0:00.000–0:06.800
meta just released its most capable ai model which is muse spark 1.
0:03.400–0:06.800
1 and according to zag this is a
0:06.800–0:12.860
model that can act it can watch a video use tools and
0:09.830–0:12.860
get it as done for you not just answer
0:12.860–0:16.800
questions and later in the video i'm going to talk about all of that mark zuckerberg actually
0:16.800–0:22.620
the way he announced it on the x platform i think he came back on it after a few years and look at
0:22.620–0:29.980
what he leads with a strong agentic and
0:26.300–0:29.980
coding model at a very low price through our new meta
0:29.980–0:37.780
model api a low price a paid api now think about this from a company that for years made free
0:37.780–0:44.840
and open source ai its entire identity so
0:41.310–0:44.840
before i get into what spark can actually do i think it'll
0:44.840–0:51.300
be interesting to understand what pushed meta to change the strategy like this so let's go back
0:51.300–0:56.700
to history and rewind right because
0:54.000–0:56.700
the backstory explains everything for years meta was the champion
0:56.700–1:04.820
of open source ai back in july 2024 zuck published an essay titled literally open source ai is the
1:04.820–1:10.880
path forward free downloadable models were meta's whole identity and
1:07.850–1:10.880
he positioned the company as
1:10.880–1:19.080
the open alternator to closed labs like open ai then
1:14.980–1:19.080
it changed fast meta brought in scale ai's
1:19.080–1:26.340
alexander van as their chief ai officer a 14 billion dollar deal they stood up a brand new super
1:26.340–1:33.380
intelligence lab and
1:28.687–1:31.033
back in april shippe
1:31.033–1:33.380
d the first muke spark model which was closed no waits no
1:33.380–1:39.720
download available that was a real break from open source for the first time right and
1:36.550–1:39.720
meta's frontier
1:39.720–1:46.200
model were locked behind its own doors running inside its apps instead
1:42.960–1:46.200
of out in the open and this week
1:46.200–1:54.240
meta took the next step a big upgrade muse spark 1.
1:50.220–1:54.240
1 and for the first time ever a public paid api
1:54.240–2:02.600
with pricing zuckerberg called very aggressive so
1:58.420–2:02.600
meta didn't just close its best model it put a price
2:02.600–2:08.180
tag on it and open it up for anyone to build it meta didn't just undercut everyone on price
2:08.180–2:15.560
it's about a quarter of what gpt and claude cost they also
2:11.870–2:15.560
made mu spark drop straight into the tools
2:15.560–2:22.400
developers already use open ai setup sure but also
2:18.980–2:22.400
clawed code anthropic's own coding tool you pointed
2:22.400–2:28.520
at meta's model change basically one line and
2:25.460–2:28.520
now you're suddenly running on muse spark at a quarter of
2:28.520–2:35.960
the price meta built a side door out of open ai and
2:32.240–2:35.960
anthropic and hung a discount sign on it this
2:35.960–2:42.100
isn't just a new model it's a direct play for their rivals customers now two things to keep honest one
2:42.100–2:47.860
llama isn't dead it's still open still downloadable what changes matters best models are closed now
2:47.860–2:54.020
and two as wild as the flip feels it makes sense right matters spending north of a hundred billion
2:54.020–2:59.300
dollars a year on ai you don't spend that and
2:56.660–2:59.300
give your best model away and if you want a personal
2:59.300–3:05.740
agent in the hands of three billion people you want to own the model doing the work so what did they
3:05.740–3:11.900
actually build and
3:07.793–3:09.847
is it worth your t
3:09.847–3:11.900
ime i started using it for the last couple of days and i also
3:11.900–3:18.460
built a little app on meta's model i handed it a video of anything that you probably want to sell and
3:18.460–3:23.340
asked it to give me a full price ready to post listing on facebook's marketplace so that's one
3:23.340–3:28.940
demo that we will see second i also ran it on clock course and
3:26.140–3:28.940
tropic so that will be an interesting one
3:28.940–3:34.060
and third i wanted to test the real multi-modality so
3:31.500–3:34.060
i gave it a picture of the ingredients in the
3:34.060–3:38.700
fridge and it was able to pinpoint the price tag on it right so
3:36.380–3:38.700
i'll show you all of the fun parts as
3:38.700–3:45.260
we go through but then
3:41.980–3:45.260
interesting part as i mentioned earlier is meta shift to the paid
3:45.260–3:50.540
strategy which is just keeping it open source all right before we dive deep into it a quick disclaimer
3:50.540–3:55.740
all opinions are my own and
3:53.140–3:55.740
do not belong to my employer all right with that let's get into it
3:56.860–4:02.300
all right so one of the fastest ways to experience spark is to actually come to meta.ai
4:02.300–4:07.340
and change the mode from instant to thinking mode so
4:04.820–4:07.340
if you ask any question like for example here i'm
4:07.340–4:12.700
asking which particular meta model are you using you would see that it is going to confirm that it is
4:12.700–4:19.180
using spark 1.1 so you just need to change the mode to thinking and
4:15.940–4:19.180
you will get to leverage the spark
4:19.180–4:25.180
1.1 model right so it was launched literally a couple of days back okay but
4:22.180–4:25.180
then from a developer
4:25.180–4:32.460
perspective if you want to use it you need to go into dev.
4:28.820–4:32.460
meta.ai so here i am in dev.meta.ai and i
4:32.460–4:39.100
was able to create an api key and
4:35.780–4:39.100
you can already see detailed docs and how you're able to use this
4:39.100–4:44.220
in plot code and these are all the details that they have provided codecs then
4:41.660–4:44.220
if you want to just call
4:44.220–4:49.260
it via python as well as curl right these are some of the details that have been given so what you could
4:49.260–4:54.380
do is you can create an api key so
4:51.820–4:54.380
what i did was as i mentioned earlier i'm going to be doing three
4:54.380–5:00.860
different demos so
4:56.540–4:58.700
the first demo her
4:58.700–5:00.860
e is think of it like a facebook marketplace listing generator so
5:00.860–5:06.300
here i would be using muse park so
5:03.580–5:06.300
what i'm going to do here is upload a video here so you can see the
5:06.300–5:13.020
video very clearly this is a video of a bike of a kid where i've just taken like a very short nine
5:13.020–5:18.460
seconds i just wanted to make it a little bit tough and
5:15.740–5:18.460
i'm going to ask it to generate listing right so
5:18.460–5:22.700
there are multiple things which are happening here right so
5:20.580–5:22.700
obviously it is calling the muse
5:22.700–5:28.380
spark 1.1 model so
5:24.593–5:26.487
the first thing wh
5:26.487–5:28.380
at's happening here is it is going to be identifying the item
5:28.380–5:34.220
so this is the response back from the first step so
5:31.300–5:34.220
it has correctly identified that this is a global
5:34.220–5:39.660
primo pink three wheel kit scooter and
5:36.940–5:39.660
obviously it is used so it has understood that as well so very
5:39.660–5:45.340
good job done very quickly you can see the amount of time it took was very short then the agent is
5:45.340–5:51.260
going to now do a pricing research so
5:48.300–5:51.260
it's going to look into some other prices so that it can give
5:51.260–5:57.340
us some comparison pricing right once you have that idea of the comparison pricing then it will
5:57.340–6:02.860
consolidate all of that and create the listing so
6:00.100–6:02.860
you can see that i'll search current resale listing
6:02.860–6:08.620
for your global this in the us market it is around this and
6:05.740–6:08.620
you can see all the pricing over here
6:08.620–6:13.900
in uk it is something of this sort right in canada it is this right and
6:11.260–6:13.900
because this is coming from
6:13.900–6:19.580
meta and facebook so and this is marketplace api so
6:16.740–6:19.580
you will be able to actually get the right
6:19.580–6:24.220
because they already have the data so
6:21.900–6:24.220
now it has already created the listing so you can clearly see
6:24.220–6:30.300
that global primo pink three wheel kit scooter hot pink and
6:27.260–6:30.300
black it also identified this the best for
6:30.300–6:35.900
scooter for toddlers and
6:32.167–6:34.033
kids this is condition i
6:34.033–6:35.900
s very good the retails for this but then here we
6:35.900–6:41.020
are charging it at this particular price right so
6:38.460–6:41.020
that's what it was able to do and you can see that
6:41.020–6:46.220
it did a pretty good job and you were able to then
6:43.620–6:46.220
copy this and just take it and paste it in in
6:46.220–6:52.860
marketplace right so
6:48.433–6:50.647
the reason i wanted
6:50.647–6:52.860
to do this was to show you how you're able to create something
6:52.860–6:58.140
like this very quickly using new spark and
6:55.500–6:58.140
it really demonstrates an agentic behavior understanding of
6:58.140–7:03.980
a multimodal input in this case video and also
7:01.060–7:03.980
doing a quick search and then not only just limited to the
7:03.980–7:09.340
us but also search across the board and then
7:06.660–7:09.340
providing you like a competitive pricing right so i was very
7:09.340–7:14.780
impressed with what i saw okay so
7:12.060–7:14.780
that's that's the first demo i hope you enjoyed it now what i want
7:14.780–7:20.940
to do here is i actually want to show you how you are also
7:17.860–7:20.940
able to use it directly in claw code so for
7:20.940–7:29.020
that let's just open clawed right so
7:24.980–7:29.020
in this case i've already configured clods connector so if i show you
7:29.020–7:36.140
this you can see that here i have got the new spark 1.
7:32.580–7:36.140
1 agent here right so the way i was able to do this
7:36.140–7:41.900
was i basically followed this specific instruction here i provided the api key which i already showed
7:41.900–7:48.460
you and i ran this in powershell first right so
7:45.180–7:48.460
before actually running clawed i basically ran this
7:48.460–7:55.100
right so once i have this now clawed is being forced to use this particular model which is new spark 1.1
7:55.100–8:00.860
right so you are welcome to try this and
7:57.980–8:00.860
see what kind of results you're getting all right so for the
8:00.860–8:06.220
third demo i decided to actually use one of meta's cookbooks which they have given and
8:03.540–8:06.220
they have really
8:06.220–8:12.300
done a great job and provided 10 actually 13 different use cases so
8:09.260–8:12.300
here in this one i actually
8:12.300–8:19.020
decided to use this perception grounding right so
8:15.660–8:19.020
the use case here is you have your fridge filled with
8:19.020–8:25.900
some food objects and can meta's multimodal model be able to get and
8:22.460–8:25.900
identify each one of the food
8:25.900–8:32.540
objects and provide some sort of a score right so
8:29.220–8:32.540
if i go in detail like this is the original image
8:32.540–8:37.580
you can see all the different food items over here and
8:35.060–8:37.580
once you basically run this particular program
8:37.580–8:44.540
and provide your api key it should be able to identify each one of these items and provide the
8:44.540–8:49.740
score right so what i did was i ran this particular prompt right i'm a pescetarian with high cholesterol
8:49.740–8:55.740
put green dots on recommended food and
8:52.740–8:55.740
now that it has run this particular output once i give this
8:55.740–9:02.780
particular command i will be able to see the output right you can see it is now generating the html overlay
9:02.780–9:08.300
so that it will be able to identify that right so
9:05.540–9:08.300
i want to also show you the output of what it
9:08.300–9:14.780
produces so again just for the reference this is how the original picture looks like right so if you look
9:14.780–9:20.460
at this is how the picture looks like now what we will be able to generate is something like this which
9:20.460–9:26.220
is after after it has generated the output right so
9:23.340–9:26.220
pescetarian plus high cholesterol fridge guide
9:26.220–9:31.500
recommended and not recommended and
9:28.860–9:31.500
you can see the scores pretty much well done so this is the output
9:31.500–9:37.020
and again thought like this was amazing because
9:34.260–9:37.020
you can see orange juice we all think that it is great but
9:37.020–9:43.420
for some reason it is giving not a great score so
9:40.220–9:43.420
let's see if i eat something which is not recommended
9:43.420–9:48.300
so cheese pack is not recommended yellow cheese is not recommended whereas
9:45.860–9:48.300
this butter is definitely
9:48.300–9:53.420
not recommended so pretty cool right so
9:50.860–9:53.420
you are able to use this out of the box and able to get this
9:53.420–9:59.900
label and it's fast it's cheaper as well so
9:56.660–9:59.900
this is what i wanted to cover the main idea is a make you
9:59.900–10:07.260
aware of this is a brand new agentic model out there so
10:03.580–10:07.260
definitely give it a try you can see there like
10:07.260–10:12.380
this is a good playground and a dashboard so
10:09.820–10:12.380
i've been playing with it it provides like the usage and stuff
10:12.380–10:18.540
like that very well etc all of those things and then
10:15.460–10:18.540
they have also given like 20 free so so far
10:18.540–10:23.180
i've ran it a few times and i only spent less than a dollar and
10:20.860–10:23.180
there's another way for you to try which
10:23.180–10:27.740
is playground so you can upload an image or ask certain questions you can also
10:25.460–10:27.740
change some of the
10:27.740–10:32.220
settings over here and then
10:29.233–10:30.727
there are some advanced set
10:30.727–10:32.220
tings as well right you can add some json schema
10:32.220–10:37.580
and stuff like that so what i would recommend is give it a shot also
10:34.900–10:37.580
try it in a cop in combination with
10:37.580–10:43.580
clock code the app that i built i asked anti-gravity to actually build the app but
10:40.580–10:43.580
leveraging the spark
10:43.580–10:48.380
api you could do like all of these types of combination where you use your id and build an
10:48.380–10:53.580
app like this and see for yourself what you feel as the performance right and
10:50.980–10:53.580
obviously on the benchmarks
10:53.580–10:59.180
they have talked about the benchmarks over here they've compared themselves against gemini 3.1 4.8
10:59.180–11:04.780
these are here for your reading and
11:01.980–11:04.780
i will share this as well of course does all different types of
11:04.780–11:11.100
use cases from an agentic perspective it also
11:07.940–11:11.100
does computer use it definitely writes code all of these
11:11.100–11:15.340
things are something which which they have actually explained right so
11:13.220–11:15.340
multimodal is something which we
11:15.340–11:21.340
saw live again hopefully this was helpful it added some extra knowledge to your existing knowledge base
11:21.340–11:25.420
let me know if you guys have any questions and
11:23.380–11:25.420
what do you feel after trying this thank you very much
11:25.420–11:29.340
for your time if you like the video please hit that like button and
11:27.380–11:29.340
if you're new here please hit that
11:29.340–11:34.000
subscribe button as well thank you for watching and

0:00.000–0:06.800
梅塔（Meta）剛剛發布了其最強大的 AI 模型，即 Muse Spark 1.1。
0:03.400–0:06.800
1.1，根據扎克伯格（Zuck）的說法，這是一款
0:06.800–0:12.860
能夠執行行動、觀看影片、使用工具，並
0:09.830–0:12.860
為你完成任務的模型，而不僅僅是回答
0:12.860–0:16.800
問題。稍後在影片中，我會談論所有這些細節。馬克·扎克伯格實際上
0:16.800–0:22.620
在 X 平台上宣布它的方式，我認為他在幾年後重新關注了這一點，看看
0:22.620–0:29.980
他首先強調的是強大的代理（agentic）和
0:26.300–0:29.980
編碼模型，價格非常低廉，通過我們新的梅塔
0:29.980–0:37.780
模型 API，一個低價格的付費 API。現在請從一家多年來提供免費
0:37.780–0:44.840
且開源 AI 的公司角度思考，這構成了它們整個身份，所以
0:41.310–0:44.840
在我深入探討 Spark 實際能做什麼之前，我認為
0:44.840–0:51.300
了解是什麼推動梅塔做出這樣的策略改變會很有趣，所以讓我們回到
0:51.300–0:56.700
歷史，倒帶回去，因為
0:54.000–0:56.700
背景故事解釋了一切。多年來，梅塔一直是開源 AI 的冠軍。
0:56.700–1:04.820
早在 2024 年 7 月，扎克發布了一篇標題為「開源 AI 是
1:04.820–1:10.880
前進之路」的文章。免費可下載的模型是梅塔的全部身份，
1:07.850–1:10.880
他將公司定位為
1:10.880–1:19.080
像 OpenAI 這樣的封閉實驗室
1:14.980–1:19.080
的開源替代方案。然後，變化迅速發生。梅塔引入了 Scale AI 的
1:19.080–1:26.340
亞歷山大·萬（Alexander Wang）擔任首席 AI 官，這是一筆 140 億美元的交易。他們建立了一個全新的超級
1:26.340–1:33.380
智能實驗室，並且
1:28.687–1:31.033
早在四月，Shippe
1:31.033–1:33.380
德（Shippe）發布了第一個 Muse Spark 模型，它是封閉的，沒有等待，沒有
1:33.380–1:39.720
下載，不可用。這確實是首次與開源脫鉤，對吧？
1:36.550–1:39.720
梅塔的邊界
1:39.720–1:46.200
模型被鎖定在它們自己的門後，運行在它們的應用程序內部，而不是
1:42.960–1:46.200
在公開場合。而這週
1:46.200–1:54.240
梅塔邁出了下一步，一個重大升級，Muse Spark 1.
1:50.220–1:54.240
1，並且有史以來第一次，公開的付費 API
1:54.240–2:02.600
帶有定價，扎克伯格稱之為「非常激進」，所以
1:58.420–2:02.600
梅塔不僅關閉了它們最好的模型，還給它貼上了價格
2:02.600–2:08.180
標籤，並向任何人開放，讓任何人來構建它。梅塔不僅在價格上壓倒所有人，
2:08.180–2:15.560
它大約只有 GPT 和 Claude 成本的三分之一。它們還
2:11.870–2:15.560
讓 Muse Spark 直接進入開發人員
2:15.560–2:22.400
已經使用的工具。OpenAI 設置確實如此，但也
2:18.980–2:22.400
包括 Clawed Code、Anthropic 自己的編碼工具。你指出
2:22.400–2:28.520
梅塔模型改變基本上只有一行，
2:25.460–2:28.520
現在你突然在 Muse Spark 上運行，價格只有四分之一，
2:28.520–2:35.960
梅塔在 OpenAI 和
2:32.240–2:35.960
Anthropic 之外建立了一條側門，並在它上面掛了一個折扣標誌
2:35.960–2:42.100
這不僅是一個新模型，更是直接針對其競爭對手客戶的搶攻。現在有兩件事要誠實說明：
2:42.100–2:47.860
LLama 還沒死，它仍然開放且仍可下載。但關鍵在於，最好的模型現在是封閉的。
2:47.860–2:54.020
第二，儘管這次轉變感覺很瘋狂，但這很合理，對吧？Matters 每年在 AI 上花費超過一千億美元，
2:54.020–2:59.300
你不會花這麼多錢，
2:56.660–2:59.300
卻把你最好的模型免費送人。如果你希望三十億人手中都有個人
2:59.300–3:05.740
助理，你就需要擁有執行這項工作的模型。那麼他們到底
3:05.740–3:11.900
建構了什麼，以及
3:07.793–3:09.847
它是否值得你花
3:09.847–3:11.900
時間？我過去幾天開始使用它，我也
3:11.900–3:18.460
在 Meta 的模型上建構了一個小應用程式。我給它一段影片，內容是你可能想出售的任何東西，並
3:18.460–3:23.340
要求它提供完整的價格，準備好發布在 Facebook 的 Marketplace 上。所以這是第一個
3:23.340–3:28.940
我們將看到的示範。第二，我也在 Clock Course 和
3:26.140–3:28.940
Tropic 上運行它，這將是一個有趣的示範。
3:28.940–3:34.060
第三，我想測試真正的多模態能力，所以
3:31.500–3:34.060
我給它一張冰箱裡食材的照片，
3:34.060–3:38.700
它能夠準確指出上面的價格標籤。所以
3:36.380–3:38.700
我會展示所有有趣的部分，
3:38.700–3:45.260
但接著
3:41.980–3:45.260
正如我 earlier 提到的，有趣的部分是 Meta 轉向付費
3:45.260–3:50.540
策略，也就是保持開源。好的，在深入探討之前，先做個快速免責聲明：
3:50.540–3:55.740
所有意見皆為我個人觀點，
3:53.140–3:55.740
不代表我的僱主。好的，那麼讓我們開始吧。
3:56.860–4:02.300
好的，體驗 Spark 最快的方法之一是實際前往 meta.ai，
4:02.300–4:07.340
並將模式從即時（instant）切換為思考模式（thinking mode）。所以
4:04.820–4:07.340
如果你問任何問題，例如這裡我
4:07.340–4:12.700
問你使用的是哪個特定的 Meta 模型，你會看到它確認正在
4:12.700–4:19.180
使用 Spark 1.1。所以你只需要將模式切換為思考，
4:15.940–4:19.180
你就能夠利用 Spark
4:19.180–4:25.180
1.1 模型。對吧？它就在幾天前剛推出，好吧，但
4:22.180–4:25.180
從開發者
4:25.180–4:32.460
的角度來看，如果你想使用它，你需要進入 dev.
4:28.820–4:32.460
meta.ai。所以這裡我在 dev.meta.ai，我
4:32.460–4:39.100
能夠建立一個 API 金鑰，
4:35.780–4:39.100
你已經可以看到詳細的文件，以及你如何能夠在
4:39.100–4:44.220
Plot Code 中使用它，這些都是他們提供的 Codec 詳細資訊。然後
4:41.660–4:44.220
如果你只想透過
4:44.220–4:49.260
Python 以及 curl 呼叫它，對吧？這些是一些提供的詳細資訊。所以你可以
4:49.260–4:54.380
做的是建立一個 API 金鑰，所以
4:51.820–4:54.380
我所做的是，正如我 earlier 提到的，我將進行三個
4:54.380–5:00.860
不同的示範，所以
4:56.540–4:58.700
第一個示範
4:58.700–5:00.860
這裡可以把它想像成一個 Facebook 市場（Marketplace）的 listings 生成器，所以
5:00.860–5:06.300
這裡我會使用 Muse Park，所以
5:03.580–5:06.300
我接下來要做的是上傳一段影片，這樣你可以清楚地看到
5:06.300–5:13.020
這段影片非常清晰，這是一段關於小孩腳踏車的影片，我只截取了短短九
5:13.020–5:18.460
秒，我只是想讓它稍微困難一點，然後
5:15.740–5:18.460
我會要求它生成 listings，對吧
5:18.460–5:22.700
這裡發生了多件事情，對吧
5:20.580–5:22.700
顯然它正在呼叫 Muse
5:22.700–5:28.380
Spark 1.1 模型，所以
5:24.593–5:26.487
這裡發生的第一件
5:26.487–5:28.380
事是它會識別物品
5:28.380–5:34.220
所以這是來自第一步的回應，所以
5:31.300–5:34.220
它正確地識別出這是 Global
5:34.220–5:39.660
Primo 粉紅色的三輪滑板車，而且
5:36.940–5:39.660
顯然它是二手的，它也理解了這一點，所以做得
5:39.660–5:45.340
很好，速度非常快，你可以看到所花時間非常短，然後代理程式
5:45.340–5:51.260
現在將進行定價研究，所以
5:48.300–5:51.260
它會查看其他一些價格，以便它能提供
5:51.260–5:57.340
我們一些比較定價，對吧，一旦你有了比較定價的概念，它就會
5:57.340–6:02.860
整合所有這些資訊並建立 listings，所以
6:00.100–6:02.860
你可以看到我搜尋目前的二手 listings
6:02.860–6:08.620
針對你的 Global，在美國市場大約是這個價格，而且
6:05.740–6:08.620
你可以看到這裡所有的定價
6:08.620–6:13.900
在英國大約是這種價格，對吧，在加拿大是這個價格，對吧，而且
6:11.260–6:13.900
因為這是來自
6:13.900–6:19.580
Meta 和 Facebook，所以這是 Marketplace API，所以
6:16.740–6:19.580
你實際上能夠獲得正確的
6:19.580–6:24.220
資訊，因為他們已經擁有數據，所以
6:21.900–6:24.220
現在它已經建立了 listings，你可以清楚地看到
6:24.220–6:30.300
Global Primo 粉紅色的三輪滑板車，亮粉紅色和
6:27.260–6:30.300
黑色，它也識別出這是最佳
6:30.300–6:35.900
的學步車，適合
6:32.167–6:34.033
幼兒和
6:34.033–6:35.900
兒童，這是狀況，
6:35.900–6:41.020
非常良好，零售價是這樣，但這裡我們
6:38.460–6:41.020
以這個特定價格收費，對吧，所以
6:41.020–6:46.220
這就是它能夠做到的，你可以看到
6:43.620–6:46.220
它做得相當不錯，然後你能夠
6:46.220–6:52.860
複製這個，直接把它貼到
6:48.433–6:50.647
我想要這樣做的理由
6:50.647–6:52.860
是想向您展示，您能夠多麼快速地使用 New Spark 建立類似這樣的內容
6:52.860–6:58.140
而且這確實展現了對多模態輸入（在此案例中為影片）的代理行為理解
6:55.500–6:58.140
以及進行快速搜尋，然後不僅限於美國
6:58.140–7:03.980
而是進行全面搜尋，接著為您提供競爭性定價，所以我对所見印象深刻
7:01.060–7:03.980
好的，這就是第一個示範，希望您喜歡。現在我想做的是，我實際上想向您展示，您也可以直接在 Claw Code 中使用它
7:03.980–7:09.340
為此，讓我們直接打開 Clawed，所以
7:06.660–7:09.340
在這種情況下，我已經配置了 Clods 連接器，所以如果我向您展示
7:09.340–7:14.780
這裡，您可以看到這裡有 New Spark 1.1 代理，對吧？所以我能夠這樣做的
7:12.060–7:14.780
方式是我基本上遵循了這裡的特定說明，我提供了 API 金鑰，這是我之前向您展示過的，並且我首先在 PowerShell 中運行了它，對吧？所以
7:14.780–7:20.940
在實際運行 Clawed 之前，我基本上運行了這個
7:17.860–7:20.940
指令，所以一旦我有了這個，現在 Clawed 就被強制使用這個特定模型，也就是 New Spark 1.1
7:20.940–7:29.020
對吧？所以歡迎您嘗試這個，
7:24.980–7:29.020
看看您會得到什麼樣的結果。好的，對於第三個示範，我決定實際使用 Meta 提供的一個 Cookbooks
7:29.020–7:36.140
他們真的
7:32.580–7:36.140
做得很好，並提供了 10 個，實際上 13 個不同的使用案例，所以
7:36.140–7:41.900
在這裡，我實際上
7:41.900–7:48.460
決定使用這個感知錨定，所以
7:45.180–7:48.460
這裡的使用案例是，您的冰箱裡裝滿了
7:48.460–7:55.100
一些食物物品，Meta 的多模態模型能否獲取並
7:55.100–8:00.860
識別每個食物
7:57.980–8:00.860
物品，並提供某種分數？對吧？所以如果我詳細查看，這是原始影像
8:00.860–8:06.220
您可以看到這裡有不同的食物項目，並且
8:03.540–8:06.220
一旦您基本上運行這個特定程式
8:06.220–8:12.300
並提供您的 API 金鑰，它應該能夠識別每個這些項目，並提供
8:09.260–8:12.300
分數，對吧？所以我做的是，我運行了這個特定的提示，對吧？我是一名高膽固醇的魚素主義者
8:12.300–8:19.020
在推薦的食物上標記綠點，並且
8:15.660–8:19.020
現在它運行了這個特定的輸出，一旦我給予這個
8:19.020–8:25.900
特定指令，我將能夠看到輸出，對吧？您可以看到它現在正在生成 HTML 疊加層
8:22.460–8:25.900
這樣它就能夠正確識別，所以
8:25.900–8:32.540
物件，並提供某種分數，對吧？所以
8:29.220–8:32.540
如果我詳細說明，這是原始影像
8:32.540–8:37.580
你可以看到這裡有各種不同的食物項目，以及
8:35.060–8:37.580
一旦你基本上執行這個特定的程式
8:37.580–8:44.540
並提供你的 API 金鑰，它應該能夠識別每一個這些項目，並提供
8:44.540–8:49.740
分數，對吧？所以我做的是，我執行了這個特定的提示詞，對吧？我是吃魚的素食者，且膽固醇偏高
8:49.740–8:55.740
在推薦的食物上標記綠點，並
8:52.740–8:55.740
現在它執行了這個特定的輸出，一旦我給予這個
8:55.740–9:02.780
特定的指令，我就能夠看到輸出結果，對吧？你可以看到它現在正在生成 HTML 覆蓋層
9:02.780–9:08.300
這樣它就能夠正確識別，對吧？所以
9:05.540–9:08.300
我也想展示它輸出的結果
9:08.300–9:14.780
所以再次作為參考，這是原始圖片的樣子，對吧？如果你看
9:14.780–9:20.460
這是圖片現在的樣子，而我們能夠生成的是像這樣的東西，
9:20.460–9:26.220
這是在它生成輸出之後的結果，對吧？所以
9:23.340–9:26.220
素食者加高膽固醇冰箱指南
9:26.220–9:31.500
推薦與不推薦項目，以及
9:28.860–9:31.500
你可以看到分數相當不錯，做得很好，所以這就是輸出結果
9:31.500–9:37.020
再次強調，我認為這非常驚人，因為
9:34.260–9:37.020
你可以看到柳橙汁，我們都認為它很好，但
9:37.020–9:43.420
出於某些原因，它給出的分數並不高，所以
9:40.220–9:43.420
讓我們看看如果我吃不推薦的食物會怎樣
9:43.420–9:48.300
所以奶酪包是不推薦的，黃色奶酪不推薦，而
9:45.860–9:48.300
這黃油絕對是
9:48.300–9:53.420
不推薦的，所以相當酷，對吧？所以
9:50.860–9:53.420
你能夠開箱即用，並獲得這個
9:53.420–9:59.900
標籤，而且它很快，也更便宜，所以
9:56.660–9:59.900
這就是我想要涵蓋的內容，主要觀念是要讓你
9:59.900–10:07.260
知道這是一個全新的代理型模型，所以
10:03.580–10:07.260
絕對要試試看，你可以看到這裡像
10:07.260–10:12.380
這樣是一個很好的遊樂場和儀表板，所以
10:09.820–10:12.380
我一直在玩它，它提供了使用情況等資訊
10:12.380–10:18.540
諸如此類，非常好等等，所有這些東西，然後
10:15.460–10:18.540
他們還提供了20個免費額度，所以到目前為止
10:18.540–10:23.180
我運行過幾次，花費不到一美元，
10:20.860–10:23.180
還有另一種嘗試方式是
10:23.180–10:27.740
遊樂場，你可以上傳圖片或提出特定問題，你也可以
10:25.460–10:27.740
更改一些
10:27.740–10:32.220
這裡的設定，然後
10:29.233–10:30.727
還有一些進階設
10:30.727–10:32.220
定，對吧？你可以添加一些 JSON 架構
10:32.220–10:37.580
之類的東西，所以我的建議是試試看，也
10:34.900–10:37.580
嘗試在組合中使用
10:37.580–10:43.580
Clock Code，我建構的應用程式，我要求 Anti-Gravity 實際建構該應用程式，但
10:40.580–10:43.580
利用 Spark
10:43.580–10:48.380
API，你可以進行所有這些類型的組合，使用你的 ID 並建構一個
10:48.380–10:53.580
這樣的應用程式，並親自看看你對效能的感受，對吧？而且
10:50.980–10:53.580
顯然在基準測試方面
10:53.580–10:59.180
他們在這裡談論了基準測試，他們將自己與 Gemini 3.1 4.8
10:59.180–11:04.780
這些是供你閱讀的，以及
11:01.980–11:04.780
我當然也會分享這個，它處理各種不同類型的
11:04.780–11:11.100
從代理人的角度來看，它也能應用於各種用例，而且
11:07.940–11:11.100
它也能操作電腦，確實能編寫程式碼，這些
11:11.100–11:15.340
功能都是他們實際說明過的，對吧？所以
11:13.220–11:15.340
多模態是我們
11:15.340–11:21.340
剛才現場看到的，希望這有幫助，它為你們既有的知識庫增添了額外知識
11:21.340–11:25.420
如果大家有任何問題，請告訴我，並且
11:23.380–11:25.420
嘗試之後你們有什麼感想？非常感謝
11:25.420–11:29.340
撥冗觀看，如果喜歡這部影片，請按讚，並且
11:27.380–11:29.340
如果你是新觀眾，也請點擊
11:29.340–11:34.000
訂閱按鈕，感謝觀看，並且

0:00.000–0:06.800
meta just released its most capable ai model which is muse spark 1.
梅塔（Meta）剛剛發布了其最強大的 AI 模型，即 Muse Spark 1.1。
0:03.400–0:06.800
1 and according to zag this is a
1.1，根據扎克伯格（Zuck）的說法，這是一款
0:06.800–0:12.860
model that can act it can watch a video use tools and
能夠執行行動、觀看影片、使用工具，並
0:09.830–0:12.860
get it as done for you not just answer
為你完成任務的模型，而不僅僅是回答
0:12.860–0:16.800
questions and later in the video i'm going to talk about all of that mark zuckerberg actually
問題。稍後在影片中，我會談論所有這些細節。馬克·扎克伯格實際上
0:16.800–0:22.620
the way he announced it on the x platform i think he came back on it after a few years and look at
在 X 平台上宣布它的方式，我認為他在幾年後重新關注了這一點，看看
0:22.620–0:29.980
what he leads with a strong agentic and
他首先強調的是強大的代理（agentic）和
0:26.300–0:29.980
coding model at a very low price through our new meta
編碼模型，價格非常低廉，通過我們新的梅塔
0:29.980–0:37.780
model api a low price a paid api now think about this from a company that for years made free
模型 API，一個低價格的付費 API。現在請從一家多年來提供免費
0:37.780–0:44.840
and open source ai its entire identity so
且開源 AI 的公司角度思考，這構成了它們整個身份，所以
0:41.310–0:44.840
before i get into what spark can actually do i think it'll
在我深入探討 Spark 實際能做什麼之前，我認為
0:44.840–0:51.300
be interesting to understand what pushed meta to change the strategy like this so let's go back
了解是什麼推動梅塔做出這樣的策略改變會很有趣，所以讓我們回到
0:51.300–0:56.700
to history and rewind right because
歷史，倒帶回去，因為
0:54.000–0:56.700
the backstory explains everything for years meta was the champion
背景故事解釋了一切。多年來，梅塔一直是開源 AI 的冠軍。
0:56.700–1:04.820
of open source ai back in july 2024 zuck published an essay titled literally open source ai is the
早在 2024 年 7 月，扎克發布了一篇標題為「開源 AI 是
1:04.820–1:10.880
path forward free downloadable models were meta's whole identity and
前進之路」的文章。免費可下載的模型是梅塔的全部身份，
1:07.850–1:10.880
he positioned the company as
他將公司定位為
1:10.880–1:19.080
the open alternator to closed labs like open ai then
像 OpenAI 這樣的封閉實驗室
1:14.980–1:19.080
it changed fast meta brought in scale ai's
的開源替代方案。然後，變化迅速發生。梅塔引入了 Scale AI 的
1:19.080–1:26.340
alexander van as their chief ai officer a 14 billion dollar deal they stood up a brand new super
亞歷山大·萬（Alexander Wang）擔任首席 AI 官，這是一筆 140 億美元的交易。他們建立了一個全新的超級
1:26.340–1:33.380
intelligence lab and
智能實驗室，並且
1:28.687–1:31.033
back in april shippe
早在四月，Shippe
1:31.033–1:33.380
d the first muke spark model which was closed no waits no
德（Shippe）發布了第一個 Muse Spark 模型，它是封閉的，沒有等待，沒有
1:33.380–1:39.720
download available that was a real break from open source for the first time right and
下載，不可用。這確實是首次與開源脫鉤，對吧？
1:36.550–1:39.720
meta's frontier
梅塔的邊界
1:39.720–1:46.200
model were locked behind its own doors running inside its apps instead
模型被鎖定在它們自己的門後，運行在它們的應用程序內部，而不是
1:42.960–1:46.200
of out in the open and this week
在公開場合。而這週
1:46.200–1:54.240
meta took the next step a big upgrade muse spark 1.
梅塔邁出了下一步，一個重大升級，Muse Spark 1.
1:50.220–1:54.240
1 and for the first time ever a public paid api
1，並且有史以來第一次，公開的付費 API
1:54.240–2:02.600
with pricing zuckerberg called very aggressive so
帶有定價，扎克伯格稱之為「非常激進」，所以
1:58.420–2:02.600
meta didn't just close its best model it put a price
梅塔不僅關閉了它們最好的模型，還給它貼上了價格
2:02.600–2:08.180
tag on it and open it up for anyone to build it meta didn't just undercut everyone on price
標籤，並向任何人開放，讓任何人來構建它。梅塔不僅在價格上壓倒所有人，
2:08.180–2:15.560
it's about a quarter of what gpt and claude cost they also
它大約只有 GPT 和 Claude 成本的三分之一。它們還
2:11.870–2:15.560
made mu spark drop straight into the tools
讓 Muse Spark 直接進入開發人員
2:15.560–2:22.400
developers already use open ai setup sure but also
已經使用的工具。OpenAI 設置確實如此，但也
2:18.980–2:22.400
clawed code anthropic's own coding tool you pointed
包括 Clawed Code、Anthropic 自己的編碼工具。你指出
2:22.400–2:28.520
at meta's model change basically one line and
梅塔模型改變基本上只有一行，
2:25.460–2:28.520
now you're suddenly running on muse spark at a quarter of
現在你突然在 Muse Spark 上運行，價格只有四分之一，
2:28.520–2:35.960
the price meta built a side door out of open ai and
梅塔在 OpenAI 和
2:32.240–2:35.960
anthropic and hung a discount sign on it this
Anthropic 之外建立了一條側門，並在它上面掛了一個折扣標誌
2:35.960–2:42.100
isn't just a new model it's a direct play for their rivals customers now two things to keep honest one
這不僅是一個新模型，更是直接針對其競爭對手客戶的搶攻。現在有兩件事要誠實說明：
2:42.100–2:47.860
llama isn't dead it's still open still downloadable what changes matters best models are closed now
LLama 還沒死，它仍然開放且仍可下載。但關鍵在於，最好的模型現在是封閉的。
2:47.860–2:54.020
and two as wild as the flip feels it makes sense right matters spending north of a hundred billion
第二，儘管這次轉變感覺很瘋狂，但這很合理，對吧？Matters 每年在 AI 上花費超過一千億美元，
2:54.020–2:59.300
dollars a year on ai you don't spend that and
你不會花這麼多錢，
2:56.660–2:59.300
give your best model away and if you want a personal
卻把你最好的模型免費送人。如果你希望三十億人手中都有個人
2:59.300–3:05.740
agent in the hands of three billion people you want to own the model doing the work so what did they
助理，你就需要擁有執行這項工作的模型。那麼他們到底
3:05.740–3:11.900
actually build and
建構了什麼，以及
3:07.793–3:09.847
is it worth your t
它是否值得你花
3:09.847–3:11.900
ime i started using it for the last couple of days and i also
時間？我過去幾天開始使用它，我也
3:11.900–3:18.460
built a little app on meta's model i handed it a video of anything that you probably want to sell and
在 Meta 的模型上建構了一個小應用程式。我給它一段影片，內容是你可能想出售的任何東西，並
3:18.460–3:23.340
asked it to give me a full price ready to post listing on facebook's marketplace so that's one
要求它提供完整的價格，準備好發布在 Facebook 的 Marketplace 上。所以這是第一個
3:23.340–3:28.940
demo that we will see second i also ran it on clock course and
我們將看到的示範。第二，我也在 Clock Course 和
3:26.140–3:28.940
tropic so that will be an interesting one
Tropic 上運行它，這將是一個有趣的示範。
3:28.940–3:34.060
and third i wanted to test the real multi-modality so
第三，我想測試真正的多模態能力，所以
3:31.500–3:34.060
i gave it a picture of the ingredients in the
我給它一張冰箱裡食材的照片，
3:34.060–3:38.700
fridge and it was able to pinpoint the price tag on it right so
它能夠準確指出上面的價格標籤。所以
3:36.380–3:38.700
i'll show you all of the fun parts as
我會展示所有有趣的部分，
3:38.700–3:45.260
we go through but then
但接著
3:41.980–3:45.260
interesting part as i mentioned earlier is meta shift to the paid
正如我 earlier 提到的，有趣的部分是 Meta 轉向付費
3:45.260–3:50.540
strategy which is just keeping it open source all right before we dive deep into it a quick disclaimer
策略，也就是保持開源。好的，在深入探討之前，先做個快速免責聲明：
3:50.540–3:55.740
all opinions are my own and
所有意見皆為我個人觀點，
3:53.140–3:55.740
do not belong to my employer all right with that let's get into it
不代表我的僱主。好的，那麼讓我們開始吧。
3:56.860–4:02.300
all right so one of the fastest ways to experience spark is to actually come to meta.ai
好的，體驗 Spark 最快的方法之一是實際前往 meta.ai，
4:02.300–4:07.340
and change the mode from instant to thinking mode so
並將模式從即時（instant）切換為思考模式（thinking mode）。所以
4:04.820–4:07.340
if you ask any question like for example here i'm
如果你問任何問題，例如這裡我
4:07.340–4:12.700
asking which particular meta model are you using you would see that it is going to confirm that it is
問你使用的是哪個特定的 Meta 模型，你會看到它確認正在
4:12.700–4:19.180
using spark 1.1 so you just need to change the mode to thinking and
使用 Spark 1.1。所以你只需要將模式切換為思考，
4:15.940–4:19.180
you will get to leverage the spark
你就能夠利用 Spark
4:19.180–4:25.180
1.1 model right so it was launched literally a couple of days back okay but
1.1 模型。對吧？它就在幾天前剛推出，好吧，但
4:22.180–4:25.180
then from a developer
從開發者
4:25.180–4:32.460
perspective if you want to use it you need to go into dev.
的角度來看，如果你想使用它，你需要進入 dev.
4:28.820–4:32.460
meta.ai so here i am in dev.meta.ai and i
meta.ai。所以這裡我在 dev.meta.ai，我
4:32.460–4:39.100
was able to create an api key and
能夠建立一個 API 金鑰，
4:35.780–4:39.100
you can already see detailed docs and how you're able to use this
你已經可以看到詳細的文件，以及你如何能夠在
4:39.100–4:44.220
in plot code and these are all the details that they have provided codecs then
Plot Code 中使用它，這些都是他們提供的 Codec 詳細資訊。然後
4:41.660–4:44.220
if you want to just call
如果你只想透過
4:44.220–4:49.260
it via python as well as curl right these are some of the details that have been given so what you could
Python 以及 curl 呼叫它，對吧？這些是一些提供的詳細資訊。所以你可以
4:49.260–4:54.380
do is you can create an api key so
做的是建立一個 API 金鑰，所以
4:51.820–4:54.380
what i did was as i mentioned earlier i'm going to be doing three
我所做的是，正如我 earlier 提到的，我將進行三個
4:54.380–5:00.860
different demos so
不同的示範，所以
4:56.540–4:58.700
the first demo her
第一個示範
4:58.700–5:00.860
e is think of it like a facebook marketplace listing generator so
這裡可以把它想像成一個 Facebook 市場（Marketplace）的 listings 生成器，所以
5:00.860–5:06.300
here i would be using muse park so
這裡我會使用 Muse Park，所以
5:03.580–5:06.300
what i'm going to do here is upload a video here so you can see the
我接下來要做的是上傳一段影片，這樣你可以清楚地看到
5:06.300–5:13.020
video very clearly this is a video of a bike of a kid where i've just taken like a very short nine
這段影片非常清晰，這是一段關於小孩腳踏車的影片，我只截取了短短九
5:13.020–5:18.460
seconds i just wanted to make it a little bit tough and
秒，我只是想讓它稍微困難一點，然後
5:15.740–5:18.460
i'm going to ask it to generate listing right so
我會要求它生成 listings，對吧
5:18.460–5:22.700
there are multiple things which are happening here right so
這裡發生了多件事情，對吧
5:20.580–5:22.700
obviously it is calling the muse
顯然它正在呼叫 Muse
5:22.700–5:28.380
spark 1.1 model so
Spark 1.1 模型，所以
5:24.593–5:26.487
the first thing wh
這裡發生的第一件
5:26.487–5:28.380
at's happening here is it is going to be identifying the item
事是它會識別物品
5:28.380–5:34.220
so this is the response back from the first step so
所以這是來自第一步的回應，所以
5:31.300–5:34.220
it has correctly identified that this is a global
它正確地識別出這是 Global
5:34.220–5:39.660
primo pink three wheel kit scooter and
Primo 粉紅色的三輪滑板車，而且
5:36.940–5:39.660
obviously it is used so it has understood that as well so very
顯然它是二手的，它也理解了這一點，所以做得
5:39.660–5:45.340
good job done very quickly you can see the amount of time it took was very short then the agent is
很好，速度非常快，你可以看到所花時間非常短，然後代理程式
5:45.340–5:51.260
going to now do a pricing research so
現在將進行定價研究，所以
5:48.300–5:51.260
it's going to look into some other prices so that it can give
它會查看其他一些價格，以便它能提供
5:51.260–5:57.340
us some comparison pricing right once you have that idea of the comparison pricing then it will
我們一些比較定價，對吧，一旦你有了比較定價的概念，它就會
5:57.340–6:02.860
consolidate all of that and create the listing so
整合所有這些資訊並建立 listings，所以
6:00.100–6:02.860
you can see that i'll search current resale listing
你可以看到我搜尋目前的二手 listings
6:02.860–6:08.620
for your global this in the us market it is around this and
針對你的 Global，在美國市場大約是這個價格，而且
6:05.740–6:08.620
you can see all the pricing over here
你可以看到這裡所有的定價
6:08.620–6:13.900
in uk it is something of this sort right in canada it is this right and
在英國大約是這種價格，對吧，在加拿大是這個價格，對吧，而且
6:11.260–6:13.900
because this is coming from
因為這是來自
6:13.900–6:19.580
meta and facebook so and this is marketplace api so
Meta 和 Facebook，所以這是 Marketplace API，所以
6:16.740–6:19.580
you will be able to actually get the right
你實際上能夠獲得正確的
6:19.580–6:24.220
because they already have the data so
資訊，因為他們已經擁有數據，所以
6:21.900–6:24.220
now it has already created the listing so you can clearly see
現在它已經建立了 listings，你可以清楚地看到
6:24.220–6:30.300
that global primo pink three wheel kit scooter hot pink and
Global Primo 粉紅色的三輪滑板車，亮粉紅色和
6:27.260–6:30.300
black it also identified this the best for
黑色，它也識別出這是最佳
6:30.300–6:35.900
scooter for toddlers and
的學步車，適合
6:32.167–6:34.033
kids this is condition i
幼兒和
6:34.033–6:35.900
s very good the retails for this but then here we
兒童，這是狀況，
6:35.900–6:41.020
are charging it at this particular price right so
非常良好，零售價是這樣，但這裡我們
6:38.460–6:41.020
that's what it was able to do and you can see that
以這個特定價格收費，對吧，所以
6:41.020–6:46.220
it did a pretty good job and you were able to then
這就是它能夠做到的，你可以看到
6:43.620–6:46.220
copy this and just take it and paste it in in
它做得相當不錯，然後你能夠
6:46.220–6:52.860
marketplace right so
複製這個，直接把它貼到
6:48.433–6:50.647
the reason i wanted
我想要這樣做的理由
6:50.647–6:52.860
to do this was to show you how you're able to create something
是想向您展示，您能夠多麼快速地使用 New Spark 建立類似這樣的內容
6:52.860–6:58.140
like this very quickly using new spark and
而且這確實展現了對多模態輸入（在此案例中為影片）的代理行為理解
6:55.500–6:58.140
it really demonstrates an agentic behavior understanding of
以及進行快速搜尋，然後不僅限於美國
6:58.140–7:03.980
a multimodal input in this case video and also
而是進行全面搜尋，接著為您提供競爭性定價，所以我对所見印象深刻
7:01.060–7:03.980
doing a quick search and then not only just limited to the
好的，這就是第一個示範，希望您喜歡。現在我想做的是，我實際上想向您展示，您也可以直接在 Claw Code 中使用它
7:03.980–7:09.340
us but also search across the board and then
為此，讓我們直接打開 Clawed，所以
7:06.660–7:09.340
providing you like a competitive pricing right so i was very
在這種情況下，我已經配置了 Clods 連接器，所以如果我向您展示
7:09.340–7:14.780
impressed with what i saw okay so
這裡，您可以看到這裡有 New Spark 1.1 代理，對吧？所以我能夠這樣做的
7:12.060–7:14.780
that's that's the first demo i hope you enjoyed it now what i want
方式是我基本上遵循了這裡的特定說明，我提供了 API 金鑰，這是我之前向您展示過的，並且我首先在 PowerShell 中運行了它，對吧？所以
7:14.780–7:20.940
to do here is i actually want to show you how you are also
在實際運行 Clawed 之前，我基本上運行了這個
7:17.860–7:20.940
able to use it directly in claw code so for
指令，所以一旦我有了這個，現在 Clawed 就被強制使用這個特定模型，也就是 New Spark 1.1
7:20.940–7:29.020
that let's just open clawed right so
對吧？所以歡迎您嘗試這個，
7:24.980–7:29.020
in this case i've already configured clods connector so if i show you
看看您會得到什麼樣的結果。好的，對於第三個示範，我決定實際使用 Meta 提供的一個 Cookbooks
7:29.020–7:36.140
this you can see that here i have got the new spark 1.
他們真的
7:32.580–7:36.140
1 agent here right so the way i was able to do this
做得很好，並提供了 10 個，實際上 13 個不同的使用案例，所以
7:36.140–7:41.900
was i basically followed this specific instruction here i provided the api key which i already showed
在這裡，我實際上
7:41.900–7:48.460
you and i ran this in powershell first right so
決定使用這個感知錨定，所以
7:45.180–7:48.460
before actually running clawed i basically ran this
這裡的使用案例是，您的冰箱裡裝滿了
7:48.460–7:55.100
right so once i have this now clawed is being forced to use this particular model which is new spark 1.1
一些食物物品，Meta 的多模態模型能否獲取並
7:55.100–8:00.860
right so you are welcome to try this and
識別每個食物
7:57.980–8:00.860
see what kind of results you're getting all right so for the
物品，並提供某種分數？對吧？所以如果我詳細查看，這是原始影像
8:00.860–8:06.220
third demo i decided to actually use one of meta's cookbooks which they have given and
您可以看到這裡有不同的食物項目，並且
8:03.540–8:06.220
they have really
一旦您基本上運行這個特定程式
8:06.220–8:12.300
done a great job and provided 10 actually 13 different use cases so
並提供您的 API 金鑰，它應該能夠識別每個這些項目，並提供
8:09.260–8:12.300
here in this one i actually
分數，對吧？所以我做的是，我運行了這個特定的提示，對吧？我是一名高膽固醇的魚素主義者
8:12.300–8:19.020
decided to use this perception grounding right so
在推薦的食物上標記綠點，並且
8:15.660–8:19.020
the use case here is you have your fridge filled with
現在它運行了這個特定的輸出，一旦我給予這個
8:19.020–8:25.900
some food objects and can meta's multimodal model be able to get and
特定指令，我將能夠看到輸出，對吧？您可以看到它現在正在生成 HTML 疊加層
8:22.460–8:25.900
identify each one of the food
這樣它就能夠正確識別，所以
8:25.900–8:32.540
objects and provide some sort of a score right so
物件，並提供某種分數，對吧？所以
8:29.220–8:32.540
if i go in detail like this is the original image
如果我詳細說明，這是原始影像
8:32.540–8:37.580
you can see all the different food items over here and
你可以看到這裡有各種不同的食物項目，以及
8:35.060–8:37.580
once you basically run this particular program
一旦你基本上執行這個特定的程式
8:37.580–8:44.540
and provide your api key it should be able to identify each one of these items and provide the
並提供你的 API 金鑰，它應該能夠識別每一個這些項目，並提供
8:44.540–8:49.740
score right so what i did was i ran this particular prompt right i'm a pescetarian with high cholesterol
分數，對吧？所以我做的是，我執行了這個特定的提示詞，對吧？我是吃魚的素食者，且膽固醇偏高
8:49.740–8:55.740
put green dots on recommended food and
在推薦的食物上標記綠點，並
8:52.740–8:55.740
now that it has run this particular output once i give this
現在它執行了這個特定的輸出，一旦我給予這個
8:55.740–9:02.780
particular command i will be able to see the output right you can see it is now generating the html overlay
特定的指令，我就能夠看到輸出結果，對吧？你可以看到它現在正在生成 HTML 覆蓋層
9:02.780–9:08.300
so that it will be able to identify that right so
這樣它就能夠正確識別，對吧？所以
9:05.540–9:08.300
i want to also show you the output of what it
我也想展示它輸出的結果
9:08.300–9:14.780
produces so again just for the reference this is how the original picture looks like right so if you look
所以再次作為參考，這是原始圖片的樣子，對吧？如果你看
9:14.780–9:20.460
at this is how the picture looks like now what we will be able to generate is something like this which
這是圖片現在的樣子，而我們能夠生成的是像這樣的東西，
9:20.460–9:26.220
is after after it has generated the output right so
這是在它生成輸出之後的結果，對吧？所以
9:23.340–9:26.220
pescetarian plus high cholesterol fridge guide
素食者加高膽固醇冰箱指南
9:26.220–9:31.500
recommended and not recommended and
推薦與不推薦項目，以及
9:28.860–9:31.500
you can see the scores pretty much well done so this is the output
你可以看到分數相當不錯，做得很好，所以這就是輸出結果
9:31.500–9:37.020
and again thought like this was amazing because
再次強調，我認為這非常驚人，因為
9:34.260–9:37.020
you can see orange juice we all think that it is great but
你可以看到柳橙汁，我們都認為它很好，但
9:37.020–9:43.420
for some reason it is giving not a great score so
出於某些原因，它給出的分數並不高，所以
9:40.220–9:43.420
let's see if i eat something which is not recommended
讓我們看看如果我吃不推薦的食物會怎樣
9:43.420–9:48.300
so cheese pack is not recommended yellow cheese is not recommended whereas
所以奶酪包是不推薦的，黃色奶酪不推薦，而
9:45.860–9:48.300
this butter is definitely
這黃油絕對是
9:48.300–9:53.420
not recommended so pretty cool right so
不推薦的，所以相當酷，對吧？所以
9:50.860–9:53.420
you are able to use this out of the box and able to get this
你能夠開箱即用，並獲得這個
9:53.420–9:59.900
label and it's fast it's cheaper as well so
標籤，而且它很快，也更便宜，所以
9:56.660–9:59.900
this is what i wanted to cover the main idea is a make you
這就是我想要涵蓋的內容，主要觀念是要讓你
9:59.900–10:07.260
aware of this is a brand new agentic model out there so
知道這是一個全新的代理型模型，所以
10:03.580–10:07.260
definitely give it a try you can see there like
絕對要試試看，你可以看到這裡像
10:07.260–10:12.380
this is a good playground and a dashboard so
這樣是一個很好的遊樂場和儀表板，所以
10:09.820–10:12.380
i've been playing with it it provides like the usage and stuff
我一直在玩它，它提供了使用情況等資訊
10:12.380–10:18.540
like that very well etc all of those things and then
諸如此類，非常好等等，所有這些東西，然後
10:15.460–10:18.540
they have also given like 20 free so so far
他們還提供了20個免費額度，所以到目前為止
10:18.540–10:23.180
i've ran it a few times and i only spent less than a dollar and
我運行過幾次，花費不到一美元，
10:20.860–10:23.180
there's another way for you to try which
還有另一種嘗試方式是
10:23.180–10:27.740
is playground so you can upload an image or ask certain questions you can also
遊樂場，你可以上傳圖片或提出特定問題，你也可以
10:25.460–10:27.740
change some of the
更改一些
10:27.740–10:32.220
settings over here and then
這裡的設定，然後
10:29.233–10:30.727
there are some advanced set
還有一些進階設
10:30.727–10:32.220
tings as well right you can add some json schema
定，對吧？你可以添加一些 JSON 架構
10:32.220–10:37.580
and stuff like that so what i would recommend is give it a shot also
之類的東西，所以我的建議是試試看，也
10:34.900–10:37.580
try it in a cop in combination with
嘗試在組合中使用
10:37.580–10:43.580
clock code the app that i built i asked anti-gravity to actually build the app but
Clock Code，我建構的應用程式，我要求 Anti-Gravity 實際建構該應用程式，但
10:40.580–10:43.580
leveraging the spark
利用 Spark
10:43.580–10:48.380
api you could do like all of these types of combination where you use your id and build an
API，你可以進行所有這些類型的組合，使用你的 ID 並建構一個
10:48.380–10:53.580
app like this and see for yourself what you feel as the performance right and
這樣的應用程式，並親自看看你對效能的感受，對吧？而且
10:50.980–10:53.580
obviously on the benchmarks
顯然在基準測試方面
10:53.580–10:59.180
they have talked about the benchmarks over here they've compared themselves against gemini 3.1 4.8
他們在這裡談論了基準測試，他們將自己與 Gemini 3.1 4.8
10:59.180–11:04.780
these are here for your reading and
這些是供你閱讀的，以及
11:01.980–11:04.780
i will share this as well of course does all different types of
我當然也會分享這個，它處理各種不同類型的
11:04.780–11:11.100
use cases from an agentic perspective it also
從代理人的角度來看，它也能應用於各種用例，而且
11:07.940–11:11.100
does computer use it definitely writes code all of these
它也能操作電腦，確實能編寫程式碼，這些
11:11.100–11:15.340
things are something which which they have actually explained right so
功能都是他們實際說明過的，對吧？所以
11:13.220–11:15.340
multimodal is something which we
多模態是我們
11:15.340–11:21.340
saw live again hopefully this was helpful it added some extra knowledge to your existing knowledge base
剛才現場看到的，希望這有幫助，它為你們既有的知識庫增添了額外知識
11:21.340–11:25.420
let me know if you guys have any questions and
如果大家有任何問題，請告訴我，並且
11:23.380–11:25.420
what do you feel after trying this thank you very much
嘗試之後你們有什麼感想？非常感謝
11:25.420–11:29.340
for your time if you like the video please hit that like button and
撥冗觀看，如果喜歡這部影片，請按讚，並且
11:27.380–11:29.340
if you're new here please hit that
如果你是新觀眾，也請點擊
11:29.340–11:34.000
subscribe button as well thank you for watching and
訂閱按鈕，感謝觀看，並且

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

一句話總結

Meta 發布了首款頂級閉源付費模型 Muse Spark 1.1，標誌著其從開源倡導者轉向商業化策略，該模型具備強大的 Agentic（代理）能力與多模態感知錨定功能，並透過相容現有開發工具（如 OpenAI 設定）以低於競爭對手四分之一的價格提供 API 服務。

核心重點

策略重大轉變：Meta 從堅持開源（Llama）轉向推出頂級閉源模型 Muse Spark 1.1，並首次提供公開付費 API。
定價策略：扎克伯格稱其定價為「極具侵略性」，約為 GPT 和 Claude 價格的四分之一。
Agentic 能力：模型不僅是回答問題，還能「行動」（act）、觀看影片、使用工具並完成任務。
開發者整合：透過相容現有工具（如 OpenAI 設定、Claude Code），建立「側門」讓開發者以低成本切換模型。
實測演示：

Facebook Marketplace 列表生成：透過影片識別物品、進行價格研究並生成上架列表。
Claude Code 整合：強制 Claude Code 使用 Spark 1.1 進行編碼。
多模態感知錨定：針對冰箱食材圖片，根據用戶健康狀況生成推薦與不推薦的 HTML 疊加層。

詳細大綱

I. Meta 策略轉變的背景與歷史

過去：開源倡導者
2024年7月，扎克伯格發布文章《開源 AI 是前進的道路》。
免費可下載模型曾是 Meta 的核心身份，定位為開放實驗室（Open Labs）對抗封閉實驗室（如 OpenAI）。
轉變：策略翻轉
引入 Scale AI 的 Alexander van 擔任首席 AI 官。
達成 140 億美元的交易。
建立全新的 Super Intelligence Lab（超級智能實驗室）。
4月發布首批 Muse Spark 模型（閉源、無下載、無等待）。
頂級模型被鎖定在 Meta 自家應用內運行。
現在：付費與開放 API
發布 Muse Spark 1.1 重大升級。
首次提供公開付費 API 及定價。
價格約為 GPT 和 Claude 的四分之一。
直接競爭對手為 OpenAI 和 Anthropic。

II. 策略轉變的商業邏輯

Llama 並未死亡：Llama 仍為開源且可下載，但「最佳模型」現在是閉源的。
成本與控制：
Meta 每年在 AI 上花費超過 1000 億美元。
不願將最佳模型免費贈送。
若希望擁有服務 30 億人的個人代理，必須擁有執行工作的模型。

III. 實測演示與功能驗證

演示一：Facebook Marketplace 列表生成器（Agentic 行為）
輸入：一段 9 秒的影片（兒童三輪滑板車）。
過程：

識別物品：Global Primo Pink 三輪滑板車，狀態為二手。
價格研究：搜尋美國、英國、加拿大市場的當前轉售價格。
整合資訊：利用 Marketplace API 數據生成完整上架列表。

結果：快速生成包含價格、描述、條件的上架內容，展示多模態輸入理解與代理行為。
演示二：Claude Code 整合
方法：在 dev.meta.ai 創建 API Key。
操作：在 PowerShell 中運行指令，配置 Claude Code 連接器。
結果：強制 Claude Code 使用 Muse Spark 1.1 模型進行編碼工作。
演示三：多模態感知錨定（Perception Grounding）
來源：Meta 提供的 Cookbooks 中的用例（13種用例之一）。
輸入：冰箱內食物圖片。
提示詞：「我是素食主義者且患有高膽固醇，在推薦食物上標記綠點。」
過程：識別食物項目並評分。
結果：生成 HTML 疊加層，標示推薦（綠色）與不推薦（如橙汁、奶酪、黃油）項目及評分。

IV. 開發者資源與使用方式

Meta.ai：
將模式從「Instant」切換為「Thinking Mode」即可體驗 Spark 1.1。
開發者入口：
dev.meta.ai：創建 API Key，查看詳細文檔（Plot Code, Python, Curl 呼叫範例）。
Playground：上傳圖片、提問、調整進階設定（如 JSON Schema）。
Dashboard：查看使用情況與儀表板。
成本與贈額：
提供 20 美元免費額度。
講者測試花費少於 1 美元。
基準測試（Benchmarks）：
與 Gemini 3.1 4.8 等進行比較（講者表示會分享相關資料）。

工具 / 模型 / 名詞整理

模型名稱：
Muse Spark 1.1（逐字稿中亦出現 "muse spark"、"muke spark"、"new spark"）
Llama（提及仍為開源）
GPT（提及價格對比）
Claude（提及價格對比）
Gemini 3.1 4.8（提及於基準測試中）
平台與服務：
Meta.ai
dev.meta.ai
Facebook's Marketplace / Marketplace API
X platform（扎克伯格發布平台）
Open AI（提及設定相容性）
Anthropic（提及競爭對手）
開發工具與環境：
Claude Code（逐字稿中亦出現 "claw code"、"clods connector"）
Plot Code（逐字稿中出現，疑為特定工具或口誤）
Python
Curl
PowerShell
HTML overlay（生成結果格式）
JSON Schema（Playground 進階設定）
其他專有名詞：
Scale AI
Alexander van（Scale AI 首席 AI 官）
Super Intelligence Lab（超級智能實驗室）
Agentic / Agentic behavior（代理行為）
Multimodal input（多模態輸入）
Perception grounding（感知錨定）
Cookbooks（Meta 提供的用例手冊）

操作流程整理

體驗 Spark 1.1：

進入 Meta.ai。
將模式從「Instant」切換為「Thinking Mode」。

開發者 API 整合：

前往 dev.meta.ai 創建 API Key。
查看詳細文檔（包含 Plot Code, Python, Curl 呼叫範例）。
在開發環境（如 PowerShell）中運行指令，配置連接器（如 Claude Code）。
強制指定使用 Muse Spark 1.1 模型進行任務。

使用 Playground：

上傳圖片。
提問並調整進階設定（如 JSON Schema）。

監控與成本管理：

查看 Dashboard 儀表板。
利用提供的 20 美元免費額度進行測試。

值得注意的限制或風險

閉源策略：頂級模型 Muse Spark 1.1 為閉源，無下載權限，僅能透過 API 使用。
定價對比：講者主觀對比稱其價格為 GPT 和 Claude 的四分之一，實際定價結構需查證。
技術依賴：整合第三方工具（如 Claude Code）可能需要特定的配置步驟，存在技術門檻。
基準測試數據：提及的 "Gemini 3.1 4.8" 版本號格式特殊，需確認其準確性與比較基準。

逐字稿辨識疑點

模型名稱不一致：逐字稿中交替出現 "muse spark"、"muse spark 1.1"、"muke spark"、"new spark"。
工具名稱聽似錯誤：
"claw code" / "clods connector"：應指 "Claude Code" 或相關連接器，但逐字稿聽寫為 claw/clods。
"plot code"：在描述 API 文檔時出現，疑為 "Playground" 或 "Code" 的口誤或聽寫錯誤。
"tropics"：在提及運行環境時出現 "clock course and tropics"，疑為 "Claude Code" 或其他工具的聽寫錯誤。
價格描述：逐字稿提到 "a quarter of what gpt and claude cost"（GPT 和 Claude 價格的四分之一），此為講者主觀對比，需查證實際定價結構。
基準測試數據：提及 "gemini 3.1 4.8"，版本號格式較為特殊，需查證是否為準確的模型版本標記。
人物名稱："alexander van" 為 Scale AI 首席 AI 官，需查證其正確姓名拼寫（疑為 Alexander Wang 或其他）。
產品功能描述：講者提到 "meta didn't just close its best model it put a price tag on it and open it up"，這與傳統開源定義有衝突，需理解為「開源模型（Llama）與頂級閉源模型（Spark）並存」的策略。

可延伸追問

Muse Spark 1.1 的具體定價表單與計費單位為何？
"Gemini 3.1 4.8" 的基準測試詳細數據與比較維度為何？
Scale AI 首席 AI 官 "Alexander van" 的正確姓名與背景為何？
Meta 如何確保 Muse Spark 1.1 在 Agentic 任務中的安全性與可靠性？
開發者如何具體在現有專案中切換至 Muse Spark 1.1 API？

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