）的代理行為理解。
7:00.780–7:02.240
以及進行快速搜尋。
7:02.240–7:08.780
並且不僅限於美國，還能夠
7:04.420–7:06.600
進行全面搜尋，然後為你提供類似
7:06.600–7:08.780
具有競爭力的定價，對吧？
7:08.820–7:11.040
所以我對我所看到的印象非常深刻。
7:11.260–7:12.960
好的，這是第一個示範。
7:13.080–7:13.960
我希望你喜歡它。
7:14.280–7:20.600
現在，我想要在這裡做的是，我實際上想向你展示你如何也能
7:17.440–7:20.600
直接在 CloudCode 中使用它。
7:20.720–7:23.560
為此，讓我們打開 Cloud，對吧？
7:23.780–7:28.140
在這種情況下，我已經配置了 Cloud 的連接器。
7:28.140–7:35.020
如果我向你展示這個，你可以看到這裡我有新的 Spark 1.
7:31.580–7:35.020
1 代理，對吧？
7:35.100–7:38.440
所以我能夠這樣做的做法是，我基本上遵循了這個特定的指示。
7:38.800–7:42.240
在這裡我提供了 API 金鑰，這是我之前已經向你展示過的。
7:42.380–7:45.360
並且我首先在 PowerShell 中執行了它，對吧？
7:45.420–7:48.680
所以在實際運行 Cloud 之前，我基本上執行了這個，對吧？
7:48.720–7:55.400
所以一旦我有了這個，現在 Cloud 就被強制使用這個特定模型，
7:52.060–7:55.400
也就是新的 Spark 1.1，對吧？
7:55.400–7:59.400
所以歡迎你嘗試這個並
7:57.400–7:59.400
看看你能得到什麼樣的結果。
8:00.280–8:05.340
好的，那麼第三個示範，
8:01.967–8:03.653
我決定實際上使用
8:03.653–8:05.340
Meta 提供的食譜之一，
8:05.600–8:10.160
他們確實做得很好，並提供了 10 個，
8:07.880–8:10.160
實際上 13 個不同的使用案例。
8:10.720–8:14.560
所以在這裡，
8:12.640–8:14.560
我實際上決定使用這個感知錨定，對吧？
8:14.560–8:28.420
所以這裡的使用案例是，你的冰箱裡裝滿了某些食物物品，
8:19.180–8:23.800
Meta 的多模態模型能否獲取並識別每個食物物
8:23.800–8:28.420
品，並提供某種分數，對吧？
8:28.800–8:34.840
所以如果我詳細說明，這是原始圖片，
8:31.820–8:34.840
你可以看到這裡有不同的食物項目。
8:34.840–8:45.120
一旦你基本上運行這個特定的程式並
8:38.267–8:41.693
提供你的 API 金鑰，它應該能夠識別每
8:41.693–8:45.120
個項目並提供分數，對吧？
8:45.180–8:47.700
所以我做的是我運行了這個特定的提示，對吧？
8:47.760–8:51.700
我是吃魚的素食者，且膽固醇偏高，
8:49.730–8:51.700
在推薦的食物上標記綠點。
8:51.700–8:59.400
現在它運行了這個特定的輸出，
8:54.267–8:56.833
一旦我發出這個特定的指令，我將
8:56.833–8:59.400
能夠看到輸出結果，對吧？
8:59.700–9:05.220
你可以看到它現在正在生成 HTML 疊加層，
9:02.460–9:05.220
這樣它就能夠識別出來，對吧？
9:05.220–9:08.840
所以我也想向你展示它產生的輸出結果。
9:09.240–9:14.100
所以再次，僅供參考，
9:11.670–9:14.100
這是原始圖片的樣子，對吧？
9:14.200–9:16.500
所以如果你看，這就是圖片的樣子。
9:16.500–9:23.560
現在，我們能夠生成的東西就像這樣，
9:20.030–9:23.560
這是在它生成輸出之後，對吧？
9:23.620–9:27.520
所以吃魚素食者加上高膽固醇，
9:25.570–9:27.520
冰箱指南推薦與不推薦。
9:27.600–9:30.060
你可以看到分數相當不錯。
9:30.180–9:31.260
所以這就是輸出結果。
9:31.760–9:39.480
我再次覺得這很驚人，因為
9:34.333–9:36.907
你可以看到柳橙汁，我們都認為它
9:36.907–9:39.480
很棒，但出於某種原因，它給出的分數不高。
9:39.720–9:43.420
所以讓我們看看如果我吃不推薦的東西會怎樣。
9:43.540–9:45.040
所以起司包不推薦。
9:45.040–9:46.840
黃起司不推薦。
9:47.080–9:48.940
而這款奶油絕對不推薦。
9:49.100–9:50.200
所以很酷吧？
9:50.240–9:55.820
因此你可以直接上手使用，
9:53.030–9:55.820
並獲得這個標籤，它速度快，也更便宜。
9:56.360–9:57.880
這就是我想介紹的內容。
9:58.120–10:04.020
主要觀念是要讓你意識到，市面上出現了一個全新的代理型（agentic）模型。
10:04.120–10:05.420
所以絕對建議你試試看。
10:05.540–10:09.180
你可以看到這裡有個很好的遊樂場（playground）和儀表板。
10:09.660–10:10.660
所以我一直在玩這個。
10:10.740–10:14.120
它提供了很好的使用量統計等資訊。
10:14.120–10:15.300
所有這些功能。
10:15.380–10:17.640
然後他們還提供了20美元的免費額度。
10:18.140–10:21.580
所以到目前為止，我運行過幾次，
10:19.860–10:21.580
花費不到一美元。
10:21.740–10:23.840
還有另一種嘗試方式，就是遊樂場（playground）。
10:24.020–10:26.720
你可以上傳圖片或提出特定問題。
10:26.720–10:28.680
你也可以在這裡更改一些設定。
10:28.880–10:30.720
此外還有一些進階設定。
10:30.760–10:30.900
沒錯。
10:30.900–10:33.000
你可以加入一些 JSON schema 等內容。
10:33.000–10:35.120
所以我的建議是去試試看。
10:35.240–10:38.280
也可以嘗試與 Cloud Code 搭配使用。
10:38.460–10:44.020
我建構的應用程式，我要求 Antigravity 實際建構該應用程式，
10:41.240–10:44.020
但利用的是 Spark API。
10:44.220–10:51.440
你可以進行各種類型的組合，使用你的 ID，
10:47.830–10:51.440
建構這樣的應用程式，並親自體驗你感受到的效能。
10:51.620–10:51.740
沒錯。
10:51.740–10:55.600
當然在基準測試方面，
10:53.670–10:55.600
他們在這裡討論了基準測試。
10:55.680–10:58.960
他們將自己與 Gemini 3.1、4.8 進行了比較。
10:59.280–11:00.780
這些供你閱讀參考。
11:00.940–11:02.540
我也會分享這些內容。
11:02.900–11:06.820
當然，從代理型（agentic）的角度來看，它確實處理各種不同類型的用例。
11:07.340–11:08.700
它也支援電腦操作（computer use）。
11:08.840–11:10.440
它絕對能撰寫程式碼。
11:10.800–11:13.780
所有這些功能都是他們實際解釋過的。
11:13.960–11:14.000
沒錯。
11:14.020–11:15.740
所以多模態（multimodal）是我們現場看到的。
11:16.620–11:18.220
再次希望這有幫助。
11:18.220–11:21.160
它為你的既有知識庫增添了額外知識。
11:21.500–11:24.920
如果大家有任何問題，或是嘗試後有什麼感受，歡迎告訴我
11:23.210–11:24.920
以及嘗試之後的感受
11:25.060–11:25.920
非常感謝大家撥冗觀看
11:26.000–11:27.840
如果您喜歡這部影片，請點擊讚按鈕
11:27.940–11:30.560
如果您是第一次來到這裡，也請點擊訂閱按鈕
11:30.960–11:33.180
感謝觀看，我們下一部影片見
0:00.000–0:05.160
Meta just released its most capable AI model, which is MuseSpark 1.1.
Meta 剛剛發布了它最強大的 AI 模型，即 MuseSpark 1.1。
0:05.540–0:13.300
And according to Zugg, this is a model that can act,
根據 Zugg 的說法，這是一個能夠執行動作的模型，
0:08.127–0:10.713
it can watch a video, use tools and get it as done f
它可以觀看影片、使用工具並完成任務，
0:10.713–0:13.300
or you, not just answer questions.
而不只是回答問題。
0:13.820–0:15.660
And later in the video, I'm going to talk about all of that.
在影片後面，我會詳細討論這些內容。
0:15.820–0:21.760
Mark Zuckerberg, actually, the way he announced it on the X platform,
實際上，馬克·祖克柏在 X 平台上宣布的方式，
0:18.790–0:21.760
I think he came back on it after a few years.
我認為他在幾年後又重新關注了這一點。
0:21.760–0:31.000
And look at what he leads with, a strong agentic and
看看他強調的重點：一個強大的代理（agentic）和
0:26.380–0:31.000
coding model at a very low price through our new Meta model API.
透過我們新的 Meta 模型 API，以非常低的價格提供編碼模型。
0:31.780–0:34.000
A low price, a paid API.
低廉的價格，付費 API。
0:34.680–0:40.700
Now think about this, from a company that for years made free and
現在想想這一點，來自一家多年來以免費和
0:37.690–0:40.700
open source AI its entire identity.
開源 AI 為其整個身份的公司的做法。
0:41.360–0:49.400
So before I get into what Spark can actually do,
所以在深入探討 Spark 實際能做什麼之前，
0:44.040–0:46.720
I think it'll be interesting to understand what
我認為了解什麼
0:46.720–0:49.400
pushed Meta to change the strategy like this.
促使 Meta 改變策略會很有趣。
0:49.400–0:52.420
So let's go back to history and rewind, right?
所以讓我們回到歷史，倒帶回去，對吧？
0:52.480–0:54.500
Because the backstory explains everything.
因為背景故事解釋了一切。
0:55.020–0:57.740
For years, Meta was the champion of open source AI.
多年來，Meta 一直是開源 AI 的冠軍。
0:57.980–1:05.520
Back in July 2024, Zuck published an essay titled,
早在 2024 年 7 月，祖克發布了一篇標題為
1:01.750–1:05.520
literally, Open Source AI is the Path Forward.
字面意義上「開源 AI 是未來的道路」的文章。
1:06.160–1:14.300
Free downloadable models were Meta's whole identity,
免費可下載的模型是 Meta 的全部身份，
1:08.873–1:11.587
and he positioned the company as the open alternativ
他將公司定位為像 OpenAI 這樣的閉源實驗室
1:11.587–1:14.300
e to closed labs like OpenAI.
的開源替代方案。
1:15.000–1:16.400
Then it changed fast.
然後變化很快。
1:16.400–1:24.040
Meta brought in Scale AI's Alexander Wang as their chief AI officer,
Meta 聘請了 Scale AI 的 Alexander Wang 擔任首席 AI 官，
1:20.220–1:24.040
a $14 billion deal.
這是一筆 140 億美元的交易。
1:24.660–1:34.200
They stood up a brand new super intelligence lab,
他們建立了一個全新的超級智能實驗室，
1:27.840–1:31.020
and back in April, shipped the first Mewk Spark m
並在四月推出了第一個 Mewk Spark m
1:31.020–1:34.200
odel, which was closed, no waits, no download available.
模型，這是閉源的，沒有等待，也不提供下載。
1:34.940–1:36.940
That was a real break from open source.
這確實是與開源的分離。
1:37.440–1:38.820
For the first time, right?
這是第一次，對吧？
1:38.900–1:45.260
And Meta's frontier model were locked behind its own doors,
而 Meta 的前沿模型被鎖在它們自己的門後，
1:42.080–1:45.260
running inside its apps instead of out in the open.
運行在它們的應用程式內部，而不是公開開放。
1:45.260–1:51.000
And this week, Meta took the next step, a big upgrade, Mews Spark 1.1.
本週，Meta 邁出了下一步，一次重大升級，Mews Spark 1.1。
1:51.320–1:58.260
And for the first time ever, a public paid API with pricing,
並且有史以來第一次，公開付費 API 並提供定價，
1:54.790–1:58.260
Zuckerberg called very aggressive.
祖克柏稱其為「非常具侵略性」。
1:58.880–2:05.600
So Meta didn't just close its best model,
所以 Meta 不僅封閉了它最好的模型，
2:02.240–2:05.600
it put a price tag on it and open it up for anyone to build it.
還給它貼上了價格標籤，並開放給任何人來構建它。
2:05.960–2:08.180
Meta didn't just undercut everyone on price.
Meta 不僅在價格上壓倒所有人。
2:08.180–2:11.800
It's about a quarter of what GPT and Claude cost.
它只有 GPT 和 Claude 成本的四分之一。
2:11.800–2:16.880
They also made Mews Spark drop straight into the tools developers already use.
他們還讓 Mews Spark 直接嵌入開發人員已經使用的工具中。
2:17.380–2:21.440
OpenAI setup, sure, but also Claude code, Anthropic's own coding tool.
無論是 OpenAI 的設定，還是 Claude Code（Anthropic 自家的編碼工具），都可以。
2:21.900–2:28.960
You pointed at Meta's model change basically one line,
你基本上只需針對 Meta 的模型更改一行程式碼，
2:24.253–2:26.607
and now you're suddenly running on Mews Spark at a qua
現在你突然就能以四分之一的價格運行 Mews Spark。
2:26.607–2:28.960
rter of the price.
四分之一的價格。
2:29.120–2:35.340
Meta built a side door out of OpenAI and Anthropic and
Meta 在 OpenAI 和 Anthropic 之間開了一扇側門，
2:32.230–2:35.340
hung a discount sign on it.
並在門上掛了一個折扣標籤。
2:35.340–2:39.520
This isn't just a new model,
這不僅僅是一個新模型，
2:37.430–2:39.520
it's a direct play for their rivals' customers.
這是直接針對其競爭對手客戶的搶攻。
2:40.160–2:41.740
Now, two things to keep honest.
現在，有兩件事需要誠實說明。
2:41.900–2:45.020
One, Lama isn't dead, it's still open, still downloadable.
第一，Llama 並沒有死，它仍然是開放的，仍可下載。
2:45.480–2:47.880
What changes, Meta's best models are closed now.
改變的是，Meta 的最佳模型現在是封閉的。
2:48.020–2:51.440
And two, as wild as the flip feels, it makes sense, right?
第二，雖然這個轉變感覺很瘋狂，但這是合理的，對吧？
2:51.520–2:55.200
Meta's spending north of $100 billion a year on AI.
Meta 每年在 AI 上的支出超過 1000 億美元。
2:55.660–2:58.000
You don't spend that and give your best model away.
你不會花了這麼多錢，卻把最好的模型免費送人。
2:58.360–3:04.160
And if you want a personal agent in the hands of 3 billion people,
如果你希望 30 億人手裡都有個人助理，
3:01.260–3:04.160
you want to own the model doing the work.
你就需要擁有執行這項工作的模型。
3:04.160–3:07.740
So what did they actually build, and is it worth your time?
那麼他們到底建立了什麼？這值得你花時間嗎？
3:08.260–3:13.580
I started using it for the last couple of days, and
過去幾天我開始使用它，
3:10.920–3:13.580
I also built a little app on Meta's model.
我還基於 Meta 的模型建構了一個小應用程式。
3:14.180–3:22.400
I handed it a video of anything that you probably want to sell,
我給它一段你可能想出售的商品影片，
3:16.920–3:19.660
and asked it to give me a full price, ready-to-post listing on
並要求它提供完整價格、準備好發布的
3:19.660–3:22.400
Facebook's marketplace.
Facebook 市集上架清單。
3:22.900–3:24.320
So that's one demo that we will see.
所以這將是我們看到的第一個示範。
3:24.480–3:29.000
Second, I also ran it on Claude code's Anthropic,
其次，我也在 Anthropic 的 Claude Code 上運行它，
3:26.740–3:29.000
so that'll be an interesting one.
所以這將是一個有趣的測試。
3:29.000–3:36.760
And third, I wanted to test the real multimodality,
第三，我想測試真正的多模態能力，
3:31.587–3:34.173
so I gave it a picture of the ingredients in the fr
所以我給它一張冰箱裡食材的照片，
3:34.173–3:36.760
idge, and it was able to pinpoint the price tag on it, right?
它能夠標註出上面的價格標籤，對吧？
3:37.080–3:47.280
So I'll show you all of the fun parts as we go through,
所以我們會逐步展示所有有趣的部分，
3:40.480–3:43.880
but then, interesting part, as I mentioned earlier, is
但接著，正如我之前提到的，有趣的部分在於
3:43.880–3:47.280
Meta's shift to the paid strategy versus just keeping it open source.
Meta 轉向付費策略，而非僅僅保持開源。
3:47.560–3:47.780
All right?
好嗎？
3:47.980–3:53.560
Before we dive deep into it, a quick disclaimer,
在深入探討之前，先做個快速聲明，
3:50.770–3:53.560
all opinions are my own and do not belong to my employer.
所有意見均為我個人觀點，不代表我的僱主。
3:53.820–3:54.080
All right?
好嗎？
3:54.080–3:55.820
With that, let's get into it.
有了這個聲明，讓我們開始吧。
3:56.720–4:04.860
All right, so one of the fastest ways to experience Spark is to actually come to Meta.
體驗 Spark 最快的方法之一，其實是來到 Meta。
4:00.790–4:04.860
ai and change the mode from Instant to Thinking mode.
ai 並將模式從即時（Instant）切換為思考（Thinking）模式。
4:05.200–4:09.840
So if you ask any question, like, for example,
所以如果你問任何問題，例如，
4:07.520–4:09.840
here I'm asking which particular Meta model are you using?
這裡我問的是，你正在使用哪個特定的 Meta 模型？
4:10.200–4:14.200
You would see that it is going to confirm that it is using Spark 1.1.
你會看到它確認正在使用 Spark 1.1。
4:14.640–4:20.380
So you just need to change the mode to Thinking,
所以，你只需要將模式切換為「思考模式」，
4:17.510–4:20.380
and you will get to leverage the Spark 1.1 model, right?
你就能夠利用 Spark 1.1 模型，對吧？
4:20.380–4:23.380
So it was launched literally a couple of days back, okay?
所以它就在幾天前剛剛發布，好吧？
4:24.080–4:28.820
But then, from a developer perspective,
但是，從開發者的角度來看，
4:26.450–4:28.820
if you want to use it, you need to go into dev.meta.ai.
如果你想使用它，你需要進入 dev.meta.ai。
4:29.220–4:40.060
So here I am in dev.meta.ai, and I was able to create an API key,
所以這裡我在 dev.meta.ai，並且我能夠建立一個 API 金鑰，
4:32.833–4:36.447
and you can already see detailed docs and how you're able to use
你可以看到詳細的文件，以及如何將
4:36.447–4:40.060
this in plot code.
這用於程式碼中。
4:40.060–4:43.040
And these are all the details that they have provided, codecs.
這些都是他們提供的詳細資訊，包括編解碼器。
4:43.320–4:46.180
Then if you want to just call it via Python as well as curl, right?
然後，如果你想透過 Python 以及 curl 來呼叫它，對吧？
4:46.220–4:48.460
These are some of the details that have been given.
這些是一些已提供的詳細資訊。
4:48.920–4:51.460
So what you could do is you can create an API key.
所以你可以做的是建立一個 API 金鑰。
4:51.580–4:55.060
So what I did was, as I mentioned earlier,
所以我所做的，正如我之前提到的，
4:53.320–4:55.060
I'm going to be doing three different demos.
我將進行三個不同的示範。
4:55.060–5:00.500
So the first demo here is,
所以這裡的第一個示範是，
4:57.780–5:00.500
think of it like a Facebook Marketplace listing generator.
把它想像成一個 Facebook Marketplace 的清單生成器。
5:00.820–5:02.720
So here I would be using MuseSpark.
所以這裡我會使用 MuseSpark。
5:03.280–5:05.740
So what I'm going to do here is upload a video here.
所以我要在這裡上傳一段影片。
5:05.860–5:07.200
So you can see the video very clearly.
所以你可以很清楚地看到這段影片。
5:07.780–5:13.400
This is a video of a bike of a kid where I've just taken like a very short nine seconds.
這是一段關於小孩腳踏車的影片，我只截取了一小段九秒鐘。
5:13.520–5:18.440
I just wanted to make it a little bit tough, and
我只是想讓它稍微困難一點，並且
5:15.980–5:18.440
I'm going to ask it to generate listing, right?
我會要求它生成清單，對吧？
5:18.480–5:20.840
So there are multiple things which are happening here, right?
所以這裡發生了多件事情，對吧？
5:20.840–5:24.040
So obviously it is calling the MuseSpark 1.1 model.
所以顯然它正在呼叫 MuseSpark 1.1 模型。
5:24.500–5:28.200
So the first thing what's happening here is it is going to be identifying the item.
所以這裡發生的第一件事是它將識別該物品。
5:28.560–5:31.540
So this is the response back from the first step.
所以這是來自第一步的回應。
5:31.940–5:36.220
So it has correctly identified that this is a Global Primo Ping three-wheel kit scooter.
所以它正確地識別出這是一輛 Global Primo Ping 三輪滑板車。
5:36.500–5:39.080
And obviously it is used, so it has understood that as well.
而且顯然它是二手的，所以它也理解了這一點。
5:39.420–5:40.580
So very good job done.
所以做得非常好。
5:40.700–5:44.000
Very quickly, you can see the amount of time it took was very short.
非常快，你可以看到所花費的時間非常短。
5:44.520–5:47.900
Then the agent is going to now do a pricing research.
然後，代理程式現在將進行定價研究。
5:47.900–5:53.280
So it's going to look into some other prices so
所以它將查看其他一些價格，以便
5:50.590–5:53.280
that it can give us some comparison pricing, right?
它可以給我們一些比較價格，對吧？
5:53.660–5:59.960
Once you have that idea of the comparison pricing,
一旦你有了比較價格的概念，
5:56.810–5:59.960
then it will consolidate all of that and create the listing.
它將整合所有這些資訊並生成清單。
6:00.080–6:00.860
So you can see that.
所以你可以看到。
6:01.240–6:04.640
I'll search current resale listing for your global this.
我將搜尋你這輛 Global 的目前二手清單。
6:04.760–6:06.660
In the US market, it is around this.
在美國市場，大約是這個價格。
6:06.760–6:08.380
And you can see all the pricing over here.
你可以在這裡看到所有的定價。
6:08.540–6:10.580
In UK, it is something of this sort, right?
在英國，價格大約是這樣，對吧？
6:10.600–6:11.820
In Canada, it is this, right?
在加拿大，價格是這樣，對吧？
6:11.820–6:16.940
And because this is coming from Meta and Facebook,
因為這是來自 Meta 和 Facebook 的資料，
6:14.380–6:16.940
so this is Marketplace API.
所以這是 Marketplace API。
6:17.240–6:21.100
So you will be able to actually get the ride because
因此，你實際上能夠獲得這個騎乘服務，因為
6:19.170–6:21.100
they already have the data.
他們已經擁有這些資料。
6:21.200–6:23.340
So now it has already created the listing.
所以現在它已經建立了這個上架項目。
6:23.500–6:28.460
So you can clearly see that Global Primo Ping three-wheel kit scooter,
所以你可以清楚地看到 Global Primo Ping 三輪車套件滑板車，
6:25.980–6:28.460
hot pink and black.
熱粉紅色和黑色。
6:28.540–6:31.700
It also identified this, the best for scooter for toddlers and kids.
它也識別出這是適合幼兒和兒童的滑板車。
6:32.100–6:33.540
This condition is very good.
這個狀況非常好。
6:33.800–6:34.780
The retail is for this.
零售價是這樣。
6:35.140–6:38.000
But then here we are charging it at this particular price, right?
但我們在這裡以這個特定價格收費，對吧？
6:38.000–6:39.920
So that's what it was able to do.
所以這就是它能夠做到的事情。
6:40.060–6:42.320
And you can see that it did a pretty good job.
你可以看到它做得相當不錯。
6:42.420–6:47.080
And you were able to then copy this and just take it and
然後你可以複製這個，並直接將其
6:44.750–6:47.080
paste it in Marketplace, right?
貼上到 Marketplace，對吧？
6:47.500–6:54.900
So the reason I wanted to do this was to show you how you're able to create something like this very quickly using new Spark.
所以我想要這樣做的理由，是要向你展示如何能夠使用新的 Spark 非常快速地建立類似這樣的內容。
6:55.080–7:00.420
And it really demonstrates an agentic behavioral understanding of a multimodal input in this case video.
這確實展示了對多模態輸入（在此情況下為影片）的代理行為理解。
7:00.780–7:02.240
And also doing a quick search.
以及進行快速搜尋。
7:02.240–7:08.780
And then not only just limited to the US, but also
並且不僅限於美國，還能夠
7:04.420–7:06.600
search across the board and then providing you lik
進行全面搜尋，然後為你提供類似
7:06.600–7:08.780
e a competitive pricing, right?
具有競爭力的定價，對吧？
7:08.820–7:11.040
So I was very impressed with what I saw.
所以我對我所看到的印象非常深刻。
7:11.260–7:12.960
Okay, so that's the first demo.
好的，這是第一個示範。
7:13.080–7:13.960
I hope you enjoyed it.
我希望你喜歡它。
7:14.280–7:20.600
Now, what I want to do here is I actually want to show you how you are also
現在，我想要在這裡做的是，我實際上想向你展示你如何也能
7:17.440–7:20.600
able to use it directly in CloudCode.
直接在 CloudCode 中使用它。
7:20.720–7:23.560
So for that, let's just open Cloud, right?
為此，讓我們打開 Cloud，對吧？
7:23.780–7:28.140
So in this case, I've already configured Cloud's connector.
在這種情況下，我已經配置了 Cloud 的連接器。
7:28.140–7:35.020
So if I show you this, you can see that here I have got the new Spark 1.
如果我向你展示這個，你可以看到這裡我有新的 Spark 1.
7:31.580–7:35.020
1 agent here, right?
1 代理，對吧？
7:35.100–7:38.440
So the way I was able to do this was I basically followed this specific instruction.
所以我能夠這樣做的做法是，我基本上遵循了這個特定的指示。
7:38.800–7:42.240
Here I provided the API key, which I already showed you.
在這裡我提供了 API 金鑰，這是我之前已經向你展示過的。
7:42.380–7:45.360
And I ran this in PowerShell first, right?
並且我首先在 PowerShell 中執行了它，對吧？
7:45.420–7:48.680
So before actually running Cloud, I basically ran this, right?
所以在實際運行 Cloud 之前，我基本上執行了這個，對吧？
7:48.720–7:55.400
So once I have this, now Cloud is being forced to use this particular model,
所以一旦我有了這個，現在 Cloud 就被強制使用這個特定模型，
7:52.060–7:55.400
which is new Spark 1.1, right?
也就是新的 Spark 1.1，對吧？
7:55.400–7:59.400
So you are welcome to try this and
所以歡迎你嘗試這個並
7:57.400–7:59.400
see what kind of results you're getting.
看看你能得到什麼樣的結果。
8:00.280–8:05.340
All right, so for the third demo,
好的，那麼第三個示範，
8:01.967–8:03.653
I decided to actually use one of
我決定實際上使用
8:03.653–8:05.340
Meta's cookbooks, which they have given.
Meta 提供的食譜之一，
8:05.600–8:10.160
And they have really done a great job and provided 10,
他們確實做得很好，並提供了 10 個，
8:07.880–8:10.160
actually 13 different use cases.
實際上 13 個不同的使用案例。
8:10.720–8:14.560
So here in this one,
所以在這裡，
8:12.640–8:14.560
I actually decided to use this perception grounding, right?
我實際上決定使用這個感知錨定，對吧？
8:14.560–8:28.420
So the use case here is you have your fridge filled with some food objects and
所以這裡的使用案例是，你的冰箱裡裝滿了某些食物物品，
8:19.180–8:23.800
can Meta's multimodal model be able to get and identify each one of the food o
Meta 的多模態模型能否獲取並識別每個食物物
8:23.800–8:28.420
bjects and provide some sort of a score, right?
品，並提供某種分數，對吧？
8:28.800–8:34.840
So if I go in detail, like this is the original image,
所以如果我詳細說明，這是原始圖片，
8:31.820–8:34.840
you can see all the different food items over here.
你可以看到這裡有不同的食物項目。
8:34.840–8:45.120
And once you basically run this particular program and
一旦你基本上運行這個特定的程式並
8:38.267–8:41.693
provide your API key, it should be able to identify ea
提供你的 API 金鑰，它應該能夠識別每
8:41.693–8:45.120
ch one of these items and provide the score, right?
個項目並提供分數，對吧？
8:45.180–8:47.700
So what I did was I ran this particular prompt, right?
所以我做的是我運行了這個特定的提示，對吧？
8:47.760–8:51.700
I'm a pescetarian with high cholesterol,
我是吃魚的素食者，且膽固醇偏高，
8:49.730–8:51.700
put green dots on recommended food.
在推薦的食物上標記綠點。
8:51.700–8:59.400
And now that it has run this particular output,
現在它運行了這個特定的輸出，
8:54.267–8:56.833
once I give this particular command, I will be
一旦我發出這個特定的指令，我將
8:56.833–8:59.400
able to see the output, right?
能夠看到輸出結果，對吧？
8:59.700–9:05.220
You can see it is now generating the HTML overlay so
你可以看到它現在正在生成 HTML 疊加層，
9:02.460–9:05.220
that it will be able to identify that, right?
這樣它就能夠識別出來，對吧？
9:05.220–9:08.840
So I want to also show you the output of what it produces.
所以我也想向你展示它產生的輸出結果。
9:09.240–9:14.100
So again, just for the reference,
所以再次，僅供參考，
9:11.670–9:14.100
this is how the original picture looks like, right?
這是原始圖片的樣子，對吧？
9:14.200–9:16.500
So if you look at this is how the picture looks like.
所以如果你看，這就是圖片的樣子。
9:16.500–9:23.560
Now, what we will be able to generate is something like this,
現在，我們能夠生成的東西就像這樣，
9:20.030–9:23.560
which is after it has generated the output, right?
這是在它生成輸出之後，對吧？
9:23.620–9:27.520
So pescetarian plus high cholesterol,
所以吃魚素食者加上高膽固醇，
9:25.570–9:27.520
fridge guide recommended and non-recommended.
冰箱指南推薦與不推薦。
9:27.600–9:30.060
And you can see the scores pretty much well done.
你可以看到分數相當不錯。
9:30.180–9:31.260
So this is the output.
所以這就是輸出結果。
9:31.760–9:39.480
And I again thought like this was amazing because
我再次覺得這很驚人，因為
9:34.333–9:36.907
you can see orange juice, we all think that it is
你可以看到柳橙汁，我們都認為它
9:36.907–9:39.480
great, but for some reason it is giving not a great score.
很棒，但出於某種原因，它給出的分數不高。
9:39.720–9:43.420
So let's see if I eat something which is not recommended.
所以讓我們看看如果我吃不推薦的東西會怎樣。
9:43.540–9:45.040
So cheese pack is not recommended.
所以起司包不推薦。
9:45.040–9:46.840
Yellow cheese is not recommended.
黃起司不推薦。
9:47.080–9:48.940
Whereas this butter is definitely not recommended.
而這款奶油絕對不推薦。
9:49.100–9:50.200
So pretty cool, right?
所以很酷吧？
9:50.240–9:55.820
So you are able to use this out of the box and
因此你可以直接上手使用，
9:53.030–9:55.820
able to get this label and it's fast, it's cheaper as well.
並獲得這個標籤，它速度快，也更便宜。
9:56.360–9:57.880
So this is what I wanted to cover.
這就是我想介紹的內容。
9:58.120–10:04.020
The main idea is a make you aware of this is a brand new agentic model out there.
主要觀念是要讓你意識到，市面上出現了一個全新的代理型（agentic）模型。
10:04.120–10:05.420
So definitely give it a try.
所以絕對建議你試試看。
10:05.540–10:09.180
You can see there like this is a good playground and a dashboard.
你可以看到這裡有個很好的遊樂場（playground）和儀表板。
10:09.660–10:10.660
So I've been playing with it.
所以我一直在玩這個。
10:10.740–10:14.120
It provides like the usage and stuff like that very well, etc.
它提供了很好的使用量統計等資訊。
10:14.120–10:15.300
All of those things.
所有這些功能。
10:15.380–10:17.640
And then they have also given like $20 free.
然後他們還提供了20美元的免費額度。
10:18.140–10:21.580
So, so far I've ran it a few times and
所以到目前為止，我運行過幾次，
10:19.860–10:21.580
I only spent less than a dollar.
花費不到一美元。
10:21.740–10:23.840
And there's another way for you to try, which is playground.
還有另一種嘗試方式，就是遊樂場（playground）。
10:24.020–10:26.720
So you can upload an image or ask certain questions.
你可以上傳圖片或提出特定問題。
10:26.720–10:28.680
You can also change some of the settings over here.
你也可以在這裡更改一些設定。
10:28.880–10:30.720
And then there are some advanced settings as well.
此外還有一些進階設定。
10:30.760–10:30.900
Right.
沒錯。
10:30.900–10:33.000
You can add some JSON schema and stuff like that.
你可以加入一些 JSON schema 等內容。
10:33.000–10:35.120
So what I would recommend is give it a shot.
所以我的建議是去試試看。
10:35.240–10:38.280
Also try it in combination with Cloud Code.
也可以嘗試與 Cloud Code 搭配使用。
10:38.460–10:44.020
The app that I built, I asked Antigravity to actually build the app,
我建構的應用程式，我要求 Antigravity 實際建構該應用程式，
10:41.240–10:44.020
but leveraging the Spark API.
但利用的是 Spark API。
10:44.220–10:51.440
You could do like all of these types of combination where you use your ID and
你可以進行各種類型的組合，使用你的 ID，
10:47.830–10:51.440
build an app like this and see for yourself what you feel as the performance.
建構這樣的應用程式，並親自體驗你感受到的效能。
10:51.620–10:51.740
Right.
沒錯。
10:51.740–10:55.600
And obviously on the benchmarks,
當然在基準測試方面，
10:53.670–10:55.600
they have talked about the benchmarks over here.
他們在這裡討論了基準測試。
10:55.680–10:58.960
They've compared themselves against Gemini 3.1, 4.8.
他們將自己與 Gemini 3.1、4.8 進行了比較。
10:59.280–11:00.780
These are here for your reading.
這些供你閱讀參考。
11:00.940–11:02.540
And I will share this as well.
我也會分享這些內容。
11:02.900–11:06.820
Of course, it does all different types of use cases from an agentic perspective.
當然，從代理型（agentic）的角度來看，它確實處理各種不同類型的用例。
11:07.340–11:08.700
It also does computer use.
它也支援電腦操作（computer use）。
11:08.840–11:10.440
It definitely writes code.
它絕對能撰寫程式碼。
11:10.800–11:13.780
All of these things are something which they have actually explained.
所有這些功能都是他們實際解釋過的。
11:13.960–11:14.000
Right.
沒錯。
11:14.020–11:15.740
So multimodal is something which we saw live.
所以多模態（multimodal）是我們現場看到的。
11:16.620–11:18.220
Again, hopefully this was helpful.
再次希望這有幫助。
11:18.220–11:21.160
It added some extra knowledge to your existing knowledge base.
它為你的既有知識庫增添了額外知識。
11:21.500–11:24.920
Let me know if you guys have any questions and
如果大家有任何問題，或是嘗試後有什麼感受，歡迎告訴我
11:23.210–11:24.920
what you feel after trying this.
以及嘗試之後的感受
11:25.060–11:25.920
Thank you very much for your time.
非常感謝大家撥冗觀看
11:26.000–11:27.840
If you like the video, please hit that like button.
如果您喜歡這部影片，請點擊讚按鈕
11:27.940–11:30.560
And if you're new here, please hit that subscribe button as well.
如果您是第一次來到這裡，也請點擊訂閱按鈕
11:30.960–11:33.180
Thank you for watching and I will see you in the next one.
感謝觀看，我們下一部影片見

影片筆記：Muse Spark 1.1: Meta's First Paid Model (No Longer Open ??)

一句話總結

Meta 宣布其最佳模型 Muse Spark 1.1 轉為閉源並透過付費 API 開放，定價具侵略性（約為競爭對手的四分之一），並演示了其具備代理（Agentic）行為、多模態感知及與現有開發工具整合的能力。

核心重點

  • 策略重大轉變：Meta 從長期堅持的「開源 AI」策略轉向「付費閉源 API」。過去 Meta 以開源為核心身份（如 Zuck 曾發表《Open Source AI is the Path Forward》），但近期引入 Scale AI 的高管 Alexander Wang 為首席 AI 官，並建立新的超級智能實驗室。
  • Muse Spark 1.1 定位：這是 Meta 的第一個付費模型，被描述為具備「代理（Agentic）」行為，不僅是回答問題，還能觀看影片、使用工具並完成任務。
  • 定價策略：定價被稱為「具侵略性」，價格約為 GPT 和 Claude 的四分之一，意圖直接搶奪競爭對手的客戶。
  • 多模態與代理能力：
  • 支援影片、圖片、文字輸入。
  • 具備自動識別物品、進行跨國價格研究、生成銷售列表等自動化任務能力。
  • 具備「多模態感知錨定（Perception Grounding）」能力，能根據用戶健康狀況生成推薦與視覺化疊加。
  • 開發者生態整合：
  • 提供 dev.meta.ai 生成 API Key。
  • Playground 提供免費額度（$20），測試成本極低。
  • 可透過 API Key 配置，強制第三方工具（如 Claude Code）使用 Muse Spark 1.1 進行編碼。

詳細大綱

1. Meta 的戰略翻轉

  • 背景：Zuck 曾發表《Open Source AI is the Path Forward》，強調開源。
  • 變化：
  • 引入 Scale AI 高管 Alexander Wang 為首席 AI 官。
  • 建立新的超級智能實驗室。
  • 發布閉源模型 Muse Spark 1.1。
  • 市場影響：以低於競爭對手（約為 GPT/Claude 的 1/4）的價格，直接搶奪競爭對手客戶。

2. Muse Spark 1.1 核心特性

  • 代理（Agentic）能力：能行動、觀看影片、使用工具。
  • 多模態輸入處理：支援影片、圖片、文字。
  • 開發者友好：整合現有工具鏈（OpenAI setup, Claude Code）。

3. 實測演示（Demos）

  • Demo 1：自動生成 Facebook Marketplace 銷售列表
  • 輸入：一段兒童滑板車影片。
  • 過程：模型自動識別物品、進行跨國價格研究（美國、英國、加拿大）。
  • 輸出：生成可直接貼上 Marketplace 的完整銷售列表。
  • Demo 2：在 Claude Code 中強制使用 Muse Spark 1.1 API
  • 演示如何透過 API Key 配置，強制 Claude Code 使用 Muse Spark 1.1 模型進行編碼工作。
  • Demo 3：冰箱食材健康評分與視覺化疊加
  • 輸入：冰箱食材圖片。
  • 過程：模型識別食物，並根據用戶健康狀況（素食、高膽固醇）生成推薦與不推薦的評分。
  • 輸出：生成 HTML 疊加層（HTML overlay）。

4. 開發者指南與資源

  • 獲取 API Key：前往 dev.meta.ai。
  • Playground 使用：提供免費額度（$20），目前測試成本極低。
  • 第三方整合：
  • 與 Antigravity（疑點）或 Claude Code 結合開發應用。
  • 提供 Python 和 curl 調用範例。
  • 基準測試（Benchmarks）：參考影片提及的比較數據。

工具 / 模型 / 名詞整理

  • 模型/產品名稱：
  • Muse Spark 1.1（影片主要討論對象，亦被口誤為 Spark 1.1 或 Mewk Spark）。
  • Meta AI / Meta model API：Meta 的模型 API 服務。
  • Meta.ai：用戶端介面。
  • dev.meta.ai：開發者介面，用於生成 API Key。
  • Facebook Marketplace API：用於生成銷售列表的相關 API。
  • Meta's cookbooks：Meta 提供的範例程式碼庫。
  • 競爭對手/相關模型：
  • GPT：OpenAI 的模型系列。
  • Claude：Anthropic 的模型系列。
  • Anthropic：Claude 的開發公司。
  • OpenAI：GPT 的開發公司。
  • Gemini 3.1：Google 的模型，影片提及作為基準比較對象。
  • 4.8：影片提及的基準測試數字或模型版本（疑點）。
  • Lama：影片提及仍為開源（疑點為 Llama）。
  • 開發工具/環境：
  • Claude Code：Anthropic 的編碼工具，影片演示強制其使用 Muse Spark 1.1。
  • Python：程式語言，用於調用 API。
  • curl：命令列工具，用於調用 API。
  • PowerShell：命令列工具。
  • HTML overlay：HTML 疊加層，用於視覺化健康評分。
  • JSON schema：資料格式規範。
  • Antigravity：影片提及用於構建應用（疑點為某款 AI 編碼工具名稱）。
  • Scale AI：AI 數據與模型公司，其高管 Alexander Wang 被 Meta 聘為首席 AI 官。
  • Zugg：影片提及來源（疑點為 Zuck 或特定評論者）。
  • Global Primo Ping three-wheel kit scooter：影片識別出的物品名稱（疑點）。
  • Pescetarian：素食者（吃魚的素食者），影片正確拼寫。
  • Codecs：影片提及（疑點為 Code 或 Codecs 的聽寫錯誤）。
  • Plot code：影片提及（疑點為 Python code 或 Code 的聽寫錯誤）。

操作流程整理

  1. 獲取開發資源：
  • 前往 dev.meta.ai 生成 API Key。
  • 在 Playground 中測試，利用免費額度（$20）進行低成本測試。
  1. 整合第三方開發工具（以 Claude Code 為例）：
  • 配置 API Key。
  • 強制 Claude Code 使用 Muse Spark 1.1 模型進行編碼工作。
  1. 執行多模態任務（以 Marketplace 列表生成為例）：
  • 上傳影片（如兒童滑板車影片）。
  • 模型自動識別物品。
  • 模型進行跨國價格研究（美、英、加）。
  • 模型生成可直接貼上 Facebook Marketplace 的完整銷售列表。
  1. 執行健康評估任務（以冰箱食材為例）：
  • 輸入冰箱食材圖片。
  • 模型識別食物。
  • 輸入用戶健康狀況（如素食、高膽固醇）。
  • 模型生成推薦與不推薦的評分。
  • 模型生成 HTML 疊加層進行視覺化展示。

值得注意的限制或風險

  • 策略轉變的不確定性：Meta 從開源轉向閉源付費，可能影響開發者生態與信任度。
  • 模型名稱混亂：影片中正負交替出現 "Muse Spark 1.1"、"Spark 1.1"、"Mewk Spark" 等名稱，可能反映內部命名尚未穩定或口誤頻繁。
  • 技術整合複雜度：強制第三方工具（如 Claude Code）使用特定 API 可能需要額外的配置與維護成本。
  • 基準測試數據解讀：影片提及與 Gemini 3.1 及 "4.8" 的比較，但數據來源與具體含義不明確（疑點）。

逐字稿辨識疑點

  • Muse Spark / Spark / Mewk Spark：逐字稿中交替出現 "MuseSpark 1.1"、"Spark 1.1"、"Mewk Spark"、"new Spark"。根據 Meta 官方命名慣例，極大機率是指 Llama 3 或 Meta 新發布的模型，但影片講者口誤或聽寫錯誤導致名稱混亂。
  • Lama：影片提到 "Lama isn't dead"，疑點為 Llama（Meta 的開源模型系列）。
  • Zugg：影片提到 "according to Zugg"，疑點為 Zuck（Mark Zuckerberg 的暱稱）或特定評論者名稱。
  • 4.8：影片提到 "compared themselves against Gemini 3.1, 4.8"，疑點為基準測試的分數或另一個模型版本號，聽寫不清。
  • Antigravity：影片提到 "I asked Antigravity to actually build the app"，疑點為某款 AI 輔助編碼工具的名稱（如 Cursor, Windsurf, 或特定內部工具），需查證。
  • Global Primo Ping three-wheel kit scooter：影片識別出的物品名稱，疑點為 Global Primo Pingo 或類似品牌/型號的聽寫錯誤。
  • Pescetarian：影片正確拼寫為素食者（吃魚的素食者），逐字稿聽寫正確，但需注意此為健康術語。
  • Codecs：影片提到 "docs and how you're able to use this in plot code... codecs"，疑點為 Code 或 Codecs 的聽寫錯誤，上下文似指編碼或代碼範例。
  • Plot code：影片提到 "use this in plot code"，疑點為 Python code 或 Code 的聽寫錯誤。

可延伸追問

  • Meta 為何在堅持開源多年後，突然轉向付費閉源策略？背後的商業考量為何？
  • Muse Spark 1.1 的「代理（Agentic）」能力具體表現為何？與現有開源模型相比有何優勢？
  • 如何準確評估 Muse Spark 1.1 在實際開發場景中的效能與成本效益？
  • Meta 如何平衡開源社區與閉源付費模型之間的關係？
  • 第三方工具（如 Claude Code）強制使用 Muse Spark 1.1 的技術實現細節與潛在問題為何？

尚未產生學習筆記

請在 Telegram 指令最後加上「學習」，例如：videonote 網址 英文 雙語 學習