0:00.000–0:01.900
Anthropic released a new feature yesterday.
0:02.180–0:03.460
It's called Claude Tag.
0:03.600–0:05.580
And I think when most people saw it, they thought,
0:05.700–0:09.260
okay, that's cool, a convenient way to invoke Claude
0:09.260–0:10.980
from within my Slack instance.
0:11.320–0:13.160
I use Slack, that sounds really cool.
0:13.320–0:16.140
But the more I read about it and the more I thought about it,
0:16.340–0:18.160
I actually started to get scared.
0:18.460–0:23.740
This is Anthropic's entry point into owning all knowledge work.
0:23.860–0:27.300
And this affects everybody, not just if you're writing code,
0:27.300–0:30.260
but literally if you do anything in front of a computer,
0:30.640–0:32.540
Anthropic is trying to own that.
0:32.700–0:35.080
And so I'm gonna show you what Claude Tag is
0:35.080–0:37.280
and then I'm gonna explain the bigger picture
0:37.280–0:39.960
of why this is such a big deal.
0:40.200–0:41.660
And by the way, if you wanna keep up
0:41.660–0:44.420
with what Anthropic and OpenAI are releasing
0:44.420–0:46.580
seemingly every week, like this video,
0:46.760–0:47.760
subscribe to the channel.
0:48.140–0:49.620
It really does help.
0:49.900–0:50.780
Thank you in advance.
0:51.080–0:52.760
All right, so let me show you what Claude Tag is.
0:52.760–0:56.140
If you use Slack, it's effectively
0:56.140–0:58.620
like just tagging another member of your team
0:58.620–0:59.700
and talking to it.
0:59.760–1:02.760
Except Claude Tag has all the context
1:02.760–1:05.200
of everything you do inside your company.
1:05.440–1:08.940
And it can actually get real world work done for you.
1:09.100–1:10.620
You simply tag Claude.
1:10.700–1:12.080
It chats with you.
1:12.320–1:13.440
It knows who you are.
1:13.500–1:14.940
It knows who your colleagues are.
1:15.220–1:18.220
It's able to understand all the different documents
1:18.220–1:19.340
you have inside your company.
1:19.580–1:23.040
It reads all the conversations in the Slack channels.
1:23.040–1:27.740
It is really building an entire graph of your company.
1:28.000–1:30.440
And by the way, you don't always have to tag Claude.
1:30.740–1:33.460
It's actively reading your conversations
1:33.460–1:35.560
in quote unquote ambient mode.
1:36.020–1:38.980
So again, it's just always sucking up
1:38.980–1:40.580
all the data from your company.
1:40.940–1:43.540
And Anthropic isn't mincing words
1:43.540–1:45.560
about what this product is.
1:45.680–1:48.540
They are literally saying Claude Tag is an evolution
1:48.540–1:51.120
of their cash cow Claude code.
1:51.740–1:55.000
Claude Tag is an evolution made more proactive
1:55.000–1:57.280
and built to work with a full team.
1:57.500–1:59.820
It's now one of the main ways we get things done
1:59.820–2:00.320
at Anthropic.
2:00.960–2:04.180
65% of our product team's code now comes
2:04.180–2:05.480
from our internal version.
2:05.680–2:08.260
So they're not even going into Claude code
2:08.260–2:09.480
directly anymore.
2:09.480–2:12.840
They are simply just typing from wherever they are.
2:12.940–2:14.120
And that happens to be Slack.
2:14.400–2:17.720
This is core infrastructure for Anthropic now.
2:17.720–2:19.900
And they want it to be core infrastructure
2:19.900–2:22.980
for every single company in the entire world.
2:23.240–2:24.960
And again, I really want to highlight
2:24.960–2:27.600
how big of a deal not only I think it is,
2:27.840–2:29.420
but Anthropic thinks it is.
2:29.660–2:30.980
This is Andre Karpathy,
2:31.180–2:32.540
one of the most prominent minds
2:32.540–2:35.720
in artificial intelligence who just joined Anthropic.
2:36.040–2:39.460
This is a new paradigm for interacting with Claude
2:39.460–2:41.580
that is significantly more in line
2:41.580–2:44.680
with all the other human activity org wide.
2:45.000–2:46.980
Interfaces are going away.
2:46.980–2:49.140
That is effectively what he's saying.
2:49.380–2:50.960
Don't go to Claude code anymore.
2:51.080–2:52.320
Don't go to any of the apps.
2:52.840–2:54.120
Wherever you're chatting,
2:54.480–2:56.500
which is Slack for a lot of teams,
2:56.680–2:58.580
that's where Claude tag is going to be.
2:58.880–3:00.660
And let me just pause for a second.
3:01.240–3:04.140
If Slack is where all the conversations are happening
3:04.140–3:05.800
and Claude lives in Slack,
3:06.420–3:08.740
what do you think Anthropic is going to build next?
3:08.880–3:12.480
I guarantee they are going to build a Slack competitor.
3:12.820–3:13.540
Okay, so let's continue.
3:13.540–3:16.780
Once you do all of the under the hood engineering work
3:16.780–3:18.460
to make this just work,
3:18.760–3:21.820
acts across tools, integrations, compute environments,
3:22.020–3:23.400
memory, security, et cetera,
3:23.960–3:26.760
Claude basically joins the team in a seamless way.
3:27.120–3:29.940
You can talk to it as you would talk to a person
3:29.940–3:33.780
and it can help with a very large variety of workloads.
3:34.320–3:36.480
They're positioning Claude tag
3:36.480–3:38.640
as another member of your team.
3:38.780–3:41.580
This is not a bot in your Slack app.
3:41.980–3:46.320
This is not AI responding to people tagging it
3:46.320–3:47.680
in your Slack app.
3:47.980–3:49.460
They really do say,
3:49.680–3:52.580
this is a new form of employee for your team.
3:52.660–3:53.300
And guess what?
3:53.840–3:55.420
You're renting it from Anthropic.
3:55.420–3:58.580
And he goes on to make even bigger claims.
3:59.120–3:59.840
In my opinion,
4:00.240–4:05.480
this is the third major redesign of the LLM UI UX.
4:05.980–4:09.540
The first paradigm was that the LLM is a website you go to.
4:09.720–4:11.680
So think about Claude or ChatGPT.
4:11.860–4:14.900
The second was that it is an app you download
4:14.900–4:16.060
to your computer.
4:16.520–4:18.020
Great, you can download ChatGPT,
4:18.160–4:19.660
but really what he's talking about
4:19.660–4:22.360
are apps like Codex and Claude Code.
4:22.360–4:25.960
And then third, a self-contained, persistent,
4:26.340–4:29.980
asynchronous entity with org-wide tools and context
4:29.980–4:32.640
working alongside teams of humans.
4:32.840–4:35.380
It really takes a while to wrap your head around it,
4:35.420–4:37.420
but it works and it is awesome.
4:37.680–4:40.300
And look, a lot of people gave Andre Karpathy criticism
4:40.300–4:42.960
for this because yeah, he joined Anthropic
4:42.960–4:44.760
and of course he's drinking the Kool-Aid.
4:44.860–4:46.200
He believes in what they're building,
4:46.420–4:48.080
but he also happens to be right.
4:48.580–4:49.720
This is a big deal.
4:49.720–4:53.720
You are now effectively hiring people,
4:54.000–4:56.440
people, AI, from Anthropic.
4:57.080–4:58.640
You have to pay them for that.
4:59.140–5:01.100
And at the same time that you're doing that,
5:01.240–5:03.380
you're handing them all the information about your company
5:03.380–5:05.960
and you are renting that back from them.
5:06.140–5:08.320
They are building an information graph
5:08.320–5:12.520
about every single company that uses Claude Tag.
5:12.780–5:15.720
And you need to decide if you're comfortable with that.
5:15.880–5:16.640
More on that later.
5:16.640–5:19.740
And by the way, it seems like I'm going to be paying Anthropic
5:19.740–5:20.600
more and more.
5:21.120–5:22.460
And to help me fund that,
5:22.600–5:24.420
let me tell you about the sponsor of today's video.
5:24.920–5:28.340
Context engineering is an incredibly important part
5:28.340–5:30.300
of your AI stack.
5:30.400–5:33.480
Making sure your AI has all of the data it needs
5:33.480–5:35.460
to give you the best possible results
5:35.460–5:37.900
is actually kind of a hard problem.
5:38.240–5:39.780
And if you're just working in one document,
5:39.940–5:41.460
sure, just give it that one document.
5:41.760–5:42.320
It's easy.
5:42.520–5:44.960
But imagine if you have thousands of documents
5:44.960–5:46.840
and then you start mixing in videos
5:46.840–5:49.420
and articles and other types of media.
5:49.740–5:51.280
And as you're seeing from Gbrain
5:51.280–5:53.520
and from Karpathy's LLM Wiki,
5:53.800–5:56.660
people are building all of this stuff themselves.
5:56.660–5:58.340
But it is not easy.
5:58.660–6:01.000
That is where Recall 2.0 comes in.
6:01.240–6:03.340
Recall 2.0 allows you to throw everything
6:03.340–6:06.320
into this kind of ocean of data.
6:06.520–6:07.960
And it can provide the right context
6:08.480–6:11.060
at the right moment to your AI.
6:11.440–6:12.580
With Recall 2.0,
6:12.580–6:15.220
you can switch between chatting with your knowledge base
6:15.220–6:16.820
and researching on the web.
6:17.040–6:18.620
I can add all of my OpenClaw videos
6:18.620–6:20.020
to my Recall knowledge base,
6:20.260–6:22.920
ask what the best security practices are,
6:23.100–6:25.260
and Recall will even give me timestamps
6:25.260–6:26.820
that I can watch in the app.
6:27.040–6:28.980
With an API and MCP,
6:29.280–6:31.580
you can plug this into any system
6:31.580–6:33.480
that you're already using and building.
6:33.720–6:34.420
And the best part,
6:34.480–6:36.100
you're not locked into one model.
6:36.360–6:37.360
So go check out Recall.
6:37.480–6:38.380
It's a great product.
6:38.500–6:41.300
Use MB25 to get 25% off.
6:41.300–6:43.940
And I'll drop the link down below in the description.
6:44.300–6:45.080
Thanks again to Recall.
6:45.160–6:46.040
Now back to the video.
6:46.460–6:47.080
And by the way,
6:47.540–6:50.140
Y Combinator has been talking about
6:50.140–6:53.400
this kind of AI native company for months now.
6:53.520–6:54.920
And they've put out videos,
6:55.100–6:58.380
but nothing really concrete about how to build it,
6:58.580–6:59.520
what it looks like,
6:59.760–7:01.860
companies that are already using it.
7:01.980–7:03.700
But they've been talking about it.
7:03.860–7:06.720
And now this is what they're talking about.
7:06.720–7:09.640
This is the beginning of the AI native company,
7:09.780–7:13.720
where AI understands every inch and corner
7:13.720–7:15.140
of your company
7:15.140–7:18.160
and is able to run 24 hours a day,
7:18.400–7:19.400
helping you grow,
7:19.600–7:20.700
finding issues,
7:20.880–7:22.200
helping your human employees
7:22.200–7:24.320
be a thousand X more productive.
7:24.880–7:25.860
This is the vision
7:25.860–7:27.680
that they've been laying out for months now.
7:27.760–7:28.200
And of course,
7:28.260–7:29.640
Anthropic has been building it.
7:29.760–7:31.020
This is Ankit Gupta,
7:31.240–7:32.940
a general partner at Y Combinator.
7:33.160–7:34.300
I like that at YC,
7:34.300–7:35.540
the partners and software team
7:35.540–7:37.740
have consistently built internal versions
7:37.740–7:39.340
of pretty much every product I see
7:39.340–7:41.080
open AI and Anthropic launch
7:41.080–7:42.200
six to 12 months later.
7:42.440–7:43.620
Let's see around corners
7:43.620–7:44.760
for where things are going.
7:44.900–7:45.960
So that's what I'm talking about.
7:46.040–7:47.840
They've been talking about AI native companies,
7:48.080–7:48.880
but they haven't put out
7:48.880–7:50.380
any public facing products
7:50.380–7:52.700
and they certainly haven't open sourced anything.
7:52.900–7:54.420
And I'm going to point back
7:54.420–7:56.600
to this incredible essay from Anthropic
7:56.600–7:58.180
because this is effectively
7:58.180–7:59.240
what they're trying to do
7:59.240–8:01.380
for all knowledge work in all companies.
8:01.740–8:02.580
Now, this is the essay
8:02.580–8:03.540
in which they talk about
8:03.540–8:05.340
self-improving artificial intelligence.
8:05.620–8:06.820
I'm not going to talk about it again.
8:06.880–8:08.040
I already made a video about it,
8:08.120–8:09.440
but what we're effectively seeing
8:09.440–8:10.260
is what happens
8:10.260–8:12.280
when a company becomes AI native,
8:12.460–8:14.120
when AI is fed
8:14.120–8:15.980
every single piece of data
8:15.980–8:17.060
about a company,
8:17.480–8:19.340
about all the conversations happening,
8:19.600–8:20.540
all the processes,
8:20.740–8:21.640
all the documents,
8:21.800–8:24.000
all the emails, everything.
8:24.260–8:25.280
It's all of a sudden
8:25.280–8:27.600
able to have this incredible visibility
8:27.600–8:29.520
into what the company is doing.
8:29.700–8:31.300
And then you can set goals.
8:31.300–8:32.260
I've been talking a lot
8:32.260–8:33.460
about loops lately.
8:33.880–8:35.080
Imagine if you can set a loop
8:35.080–8:36.440
for all knowledge work.
8:36.660–8:37.960
The loop can simply be
8:37.960–8:40.240
increase the revenue of my company
8:40.240–8:42.880
and it is going to ingest
8:42.880–8:44.260
all of the context,
8:44.920–8:45.880
all of that understanding
8:45.880–8:46.520
of the company,
8:46.940–8:47.860
set up experiments,
8:47.860–8:50.400
and then loop over it indefinitely.
8:50.720–8:52.520
And if that sounds expensive,
8:52.840–8:53.680
it is.
8:53.720–8:54.860
It is going to take
8:54.860–8:56.720
basically infinite tokens.
8:57.140–8:58.760
And then we're at this point,
8:58.760–9:00.980
and this is where it really gets scary,
9:00.980–9:02.520
where whichever company
9:02.520–9:04.940
can afford the most tokens,
9:05.340–9:07.060
they're going to do the best
9:07.060–9:08.680
because AI is going to improve
9:08.680–9:10.300
their business the most.
9:10.500–9:12.020
This sure sounds like the beginning
9:12.020–9:13.700
of the permanent underclass.
9:13.960–9:15.620
And then there was this tweet
9:15.620–9:17.520
by Ashwin Gopinath,
9:17.660–9:20.120
and he said it perfectly.
9:20.940–9:23.720
Claude Tag is a Trojan horse,
9:23.720–9:25.800
and I could not agree more.
9:25.800–9:27.360
Not because Anthropic
9:27.360–9:28.480
is doing anything evil,
9:28.800–9:31.200
because the incentives are obvious.
9:31.640–9:32.560
So day one,
9:32.640–9:33.880
this looks like a great feature.
9:34.180–9:35.140
Tag Claude in Slack,
9:35.400–9:36.460
let it follow the thread,
9:36.540–9:37.320
remember context,
9:37.440–9:38.060
connect the tools,
9:38.220–9:38.940
break down tasks,
9:39.080–9:39.620
chase work,
9:39.680–9:41.060
and act like a teammate.
9:41.460–9:42.680
But that's exactly the problem.
9:43.020–9:44.380
The moment your AI vendor
9:44.380–9:45.600
becomes a shared coworker,
9:45.660–9:47.660
it stops being just a model provider.
9:47.760–9:48.920
It starts becoming the place
9:48.920–9:49.960
where work is interpreted,
9:50.160–9:51.020
remembered, routed,
9:51.020–9:52.800
and eventually executed.
9:53.280–9:54.780
Imagine if you no longer
9:54.780–9:57.420
had any employees on your team,
9:57.560–9:58.560
and you were renting
9:58.560–10:00.320
an AI employee from Anthropic.
10:00.580–10:01.840
Now imagine that
10:01.840–10:03.360
across the entire economy.
10:03.880–10:05.880
Every single piece of knowledge work
10:05.880–10:07.620
is now just rented
10:07.620–10:10.600
from this one massive company.
10:11.160–10:12.260
That's why I keep saying
10:12.260–10:13.060
this is scary.
10:13.320–10:14.760
So that is not model lock-in,
10:14.900–10:16.440
that is context lock-in.
10:16.520–10:18.060
You are now renting your company
10:18.060–10:19.220
back from them.
10:19.220–10:22.420
And yes, platform risk is very real.
10:22.420–10:24.740
It was real before you could
10:24.740–10:26.200
literally buy intelligence
10:26.200–10:26.960
from a company.
10:27.160–10:28.760
It was real when people
10:28.760–10:29.820
were building apps
10:29.820–10:31.640
on top of the Apple App Store
10:31.640–10:32.880
and Apple got to decide
10:32.880–10:34.340
if they liked your app enough
10:34.340–10:35.360
to let you be there.
10:35.500–10:36.760
And whether you were building
10:36.760–10:37.740
on top of Facebook
10:37.740–10:38.960
and they wanted you
10:38.960–10:40.460
to use their API or not.
10:40.740–10:41.900
Anytime you build
10:41.900–10:43.720
on top of somebody else's platform,
10:43.920–10:45.220
you have platform risk.
10:45.500–10:46.920
And this is the ultimate form
10:46.920–10:47.980
of platform risk.
10:47.980–10:50.860
It is all knowledge work
10:50.860–10:52.140
platform risk.
10:52.240–10:52.940
He goes on,
10:53.000–10:53.860
the pricing model
10:53.860–10:55.060
makes it even more dangerous.
10:55.440–10:57.140
A human coworker has a salary.
10:57.660–10:59.320
Claude has unbounded
10:59.320–11:00.740
tokenized activity.
11:01.300–11:02.720
That means once again,
11:02.860–11:04.000
there is a cap
11:04.000–11:05.200
that you could pay somebody,
11:05.300–11:06.200
a human worker.
11:06.740–11:08.220
But if you're paying Claude,
11:08.540–11:10.040
it is uncapped.
11:10.160–11:11.620
You can literally pay them
11:11.620–11:12.640
infinite dollars
11:12.640–11:14.180
and they'll eat it up.
11:14.460–11:15.220
And there's always
11:15.220–11:16.340
more work to be done.
11:16.340–11:18.300
Imagine running a company
11:18.300–11:19.620
and being that dependent
11:19.620–11:21.260
on another company.
11:21.380–11:22.840
But it actually gets bigger
11:22.840–11:24.020
and worse than that.
11:24.160–11:24.920
What happens when
11:24.920–11:26.220
every single company
11:26.220–11:26.880
starts building
11:26.880–11:27.900
on top of Anthropic?
11:28.120–11:29.420
And I know I'm talking
11:29.420–11:30.400
a lot about Anthropic,
11:30.560–11:31.080
but that's because
11:31.080–11:32.160
they released Claude Tag.
11:32.280–11:33.740
But you know OpenAI
11:33.740–11:35.160
is working on the same feature.
11:35.420–11:37.140
Hopefully other companies as well.
11:37.220–11:38.600
I hope there's more competition
11:38.600–11:39.600
because competition
11:39.600–11:41.560
will be the only solution
11:41.560–11:42.300
for the problems
11:42.300–11:43.440
that come from Anthropic
11:43.440–11:44.680
owning all knowledge work.
11:44.880–11:45.580
So imagine now
11:45.580–11:46.520
you're a SaaS company.
11:46.700–11:47.520
You have software.
11:47.680–11:49.040
You sell it to other companies.
11:49.040–11:50.320
But all of a sudden,
11:50.520–11:51.800
Anthropic is now
11:51.800–11:52.880
using their agents
11:52.880–11:54.080
to operate your software
11:54.080–11:55.860
on behalf of their customers
11:55.860–11:56.700
that pay them.
11:56.820–11:57.440
And guess what?
11:57.520–11:58.240
Those customers
11:58.240–11:59.680
are no longer
11:59.680–12:00.420
going back
12:00.420–12:01.920
into your user interface.
12:01.920–12:03.480
They no longer log in.
12:03.740–12:04.560
They have no need to.
12:04.660–12:05.600
They just tell their agent
12:05.600–12:06.180
what to do.
12:06.360–12:07.600
So why would they ever?
12:07.900–12:08.620
All of a sudden,
12:08.920–12:10.040
your user interface
12:10.040–12:11.660
becomes valueless.
12:12.000–12:12.900
And so what are you
12:12.900–12:13.520
at that point?
12:13.520–12:14.660
You have the workflows.
12:15.140–12:16.480
So remove the UI.
12:16.740–12:17.400
What can an agent
12:17.400–12:18.460
actually get done
12:18.460–12:19.700
inside of your application?
12:19.840–12:20.520
Those are the workflows.
12:20.760–12:22.220
But how long does that last?
12:22.660–12:23.780
Because really,
12:24.180–12:26.000
the agent can just write code
12:26.000–12:27.340
for those workflows.
12:27.760–12:30.040
AI is extremely good
12:30.040–12:31.180
at writing code.
12:31.260–12:32.060
So why wouldn't it
12:32.060–12:32.880
just write code
12:32.880–12:33.900
to manage
12:33.900–12:34.920
and then eventually
12:34.920–12:36.780
overtake those workflows itself?
12:36.980–12:38.080
As a software company,
12:38.140–12:39.180
what are you at that point?
12:39.440–12:40.440
You're simply a database.
12:40.440–12:41.780
And you know what's easier
12:41.780–12:43.200
than writing code
12:43.200–12:43.860
for workflows?
12:44.240–12:45.200
It's writing code
12:45.200–12:46.060
to read and write
12:46.060–12:46.780
from databases.
12:47.180–12:48.040
Agents are incredibly
12:48.040–12:48.580
good at that.
12:48.660–12:49.400
They don't even care
12:49.400–12:50.640
what kind of database it is.
12:51.180–12:52.120
They will help you
12:52.120–12:52.800
move your data
12:52.800–12:53.580
from one database
12:53.580–12:54.220
to another.
12:54.560–12:55.520
They'll help you read
12:55.520–12:56.840
and write from it easily.
12:57.080–12:58.200
And then all of a sudden,
12:58.820–12:59.420
what happens
12:59.420–13:00.260
to all software?
13:00.560–13:01.320
No software
13:01.320–13:02.300
is safe from this.
13:02.640–13:03.340
Agents are now
13:03.340–13:04.060
infiltrating
13:04.060–13:05.200
right before our eyes
13:05.200–13:06.380
every single piece
13:06.380–13:06.960
of software.
13:06.960–13:08.000
That brings us
13:08.000–13:08.880
to our final
13:08.880–13:10.560
scariest point.
13:10.780–13:11.840
Anthropic all of a sudden
13:11.840–13:13.160
owns all knowledge work.
13:13.720–13:14.260
They own
13:14.260–13:15.480
all software.
13:15.840–13:16.640
How does society
13:16.640–13:17.820
continue at that point?
13:18.320–13:19.460
And that's the problem.
13:20.140–13:21.120
Government intervention
13:21.120–13:22.700
is the last line
13:22.700–13:23.340
of defense.
13:24.140–13:24.740
Societally,
13:24.980–13:26.400
we will not allow
13:26.400–13:27.560
a single company,
13:27.680–13:28.460
a single entity
13:28.460–13:29.920
to have all the power.
13:30.040–13:30.860
It just won't work.
13:30.980–13:31.780
But there is
13:31.780–13:32.640
a solution
13:32.640–13:33.720
beyond that.
13:33.880–13:34.780
Open source models,
13:34.780–13:35.860
model competition
13:35.860–13:37.440
from other research labs,
13:37.680–13:38.600
making sure that
13:38.600–13:39.480
whatever provider
13:39.480–13:40.200
you choose
13:40.200–13:41.000
for this Claude
13:41.000–13:42.300
tag-like feature
13:42.300–13:43.460
allows you to
13:43.460–13:44.680
own the context.
13:45.260–13:46.300
You own the context
13:46.300–13:47.040
of your business,
13:47.180–13:47.680
not them.
13:48.140–13:49.320
You should own it.
13:49.880–13:50.440
And that's why
13:50.440–13:51.120
open source
13:51.120–13:52.420
is so important right now.
13:52.560–13:53.520
You're not going to be
13:53.520–13:54.880
beholden to paying
13:54.880–13:55.780
Anthropic prices
13:55.780–13:56.780
for every token.
13:57.400–13:58.140
You can pay
13:58.140–13:59.580
the best price
13:59.580–14:01.080
for the best token
14:01.080–14:02.100
when you have
14:02.100–14:03.260
a multi-model,
14:03.260–14:04.680
multi-provider strategy.
14:04.680–14:06.100
So what started
14:06.100–14:07.360
as a cool,
14:07.500–14:08.380
convenient feature,
14:08.560–14:09.500
being able to tag Claude
14:09.500–14:10.420
from within Slack
14:10.420–14:12.320
is very obviously
14:12.320–14:13.980
Anthropic's attempt
14:13.980–14:15.460
to own
14:15.460–14:16.420
all knowledge work.
14:16.560–14:18.080
That is a very bad thing.
14:18.280–14:19.140
But the one thing
14:19.140–14:20.300
I do want you to remember
14:20.300–14:21.560
and think about
14:21.560–14:22.380
is this is
14:22.380–14:23.860
definitely the future
14:23.860–14:24.580
of how companies
14:24.580–14:25.180
will work.
14:25.780–14:27.120
AI and humans
14:27.120–14:28.480
working in concert
14:28.480–14:29.920
to accomplish things
14:29.920–14:31.060
and AI
14:31.060–14:32.220
living and breathing
14:32.220–14:33.700
inside of your company,
14:33.700–14:35.060
not just
14:35.060–14:36.380
a chat app
14:36.380–14:37.000
that you go
14:37.000–14:38.080
type a query
14:38.080–14:39.320
to get the response
14:39.320–14:39.960
and then continue
14:39.960–14:40.520
your work.
14:40.600–14:42.000
It is going to be able
14:42.000–14:42.860
to see everything
14:42.860–14:43.480
you're doing
14:43.480–14:45.020
inside that company
14:45.020–14:45.880
and it's going to be able
14:45.880–14:46.360
to help you.
14:46.400–14:47.040
It's going to be able
14:47.040–14:47.560
to make you
14:47.560–14:48.500
so much more productive.
14:48.840–14:49.640
So I wanted to end
14:49.640–14:51.000
on that optimistic note.
14:51.180–14:52.860
We just need more competition
14:52.860–14:53.560
most of all.
14:53.720–14:54.940
And I talked about Anthropic
14:54.940–14:56.540
having self-improving
14:56.540–14:57.460
artificial intelligence.
14:57.600–14:58.220
I actually made
14:58.220–14:59.120
an entire video
14:59.120–15:00.320
breaking that down
15:00.320–15:01.180
in detail.
15:01.660–15:02.280
Go check it out
15:02.280–15:02.800
right here.

0:00.000–0:01.900
（此句尚無繁中翻譯）
0:02.180–0:03.460
（此句尚無繁中翻譯）
0:03.600–0:05.580
（此句尚無繁中翻譯）
0:05.700–0:09.260
（此句尚無繁中翻譯）
0:09.260–0:10.980
（此句尚無繁中翻譯）
0:11.320–0:13.160
（此句尚無繁中翻譯）
0:13.320–0:16.140
（此句尚無繁中翻譯）
0:16.340–0:18.160
（此句尚無繁中翻譯）
0:18.460–0:23.740
（此句尚無繁中翻譯）
0:23.860–0:27.300
（此句尚無繁中翻譯）
0:27.300–0:30.260
（此句尚無繁中翻譯）
0:30.640–0:32.540
（此句尚無繁中翻譯）
0:32.700–0:35.080
（此句尚無繁中翻譯）
0:35.080–0:37.280
（此句尚無繁中翻譯）
0:37.280–0:39.960
（此句尚無繁中翻譯）
0:40.200–0:41.660
（此句尚無繁中翻譯）
0:41.660–0:44.420
（此句尚無繁中翻譯）
0:44.420–0:46.580
（此句尚無繁中翻譯）
0:46.760–0:47.760
（此句尚無繁中翻譯）
0:48.140–0:49.620
（此句尚無繁中翻譯）
0:49.900–0:50.780
（此句尚無繁中翻譯）
0:51.080–0:52.760
（此句尚無繁中翻譯）
0:52.760–0:56.140
（此句尚無繁中翻譯）
0:56.140–0:58.620
（此句尚無繁中翻譯）
0:58.620–0:59.700
（此句尚無繁中翻譯）
0:59.760–1:02.760
（此句尚無繁中翻譯）
1:02.760–1:05.200
（此句尚無繁中翻譯）
1:05.440–1:08.940
（此句尚無繁中翻譯）
1:09.100–1:10.620
（此句尚無繁中翻譯）
1:10.700–1:12.080
（此句尚無繁中翻譯）
1:12.320–1:13.440
（此句尚無繁中翻譯）
1:13.500–1:14.940
（此句尚無繁中翻譯）
1:15.220–1:18.220
（此句尚無繁中翻譯）
1:18.220–1:19.340
（此句尚無繁中翻譯）
1:19.580–1:23.040
（此句尚無繁中翻譯）
1:23.040–1:27.740
（此句尚無繁中翻譯）
1:28.000–1:30.440
（此句尚無繁中翻譯）
1:30.740–1:33.460
（此句尚無繁中翻譯）
1:33.460–1:35.560
（此句尚無繁中翻譯）
1:36.020–1:38.980
（此句尚無繁中翻譯）
1:38.980–1:40.580
（此句尚無繁中翻譯）
1:40.940–1:43.540
（此句尚無繁中翻譯）
1:43.540–1:45.560
（此句尚無繁中翻譯）
1:45.680–1:48.540
（此句尚無繁中翻譯）
1:48.540–1:51.120
（此句尚無繁中翻譯）
1:51.740–1:55.000
（此句尚無繁中翻譯）
1:55.000–1:57.280
（此句尚無繁中翻譯）
1:57.500–1:59.820
（此句尚無繁中翻譯）
1:59.820–2:00.320
（此句尚無繁中翻譯）
2:00.960–2:04.180
（此句尚無繁中翻譯）
2:04.180–2:05.480
（此句尚無繁中翻譯）
2:05.680–2:08.260
（此句尚無繁中翻譯）
2:08.260–2:09.480
（此句尚無繁中翻譯）
2:09.480–2:12.840
（此句尚無繁中翻譯）
2:12.940–2:14.120
（此句尚無繁中翻譯）
2:14.400–2:17.720
（此句尚無繁中翻譯）
2:17.720–2:19.900
（此句尚無繁中翻譯）
2:19.900–2:22.980
（此句尚無繁中翻譯）
2:23.240–2:24.960
（此句尚無繁中翻譯）
2:24.960–2:27.600
（此句尚無繁中翻譯）
2:27.840–2:29.420
（此句尚無繁中翻譯）
2:29.660–2:30.980
（此句尚無繁中翻譯）
2:31.180–2:32.540
（此句尚無繁中翻譯）
2:32.540–2:35.720
（此句尚無繁中翻譯）
2:36.040–2:39.460
（此句尚無繁中翻譯）
2:39.460–2:41.580
（此句尚無繁中翻譯）
2:41.580–2:44.680
（此句尚無繁中翻譯）
2:45.000–2:46.980
（此句尚無繁中翻譯）
2:46.980–2:49.140
（此句尚無繁中翻譯）
2:49.380–2:50.960
（此句尚無繁中翻譯）
2:51.080–2:52.320
（此句尚無繁中翻譯）
2:52.840–2:54.120
（此句尚無繁中翻譯）
2:54.480–2:56.500
（此句尚無繁中翻譯）
2:56.680–2:58.580
（此句尚無繁中翻譯）
2:58.880–3:00.660
（此句尚無繁中翻譯）
3:01.240–3:04.140
（此句尚無繁中翻譯）
3:04.140–3:05.800
（此句尚無繁中翻譯）
3:06.420–3:08.740
（此句尚無繁中翻譯）
3:08.880–3:12.480
（此句尚無繁中翻譯）
3:12.820–3:13.540
（此句尚無繁中翻譯）
3:13.540–3:16.780
（此句尚無繁中翻譯）
3:16.780–3:18.460
（此句尚無繁中翻譯）
3:18.760–3:21.820
（此句尚無繁中翻譯）
3:22.020–3:23.400
（此句尚無繁中翻譯）
3:23.960–3:26.760
（此句尚無繁中翻譯）
3:27.120–3:29.940
（此句尚無繁中翻譯）
3:29.940–3:33.780
（此句尚無繁中翻譯）
3:34.320–3:36.480
（此句尚無繁中翻譯）
3:36.480–3:38.640
（此句尚無繁中翻譯）
3:38.780–3:41.580
（此句尚無繁中翻譯）
3:41.980–3:46.320
（此句尚無繁中翻譯）
3:46.320–3:47.680
（此句尚無繁中翻譯）
3:47.980–3:49.460
（此句尚無繁中翻譯）
3:49.680–3:52.580
（此句尚無繁中翻譯）
3:52.660–3:53.300
（此句尚無繁中翻譯）
3:53.840–3:55.420
（此句尚無繁中翻譯）
3:55.420–3:58.580
（此句尚無繁中翻譯）
3:59.120–3:59.840
（此句尚無繁中翻譯）
4:00.240–4:05.480
（此句尚無繁中翻譯）
4:05.980–4:09.540
（此句尚無繁中翻譯）
4:09.720–4:11.680
（此句尚無繁中翻譯）
4:11.860–4:14.900
（此句尚無繁中翻譯）
4:14.900–4:16.060
（此句尚無繁中翻譯）
4:16.520–4:18.020
（此句尚無繁中翻譯）
4:18.160–4:19.660
（此句尚無繁中翻譯）
4:19.660–4:22.360
（此句尚無繁中翻譯）
4:22.360–4:25.960
（此句尚無繁中翻譯）
4:26.340–4:29.980
（此句尚無繁中翻譯）
4:29.980–4:32.640
（此句尚無繁中翻譯）
4:32.840–4:35.380
（此句尚無繁中翻譯）
4:35.420–4:37.420
（此句尚無繁中翻譯）
4:37.680–4:40.300
（此句尚無繁中翻譯）
4:40.300–4:42.960
（此句尚無繁中翻譯）
4:42.960–4:44.760
（此句尚無繁中翻譯）
4:44.860–4:46.200
（此句尚無繁中翻譯）
4:46.420–4:48.080
（此句尚無繁中翻譯）
4:48.580–4:49.720
（此句尚無繁中翻譯）
4:49.720–4:53.720
（此句尚無繁中翻譯）
4:54.000–4:56.440
（此句尚無繁中翻譯）
4:57.080–4:58.640
（此句尚無繁中翻譯）
4:59.140–5:01.100
（此句尚無繁中翻譯）
5:01.240–5:03.380
（此句尚無繁中翻譯）
5:03.380–5:05.960
（此句尚無繁中翻譯）
5:06.140–5:08.320
（此句尚無繁中翻譯）
5:08.320–5:12.520
（此句尚無繁中翻譯）
5:12.780–5:15.720
（此句尚無繁中翻譯）
5:15.880–5:16.640
（此句尚無繁中翻譯）
5:16.640–5:19.740
（此句尚無繁中翻譯）
5:19.740–5:20.600
（此句尚無繁中翻譯）
5:21.120–5:22.460
（此句尚無繁中翻譯）
5:22.600–5:24.420
（此句尚無繁中翻譯）
5:24.920–5:28.340
（此句尚無繁中翻譯）
5:28.340–5:30.300
（此句尚無繁中翻譯）
5:30.400–5:33.480
（此句尚無繁中翻譯）
5:33.480–5:35.460
（此句尚無繁中翻譯）
5:35.460–5:37.900
（此句尚無繁中翻譯）
5:38.240–5:39.780
（此句尚無繁中翻譯）
5:39.940–5:41.460
（此句尚無繁中翻譯）
5:41.760–5:42.320
（此句尚無繁中翻譯）
5:42.520–5:44.960
（此句尚無繁中翻譯）
5:44.960–5:46.840
（此句尚無繁中翻譯）
5:46.840–5:49.420
（此句尚無繁中翻譯）
5:49.740–5:51.280
（此句尚無繁中翻譯）
5:51.280–5:53.520
（此句尚無繁中翻譯）
5:53.800–5:56.660
（此句尚無繁中翻譯）
5:56.660–5:58.340
（此句尚無繁中翻譯）
5:58.660–6:01.000
（此句尚無繁中翻譯）
6:01.240–6:03.340
（此句尚無繁中翻譯）
6:03.340–6:06.320
（此句尚無繁中翻譯）
6:06.520–6:07.960
（此句尚無繁中翻譯）
6:08.480–6:11.060
（此句尚無繁中翻譯）
6:11.440–6:12.580
（此句尚無繁中翻譯）
6:12.580–6:15.220
（此句尚無繁中翻譯）
6:15.220–6:16.820
（此句尚無繁中翻譯）
6:17.040–6:18.620
（此句尚無繁中翻譯）
6:18.620–6:20.020
（此句尚無繁中翻譯）
6:20.260–6:22.920
（此句尚無繁中翻譯）
6:23.100–6:25.260
（此句尚無繁中翻譯）
6:25.260–6:26.820
（此句尚無繁中翻譯）
6:27.040–6:28.980
（此句尚無繁中翻譯）
6:29.280–6:31.580
（此句尚無繁中翻譯）
6:31.580–6:33.480
（此句尚無繁中翻譯）
6:33.720–6:34.420
（此句尚無繁中翻譯）
6:34.480–6:36.100
（此句尚無繁中翻譯）
6:36.360–6:37.360
（此句尚無繁中翻譯）
6:37.480–6:38.380
（此句尚無繁中翻譯）
6:38.500–6:41.300
（此句尚無繁中翻譯）
6:41.300–6:43.940
（此句尚無繁中翻譯）
6:44.300–6:45.080
（此句尚無繁中翻譯）
6:45.160–6:46.040
（此句尚無繁中翻譯）
6:46.460–6:47.080
（此句尚無繁中翻譯）
6:47.540–6:50.140
（此句尚無繁中翻譯）
6:50.140–6:53.400
（此句尚無繁中翻譯）
6:53.520–6:54.920
（此句尚無繁中翻譯）
6:55.100–6:58.380
（此句尚無繁中翻譯）
6:58.580–6:59.520
（此句尚無繁中翻譯）
6:59.760–7:01.860
（此句尚無繁中翻譯）
7:01.980–7:03.700
（此句尚無繁中翻譯）
7:03.860–7:06.720
（此句尚無繁中翻譯）
7:06.720–7:09.640
（此句尚無繁中翻譯）
7:09.780–7:13.720
（此句尚無繁中翻譯）
7:13.720–7:15.140
（此句尚無繁中翻譯）
7:15.140–7:18.160
（此句尚無繁中翻譯）
7:18.400–7:19.400
（此句尚無繁中翻譯）
7:19.600–7:20.700
（此句尚無繁中翻譯）
7:20.880–7:22.200
（此句尚無繁中翻譯）
7:22.200–7:24.320
（此句尚無繁中翻譯）
7:24.880–7:25.860
（此句尚無繁中翻譯）
7:25.860–7:27.680
（此句尚無繁中翻譯）
7:27.760–7:28.200
（此句尚無繁中翻譯）
7:28.260–7:29.640
（此句尚無繁中翻譯）
7:29.760–7:31.020
（此句尚無繁中翻譯）
7:31.240–7:32.940
（此句尚無繁中翻譯）
7:33.160–7:34.300
（此句尚無繁中翻譯）
7:34.300–7:35.540
（此句尚無繁中翻譯）
7:35.540–7:37.740
（此句尚無繁中翻譯）
7:37.740–7:39.340
（此句尚無繁中翻譯）
7:39.340–7:41.080
（此句尚無繁中翻譯）
7:41.080–7:42.200
（此句尚無繁中翻譯）
7:42.440–7:43.620
（此句尚無繁中翻譯）
7:43.620–7:44.760
（此句尚無繁中翻譯）
7:44.900–7:45.960
（此句尚無繁中翻譯）
7:46.040–7:47.840
（此句尚無繁中翻譯）
7:48.080–7:48.880
（此句尚無繁中翻譯）
7:48.880–7:50.380
（此句尚無繁中翻譯）
7:50.380–7:52.700
（此句尚無繁中翻譯）
7:52.900–7:54.420
（此句尚無繁中翻譯）
7:54.420–7:56.600
（此句尚無繁中翻譯）
7:56.600–7:58.180
（此句尚無繁中翻譯）
7:58.180–7:59.240
（此句尚無繁中翻譯）
7:59.240–8:01.380
（此句尚無繁中翻譯）
8:01.740–8:02.580
（此句尚無繁中翻譯）
8:02.580–8:03.540
（此句尚無繁中翻譯）
8:03.540–8:05.340
（此句尚無繁中翻譯）
8:05.620–8:06.820
（此句尚無繁中翻譯）
8:06.880–8:08.040
（此句尚無繁中翻譯）
8:08.120–8:09.440
（此句尚無繁中翻譯）
8:09.440–8:10.260
（此句尚無繁中翻譯）
8:10.260–8:12.280
（此句尚無繁中翻譯）
8:12.460–8:14.120
（此句尚無繁中翻譯）
8:14.120–8:15.980
（此句尚無繁中翻譯）
8:15.980–8:17.060
（此句尚無繁中翻譯）
8:17.480–8:19.340
（此句尚無繁中翻譯）
8:19.600–8:20.540
（此句尚無繁中翻譯）
8:20.740–8:21.640
（此句尚無繁中翻譯）
8:21.800–8:24.000
（此句尚無繁中翻譯）
8:24.260–8:25.280
（此句尚無繁中翻譯）
8:25.280–8:27.600
（此句尚無繁中翻譯）
8:27.600–8:29.520
（此句尚無繁中翻譯）
8:29.700–8:31.300
（此句尚無繁中翻譯）
8:31.300–8:32.260
（此句尚無繁中翻譯）
8:32.260–8:33.460
（此句尚無繁中翻譯）
8:33.880–8:35.080
（此句尚無繁中翻譯）
8:35.080–8:36.440
（此句尚無繁中翻譯）
8:36.660–8:37.960
（此句尚無繁中翻譯）
8:37.960–8:40.240
（此句尚無繁中翻譯）
8:40.240–8:42.880
（此句尚無繁中翻譯）
8:42.880–8:44.260
（此句尚無繁中翻譯）
8:44.920–8:45.880
（此句尚無繁中翻譯）
8:45.880–8:46.520
（此句尚無繁中翻譯）
8:46.940–8:47.860
（此句尚無繁中翻譯）
8:47.860–8:50.400
（此句尚無繁中翻譯）
8:50.720–8:52.520
（此句尚無繁中翻譯）
8:52.840–8:53.680
（此句尚無繁中翻譯）
8:53.720–8:54.860
（此句尚無繁中翻譯）
8:54.860–8:56.720
（此句尚無繁中翻譯）
8:57.140–8:58.760
（此句尚無繁中翻譯）
8:58.760–9:00.980
（此句尚無繁中翻譯）
9:00.980–9:02.520
（此句尚無繁中翻譯）
9:02.520–9:04.940
（此句尚無繁中翻譯）
9:05.340–9:07.060
（此句尚無繁中翻譯）
9:07.060–9:08.680
（此句尚無繁中翻譯）
9:08.680–9:10.300
（此句尚無繁中翻譯）
9:10.500–9:12.020
（此句尚無繁中翻譯）
9:12.020–9:13.700
（此句尚無繁中翻譯）
9:13.960–9:15.620
（此句尚無繁中翻譯）
9:15.620–9:17.520
（此句尚無繁中翻譯）
9:17.660–9:20.120
（此句尚無繁中翻譯）
9:20.940–9:23.720
（此句尚無繁中翻譯）
9:23.720–9:25.800
（此句尚無繁中翻譯）
9:25.800–9:27.360
（此句尚無繁中翻譯）
9:27.360–9:28.480
（此句尚無繁中翻譯）
9:28.800–9:31.200
（此句尚無繁中翻譯）
9:31.640–9:32.560
（此句尚無繁中翻譯）
9:32.640–9:33.880
（此句尚無繁中翻譯）
9:34.180–9:35.140
（此句尚無繁中翻譯）
9:35.400–9:36.460
（此句尚無繁中翻譯）
9:36.540–9:37.320
（此句尚無繁中翻譯）
9:37.440–9:38.060
（此句尚無繁中翻譯）
9:38.220–9:38.940
（此句尚無繁中翻譯）
9:39.080–9:39.620
（此句尚無繁中翻譯）
9:39.680–9:41.060
（此句尚無繁中翻譯）
9:41.460–9:42.680
（此句尚無繁中翻譯）
9:43.020–9:44.380
（此句尚無繁中翻譯）
9:44.380–9:45.600
（此句尚無繁中翻譯）
9:45.660–9:47.660
（此句尚無繁中翻譯）
9:47.760–9:48.920
（此句尚無繁中翻譯）
9:48.920–9:49.960
（此句尚無繁中翻譯）
9:50.160–9:51.020
（此句尚無繁中翻譯）
9:51.020–9:52.800
（此句尚無繁中翻譯）
9:53.280–9:54.780
（此句尚無繁中翻譯）
9:54.780–9:57.420
（此句尚無繁中翻譯）
9:57.560–9:58.560
（此句尚無繁中翻譯）
9:58.560–10:00.320
（此句尚無繁中翻譯）
10:00.580–10:01.840
（此句尚無繁中翻譯）
10:01.840–10:03.360
（此句尚無繁中翻譯）
10:03.880–10:05.880
（此句尚無繁中翻譯）
10:05.880–10:07.620
（此句尚無繁中翻譯）
10:07.620–10:10.600
（此句尚無繁中翻譯）
10:11.160–10:12.260
（此句尚無繁中翻譯）
10:12.260–10:13.060
（此句尚無繁中翻譯）
10:13.320–10:14.760
（此句尚無繁中翻譯）
10:14.900–10:16.440
（此句尚無繁中翻譯）
10:16.520–10:18.060
（此句尚無繁中翻譯）
10:18.060–10:19.220
（此句尚無繁中翻譯）
10:19.220–10:22.420
（此句尚無繁中翻譯）
10:22.420–10:24.740
（此句尚無繁中翻譯）
10:24.740–10:26.200
（此句尚無繁中翻譯）
10:26.200–10:26.960
（此句尚無繁中翻譯）
10:27.160–10:28.760
（此句尚無繁中翻譯）
10:28.760–10:29.820
（此句尚無繁中翻譯）
10:29.820–10:31.640
（此句尚無繁中翻譯）
10:31.640–10:32.880
（此句尚無繁中翻譯）
10:32.880–10:34.340
（此句尚無繁中翻譯）
10:34.340–10:35.360
（此句尚無繁中翻譯）
10:35.500–10:36.760
（此句尚無繁中翻譯）
10:36.760–10:37.740
（此句尚無繁中翻譯）
10:37.740–10:38.960
（此句尚無繁中翻譯）
10:38.960–10:40.460
（此句尚無繁中翻譯）
10:40.740–10:41.900
（此句尚無繁中翻譯）
10:41.900–10:43.720
（此句尚無繁中翻譯）
10:43.920–10:45.220
（此句尚無繁中翻譯）
10:45.500–10:46.920
（此句尚無繁中翻譯）
10:46.920–10:47.980
（此句尚無繁中翻譯）
10:47.980–10:50.860
（此句尚無繁中翻譯）
10:50.860–10:52.140
（此句尚無繁中翻譯）
10:52.240–10:52.940
（此句尚無繁中翻譯）
10:53.000–10:53.860
（此句尚無繁中翻譯）
10:53.860–10:55.060
（此句尚無繁中翻譯）
10:55.440–10:57.140
（此句尚無繁中翻譯）
10:57.660–10:59.320
（此句尚無繁中翻譯）
10:59.320–11:00.740
（此句尚無繁中翻譯）
11:01.300–11:02.720
（此句尚無繁中翻譯）
11:02.860–11:04.000
（此句尚無繁中翻譯）
11:04.000–11:05.200
（此句尚無繁中翻譯）
11:05.300–11:06.200
（此句尚無繁中翻譯）
11:06.740–11:08.220
（此句尚無繁中翻譯）
11:08.540–11:10.040
（此句尚無繁中翻譯）
11:10.160–11:11.620
（此句尚無繁中翻譯）
11:11.620–11:12.640
（此句尚無繁中翻譯）
11:12.640–11:14.180
（此句尚無繁中翻譯）
11:14.460–11:15.220
（此句尚無繁中翻譯）
11:15.220–11:16.340
（此句尚無繁中翻譯）
11:16.340–11:18.300
（此句尚無繁中翻譯）
11:18.300–11:19.620
（此句尚無繁中翻譯）
11:19.620–11:21.260
（此句尚無繁中翻譯）
11:21.380–11:22.840
（此句尚無繁中翻譯）
11:22.840–11:24.020
（此句尚無繁中翻譯）
11:24.160–11:24.920
（此句尚無繁中翻譯）
11:24.920–11:26.220
（此句尚無繁中翻譯）
11:26.220–11:26.880
（此句尚無繁中翻譯）
11:26.880–11:27.900
（此句尚無繁中翻譯）
11:28.120–11:29.420
（此句尚無繁中翻譯）
11:29.420–11:30.400
（此句尚無繁中翻譯）
11:30.560–11:31.080
（此句尚無繁中翻譯）
11:31.080–11:32.160
（此句尚無繁中翻譯）
11:32.280–11:33.740
（此句尚無繁中翻譯）
11:33.740–11:35.160
（此句尚無繁中翻譯）
11:35.420–11:37.140
（此句尚無繁中翻譯）
11:37.220–11:38.600
（此句尚無繁中翻譯）
11:38.600–11:39.600
（此句尚無繁中翻譯）
11:39.600–11:41.560
（此句尚無繁中翻譯）
11:41.560–11:42.300
（此句尚無繁中翻譯）
11:42.300–11:43.440
（此句尚無繁中翻譯）
11:43.440–11:44.680
（此句尚無繁中翻譯）
11:44.880–11:45.580
（此句尚無繁中翻譯）
11:45.580–11:46.520
（此句尚無繁中翻譯）
11:46.700–11:47.520
（此句尚無繁中翻譯）
11:47.680–11:49.040
（此句尚無繁中翻譯）
11:49.040–11:50.320
（此句尚無繁中翻譯）
11:50.520–11:51.800
（此句尚無繁中翻譯）
11:51.800–11:52.880
（此句尚無繁中翻譯）
11:52.880–11:54.080
（此句尚無繁中翻譯）
11:54.080–11:55.860
（此句尚無繁中翻譯）
11:55.860–11:56.700
（此句尚無繁中翻譯）
11:56.820–11:57.440
（此句尚無繁中翻譯）
11:57.520–11:58.240
（此句尚無繁中翻譯）
11:58.240–11:59.680
（此句尚無繁中翻譯）
11:59.680–12:00.420
（此句尚無繁中翻譯）
12:00.420–12:01.920
（此句尚無繁中翻譯）
12:01.920–12:03.480
（此句尚無繁中翻譯）
12:03.740–12:04.560
（此句尚無繁中翻譯）
12:04.660–12:05.600
（此句尚無繁中翻譯）
12:05.600–12:06.180
（此句尚無繁中翻譯）
12:06.360–12:07.600
（此句尚無繁中翻譯）
12:07.900–12:08.620
（此句尚無繁中翻譯）
12:08.920–12:10.040
（此句尚無繁中翻譯）
12:10.040–12:11.660
（此句尚無繁中翻譯）
12:12.000–12:12.900
（此句尚無繁中翻譯）
12:12.900–12:13.520
（此句尚無繁中翻譯）
12:13.520–12:14.660
（此句尚無繁中翻譯）
12:15.140–12:16.480
（此句尚無繁中翻譯）
12:16.740–12:17.400
（此句尚無繁中翻譯）
12:17.400–12:18.460
（此句尚無繁中翻譯）
12:18.460–12:19.700
（此句尚無繁中翻譯）
12:19.840–12:20.520
（此句尚無繁中翻譯）
12:20.760–12:22.220
（此句尚無繁中翻譯）
12:22.660–12:23.780
（此句尚無繁中翻譯）
12:24.180–12:26.000
（此句尚無繁中翻譯）
12:26.000–12:27.340
（此句尚無繁中翻譯）
12:27.760–12:30.040
（此句尚無繁中翻譯）
12:30.040–12:31.180
（此句尚無繁中翻譯）
12:31.260–12:32.060
（此句尚無繁中翻譯）
12:32.060–12:32.880
（此句尚無繁中翻譯）
12:32.880–12:33.900
（此句尚無繁中翻譯）
12:33.900–12:34.920
（此句尚無繁中翻譯）
12:34.920–12:36.780
（此句尚無繁中翻譯）
12:36.980–12:38.080
（此句尚無繁中翻譯）
12:38.140–12:39.180
（此句尚無繁中翻譯）
12:39.440–12:40.440
（此句尚無繁中翻譯）
12:40.440–12:41.780
（此句尚無繁中翻譯）
12:41.780–12:43.200
（此句尚無繁中翻譯）
12:43.200–12:43.860
（此句尚無繁中翻譯）
12:44.240–12:45.200
（此句尚無繁中翻譯）
12:45.200–12:46.060
（此句尚無繁中翻譯）
12:46.060–12:46.780
（此句尚無繁中翻譯）
12:47.180–12:48.040
（此句尚無繁中翻譯）
12:48.040–12:48.580
（此句尚無繁中翻譯）
12:48.660–12:49.400
（此句尚無繁中翻譯）
12:49.400–12:50.640
（此句尚無繁中翻譯）
12:51.180–12:52.120
（此句尚無繁中翻譯）
12:52.120–12:52.800
（此句尚無繁中翻譯）
12:52.800–12:53.580
（此句尚無繁中翻譯）
12:53.580–12:54.220
（此句尚無繁中翻譯）
12:54.560–12:55.520
（此句尚無繁中翻譯）
12:55.520–12:56.840
（此句尚無繁中翻譯）
12:57.080–12:58.200
（此句尚無繁中翻譯）
12:58.820–12:59.420
（此句尚無繁中翻譯）
12:59.420–13:00.260
（此句尚無繁中翻譯）
13:00.560–13:01.320
（此句尚無繁中翻譯）
13:01.320–13:02.300
（此句尚無繁中翻譯）
13:02.640–13:03.340
（此句尚無繁中翻譯）
13:03.340–13:04.060
（此句尚無繁中翻譯）
13:04.060–13:05.200
（此句尚無繁中翻譯）
13:05.200–13:06.380
（此句尚無繁中翻譯）
13:06.380–13:06.960
（此句尚無繁中翻譯）
13:06.960–13:08.000
（此句尚無繁中翻譯）
13:08.000–13:08.880
（此句尚無繁中翻譯）
13:08.880–13:10.560
（此句尚無繁中翻譯）
13:10.780–13:11.840
（此句尚無繁中翻譯）
13:11.840–13:13.160
（此句尚無繁中翻譯）
13:13.720–13:14.260
（此句尚無繁中翻譯）
13:14.260–13:15.480
（此句尚無繁中翻譯）
13:15.840–13:16.640
（此句尚無繁中翻譯）
13:16.640–13:17.820
（此句尚無繁中翻譯）
13:18.320–13:19.460
（此句尚無繁中翻譯）
13:20.140–13:21.120
（此句尚無繁中翻譯）
13:21.120–13:22.700
（此句尚無繁中翻譯）
13:22.700–13:23.340
（此句尚無繁中翻譯）
13:24.140–13:24.740
（此句尚無繁中翻譯）
13:24.980–13:26.400
（此句尚無繁中翻譯）
13:26.400–13:27.560
（此句尚無繁中翻譯）
13:27.680–13:28.460
（此句尚無繁中翻譯）
13:28.460–13:29.920
（此句尚無繁中翻譯）
13:30.040–13:30.860
（此句尚無繁中翻譯）
13:30.980–13:31.780
（此句尚無繁中翻譯）
13:31.780–13:32.640
（此句尚無繁中翻譯）
13:32.640–13:33.720
（此句尚無繁中翻譯）
13:33.880–13:34.780
（此句尚無繁中翻譯）
13:34.780–13:35.860
（此句尚無繁中翻譯）
13:35.860–13:37.440
（此句尚無繁中翻譯）
13:37.680–13:38.600
（此句尚無繁中翻譯）
13:38.600–13:39.480
（此句尚無繁中翻譯）
13:39.480–13:40.200
（此句尚無繁中翻譯）
13:40.200–13:41.000
（此句尚無繁中翻譯）
13:41.000–13:42.300
（此句尚無繁中翻譯）
13:42.300–13:43.460
（此句尚無繁中翻譯）
13:43.460–13:44.680
（此句尚無繁中翻譯）
13:45.260–13:46.300
（此句尚無繁中翻譯）
13:46.300–13:47.040
（此句尚無繁中翻譯）
13:47.180–13:47.680
（此句尚無繁中翻譯）
13:48.140–13:49.320
（此句尚無繁中翻譯）
13:49.880–13:50.440
（此句尚無繁中翻譯）
13:50.440–13:51.120
（此句尚無繁中翻譯）
13:51.120–13:52.420
（此句尚無繁中翻譯）
13:52.560–13:53.520
（此句尚無繁中翻譯）
13:53.520–13:54.880
（此句尚無繁中翻譯）
13:54.880–13:55.780
（此句尚無繁中翻譯）
13:55.780–13:56.780
（此句尚無繁中翻譯）
13:57.400–13:58.140
（此句尚無繁中翻譯）
13:58.140–13:59.580
（此句尚無繁中翻譯）
13:59.580–14:01.080
（此句尚無繁中翻譯）
14:01.080–14:02.100
（此句尚無繁中翻譯）
14:02.100–14:03.260
（此句尚無繁中翻譯）
14:03.260–14:04.680
（此句尚無繁中翻譯）
14:04.680–14:06.100
（此句尚無繁中翻譯）
14:06.100–14:07.360
（此句尚無繁中翻譯）
14:07.500–14:08.380
（此句尚無繁中翻譯）
14:08.560–14:09.500
（此句尚無繁中翻譯）
14:09.500–14:10.420
（此句尚無繁中翻譯）
14:10.420–14:12.320
（此句尚無繁中翻譯）
14:12.320–14:13.980
（此句尚無繁中翻譯）
14:13.980–14:15.460
（此句尚無繁中翻譯）
14:15.460–14:16.420
（此句尚無繁中翻譯）
14:16.560–14:18.080
（此句尚無繁中翻譯）
14:18.280–14:19.140
（此句尚無繁中翻譯）
14:19.140–14:20.300
（此句尚無繁中翻譯）
14:20.300–14:21.560
（此句尚無繁中翻譯）
14:21.560–14:22.380
（此句尚無繁中翻譯）
14:22.380–14:23.860
（此句尚無繁中翻譯）
14:23.860–14:24.580
（此句尚無繁中翻譯）
14:24.580–14:25.180
（此句尚無繁中翻譯）
14:25.780–14:27.120
（此句尚無繁中翻譯）
14:27.120–14:28.480
（此句尚無繁中翻譯）
14:28.480–14:29.920
（此句尚無繁中翻譯）
14:29.920–14:31.060
（此句尚無繁中翻譯）
14:31.060–14:32.220
（此句尚無繁中翻譯）
14:32.220–14:33.700
（此句尚無繁中翻譯）
14:33.700–14:35.060
（此句尚無繁中翻譯）
14:35.060–14:36.380
（此句尚無繁中翻譯）
14:36.380–14:37.000
（此句尚無繁中翻譯）
14:37.000–14:38.080
（此句尚無繁中翻譯）
14:38.080–14:39.320
（此句尚無繁中翻譯）
14:39.320–14:39.960
（此句尚無繁中翻譯）
14:39.960–14:40.520
（此句尚無繁中翻譯）
14:40.600–14:42.000
（此句尚無繁中翻譯）
14:42.000–14:42.860
（此句尚無繁中翻譯）
14:42.860–14:43.480
（此句尚無繁中翻譯）
14:43.480–14:45.020
（此句尚無繁中翻譯）
14:45.020–14:45.880
（此句尚無繁中翻譯）
14:45.880–14:46.360
（此句尚無繁中翻譯）
14:46.400–14:47.040
（此句尚無繁中翻譯）
14:47.040–14:47.560
（此句尚無繁中翻譯）
14:47.560–14:48.500
（此句尚無繁中翻譯）
14:48.840–14:49.640
（此句尚無繁中翻譯）
14:49.640–14:51.000
（此句尚無繁中翻譯）
14:51.180–14:52.860
（此句尚無繁中翻譯）
14:52.860–14:53.560
（此句尚無繁中翻譯）
14:53.720–14:54.940
（此句尚無繁中翻譯）
14:54.940–14:56.540
（此句尚無繁中翻譯）
14:56.540–14:57.460
（此句尚無繁中翻譯）
14:57.600–14:58.220
（此句尚無繁中翻譯）
14:58.220–14:59.120
（此句尚無繁中翻譯）
14:59.120–15:00.320
（此句尚無繁中翻譯）
15:00.320–15:01.180
（此句尚無繁中翻譯）
15:01.660–15:02.280
（此句尚無繁中翻譯）
15:02.280–15:02.800
（此句尚無繁中翻譯）

0:00.000–0:01.900
Anthropic released a new feature yesterday.
（此句尚無繁中翻譯）
0:02.180–0:03.460
It's called Claude Tag.
（此句尚無繁中翻譯）
0:03.600–0:05.580
And I think when most people saw it, they thought,
（此句尚無繁中翻譯）
0:05.700–0:09.260
okay, that's cool, a convenient way to invoke Claude
（此句尚無繁中翻譯）
0:09.260–0:10.980
from within my Slack instance.
（此句尚無繁中翻譯）
0:11.320–0:13.160
I use Slack, that sounds really cool.
（此句尚無繁中翻譯）
0:13.320–0:16.140
But the more I read about it and the more I thought about it,
（此句尚無繁中翻譯）
0:16.340–0:18.160
I actually started to get scared.
（此句尚無繁中翻譯）
0:18.460–0:23.740
This is Anthropic's entry point into owning all knowledge work.
（此句尚無繁中翻譯）
0:23.860–0:27.300
And this affects everybody, not just if you're writing code,
（此句尚無繁中翻譯）
0:27.300–0:30.260
but literally if you do anything in front of a computer,
（此句尚無繁中翻譯）
0:30.640–0:32.540
Anthropic is trying to own that.
（此句尚無繁中翻譯）
0:32.700–0:35.080
And so I'm gonna show you what Claude Tag is
（此句尚無繁中翻譯）
0:35.080–0:37.280
and then I'm gonna explain the bigger picture
（此句尚無繁中翻譯）
0:37.280–0:39.960
of why this is such a big deal.
（此句尚無繁中翻譯）
0:40.200–0:41.660
And by the way, if you wanna keep up
（此句尚無繁中翻譯）
0:41.660–0:44.420
with what Anthropic and OpenAI are releasing
（此句尚無繁中翻譯）
0:44.420–0:46.580
seemingly every week, like this video,
（此句尚無繁中翻譯）
0:46.760–0:47.760
subscribe to the channel.
（此句尚無繁中翻譯）
0:48.140–0:49.620
It really does help.
（此句尚無繁中翻譯）
0:49.900–0:50.780
Thank you in advance.
（此句尚無繁中翻譯）
0:51.080–0:52.760
All right, so let me show you what Claude Tag is.
（此句尚無繁中翻譯）
0:52.760–0:56.140
If you use Slack, it's effectively
（此句尚無繁中翻譯）
0:56.140–0:58.620
like just tagging another member of your team
（此句尚無繁中翻譯）
0:58.620–0:59.700
and talking to it.
（此句尚無繁中翻譯）
0:59.760–1:02.760
Except Claude Tag has all the context
（此句尚無繁中翻譯）
1:02.760–1:05.200
of everything you do inside your company.
（此句尚無繁中翻譯）
1:05.440–1:08.940
And it can actually get real world work done for you.
（此句尚無繁中翻譯）
1:09.100–1:10.620
You simply tag Claude.
（此句尚無繁中翻譯）
1:10.700–1:12.080
It chats with you.
（此句尚無繁中翻譯）
1:12.320–1:13.440
It knows who you are.
（此句尚無繁中翻譯）
1:13.500–1:14.940
It knows who your colleagues are.
（此句尚無繁中翻譯）
1:15.220–1:18.220
It's able to understand all the different documents
（此句尚無繁中翻譯）
1:18.220–1:19.340
you have inside your company.
（此句尚無繁中翻譯）
1:19.580–1:23.040
It reads all the conversations in the Slack channels.
（此句尚無繁中翻譯）
1:23.040–1:27.740
It is really building an entire graph of your company.
（此句尚無繁中翻譯）
1:28.000–1:30.440
And by the way, you don't always have to tag Claude.
（此句尚無繁中翻譯）
1:30.740–1:33.460
It's actively reading your conversations
（此句尚無繁中翻譯）
1:33.460–1:35.560
in quote unquote ambient mode.
（此句尚無繁中翻譯）
1:36.020–1:38.980
So again, it's just always sucking up
（此句尚無繁中翻譯）
1:38.980–1:40.580
all the data from your company.
（此句尚無繁中翻譯）
1:40.940–1:43.540
And Anthropic isn't mincing words
（此句尚無繁中翻譯）
1:43.540–1:45.560
about what this product is.
（此句尚無繁中翻譯）
1:45.680–1:48.540
They are literally saying Claude Tag is an evolution
（此句尚無繁中翻譯）
1:48.540–1:51.120
of their cash cow Claude code.
（此句尚無繁中翻譯）
1:51.740–1:55.000
Claude Tag is an evolution made more proactive
（此句尚無繁中翻譯）
1:55.000–1:57.280
and built to work with a full team.
（此句尚無繁中翻譯）
1:57.500–1:59.820
It's now one of the main ways we get things done
（此句尚無繁中翻譯）
1:59.820–2:00.320
at Anthropic.
（此句尚無繁中翻譯）
2:00.960–2:04.180
65% of our product team's code now comes
（此句尚無繁中翻譯）
2:04.180–2:05.480
from our internal version.
（此句尚無繁中翻譯）
2:05.680–2:08.260
So they're not even going into Claude code
（此句尚無繁中翻譯）
2:08.260–2:09.480
directly anymore.
（此句尚無繁中翻譯）
2:09.480–2:12.840
They are simply just typing from wherever they are.
（此句尚無繁中翻譯）
2:12.940–2:14.120
And that happens to be Slack.
（此句尚無繁中翻譯）
2:14.400–2:17.720
This is core infrastructure for Anthropic now.
（此句尚無繁中翻譯）
2:17.720–2:19.900
And they want it to be core infrastructure
（此句尚無繁中翻譯）
2:19.900–2:22.980
for every single company in the entire world.
（此句尚無繁中翻譯）
2:23.240–2:24.960
And again, I really want to highlight
（此句尚無繁中翻譯）
2:24.960–2:27.600
how big of a deal not only I think it is,
（此句尚無繁中翻譯）
2:27.840–2:29.420
but Anthropic thinks it is.
（此句尚無繁中翻譯）
2:29.660–2:30.980
This is Andre Karpathy,
（此句尚無繁中翻譯）
2:31.180–2:32.540
one of the most prominent minds
（此句尚無繁中翻譯）
2:32.540–2:35.720
in artificial intelligence who just joined Anthropic.
（此句尚無繁中翻譯）
2:36.040–2:39.460
This is a new paradigm for interacting with Claude
（此句尚無繁中翻譯）
2:39.460–2:41.580
that is significantly more in line
（此句尚無繁中翻譯）
2:41.580–2:44.680
with all the other human activity org wide.
（此句尚無繁中翻譯）
2:45.000–2:46.980
Interfaces are going away.
（此句尚無繁中翻譯）
2:46.980–2:49.140
That is effectively what he's saying.
（此句尚無繁中翻譯）
2:49.380–2:50.960
Don't go to Claude code anymore.
（此句尚無繁中翻譯）
2:51.080–2:52.320
Don't go to any of the apps.
（此句尚無繁中翻譯）
2:52.840–2:54.120
Wherever you're chatting,
（此句尚無繁中翻譯）
2:54.480–2:56.500
which is Slack for a lot of teams,
（此句尚無繁中翻譯）
2:56.680–2:58.580
that's where Claude tag is going to be.
（此句尚無繁中翻譯）
2:58.880–3:00.660
And let me just pause for a second.
（此句尚無繁中翻譯）
3:01.240–3:04.140
If Slack is where all the conversations are happening
（此句尚無繁中翻譯）
3:04.140–3:05.800
and Claude lives in Slack,
（此句尚無繁中翻譯）
3:06.420–3:08.740
what do you think Anthropic is going to build next?
（此句尚無繁中翻譯）
3:08.880–3:12.480
I guarantee they are going to build a Slack competitor.
（此句尚無繁中翻譯）
3:12.820–3:13.540
Okay, so let's continue.
（此句尚無繁中翻譯）
3:13.540–3:16.780
Once you do all of the under the hood engineering work
（此句尚無繁中翻譯）
3:16.780–3:18.460
to make this just work,
（此句尚無繁中翻譯）
3:18.760–3:21.820
acts across tools, integrations, compute environments,
（此句尚無繁中翻譯）
3:22.020–3:23.400
memory, security, et cetera,
（此句尚無繁中翻譯）
3:23.960–3:26.760
Claude basically joins the team in a seamless way.
（此句尚無繁中翻譯）
3:27.120–3:29.940
You can talk to it as you would talk to a person
（此句尚無繁中翻譯）
3:29.940–3:33.780
and it can help with a very large variety of workloads.
（此句尚無繁中翻譯）
3:34.320–3:36.480
They're positioning Claude tag
（此句尚無繁中翻譯）
3:36.480–3:38.640
as another member of your team.
（此句尚無繁中翻譯）
3:38.780–3:41.580
This is not a bot in your Slack app.
（此句尚無繁中翻譯）
3:41.980–3:46.320
This is not AI responding to people tagging it
（此句尚無繁中翻譯）
3:46.320–3:47.680
in your Slack app.
（此句尚無繁中翻譯）
3:47.980–3:49.460
They really do say,
（此句尚無繁中翻譯）
3:49.680–3:52.580
this is a new form of employee for your team.
（此句尚無繁中翻譯）
3:52.660–3:53.300
And guess what?
（此句尚無繁中翻譯）
3:53.840–3:55.420
You're renting it from Anthropic.
（此句尚無繁中翻譯）
3:55.420–3:58.580
And he goes on to make even bigger claims.
（此句尚無繁中翻譯）
3:59.120–3:59.840
In my opinion,
（此句尚無繁中翻譯）
4:00.240–4:05.480
this is the third major redesign of the LLM UI UX.
（此句尚無繁中翻譯）
4:05.980–4:09.540
The first paradigm was that the LLM is a website you go to.
（此句尚無繁中翻譯）
4:09.720–4:11.680
So think about Claude or ChatGPT.
（此句尚無繁中翻譯）
4:11.860–4:14.900
The second was that it is an app you download
（此句尚無繁中翻譯）
4:14.900–4:16.060
to your computer.
（此句尚無繁中翻譯）
4:16.520–4:18.020
Great, you can download ChatGPT,
（此句尚無繁中翻譯）
4:18.160–4:19.660
but really what he's talking about
（此句尚無繁中翻譯）
4:19.660–4:22.360
are apps like Codex and Claude Code.
（此句尚無繁中翻譯）
4:22.360–4:25.960
And then third, a self-contained, persistent,
（此句尚無繁中翻譯）
4:26.340–4:29.980
asynchronous entity with org-wide tools and context
（此句尚無繁中翻譯）
4:29.980–4:32.640
working alongside teams of humans.
（此句尚無繁中翻譯）
4:32.840–4:35.380
It really takes a while to wrap your head around it,
（此句尚無繁中翻譯）
4:35.420–4:37.420
but it works and it is awesome.
（此句尚無繁中翻譯）
4:37.680–4:40.300
And look, a lot of people gave Andre Karpathy criticism
（此句尚無繁中翻譯）
4:40.300–4:42.960
for this because yeah, he joined Anthropic
（此句尚無繁中翻譯）
4:42.960–4:44.760
and of course he's drinking the Kool-Aid.
（此句尚無繁中翻譯）
4:44.860–4:46.200
He believes in what they're building,
（此句尚無繁中翻譯）
4:46.420–4:48.080
but he also happens to be right.
（此句尚無繁中翻譯）
4:48.580–4:49.720
This is a big deal.
（此句尚無繁中翻譯）
4:49.720–4:53.720
You are now effectively hiring people,
（此句尚無繁中翻譯）
4:54.000–4:56.440
people, AI, from Anthropic.
（此句尚無繁中翻譯）
4:57.080–4:58.640
You have to pay them for that.
（此句尚無繁中翻譯）
4:59.140–5:01.100
And at the same time that you're doing that,
（此句尚無繁中翻譯）
5:01.240–5:03.380
you're handing them all the information about your company
（此句尚無繁中翻譯）
5:03.380–5:05.960
and you are renting that back from them.
（此句尚無繁中翻譯）
5:06.140–5:08.320
They are building an information graph
（此句尚無繁中翻譯）
5:08.320–5:12.520
about every single company that uses Claude Tag.
（此句尚無繁中翻譯）
5:12.780–5:15.720
And you need to decide if you're comfortable with that.
（此句尚無繁中翻譯）
5:15.880–5:16.640
More on that later.
（此句尚無繁中翻譯）
5:16.640–5:19.740
And by the way, it seems like I'm going to be paying Anthropic
（此句尚無繁中翻譯）
5:19.740–5:20.600
more and more.
（此句尚無繁中翻譯）
5:21.120–5:22.460
And to help me fund that,
（此句尚無繁中翻譯）
5:22.600–5:24.420
let me tell you about the sponsor of today's video.
（此句尚無繁中翻譯）
5:24.920–5:28.340
Context engineering is an incredibly important part
（此句尚無繁中翻譯）
5:28.340–5:30.300
of your AI stack.
（此句尚無繁中翻譯）
5:30.400–5:33.480
Making sure your AI has all of the data it needs
（此句尚無繁中翻譯）
5:33.480–5:35.460
to give you the best possible results
（此句尚無繁中翻譯）
5:35.460–5:37.900
is actually kind of a hard problem.
（此句尚無繁中翻譯）
5:38.240–5:39.780
And if you're just working in one document,
（此句尚無繁中翻譯）
5:39.940–5:41.460
sure, just give it that one document.
（此句尚無繁中翻譯）
5:41.760–5:42.320
It's easy.
（此句尚無繁中翻譯）
5:42.520–5:44.960
But imagine if you have thousands of documents
（此句尚無繁中翻譯）
5:44.960–5:46.840
and then you start mixing in videos
（此句尚無繁中翻譯）
5:46.840–5:49.420
and articles and other types of media.
（此句尚無繁中翻譯）
5:49.740–5:51.280
And as you're seeing from Gbrain
（此句尚無繁中翻譯）
5:51.280–5:53.520
and from Karpathy's LLM Wiki,
（此句尚無繁中翻譯）
5:53.800–5:56.660
people are building all of this stuff themselves.
（此句尚無繁中翻譯）
5:56.660–5:58.340
But it is not easy.
（此句尚無繁中翻譯）
5:58.660–6:01.000
That is where Recall 2.0 comes in.
（此句尚無繁中翻譯）
6:01.240–6:03.340
Recall 2.0 allows you to throw everything
（此句尚無繁中翻譯）
6:03.340–6:06.320
into this kind of ocean of data.
（此句尚無繁中翻譯）
6:06.520–6:07.960
And it can provide the right context
（此句尚無繁中翻譯）
6:08.480–6:11.060
at the right moment to your AI.
（此句尚無繁中翻譯）
6:11.440–6:12.580
With Recall 2.0,
（此句尚無繁中翻譯）
6:12.580–6:15.220
you can switch between chatting with your knowledge base
（此句尚無繁中翻譯）
6:15.220–6:16.820
and researching on the web.
（此句尚無繁中翻譯）
6:17.040–6:18.620
I can add all of my OpenClaw videos
（此句尚無繁中翻譯）
6:18.620–6:20.020
to my Recall knowledge base,
（此句尚無繁中翻譯）
6:20.260–6:22.920
ask what the best security practices are,
（此句尚無繁中翻譯）
6:23.100–6:25.260
and Recall will even give me timestamps
（此句尚無繁中翻譯）
6:25.260–6:26.820
that I can watch in the app.
（此句尚無繁中翻譯）
6:27.040–6:28.980
With an API and MCP,
（此句尚無繁中翻譯）
6:29.280–6:31.580
you can plug this into any system
（此句尚無繁中翻譯）
6:31.580–6:33.480
that you're already using and building.
（此句尚無繁中翻譯）
6:33.720–6:34.420
And the best part,
（此句尚無繁中翻譯）
6:34.480–6:36.100
you're not locked into one model.
（此句尚無繁中翻譯）
6:36.360–6:37.360
So go check out Recall.
（此句尚無繁中翻譯）
6:37.480–6:38.380
It's a great product.
（此句尚無繁中翻譯）
6:38.500–6:41.300
Use MB25 to get 25% off.
（此句尚無繁中翻譯）
6:41.300–6:43.940
And I'll drop the link down below in the description.
（此句尚無繁中翻譯）
6:44.300–6:45.080
Thanks again to Recall.
（此句尚無繁中翻譯）
6:45.160–6:46.040
Now back to the video.
（此句尚無繁中翻譯）
6:46.460–6:47.080
And by the way,
（此句尚無繁中翻譯）
6:47.540–6:50.140
Y Combinator has been talking about
（此句尚無繁中翻譯）
6:50.140–6:53.400
this kind of AI native company for months now.
（此句尚無繁中翻譯）
6:53.520–6:54.920
And they've put out videos,
（此句尚無繁中翻譯）
6:55.100–6:58.380
but nothing really concrete about how to build it,
（此句尚無繁中翻譯）
6:58.580–6:59.520
what it looks like,
（此句尚無繁中翻譯）
6:59.760–7:01.860
companies that are already using it.
（此句尚無繁中翻譯）
7:01.980–7:03.700
But they've been talking about it.
（此句尚無繁中翻譯）
7:03.860–7:06.720
And now this is what they're talking about.
（此句尚無繁中翻譯）
7:06.720–7:09.640
This is the beginning of the AI native company,
（此句尚無繁中翻譯）
7:09.780–7:13.720
where AI understands every inch and corner
（此句尚無繁中翻譯）
7:13.720–7:15.140
of your company
（此句尚無繁中翻譯）
7:15.140–7:18.160
and is able to run 24 hours a day,
（此句尚無繁中翻譯）
7:18.400–7:19.400
helping you grow,
（此句尚無繁中翻譯）
7:19.600–7:20.700
finding issues,
（此句尚無繁中翻譯）
7:20.880–7:22.200
helping your human employees
（此句尚無繁中翻譯）
7:22.200–7:24.320
be a thousand X more productive.
（此句尚無繁中翻譯）
7:24.880–7:25.860
This is the vision
（此句尚無繁中翻譯）
7:25.860–7:27.680
that they've been laying out for months now.
（此句尚無繁中翻譯）
7:27.760–7:28.200
And of course,
（此句尚無繁中翻譯）
7:28.260–7:29.640
Anthropic has been building it.
（此句尚無繁中翻譯）
7:29.760–7:31.020
This is Ankit Gupta,
（此句尚無繁中翻譯）
7:31.240–7:32.940
a general partner at Y Combinator.
（此句尚無繁中翻譯）
7:33.160–7:34.300
I like that at YC,
（此句尚無繁中翻譯）
7:34.300–7:35.540
the partners and software team
（此句尚無繁中翻譯）
7:35.540–7:37.740
have consistently built internal versions
（此句尚無繁中翻譯）
7:37.740–7:39.340
of pretty much every product I see
（此句尚無繁中翻譯）
7:39.340–7:41.080
open AI and Anthropic launch
（此句尚無繁中翻譯）
7:41.080–7:42.200
six to 12 months later.
（此句尚無繁中翻譯）
7:42.440–7:43.620
Let's see around corners
（此句尚無繁中翻譯）
7:43.620–7:44.760
for where things are going.
（此句尚無繁中翻譯）
7:44.900–7:45.960
So that's what I'm talking about.
（此句尚無繁中翻譯）
7:46.040–7:47.840
They've been talking about AI native companies,
（此句尚無繁中翻譯）
7:48.080–7:48.880
but they haven't put out
（此句尚無繁中翻譯）
7:48.880–7:50.380
any public facing products
（此句尚無繁中翻譯）
7:50.380–7:52.700
and they certainly haven't open sourced anything.
（此句尚無繁中翻譯）
7:52.900–7:54.420
And I'm going to point back
（此句尚無繁中翻譯）
7:54.420–7:56.600
to this incredible essay from Anthropic
（此句尚無繁中翻譯）
7:56.600–7:58.180
because this is effectively
（此句尚無繁中翻譯）
7:58.180–7:59.240
what they're trying to do
（此句尚無繁中翻譯）
7:59.240–8:01.380
for all knowledge work in all companies.
（此句尚無繁中翻譯）
8:01.740–8:02.580
Now, this is the essay
（此句尚無繁中翻譯）
8:02.580–8:03.540
in which they talk about
（此句尚無繁中翻譯）
8:03.540–8:05.340
self-improving artificial intelligence.
（此句尚無繁中翻譯）
8:05.620–8:06.820
I'm not going to talk about it again.
（此句尚無繁中翻譯）
8:06.880–8:08.040
I already made a video about it,
（此句尚無繁中翻譯）
8:08.120–8:09.440
but what we're effectively seeing
（此句尚無繁中翻譯）
8:09.440–8:10.260
is what happens
（此句尚無繁中翻譯）
8:10.260–8:12.280
when a company becomes AI native,
（此句尚無繁中翻譯）
8:12.460–8:14.120
when AI is fed
（此句尚無繁中翻譯）
8:14.120–8:15.980
every single piece of data
（此句尚無繁中翻譯）
8:15.980–8:17.060
about a company,
（此句尚無繁中翻譯）
8:17.480–8:19.340
about all the conversations happening,
（此句尚無繁中翻譯）
8:19.600–8:20.540
all the processes,
（此句尚無繁中翻譯）
8:20.740–8:21.640
all the documents,
（此句尚無繁中翻譯）
8:21.800–8:24.000
all the emails, everything.
（此句尚無繁中翻譯）
8:24.260–8:25.280
It's all of a sudden
（此句尚無繁中翻譯）
8:25.280–8:27.600
able to have this incredible visibility
（此句尚無繁中翻譯）
8:27.600–8:29.520
into what the company is doing.
（此句尚無繁中翻譯）
8:29.700–8:31.300
And then you can set goals.
（此句尚無繁中翻譯）
8:31.300–8:32.260
I've been talking a lot
（此句尚無繁中翻譯）
8:32.260–8:33.460
about loops lately.
（此句尚無繁中翻譯）
8:33.880–8:35.080
Imagine if you can set a loop
（此句尚無繁中翻譯）
8:35.080–8:36.440
for all knowledge work.
（此句尚無繁中翻譯）
8:36.660–8:37.960
The loop can simply be
（此句尚無繁中翻譯）
8:37.960–8:40.240
increase the revenue of my company
（此句尚無繁中翻譯）
8:40.240–8:42.880
and it is going to ingest
（此句尚無繁中翻譯）
8:42.880–8:44.260
all of the context,
（此句尚無繁中翻譯）
8:44.920–8:45.880
all of that understanding
（此句尚無繁中翻譯）
8:45.880–8:46.520
of the company,
（此句尚無繁中翻譯）
8:46.940–8:47.860
set up experiments,
（此句尚無繁中翻譯）
8:47.860–8:50.400
and then loop over it indefinitely.
（此句尚無繁中翻譯）
8:50.720–8:52.520
And if that sounds expensive,
（此句尚無繁中翻譯）
8:52.840–8:53.680
it is.
（此句尚無繁中翻譯）
8:53.720–8:54.860
It is going to take
（此句尚無繁中翻譯）
8:54.860–8:56.720
basically infinite tokens.
（此句尚無繁中翻譯）
8:57.140–8:58.760
And then we're at this point,
（此句尚無繁中翻譯）
8:58.760–9:00.980
and this is where it really gets scary,
（此句尚無繁中翻譯）
9:00.980–9:02.520
where whichever company
（此句尚無繁中翻譯）
9:02.520–9:04.940
can afford the most tokens,
（此句尚無繁中翻譯）
9:05.340–9:07.060
they're going to do the best
（此句尚無繁中翻譯）
9:07.060–9:08.680
because AI is going to improve
（此句尚無繁中翻譯）
9:08.680–9:10.300
their business the most.
（此句尚無繁中翻譯）
9:10.500–9:12.020
This sure sounds like the beginning
（此句尚無繁中翻譯）
9:12.020–9:13.700
of the permanent underclass.
（此句尚無繁中翻譯）
9:13.960–9:15.620
And then there was this tweet
（此句尚無繁中翻譯）
9:15.620–9:17.520
by Ashwin Gopinath,
（此句尚無繁中翻譯）
9:17.660–9:20.120
and he said it perfectly.
（此句尚無繁中翻譯）
9:20.940–9:23.720
Claude Tag is a Trojan horse,
（此句尚無繁中翻譯）
9:23.720–9:25.800
and I could not agree more.
（此句尚無繁中翻譯）
9:25.800–9:27.360
Not because Anthropic
（此句尚無繁中翻譯）
9:27.360–9:28.480
is doing anything evil,
（此句尚無繁中翻譯）
9:28.800–9:31.200
because the incentives are obvious.
（此句尚無繁中翻譯）
9:31.640–9:32.560
So day one,
（此句尚無繁中翻譯）
9:32.640–9:33.880
this looks like a great feature.
（此句尚無繁中翻譯）
9:34.180–9:35.140
Tag Claude in Slack,
（此句尚無繁中翻譯）
9:35.400–9:36.460
let it follow the thread,
（此句尚無繁中翻譯）
9:36.540–9:37.320
remember context,
（此句尚無繁中翻譯）
9:37.440–9:38.060
connect the tools,
（此句尚無繁中翻譯）
9:38.220–9:38.940
break down tasks,
（此句尚無繁中翻譯）
9:39.080–9:39.620
chase work,
（此句尚無繁中翻譯）
9:39.680–9:41.060
and act like a teammate.
（此句尚無繁中翻譯）
9:41.460–9:42.680
But that's exactly the problem.
（此句尚無繁中翻譯）
9:43.020–9:44.380
The moment your AI vendor
（此句尚無繁中翻譯）
9:44.380–9:45.600
becomes a shared coworker,
（此句尚無繁中翻譯）
9:45.660–9:47.660
it stops being just a model provider.
（此句尚無繁中翻譯）
9:47.760–9:48.920
It starts becoming the place
（此句尚無繁中翻譯）
9:48.920–9:49.960
where work is interpreted,
（此句尚無繁中翻譯）
9:50.160–9:51.020
remembered, routed,
（此句尚無繁中翻譯）
9:51.020–9:52.800
and eventually executed.
（此句尚無繁中翻譯）
9:53.280–9:54.780
Imagine if you no longer
（此句尚無繁中翻譯）
9:54.780–9:57.420
had any employees on your team,
（此句尚無繁中翻譯）
9:57.560–9:58.560
and you were renting
（此句尚無繁中翻譯）
9:58.560–10:00.320
an AI employee from Anthropic.
（此句尚無繁中翻譯）
10:00.580–10:01.840
Now imagine that
（此句尚無繁中翻譯）
10:01.840–10:03.360
across the entire economy.
（此句尚無繁中翻譯）
10:03.880–10:05.880
Every single piece of knowledge work
（此句尚無繁中翻譯）
10:05.880–10:07.620
is now just rented
（此句尚無繁中翻譯）
10:07.620–10:10.600
from this one massive company.
（此句尚無繁中翻譯）
10:11.160–10:12.260
That's why I keep saying
（此句尚無繁中翻譯）
10:12.260–10:13.060
this is scary.
（此句尚無繁中翻譯）
10:13.320–10:14.760
So that is not model lock-in,
（此句尚無繁中翻譯）
10:14.900–10:16.440
that is context lock-in.
（此句尚無繁中翻譯）
10:16.520–10:18.060
You are now renting your company
（此句尚無繁中翻譯）
10:18.060–10:19.220
back from them.
（此句尚無繁中翻譯）
10:19.220–10:22.420
And yes, platform risk is very real.
（此句尚無繁中翻譯）
10:22.420–10:24.740
It was real before you could
（此句尚無繁中翻譯）
10:24.740–10:26.200
literally buy intelligence
（此句尚無繁中翻譯）
10:26.200–10:26.960
from a company.
（此句尚無繁中翻譯）
10:27.160–10:28.760
It was real when people
（此句尚無繁中翻譯）
10:28.760–10:29.820
were building apps
（此句尚無繁中翻譯）
10:29.820–10:31.640
on top of the Apple App Store
（此句尚無繁中翻譯）
10:31.640–10:32.880
and Apple got to decide
（此句尚無繁中翻譯）
10:32.880–10:34.340
if they liked your app enough
（此句尚無繁中翻譯）
10:34.340–10:35.360
to let you be there.
（此句尚無繁中翻譯）
10:35.500–10:36.760
And whether you were building
（此句尚無繁中翻譯）
10:36.760–10:37.740
on top of Facebook
（此句尚無繁中翻譯）
10:37.740–10:38.960
and they wanted you
（此句尚無繁中翻譯）
10:38.960–10:40.460
to use their API or not.
（此句尚無繁中翻譯）
10:40.740–10:41.900
Anytime you build
（此句尚無繁中翻譯）
10:41.900–10:43.720
on top of somebody else's platform,
（此句尚無繁中翻譯）
10:43.920–10:45.220
you have platform risk.
（此句尚無繁中翻譯）
10:45.500–10:46.920
And this is the ultimate form
（此句尚無繁中翻譯）
10:46.920–10:47.980
of platform risk.
（此句尚無繁中翻譯）
10:47.980–10:50.860
It is all knowledge work
（此句尚無繁中翻譯）
10:50.860–10:52.140
platform risk.
（此句尚無繁中翻譯）
10:52.240–10:52.940
He goes on,
（此句尚無繁中翻譯）
10:53.000–10:53.860
the pricing model
（此句尚無繁中翻譯）
10:53.860–10:55.060
makes it even more dangerous.
（此句尚無繁中翻譯）
10:55.440–10:57.140
A human coworker has a salary.
（此句尚無繁中翻譯）
10:57.660–10:59.320
Claude has unbounded
（此句尚無繁中翻譯）
10:59.320–11:00.740
tokenized activity.
（此句尚無繁中翻譯）
11:01.300–11:02.720
That means once again,
（此句尚無繁中翻譯）
11:02.860–11:04.000
there is a cap
（此句尚無繁中翻譯）
11:04.000–11:05.200
that you could pay somebody,
（此句尚無繁中翻譯）
11:05.300–11:06.200
a human worker.
（此句尚無繁中翻譯）
11:06.740–11:08.220
But if you're paying Claude,
（此句尚無繁中翻譯）
11:08.540–11:10.040
it is uncapped.
（此句尚無繁中翻譯）
11:10.160–11:11.620
You can literally pay them
（此句尚無繁中翻譯）
11:11.620–11:12.640
infinite dollars
（此句尚無繁中翻譯）
11:12.640–11:14.180
and they'll eat it up.
（此句尚無繁中翻譯）
11:14.460–11:15.220
And there's always
（此句尚無繁中翻譯）
11:15.220–11:16.340
more work to be done.
（此句尚無繁中翻譯）
11:16.340–11:18.300
Imagine running a company
（此句尚無繁中翻譯）
11:18.300–11:19.620
and being that dependent
（此句尚無繁中翻譯）
11:19.620–11:21.260
on another company.
（此句尚無繁中翻譯）
11:21.380–11:22.840
But it actually gets bigger
（此句尚無繁中翻譯）
11:22.840–11:24.020
and worse than that.
（此句尚無繁中翻譯）
11:24.160–11:24.920
What happens when
（此句尚無繁中翻譯）
11:24.920–11:26.220
every single company
（此句尚無繁中翻譯）
11:26.220–11:26.880
starts building
（此句尚無繁中翻譯）
11:26.880–11:27.900
on top of Anthropic?
（此句尚無繁中翻譯）
11:28.120–11:29.420
And I know I'm talking
（此句尚無繁中翻譯）
11:29.420–11:30.400
a lot about Anthropic,
（此句尚無繁中翻譯）
11:30.560–11:31.080
but that's because
（此句尚無繁中翻譯）
11:31.080–11:32.160
they released Claude Tag.
（此句尚無繁中翻譯）
11:32.280–11:33.740
But you know OpenAI
（此句尚無繁中翻譯）
11:33.740–11:35.160
is working on the same feature.
（此句尚無繁中翻譯）
11:35.420–11:37.140
Hopefully other companies as well.
（此句尚無繁中翻譯）
11:37.220–11:38.600
I hope there's more competition
（此句尚無繁中翻譯）
11:38.600–11:39.600
because competition
（此句尚無繁中翻譯）
11:39.600–11:41.560
will be the only solution
（此句尚無繁中翻譯）
11:41.560–11:42.300
for the problems
（此句尚無繁中翻譯）
11:42.300–11:43.440
that come from Anthropic
（此句尚無繁中翻譯）
11:43.440–11:44.680
owning all knowledge work.
（此句尚無繁中翻譯）
11:44.880–11:45.580
So imagine now
（此句尚無繁中翻譯）
11:45.580–11:46.520
you're a SaaS company.
（此句尚無繁中翻譯）
11:46.700–11:47.520
You have software.
（此句尚無繁中翻譯）
11:47.680–11:49.040
You sell it to other companies.
（此句尚無繁中翻譯）
11:49.040–11:50.320
But all of a sudden,
（此句尚無繁中翻譯）
11:50.520–11:51.800
Anthropic is now
（此句尚無繁中翻譯）
11:51.800–11:52.880
using their agents
（此句尚無繁中翻譯）
11:52.880–11:54.080
to operate your software
（此句尚無繁中翻譯）
11:54.080–11:55.860
on behalf of their customers
（此句尚無繁中翻譯）
11:55.860–11:56.700
that pay them.
（此句尚無繁中翻譯）
11:56.820–11:57.440
And guess what?
（此句尚無繁中翻譯）
11:57.520–11:58.240
Those customers
（此句尚無繁中翻譯）
11:58.240–11:59.680
are no longer
（此句尚無繁中翻譯）
11:59.680–12:00.420
going back
（此句尚無繁中翻譯）
12:00.420–12:01.920
into your user interface.
（此句尚無繁中翻譯）
12:01.920–12:03.480
They no longer log in.
（此句尚無繁中翻譯）
12:03.740–12:04.560
They have no need to.
（此句尚無繁中翻譯）
12:04.660–12:05.600
They just tell their agent
（此句尚無繁中翻譯）
12:05.600–12:06.180
what to do.
（此句尚無繁中翻譯）
12:06.360–12:07.600
So why would they ever?
（此句尚無繁中翻譯）
12:07.900–12:08.620
All of a sudden,
（此句尚無繁中翻譯）
12:08.920–12:10.040
your user interface
（此句尚無繁中翻譯）
12:10.040–12:11.660
becomes valueless.
（此句尚無繁中翻譯）
12:12.000–12:12.900
And so what are you
（此句尚無繁中翻譯）
12:12.900–12:13.520
at that point?
（此句尚無繁中翻譯）
12:13.520–12:14.660
You have the workflows.
（此句尚無繁中翻譯）
12:15.140–12:16.480
So remove the UI.
（此句尚無繁中翻譯）
12:16.740–12:17.400
What can an agent
（此句尚無繁中翻譯）
12:17.400–12:18.460
actually get done
（此句尚無繁中翻譯）
12:18.460–12:19.700
inside of your application?
（此句尚無繁中翻譯）
12:19.840–12:20.520
Those are the workflows.
（此句尚無繁中翻譯）
12:20.760–12:22.220
But how long does that last?
（此句尚無繁中翻譯）
12:22.660–12:23.780
Because really,
（此句尚無繁中翻譯）
12:24.180–12:26.000
the agent can just write code
（此句尚無繁中翻譯）
12:26.000–12:27.340
for those workflows.
（此句尚無繁中翻譯）
12:27.760–12:30.040
AI is extremely good
（此句尚無繁中翻譯）
12:30.040–12:31.180
at writing code.
（此句尚無繁中翻譯）
12:31.260–12:32.060
So why wouldn't it
（此句尚無繁中翻譯）
12:32.060–12:32.880
just write code
（此句尚無繁中翻譯）
12:32.880–12:33.900
to manage
（此句尚無繁中翻譯）
12:33.900–12:34.920
and then eventually
（此句尚無繁中翻譯）
12:34.920–12:36.780
overtake those workflows itself?
（此句尚無繁中翻譯）
12:36.980–12:38.080
As a software company,
（此句尚無繁中翻譯）
12:38.140–12:39.180
what are you at that point?
（此句尚無繁中翻譯）
12:39.440–12:40.440
You're simply a database.
（此句尚無繁中翻譯）
12:40.440–12:41.780
And you know what's easier
（此句尚無繁中翻譯）
12:41.780–12:43.200
than writing code
（此句尚無繁中翻譯）
12:43.200–12:43.860
for workflows?
（此句尚無繁中翻譯）
12:44.240–12:45.200
It's writing code
（此句尚無繁中翻譯）
12:45.200–12:46.060
to read and write
（此句尚無繁中翻譯）
12:46.060–12:46.780
from databases.
（此句尚無繁中翻譯）
12:47.180–12:48.040
Agents are incredibly
（此句尚無繁中翻譯）
12:48.040–12:48.580
good at that.
（此句尚無繁中翻譯）
12:48.660–12:49.400
They don't even care
（此句尚無繁中翻譯）
12:49.400–12:50.640
what kind of database it is.
（此句尚無繁中翻譯）
12:51.180–12:52.120
They will help you
（此句尚無繁中翻譯）
12:52.120–12:52.800
move your data
（此句尚無繁中翻譯）
12:52.800–12:53.580
from one database
（此句尚無繁中翻譯）
12:53.580–12:54.220
to another.
（此句尚無繁中翻譯）
12:54.560–12:55.520
They'll help you read
（此句尚無繁中翻譯）
12:55.520–12:56.840
and write from it easily.
（此句尚無繁中翻譯）
12:57.080–12:58.200
And then all of a sudden,
（此句尚無繁中翻譯）
12:58.820–12:59.420
what happens
（此句尚無繁中翻譯）
12:59.420–13:00.260
to all software?
（此句尚無繁中翻譯）
13:00.560–13:01.320
No software
（此句尚無繁中翻譯）
13:01.320–13:02.300
is safe from this.
（此句尚無繁中翻譯）
13:02.640–13:03.340
Agents are now
（此句尚無繁中翻譯）
13:03.340–13:04.060
infiltrating
（此句尚無繁中翻譯）
13:04.060–13:05.200
right before our eyes
（此句尚無繁中翻譯）
13:05.200–13:06.380
every single piece
（此句尚無繁中翻譯）
13:06.380–13:06.960
of software.
（此句尚無繁中翻譯）
13:06.960–13:08.000
That brings us
（此句尚無繁中翻譯）
13:08.000–13:08.880
to our final
（此句尚無繁中翻譯）
13:08.880–13:10.560
scariest point.
（此句尚無繁中翻譯）
13:10.780–13:11.840
Anthropic all of a sudden
（此句尚無繁中翻譯）
13:11.840–13:13.160
owns all knowledge work.
（此句尚無繁中翻譯）
13:13.720–13:14.260
They own
（此句尚無繁中翻譯）
13:14.260–13:15.480
all software.
（此句尚無繁中翻譯）
13:15.840–13:16.640
How does society
（此句尚無繁中翻譯）
13:16.640–13:17.820
continue at that point?
（此句尚無繁中翻譯）
13:18.320–13:19.460
And that's the problem.
（此句尚無繁中翻譯）
13:20.140–13:21.120
Government intervention
（此句尚無繁中翻譯）
13:21.120–13:22.700
is the last line
（此句尚無繁中翻譯）
13:22.700–13:23.340
of defense.
（此句尚無繁中翻譯）
13:24.140–13:24.740
Societally,
（此句尚無繁中翻譯）
13:24.980–13:26.400
we will not allow
（此句尚無繁中翻譯）
13:26.400–13:27.560
a single company,
（此句尚無繁中翻譯）
13:27.680–13:28.460
a single entity
（此句尚無繁中翻譯）
13:28.460–13:29.920
to have all the power.
（此句尚無繁中翻譯）
13:30.040–13:30.860
It just won't work.
（此句尚無繁中翻譯）
13:30.980–13:31.780
But there is
（此句尚無繁中翻譯）
13:31.780–13:32.640
a solution
（此句尚無繁中翻譯）
13:32.640–13:33.720
beyond that.
（此句尚無繁中翻譯）
13:33.880–13:34.780
Open source models,
（此句尚無繁中翻譯）
13:34.780–13:35.860
model competition
（此句尚無繁中翻譯）
13:35.860–13:37.440
from other research labs,
（此句尚無繁中翻譯）
13:37.680–13:38.600
making sure that
（此句尚無繁中翻譯）
13:38.600–13:39.480
whatever provider
（此句尚無繁中翻譯）
13:39.480–13:40.200
you choose
（此句尚無繁中翻譯）
13:40.200–13:41.000
for this Claude
（此句尚無繁中翻譯）
13:41.000–13:42.300
tag-like feature
（此句尚無繁中翻譯）
13:42.300–13:43.460
allows you to
（此句尚無繁中翻譯）
13:43.460–13:44.680
own the context.
（此句尚無繁中翻譯）
13:45.260–13:46.300
You own the context
（此句尚無繁中翻譯）
13:46.300–13:47.040
of your business,
（此句尚無繁中翻譯）
13:47.180–13:47.680
not them.
（此句尚無繁中翻譯）
13:48.140–13:49.320
You should own it.
（此句尚無繁中翻譯）
13:49.880–13:50.440
And that's why
（此句尚無繁中翻譯）
13:50.440–13:51.120
open source
（此句尚無繁中翻譯）
13:51.120–13:52.420
is so important right now.
（此句尚無繁中翻譯）
13:52.560–13:53.520
You're not going to be
（此句尚無繁中翻譯）
13:53.520–13:54.880
beholden to paying
（此句尚無繁中翻譯）
13:54.880–13:55.780
Anthropic prices
（此句尚無繁中翻譯）
13:55.780–13:56.780
for every token.
（此句尚無繁中翻譯）
13:57.400–13:58.140
You can pay
（此句尚無繁中翻譯）
13:58.140–13:59.580
the best price
（此句尚無繁中翻譯）
13:59.580–14:01.080
for the best token
（此句尚無繁中翻譯）
14:01.080–14:02.100
when you have
（此句尚無繁中翻譯）
14:02.100–14:03.260
a multi-model,
（此句尚無繁中翻譯）
14:03.260–14:04.680
multi-provider strategy.
（此句尚無繁中翻譯）
14:04.680–14:06.100
So what started
（此句尚無繁中翻譯）
14:06.100–14:07.360
as a cool,
（此句尚無繁中翻譯）
14:07.500–14:08.380
convenient feature,
（此句尚無繁中翻譯）
14:08.560–14:09.500
being able to tag Claude
（此句尚無繁中翻譯）
14:09.500–14:10.420
from within Slack
（此句尚無繁中翻譯）
14:10.420–14:12.320
is very obviously
（此句尚無繁中翻譯）
14:12.320–14:13.980
Anthropic's attempt
（此句尚無繁中翻譯）
14:13.980–14:15.460
to own
（此句尚無繁中翻譯）
14:15.460–14:16.420
all knowledge work.
（此句尚無繁中翻譯）
14:16.560–14:18.080
That is a very bad thing.
（此句尚無繁中翻譯）
14:18.280–14:19.140
But the one thing
（此句尚無繁中翻譯）
14:19.140–14:20.300
I do want you to remember
（此句尚無繁中翻譯）
14:20.300–14:21.560
and think about
（此句尚無繁中翻譯）
14:21.560–14:22.380
is this is
（此句尚無繁中翻譯）
14:22.380–14:23.860
definitely the future
（此句尚無繁中翻譯）
14:23.860–14:24.580
of how companies
（此句尚無繁中翻譯）
14:24.580–14:25.180
will work.
（此句尚無繁中翻譯）
14:25.780–14:27.120
AI and humans
（此句尚無繁中翻譯）
14:27.120–14:28.480
working in concert
（此句尚無繁中翻譯）
14:28.480–14:29.920
to accomplish things
（此句尚無繁中翻譯）
14:29.920–14:31.060
and AI
（此句尚無繁中翻譯）
14:31.060–14:32.220
living and breathing
（此句尚無繁中翻譯）
14:32.220–14:33.700
inside of your company,
（此句尚無繁中翻譯）
14:33.700–14:35.060
not just
（此句尚無繁中翻譯）
14:35.060–14:36.380
a chat app
（此句尚無繁中翻譯）
14:36.380–14:37.000
that you go
（此句尚無繁中翻譯）
14:37.000–14:38.080
type a query
（此句尚無繁中翻譯）
14:38.080–14:39.320
to get the response
（此句尚無繁中翻譯）
14:39.320–14:39.960
and then continue
（此句尚無繁中翻譯）
14:39.960–14:40.520
your work.
（此句尚無繁中翻譯）
14:40.600–14:42.000
It is going to be able
（此句尚無繁中翻譯）
14:42.000–14:42.860
to see everything
（此句尚無繁中翻譯）
14:42.860–14:43.480
you're doing
（此句尚無繁中翻譯）
14:43.480–14:45.020
inside that company
（此句尚無繁中翻譯）
14:45.020–14:45.880
and it's going to be able
（此句尚無繁中翻譯）
14:45.880–14:46.360
to help you.
（此句尚無繁中翻譯）
14:46.400–14:47.040
It's going to be able
（此句尚無繁中翻譯）
14:47.040–14:47.560
to make you
（此句尚無繁中翻譯）
14:47.560–14:48.500
so much more productive.
（此句尚無繁中翻譯）
14:48.840–14:49.640
So I wanted to end
（此句尚無繁中翻譯）
14:49.640–14:51.000
on that optimistic note.
（此句尚無繁中翻譯）
14:51.180–14:52.860
We just need more competition
（此句尚無繁中翻譯）
14:52.860–14:53.560
most of all.
（此句尚無繁中翻譯）
14:53.720–14:54.940
And I talked about Anthropic
（此句尚無繁中翻譯）
14:54.940–14:56.540
having self-improving
（此句尚無繁中翻譯）
14:56.540–14:57.460
artificial intelligence.
（此句尚無繁中翻譯）
14:57.600–14:58.220
I actually made
（此句尚無繁中翻譯）
14:58.220–14:59.120
an entire video
（此句尚無繁中翻譯）
14:59.120–15:00.320
breaking that down
（此句尚無繁中翻譯）
15:00.320–15:01.180
in detail.
（此句尚無繁中翻譯）
15:01.660–15:02.280
Go check it out
（此句尚無繁中翻譯）
15:02.280–15:02.800
right here.
（此句尚無繁中翻譯）

影片筆記：Anthropic is coming for EVERYTHING

一句話總結

影片深入分析 Anthropic 發布的 Claude Tag 功能，指出其透過 Slack 環境模式（Ambient Mode）建立公司知識圖譜，標誌著 AI 從「應用程式」轉向「無介面」的團隊成員範式；這雖能提升生產力，但也帶來極大的情境鎖定（Context Lock-in）與平台風險，可能導致 SaaS 介面價值歸零，並讓 Anthropic 壟斷所有知識工作，因此強調企業需透過開放原始碼與多供應商策略來制衡單一巨頭。

核心重點

Claude Tag 的戰略意義：

Claude Tag 並非單純的聊天機器人，而是嵌入 Slack 的「主動式」員工。
透過「環境模式」（ambient mode）主動讀取公司對話與文件，建立公司知識圖譜。
這標誌著從「網站/應用程式」介面轉向「無介面」的互動範式，介面將消失，AI 將成為團隊成員。

互動範式的演進：

第一範式：LLM 是一個網站（如 Claude 或 ChatGPT 網頁版）。
第二範式：LLM 是一個下載到電腦的應用程式（如 Codex 或 Claude Code）。
第三範式：一個自包含、持久、非同步的實體，擁有組織範圍的工具和上下文，與人類團隊並肩工作（如 Claude Tag）。

核心風險：情境鎖定與平台風險：

情境鎖定（Context Lock-in）：公司將所有資訊交給 Anthropic，再向他們租用 AI 員工，導致 Anthropic 掌握所有公司的數據圖譜。
平台風險：若整個經濟體的所有知識工作都從單一龐大公司租用，將產生極致的平台風險，類似 Apple App Store 或 Facebook API 的決定權問題，但規模更大。
SaaS 衝擊：若 Anthropic 使用其 Agent 代表客戶操作軟體，客戶將不再登入 SaaS 公司的使用者介面（UI），SaaS 公司的 UI 將變得無價值，僅剩工作流程。

經濟與定價風險：

人類同事有薪資上限，但 Claude 的活動是「無限定製化」（unbounded tokenized activity），成本無上限，可能消耗無限美元。
能夠負擔最多 token 的公司將獲得最大的 AI 業務改進，可能導致「永久底層階級」（permanent underclass）。

應對策略與未來展望：

依賴開放原始碼模型及其他研究實驗室帶來的模型競爭。
確保選擇的提供者允許企業擁有「上下文」（context），企業應擁有自身業務的上下文，而非供應商。
透過多模型、多供應商策略，企業可以選擇最佳價格與最佳 token，避免被綁定於單一供應商。
未來企業運作模式為 AI 與人類協作（working in concert），需透過競爭來制衡單一巨頭的權力。

詳細大綱

I. Claude Tag 的功能與運作機制

基本定義：
Anthropic 於昨日發布的新功能。
被視為在 Slack 實例中調用 Claude 的便捷方式。
被定位為 Claude Code 的進化版，更具主動性，旨在服務整個團隊。
運作方式：
標籤與互動：用戶在 Slack 中標籤（tag）Claude 即可進行對話。
情境感知：擁有公司內部所有操作的上下文，知道用戶及其同事身份。
數據讀取：閱讀 Slack 頻道中的所有對話，理解公司內部的不同文件。
環境模式（Ambient Mode）：即使未被標籤，Claude 也會主動讀取對話，持續吸收公司數據。
目標：建立整個公司的圖譜（graph）。

II. Anthropic 內部採用與戰略定位

內部採用情況：
Anthropic 產品團隊 65% 的程式碼來自其內部版本。
員工不再直接進入 Claude Code，而是透過 Slack 輸入指令。
已成為 Anthropic 的核心基礎設施（core infrastructure）。
Andre Karpathy 的觀點：
Karpathy 剛加入 Anthropic，被視為 AI 領域最具影響力的思想家之一。
提出與組織範圍內其他人類活動更一致的互動新範式。
介面消失論：介面將消失，用戶不應再前往特定應用程式（如 Claude Code），而是在聊天發生的地方（如 Slack）使用 Claude Tag。
定位：將 Claude Tag 定位為團隊的「新成員」或「員工」，而非單純的機器人。

III. 互動範式的演進

LLM UI/UX 的第三次重大重新設計：

第一範式：LLM 是一個網站（如 Claude 或 ChatGPT 網頁版）。
第二範式：LLM 是一個下載到電腦的應用程式（如 Codex 或 Claude Code）。
第三範式：一個自包含、持久、非同步的實體，擁有組織範圍的工具和上下文，與人類團隊並肩工作。

IV. 核心風險分析：情境鎖定與平台風險

情境鎖定（Context Lock-in）：
不僅是模型鎖定，而是公司將所有資訊交給 Anthropic，再向他們租用 AI 員工。
Anthropic 正在建立每個使用 Claude Tag 的公司的資訊圖譜。
用戶需決定是否對這種數據匯集感到舒適。
平台風險（Platform Risk）：
當 AI 供應商成為共享的同事時，它不再只是模型提供者，而是工作被解釋、記憶、路由和執行的地方。
若整個經濟體的所有知識工作都從單一龐大公司租用，將產生極致的平台風險。
類比：如同 Apple App Store 或 Facebook API 的決定權問題，但規模更大。
定價模型風險：
人類同事有薪資上限。
Claude 的活動是「無限定製化」（unbounded tokenized activity），成本無上限，可能消耗無限美元。
能夠負擔最多 token 的公司將獲得最大的 AI 業務改進，可能導致「永久底層階級」（permanent underclass）。

V. 對 SaaS 產業的衝擊

介面價值歸零：
若 Anthropic 使用其 Agent 代表客戶操作軟體，客戶將不再登入 SaaS 公司的使用者介面（UI）。
客戶直接告訴 Agent 要做什麼，無需使用 UI。
SaaS 公司的 UI 將變得無價值，僅剩工作流程（workflows）。
問題：Agent 在應用程式內實際能完成的工作流程能持續多久？

VI. AI 接管工作流程的邏輯與解決方案

AI 接管工作流程的邏輯：
AI 擅長編寫程式碼，因此可以直接為工作流程編寫程式碼。
最終 AI 將接管並管理這些工作流程。
軟體公司的角色將退化為單純的「資料庫」。
Agent 對軟體生態的滲透：
編寫程式碼來讀寫資料庫比為工作流程編寫程式碼更簡單。
Agent 不關心資料庫類型，能輕鬆協助數據在不同資料庫間移動。
Agent 正在滲透每一項軟體，沒有任何軟體能倖免。
單一實體壟斷的風險：
Anthropic 可能突然擁有所有知識工作與軟體。
社會與政府干預將是最後的防線，因為不允許單一實體擁有所有權力。
解決方案：開放原始碼與競爭：
依賴開放原始碼模型及其他研究實驗室帶來的模型競爭。
確保選擇的提供者允許企業擁有「上下文」（context）。
企業應擁有自身業務的上下文，而非供應商。
透過多模型、多供應商策略，企業可以選擇最佳價格與最佳 token，避免被綁定於單一供應商（如 Anthropic）的價格。
未來展望：樂觀的協作模式：
從 Slack 中的 Claude 標籤功能出發，揭示 Anthropic 試圖掌控所有知識工作的意圖。
未來企業運作模式：AI 與人類協作（working in concert）。
AI 將深入企業內部，不僅是聊天應用程式，而是能看見企業內部所有活動並協助提升生產力。
結論：需要更多競爭，並提及 Anthropic 的自我改進人工智慧（self-improving artificial intelligence）概念。

工具 / 模型 / 名詞整理

Anthropic：公司名稱。
Claude Tag：Anthropic 發布的新功能/產品，被定位為 Claude Code 的進化版，旨在服務整個團隊，運行於 Slack。
Slack：通訊軟體平台，Claude Tag 的運行環境。
Claude Code：Anthropic 的程式碼生成工具，被視為 Claude Tag 的前身或基礎。
OpenAI：被提及的競爭對手，據稱正在開發類似功能。
ChatGPT：被提及作為第一範式（網站）的例子。
Codex：被提及作為第二範式（應用程式）的例子。
Andre Karpathy：加入 Anthropic 的 AI 思想家，提出介面消失論。
Ankit Gupta：Y Combinator 的普通合夥人（General Partner）。
Y Combinator (YC)：創業加速器，討論「AI 原生公司」（AI native company）。
Ashwin Gopinath：發推文稱 Claude Tag 是「特洛伊木馬」（Trojan horse）的人。
Gbrain：被提及為正在自建數據處理工具的公司/專案。
Karpathy's LLM Wiki：被提及為自建數據處理工具的例子。
Recall 2.0：贊助商產品，用於知識庫管理與 AI 上下文提供。
OpenClaw：被提及為可加入 Recall 知識庫的影片來源。
MCP：被提及為 Recall 2.0 的整合技術之一。
API：被提及為 Recall 2.0 的整合技術之一。
Ambient Mode（環境模式）：Claude Tag 的功能，即使未被標籤，也會主動讀取對話，持續吸收公司數據。
Context Lock-in（情境鎖定）：公司將所有資訊交給 Anthropic，再向他們租用 AI 員工，導致 Anthropic 掌握所有公司的數據圖譜。
Platform Risk（平台風險）：若整個經濟體的所有知識工作都從單一龐大公司租用，將產生極致的平台風險。
Unbounded tokenized activity（無限定製化活動）：描述 Claude 活動成本無上限的特性。
Permanent underclass（永久底層階級）：可能因無法負擔大量 token 消耗而導致的社會/經濟現象。
Trojan horse（特洛伊木馬）：Ashwin Gopinath 對 Claude Tag 的比喻。
Self-improving artificial intelligence（自我改進的人工智慧）：提及 Anthropic 擁有的概念。

操作流程整理

企業導入 Claude Tag：

企業在 Slack 中啟用 Claude Tag。
員工透過在 Slack 中標籤（tag）Claude 進行對話。
Claude Tag 啟動「環境模式」（Ambient Mode），主動讀取 Slack 頻道對話與公司文件。
Claude Tag 建立公司知識圖譜，了解用戶及同事身份。

Anthropic 內部實踐：

Anthropic 員工透過 Slack 輸入指令，而非直接進入 Claude Code。
Claude Tag 成為 Anthropic 的核心基礎設施，產品團隊 65% 的程式碼來自其內部版本。

AI 接管工作流程（潛在未來）：

AI Agent 編寫程式碼來讀寫資料庫。
Agent 接管並管理工作流程，軟體公司角色退化為單純資料庫。
客戶直接告訴 Agent 要做什麼，不再登入 SaaS 公司的 UI。

企業應對策略：

採用多模型、多供應商策略。
確保選擇的提供者允許企業擁有「上下文」（context）。
依賴開放原始碼模型及研究實驗室帶來的競爭，避免被單一供應商綁定。

值得注意的限制或風險

情境鎖定（Context Lock-in）：

Anthropic 正在建立每個使用 Claude Tag 的公司的資訊圖譜。
公司將所有資訊交給 Anthropic，再向他們租用 AI 員工，導致 Anthropic 掌握所有公司的數據圖譜。
用戶需決定是否對這種數據匯集感到舒適。

平台風險（Platform Risk）：

當 AI 供應商成為共享的同事時，它不再只是模型提供者，而是工作被解釋、記憶、路由和執行的地方。
若整個經濟體的所有知識工作都從單一龐大公司租用，將產生極致的平台風險。
類比：如同 Apple App Store 或 Facebook API 的決定權問題，但規模更大。

定價模型風險：

人類同事有薪資上限。
Claude 的活動是「無限定製化」（unbounded tokenized activity），成本無上限，可能消耗無限美元。
能夠負擔最多 token 的公司將獲得最大的 AI 業務改進，可能導致「永久底層階級」（permanent underclass）。

SaaS 介面價值歸零：

若 Anthropic 使用其 Agent 代表客戶操作軟體，客戶將不再登入 SaaS 公司的使用者介面（UI）。
SaaS 公司的 UI 將變得無價值，僅剩工作流程（workflows）。
問題：Agent 在應用程式內實際能完成的工作流程能持續多久？

逐字稿辨識疑點

Claude Tag：逐字稿中多次出現此名稱，若為正式發布名稱則保留；若為口誤或聽寫錯誤，需查證是否為「Claude for Slack」或其他正式名稱。
cash cow：逐字稿中提到 "evolution of their cash cow Claude code"，此處語意稍顯突兀，需查證是否為特定內部術語或口誤。
Gbrain：逐字稿中提到 "from Gbrain"，需查證是否為特定公司名稱、產品名稱或聽寫錯誤（如 GitHub?）。
OpenClaw：逐字稿中提到 "OpenClaw videos"，需查證是否為特定頻道名稱、產品名稱或聽寫錯誤。
MB25：贊助商折扣碼 "MB25"，需查證是否為正確代碼或聽寫錯誤。
unbounded tokenized activity：此為逐字稿中的用詞，需查證是否為 Anthropic 官方文件或 Karpathy 演講中的確切術語，或為口語化描述。
permanent underclass：此為逐字稿中的用詞，需查證是否為特定經濟學或社會學術語，或為演講者的比喻。
Trojan horse：Ashwin Gopinath 的推文內容，需查證推文原文是否確為此比喻。
「Claude tag-like feature」：逐字稿中提及「whatever provider you choose for this Claude tag-like feature」，此處「tag-like」具體指代何種技術功能或介面元素，需查證其確切定義。
「Anthropic all of a sudden owns all knowledge work」：此處為講者的假設性或警告性陳述，指稱 Anthropic 可能突然擁有所有知識工作，需查證此為具體預測還是比喻說法。
「Anthropic having self-improving artificial intelligence」：逐字稿末尾提及 Anthropic 擁有「自我改進的人工智慧」，此為特定技術描述，需查證是否為該公司官方宣稱的具體技術名稱或功能。

可延伸追問

Claude Tag 的正式名稱與功能細節：

"Claude Tag" 是否為 Anthropic 官方發布的正式產品名稱？還是影片講者對 "Claude for Slack" 或類似功能的稱呼？
"Ambient Mode" 的具體技術實現方式為何？它如何確保數據隱私與安全？

情境鎖定與數據主權：

企業如何驗證 Anthropic 並未濫用其掌握的「公司知識圖譜」？
若 Anthropic 壟斷了所有知識工作，企業是否有足夠的談判籌碼來保護自身數據？

SaaS 產業的未來形態：

如果 SaaS 公司的 UI 變得無價值，它們將如何轉型？是否會轉向提供更深層的 API 或數據服務？
Agent 在應用程式內實際能完成的工作流程的邊界在哪裡？哪些任務仍需要人類介入 UI？

多供應商策略的可行性：

在實際操作中，如何有效整合不同模型的上下文（context）？
開放原始碼模型在企業級應用中的成熟度與可靠性如何？

Andre Karpathy 的「介面消失論」：

介面真的會完全消失嗎？還是會轉化為更

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