0:00.000–0:04.000
DeepSeq showed Wall Street that China could build cheap, powerful AI.
0:04.000–0:08.000
Now, Ji-Poo, the AI, is showing how that advantage could spread.
0:08.000–0:14.000
I've been consistently surprised by how quickly the open source has caught up.
0:14.000–0:20.000
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7
0:20.000–0:23.000
and some of these like very closed source frontier models.
0:23.000–0:29.000
What GLM 5.2 kind of shows is, you know, we might still be in this sort of three to six month territory.
0:29.000–0:34.000
Everything about the economics of AI changes depending on which outcome we're in.
0:34.000–0:38.000
We need costs to come down. That's how you drive adoption. That's how you get returns.
0:38.000–0:41.000
The AI leaderboard is starting to look incomplete.
0:41.000–0:44.000
The next fight may be over intelligence per dollar.
0:44.000–0:48.000
Models that are capable enough for real work and cheap enough to run constantly.
0:48.000–0:52.000
That is where ZAI is forcing a new conversation.
0:59.000–1:04.000
Ji-Poo's latest model, it's called GLM 5.2, comes out of China.
1:04.000–1:10.000
It has landed with a bang in Silicon Valley, blowing past all other open source models,
1:10.000–1:13.000
nearly matching the American frontier for just a fraction of the price.
1:13.000–1:19.000
Developers, they're piling in open router token traffic showing much quicker adoption for GLM 5.2
1:19.000–1:22.000
than DeepSeek's V4 launch back in April.
1:22.000–1:25.000
It was a big deal then, it's a big deal now.
1:25.000–1:28.000
This just carries the story that we've been reporting on even further.
1:28.000–1:32.000
It all matters because Ji-Poo is hitting at a different moment though.
1:32.000–1:37.000
The trillion dollar sell-off post-DeepSeek that was treated as a kind of a one-off shock,
1:37.000–1:39.000
partly because people saw it as a chat bot story.
1:39.000–1:42.000
GLM 5.2, this is different.
1:42.000–1:45.000
It's strong at agentic work and that is key.
1:45.000–1:53.000
On one agentic benchmark, it is just one percentage point away from Opus 4.8 for a fifth of the cost.
1:53.000–1:59.000
Now, Opus 4.8, that is Anthropics' most powerfully available model.
1:59.000–2:04.000
So basically, you're getting very close to the same horsepower for just a fraction of a cost.
2:04.000–2:08.000
And it really comes at a time when expensive AI is already eating into budgets.
2:08.000–2:14.000
And that is a really hard trade-off for enterprises, for companies, for Fortune 500 to ignore.
2:14.000–2:17.000
Agentic AI is only going to intensify those costs.
2:17.000–2:20.000
More complex tasks, more steps, more tokens.
2:20.000–2:25.000
So now that Ji-Poo, also known as Z.AI, is in the picture,
2:25.000–2:30.000
agentic open source, it might be the next big threat out of China.
2:30.000–2:34.000
And here's a part of the story, the AI story, that I think Wall Street is still missing here.
2:34.000–2:38.000
For the last few years, everyone has been obsessed with these AI leaderboards.
2:38.000–2:40.000
Who has the smartest model?
2:40.000–2:41.000
Who's the best at coding?
2:41.000–2:43.000
Which one is ahead on reasoning?
2:43.000–2:45.000
But that is not how companies buy software.
2:45.000–2:52.000
Companies are increasingly asking, what is good enough and what does it cost to run this thing a million times or more?
2:52.000–2:57.000
All of my employees running it a million times over the course of weeks, months, a year.
2:57.000–3:02.000
So the new metric in AI, it is intelligence per dollar.
3:02.000–3:06.000
And one reason Chinese models are pushing so hard on that metric is because of distillation,
3:06.000–3:12.000
which is basically you take a big expensive model, use it to train a smaller, cheaper model to act like it.
3:12.000–3:18.000
The American AI story, it has been built around bigger models, bigger data centers, huge spending.
3:18.000–3:25.000
But Chinese labs, they're putting these cheaper versions front and center, very, very close to the frontier performance without the cost.
3:25.000–3:27.000
That's why GPO is interesting.
3:27.000–3:30.000
Artificial analysis has this chart that looks at both sides of this.
3:30.000–3:33.000
How smart is a model and how much does it cost to run it?
3:33.000–3:38.000
Now, the most attractive quadrant, it's obvious, high performance, low cost.
3:38.000–3:40.000
That's where every company wants to be.
3:40.000–3:45.000
GPO's GLM 5.2, it's getting very close to that sweet spot.
3:45.000–3:48.000
It's not quite at the top in terms of performance.
3:48.000–3:50.000
You can see it here, very close to that green square.
3:50.000–3:58.000
So not at the very top in terms of performance, but it is close enough to the best models from OpenAI and Anthropic to make this price gap hard to ignore.
3:58.000–4:01.000
And on agentic work, that gap matters even more.
4:01.000–4:04.000
These are tasks where models don't just answer one question.
4:04.000–4:08.000
They plan, they code, test, fix mistakes, they loop, they keep going.
4:08.000–4:10.000
That is a lot more expensive.
4:10.000–4:14.000
So if you're a company, you need a great model, cheap enough to use over and over again.
4:14.000–4:15.000
And that is the setup.
4:15.000–4:18.000
I want to get into all of this with two people who are actually living this.
4:18.000–4:20.000
Aaron Levy, he's been running Box for 20 years.
4:20.000–4:23.000
He saw the move to the cloud, the move to mobile, the move to AI.
4:23.000–4:29.000
He is one of the most plugged in enterprise guys in the valley on what this stuff actually looks like when it hits a real company.
4:29.000–4:37.000
We've also got Gabe Pereira, co-founder of Harvey, which is the AI platform that half of the AM Law 100 is now running on.
4:37.000–4:45.000
Now, before that, he was a research scientist at DeepMind and Meta, so he has seen this from both sides, the model layer and the application layer.
4:45.000–4:47.000
Aaron and Gabe, it's great to have you both on.
4:47.000–4:53.000
We got Aaron too?
4:53.000–5:06.000
Okay, we're going to, I can't, let's see, do we have both their sound?
5:06.000–5:08.000
Okay, we're going to try and get this sound fixed.
5:08.000–5:11.000
So, Aaron, you're an observer now.
5:11.000–5:14.000
I'm assuming that you can hear Gabe and I, and then I'll come to you on this.
5:14.000–5:17.000
Gabe, let me start this off with what's happening with, you know, GLM.
5:17.000–5:24.000
It feels like my entire X feed, everyone I'm talking to here just cannot get enough of this new model GLM 5.2 right now.
5:24.000–5:26.000
Kind of feels like DeepSeq all over again.
5:26.000–5:30.000
People also aren't being shy and calling that maybe even bigger.
5:30.000–5:38.000
So, I guess first, how do you explain this to someone outside of tech or like a non-tech CEO who has been all in on American AI,
5:38.000–5:44.000
but sees all of this buzz around a new model that is almost as capable, but a lot cheaper?
5:44.000–5:54.000
Yeah, I would say the big question since kind of ChatGPT came out in these closed source models is just how big is the gap between the closed source and the open source?
5:54.000–6:00.000
And I think I've been consistently surprised by how quickly the open source has caught up.
6:00.000–6:10.000
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7 and some of these like very closed source frontier models.
6:10.000–6:17.000
And exactly to your point, it lets you think about the cost curves, where you need frontier intelligence, where you can use open source intelligence.
6:17.000–6:21.000
And I think this is what we're starting to figure out and a lot of companies are starting to figure out now.
6:21.000–6:25.000
Right. So explain that kind of in like just basic terms.
6:25.000–6:34.000
If you are a company CEO and you have this AI budget, but you've seen, you know, your company blowing through it, when you get something like this, what should they be thinking?
6:34.000–6:39.000
Especially when you say like it almost reaches the frontier, the very best models, but for a fraction of the price.
6:39.000–6:47.000
People you are talking to, Gabe, do you think that they're taking a second look at Chinese and these open source models or should they still be concerned about them?
6:47.000–6:56.000
Yeah, I think every company will use a mix of both, but exactly to what you said, it's starting to think about which tasks do I not need to use frontier intelligence.
6:56.000–7:03.000
And so as these models get smarter, do you need the smartest model reviewing your contracts or doing simple red lines?
7:03.000–7:10.000
And so I think for every company, the same way you organize a large company by figuring out, you know, I need different seniority.
7:10.000–7:12.000
I need to pay different people for different roles.
7:12.000–7:17.000
I think you're going to start seeing this with agents where there'll be some agents that are making critical decisions.
7:17.000–7:21.000
You want frontier intelligence and there'll be some agents that are doing simpler tasks.
7:21.000–7:26.000
And there you can use open source and you can really reduce the cost of of intelligence.
7:26.000–7:34.000
Right. And are you seeing sort of the buzz that I am as well as your feed, the people you're talking to, are they very excited about GLM five, two?
7:34.000–7:39.000
And like, can you explain why this is so interesting, why this is different than deep seek?
7:39.000–7:40.000
I tried to, but I'd love to hear it from you.
7:40.000–7:45.000
Yeah, I think we're seeing the same buzz in terms of our benchmarks.
7:45.000–7:48.000
It was kind of a big open source jump.
7:48.000–7:55.000
I think it's the first model that some of my friends at the labs were like, oh, this is like a very serious open source model.
7:55.000–7:58.000
And so I think it's kind of similar.
7:58.000–8:07.000
I would say to me, the deep seek moment was much more of I think there's always this question of how much money do you need to spend to stay on the frontier?
8:07.000–8:09.000
I don't think that's changing.
8:09.000–8:12.000
I think it's still very expensive to train these frontier models.
8:12.000–8:18.000
I think the question is just how quickly can open source catch up to that frontier as the frontier keeps moving?
8:18.000–8:22.000
Right. And that gap keeps like becoming narrower and narrower.
8:22.000–8:23.000
Aaron, I think we've got you now, right?
8:27.000–8:28.000
Yes. Okay. I can hear you.
8:28.000–8:31.000
You know, it's a live stream, Aaron. I know you're used to TV.
8:31.000–8:33.000
We kind of shoot from the hip here.
8:33.000–8:37.000
I, we, we, we need AGI as soon as we can get it.
8:37.000–8:40.000
So we will have audio figured out.
8:40.000–8:42.000
Yeah. Okay. Agreed.
8:42.000–8:43.000
Okay. Give me your thoughts.
8:43.000–8:44.000
I know you've been tweeting about this.
8:44.000–8:47.000
I know you've been tweeting a lot about open source.
8:47.000–8:50.000
You've been talking about this for years, kind of as I have as well.
8:50.000–8:52.000
What is your take on this moment?
8:52.000–8:54.000
Is it different than anything else we've seen over the last few years?
8:54.000–8:56.000
It certainly feels like that from where I sit.
8:56.000–8:59.000
Yeah. I mean, I totally concur with, with Gabe.
8:59.000–9:12.000
The, the, the, you know, there's been a big question in the Valley, which is how far behind is the open weights, kind of architecture from frontier closed sourced models.
9:12.000–9:16.000
And there's like fundamentally different outcomes in the future of AI.
9:16.000–9:24.000
If we see an, sort of an exponentially kind of increasing gap or a narrowing, or at least sort of sustaining three or six month gap.
9:24.000–9:38.000
And what, what, what GLM five two kind of shows is, you know, we might still be in this sort of three to six month territory, as opposed to the kind of fast takeoff scenario of closed weights models versus open weights.
9:38.000–9:43.000
And, and everything about the economics of AI changes depending on which outcome we're in.
9:43.000–9:53.000
I think to your, to your intro, I actually do think for the, for the next, you know, year or two years, actually frontier level intelligence still is incredibly important.
9:53.000–10:02.000
I think enterprises are only going to kind of peel off sort of workloads that are sort of good enough on, on open weights models.
10:02.000–10:09.000
When they, when they really feel like we've reached a level where you can kind of, you know, bank on, on the level of intelligence we're getting.
10:09.000–10:16.000
I still think that you, you, you need fable level kind of intelligence for a lot of the types of, of use cases that we're seeing in the enterprise.
10:16.000–10:27.000
But what this shows, and you saw the founder of, of ZAI really talk about, you know, the ability to get to fable level intelligence and, you know, within the year with an open weights model.
10:27.000–10:40.000
I think that, that does fundamentally change the architecture of AI because it means that now, whether you're using Harvey or box or factory or cognition or, you know, any, any other kind of neutral, you know, AI layer.
10:40.000–10:47.000
You can kind of route the workloads between frontier intelligence and open weights models as appropriate for, for your tasks.
10:47.000–10:53.000
So it's, it's very meaningful for what the future of the kind of AI stack looks like.
10:53.000–10:56.000
Right. So the big thing right now is model routing, right?
10:56.000–11:03.000
As Erin, as you laid out, it's use the frontier models for the hardest tasks, use the, you know, open source models for the easier ones.
11:03.000–11:07.000
But you did say for the next one to two years, you think that frontier is going to be really important.
11:07.000–11:11.000
And yes, the ZAI see our founder said that it was only a year.
11:11.000–11:13.000
So what happens after that?
11:13.000–11:17.000
Is that gap becoming even shorter?
11:17.000–11:22.000
Like you said, three to six months, does that last or does that eventually go away as a whole?
11:22.000–11:26.000
I think you'd be fine with a three to six month gap.
11:26.000–11:39.000
Because that, that's not enough time for, you know, the, the big issue is, is what is the gap where if I don't use frontier on everything that my firm or my organization loses some kind of meaningful competitive advantage.
11:39.000–11:58.000
And you can totally imagine a plausible scenario where if that gap was two to three years, you would, you would never use the kind of, you know, second best model because you just simply wouldn't be able to produce as effectively as a company in, in the code you're producing or the, the legal, you know, analysis you're doing for your clients or the, the marketing assets that you're generating.
11:58.000–12:07.000
But in a world where there's a three or six month difference, that's, that's totally within the noise of, of the, the general diffusion of AI anyway, within an organization.
12:07.000–12:10.000
So, so that, that's why you kind of see this fundamental shift.
12:10.000–12:20.000
You don't need open weights to be at the exact same, you know, kind of the same week as, as frontier intelligence, but you can't have it be two or three months, two or three years behind.
12:20.000–12:28.000
And it appears that we still are remaining that really kind of important zone where, where you're going to get very, very high quality open weights models.
12:28.000–12:34.000
And then, you know, that gets to the second part, which is now you can post train these models on just your tasks.
12:34.000–12:49.000
You know, everybody kind of looks at it as a, as a pure kind of cost of the token kind of equation, but actually just as big of a deal is the fact that now I can go and post train this model just on the purpose built use cases that my, you know, particular domain requires.
12:49.000–13:02.000
And that has a different set of performance access, axes that matter, which is like, what are the capabilities I need the model to be able to do that actually now can be like built into the model in a way that, that you don't always get from kind of general purpose intelligence.
13:02.000–13:13.000
Right. And that changes the dynamics, like once again, right, Gabe, I was going to come to you because you're actually operating in this app layer and like at the beginning, you're dependent on the frontier labs, open AI, Anthropic.
13:13.000–13:16.000
And that's what we're starting to do for kind of both costs and performance reasons.
13:16.000–13:22.000
And I think to the kind of longer term question of what is going to happen with frontier intelligence, I think the shift we're starting to see from models to agents.
13:22.000–13:27.000
I think when you just use models, it was easy to kind of train a model and then that was separate from the rest of your business.
13:27.000–13:31.000
And you could swap these out. If a model got better, that was kind of easy to replace.
13:31.000–13:34.000
And then you could swap these out. If a model got better, that was kind of easy to replace.
13:34.000–13:39.000
And then you could swap these out. If a model got better, that was kind of easy to replace.
13:39.000–13:48.000
As you start moving to the to these agentic systems, this AI, these agents are getting so embedded in your business that creating the next version of frontier intelligence.
13:48.000–13:53.000
I think a lot of that training data is going to come from the companies themselves.
13:53.000–14:19.000
And so as a simple example, when a law firm works for like a large private equity firm and they help them do M&A, like that training data and creating frontier models can do that type of work like that training data can't go into the general models.
14:19.000–14:30.000
And so I think there's going to be this really interesting question in the future of where is all of that training data eval coming from as these models start doing more and more economically valuable work.
14:30.000–14:40.000
And part of the motivation for us of training these open source models is how do we allow our customers and each of our law firms or kind of in-house teams to build these systems?
14:40.000–14:47.000
Right. And keep them for yourselves, right? You don't want to train someone else's model so that they can go and compete with you.
14:47.000–14:58.000
What I'm hearing from both of you does not sound good for Frontier Labs or a company that has built its entire business model on staying ahead. Am I reading that right, Aaron?
14:58.000–15:05.000
I feel like you're not going to say that, though, because you work with all of them. But like, what is the case for continuing to buy AI?
15:05.000–15:08.000
I feel like I'm trying to ask that. What's that? Sorry?
15:08.000–15:18.000
You know, what's the case for continuing to buy AI from Anthropic or OpenAI when that gap you essentially said is like, we're kind of there. It doesn't matter anymore. We're there.
15:18.000–15:40.000
Well, I actually didn't say that. And this is where it's very nuanced. I think what you're going to see is sort of this barbell dynamic where Frontier Intelligence, again, you know, kind of might be always at two or three, five, 10% better than whatever the open weights equivalent is at any given time.
15:40.000–15:59.000
Like that's the kind of scenario that as just a mental model is sort of, you know, you might want to just sort of, you know, kind of establish. And so what you'll have is that you'll still actually have an incredible demand for Frontier Intelligence, but often as things like the orchestrator or the planner or the reviewer of the work that's happening.
15:59.000–16:07.000
And so interestingly, you might spend the same amount of money on both Frontier Intelligence and Open weights intelligence, but the token volume might be entirely different.
16:07.000–16:17.000
You might have, you might have, you know, a third of or a fifth or a tenth of the tokens going into the Frontier Intelligence. Those will actually cost five or 10 times as much.
16:17.000–16:28.000
But the bulk of sort of the heavy document processing, heavy analysis and combing through every sort of, you know, piece of text in the repository that you're trying to analyze.
16:28.000–16:33.000
All of that might be happening on, again, a cheaper, faster model that is sort of more tuned for that workflow.
16:33.000–16:41.000
So I actually don't know that this spells anything negative for even the Frontier Labs right now, assuming you are bullish on AI.
16:41.000–16:48.000
If you're bullish on AI, what happens is we're going to be using, you know, a thousand times more tokens in 10 years from now than we are today.
16:48.000–16:55.000
And you're still going to want Frontier Intelligence for a lot of your most important sort of work that an agent is doing.
16:55.000–17:05.000
But there's just so many tokens that you now need for these agentic systems that actually you do have to then have something that is sort of cheaper and faster for a large amount of the work.
17:05.000–17:13.000
The other nuance that I would just throw out there is none of this means that a Frontier Lab won't have cheaper and faster models.
17:13.000–17:18.000
And so right now we're seeing that come from these open weights models as an example.
17:18.000–17:23.000
But, you know, OpenAI released an open weights model a year ago as an example.
17:23.000–17:30.000
And there's sort of no reason why if they became a great inference cloud as an example, that you wouldn't just run that model in their environment.
17:30.000–17:47.000
So it's not obvious that this kind of tells you really anything about the market structure at the Frontier Labs other than the market structure of the applied layer is now the very logical place to have a lot of your routing go because it is sort of independent and model agnostic.
17:47.000–17:50.000
And so it's going to just route the workload depending on what the use cases.
17:50.000–17:58.000
Right. So that, you know, three to six month gap that is going to exist is going to be worth paying for, you're saying, for some of the hardest tasks.
17:58.000–18:02.000
But, you know, Aaron, you bring up a good point. Yes.
18:02.000–18:11.000
OpenAI, Google, Gemini have trained open source models, but they don't show up in terms of the adoption data and really on the benchmarks.
18:11.000–18:15.000
Why are they letting the Chinese models run away with this?
18:15.000–18:20.000
And I mean, right now it feels like it's maybe looking for something from reflection.
18:20.000–18:26.000
But where are the Frontier Labs here? Do they turn their attention to it?
18:26.000–18:28.000
It's hard to make money from open source. Is that part of it?
18:28.000–18:30.000
Gabe, maybe you can go first.
18:30.000–18:35.000
Yeah, I think it's kind of depends on where you are in the stack.
18:35.000–18:42.000
So if you're kind of the chip providers, the cloud providers, I think you want there to be a lot of good open source models.
18:42.000–18:46.000
I think if you're a closed source, you'd rather this gap be a bit bigger.
18:46.000–18:50.000
But I think to Aaron's point, there's room for both of these.
18:50.000–19:00.000
And I think like the US versus China have just kind of taken different strategies of I think we are pushing the frontier with closed and their strategy is kind of doing more open.
19:00.000–19:03.000
But I think you've seen companies flip flop. Right.
19:03.000–19:08.000
You've seen companies catch up and then go close source or, you know, release open source.
19:08.000–19:12.000
And so I think it's not clear what the right strategy is yet.
19:12.000–19:15.000
But I think China's like open source models are getting like very good.
19:15.000–19:17.000
And that's something to like pay attention to.
19:17.000–19:22.000
Right. And I wanted to sort of make this connection to I know, Gabe, like Harvey is looking at post training.
19:22.000–19:24.000
I don't know what kind of model you're going to use.
19:24.000–19:27.000
Is it a Chinese open source model, do you think?
19:27.000–19:29.000
We need to support kind of a range of these.
19:29.000–19:33.000
So we think of it like the same way we're agnostic to closed source models.
19:33.000–19:39.000
We need to be agnostic to open source or we have some customers that can't use Chinese models for certain sensitive legal work.
19:39.000–19:42.000
We have European customers that want to use Mistral.
19:42.000–19:44.000
We have customers that want us models.
19:44.000–19:49.000
So I think you'll want kind of this healthy ecosystem of open source models by region, by country.
19:49.000–19:51.000
Yeah, that's right.
19:51.000–20:00.000
Yeah. And I would I would say that that if I if I could kind of like design the perfect strategy of, you know, kind of one of the top few labs and Google is sort of the closest to this with with Gemma.
20:00.000–20:11.000
I actually think you would want to consistently have a just close to frontier open weights model because now you kind of you get the ecosystem effects on both sides.
20:11.000–20:22.000
And and that barbell that I was talking about of you want frontier intelligence from GPD, you know, 5.5 or Fable or, you know, the rumored GPD 5.6.
20:22.000–20:32.000
But there's no reason that the the cheaper, faster version isn't in the same kind of model family from the same from the same sort of underlying weights and provider.
20:32.000–20:37.000
So I actually think from the game theory standpoint, you'd probably want to be heavily supporting open weights.
20:37.000–20:42.000
Now, there's a probably a philosophical reason why Anthropic in particular might not do that strategy.
20:42.000–20:50.000
But for open AI where they already have precedent for doing so for Google, which clearly is doing that with Gemma, which is actually pretty actively used.
20:50.000–20:56.000
It's just not not as obvious that it's at the at the frontier of a GLM or or a Kimmy.
20:56.000–21:02.000
I think actually, you know, from a game theory standpoint, they probably should be pursuing this strategy and historical. Right.
21:02.000–21:06.000
I mean, they sort of led the way on open source and only ecosystem for Android.
21:06.000–21:11.000
Why wouldn't they do that again? It surprises me, too, that we don't hear or see more.
21:11.000–21:15.000
I do think about cursor as well. Right. That was sort of laid the playbook.
21:15.000–21:21.000
Gabe, maybe trained a custom model built on Kimmy's architecture. That's Moonshot, a Chinese company.
21:21.000–21:27.000
It's now one of the hottest AI companies right now. And the whole secret sauce runs on a Chinese base model.
21:27.000–21:32.000
And it feels like nobody actually remembers that or cares that much, especially in the corporate world.
21:32.000–21:40.000
So you can see how you can sort of use these Chinese open source models to get ahead just in terms of especially if you're on the app layer.
21:40.000–21:44.000
Aaron, though, I wonder like because I love that you bring up game theory.
21:44.000–21:47.000
It feels like there's a lot of that going on right now.
21:47.000–21:52.000
And so when I think about Fable and Carson, one of our viewers wrote in with a question on this, too,
21:52.000–21:57.000
it kind of seems like a disaster for Anthropic.
21:57.000–22:03.000
I mean, right after they cut off access to Fable Mythos, their most powerful model,
22:03.000–22:11.000
you had all of these sort of policymakers around the world say, you know, how can you build on a model when your access could be cut off at any time?
22:11.000–22:14.000
And it felt like it just gave so much momentum to open source.
22:14.000–22:16.000
Do you think that that matters?
22:16.000–22:19.000
Do you think that, you know, that was misplayed by Anthropic?
22:19.000–22:21.000
What are the what's the impact of it?
22:21.000–22:29.000
Well, it's it's not obvious that well, it kind of depends on, I guess, what what, you know,
22:29.000–22:35.000
you know, either who you kind of like, you know, maybe blame for the situation we're in or kind of what transpired.
22:35.000–22:43.000
But because it might have been misplayed by the US government for the same for the same kind of economic and national security reasons long term.
22:43.000–22:54.000
If if if if you're kind of like encouraging sovereign AI in these other countries by sort of, you know, being the first to show that will will block access to an AI model through export controls.
22:54.000–23:02.000
So so, you know, I actually think from a game theory standpoint, it might be more on the US government side to to have to kind of consider the implications of.
23:02.000–23:20.000
And but absolutely, like like if you are another country right now, this was a sort of shock to the system that says, oh, actually, like, like, AI is not going to be like most other technologies where you sort of, you know, just kind of assume that software is going to work globally.
23:20.000–23:33.000
There's a few nuances where, you know, there might be some privacy requirements in different regions, but generally it's considered just like a an export that the US is always just going to put out there and and to our kind of, you know, kind of economic benefit.
23:33.000–23:52.000
AI, because the export itself is coming with with, you know, possible risks in at least in the table case, you know, the perceived risk of cybersecurity or other factors out now it's it's showing up as maybe it's something that the government wants a bit more control over and who should have access to it, at least now the precedent has been set for that.
23:52.000–24:03.000
So then, you know, by definition, what another country should at least be doing is saying, oh, actually, I need to be able to protect my my ability to have access to intelligence.
24:03.000–24:10.000
And if you were, you know, the EU and if you were, I mean, you know, China's fine because they're building their own models.
24:10.000–24:21.000
You probably would sort of say, hey, we should actually be, you know, fueling, you know, more of either open source investment or post trained, you know, models and be able to get really, really good at making these things better and better.
24:22.000–24:28.000
At a minimum is a hedge, you know, maybe you still do as much work as you can with US labs.
24:28.000–24:31.000
But now you do need some sort of, you know, backstop.
24:31.000–24:51.000
Because if for some reason, we treated access to AI as another kind of geopolitical or economic kind of weapon, you could see, you know, incredibly interesting consequences as a result of that, you know, imagine, imagine AI being used as a lever in in trade negotiations or trade deals in five years.
24:51.000–24:59.000
You would you would sort of that would that could very much threaten your your kind of economic and national security if that was the case.
24:59.000–25:11.000
So so we're, you know, it's really interesting because what's happening is we have these sort of blunt instruments, you know, export controls is a really interesting blunt instrument that I, you know, again, like very smart people probably predicted this.
25:11.000–25:18.000
But like a year ago, I would not have put that in the in the board of possible chess moves of blocking access to an AI model.
25:18.000–25:22.000
Well, now that that's that now that's a possible move, you sort of open up.
25:22.000–25:26.000
Well, well, wait a second, maybe this is like going to be treated as like nuclear arms or not.
25:26.000–25:41.000
I mean, I mean, not nuclear arms, but, you know, any kind of like weapon, in which case, you know, the set of controls that the government has of who has access to it is is way greater than what we had sort of considered before, which again, then means downstream as another country.
25:41.000–25:47.000
You have to be preparing for that. What you just laid out, though, is kind of like exactly what we saw with chips.
25:47.000–25:51.000
Right. And I'm with you. I never thought that could apply to models, but you saw that.
25:51.000–25:57.000
So this is just telling sort of every government around the world, you need sovereign AI, you need to own something.
25:57.000–26:05.000
And this process of distillation, which chips were actually not even as blunt of an instrument, though chips were even more sort of fine grained.
26:05.000–26:11.000
This was this was this was very, very blunt is like non U.S. citizens is like, wow.
26:11.000–26:15.000
OK, there's like no no real way to kind of execute that at scale.
26:15.000–26:20.000
It's even more blunt, yet the solution is like sitting right there for anyone to take.
26:20.000–26:23.000
Right. And it's just distillation. That is what the Chinese models are doing.
26:23.000–26:33.000
It's still inside baseball, but it's this idea that you take a big, expensive, powerful model and use its outputs to train a smaller, cheaper model to behave like it.
26:33.000–26:38.000
And Gabe, I wonder, like, this is how Chinese AI is spreading around the world.
26:38.000–26:41.000
And what Aaron just laid out is the case for governments.
26:41.000–26:44.000
But, you know, Harvey operates in a very sensitive industry.
26:44.000–26:51.000
Do you see the same thing happening? Have you seen the same blowback happen after what happened with Mythos and Fable?
26:51.000–26:54.000
Yeah, I would say distillation is one way.
26:54.000–26:59.000
But I I think these Chinese labs are actually just training their own models.
26:59.000–27:06.000
Like I think it's a bit like extreme for us to just say the only reason they're catching up is because they're distilling this.
27:06.000–27:15.000
They're U.S. models. I actually think they're doing incredible research and like their open source, even independent of distillation, I think is competitive or ahead of ours.
27:15.000–27:21.000
And then I think what you're seeing the labs do is making it harder to distill their models like that's what you saw with Fable.
27:21.000–27:25.000
And it's like I think that strategy will work as these models get smarter.
27:25.000–27:30.000
You can just tell the model, don't let users distill me when they use them.
27:30.000–27:33.000
And it's as they get smarter, they'll get better at preventing that.
27:33.000–27:35.000
So I think that is defensible.
27:35.000–27:48.000
But then I think the kind of the blowback we saw is kind of exactly what Aaron mentioned of even within the U.S., as these systems start becoming mission critical, like you just need to think about business continuity.
27:48.000–27:54.000
And if one of the labs that you're using runs out of compute, the government prevents you from accessing their models.
27:54.000–27:59.000
And so I think that's the way that you're using the data to be able to do that.
27:59.000–28:02.000
And so I think that's the way that you're using the data to be able to do that.
28:02.000–28:04.000
And so I think that's the way that you're using the data to be able to do that.
28:04.000–28:06.000
And so I think that's the way that you're using the data to be able to do that.
28:06.000–28:08.000
And so I think that's the way that you're using the data to be able to do that.
28:08.000–28:11.000
And so I think that's the way that you're using the data to be able to do that.
28:11.000–28:14.000
And so I think that's the way that you're using the data to be able to do that.
28:14.000–28:18.000
And so I think that's the way that you're using the data to be able to do that.
28:18.000–28:20.000
And so I think that's the way that you're using the data to be able to do that.
28:20.000–28:25.000
That's so fascinating. So you see it among sovereigns and then you also see it among corporate America.
28:25.000–28:28.000
And that's why I guess I started by saying it felt like a miscalculation.
28:28.000–28:31.000
You're right, Aaron, probably by the government, less so anthropic.
28:31.000–28:33.000
But the result is sort of the same.
28:33.000–28:40.000
Aaron, I want to give you a word on distillation, too, because, you know, I feel like the smartest people I talk to say that distillation should not be a dirty word.
28:40.000–28:48.000
Right. When you see that anthropic is suing Alibaba, that is non sanctioned distillation, but can actually be a really positive force.
28:48.000–28:52.000
Right. Turning expensive AI into cheaper products that can actually scale.
28:52.000–28:57.000
And as Gabe was saying, like, do you agree that the Chinese are really doing some stuff on the leading edge, innovative edge?
28:57.000–29:01.000
And it's too simplistic to say they're just distilling.
29:01.000–29:08.000
Yeah, I would say I'm not as close to the latest views on what the new method, what the latest methods are.
29:08.000–29:12.000
And I would probably bias toward Gabe's view of that.
29:12.000–29:19.000
And like it also is just like first principles like China has incredible, you know, smart scientists.
29:19.000–29:24.000
Some of them are the ones that came over here and are, you know, leaders at many of these labs.
29:24.000–29:29.000
So it's sort of like a it's a very tractable problem to get very good at at AI and AI training.
29:29.000–29:35.000
And really, it's just been classically like the chip problem of do you have enough compute?
29:35.000–29:37.000
And that's also tractable for the most part.
29:37.000–29:51.000
And then from a kind of philosophical standpoint on distillation, I don't know, I don't I'm not bothered by distillation because, you know, if I were bothered on distillation, I'd probably have to be bothered on the fact that these AI models are just trained on then the public Internet.
29:51.000–30:02.000
It's like we're all benefiting from the collective sort of creation of information and knowledge and everything is sort of riding on on some other set of information that's out there.
30:02.000–30:10.000
Like, you know, I would be bothered by, you know, a model being trained on Wikipedia if I if I was bothered by bothered by distillation.
30:10.000–30:16.000
So so to me, it's sort of all within the realm of, you know, you need access to data.
30:16.000–30:21.000
Obviously, the more intelligent data you can get, the better your model becomes.
30:21.000–30:29.000
And and at the same time, I think that there should be kind of a cat and mouse game because for competitive reasons, you probably want to block that if you're a frontier lab.
30:29.000–30:38.000
But I don't think of it as a as as some kind of, you know, unethical or or, you know, overly kind of, you know, complicated problem.
30:38.000–30:42.000
Yeah, it's a great point. It kind of goes back to the beginning of this, like modern AI era.
30:42.000–30:47.000
All of these models are trained on, you know, the Internet's all of the Internet's knowledge.
30:47.000–30:48.000
That's a form of distillation.
30:48.000–30:53.000
They were they were trained on my my trolling tweets from 15 years ago.
30:53.000–30:56.000
I mean, how mad am I about that? I don't know.
30:56.000–30:58.000
That's a scary, scary thought.
30:58.000–31:02.000
Aaron Levy tweets and all of Reddit.
31:02.000–31:07.000
I don't know, Gabe, I haven't gone too deep into your X feed, but it's trained on that, too, I'm sure.
31:07.000–31:11.000
Last one I wanted to ask you guys about because this is a little bit of a turn.
31:11.000–31:15.000
I want to ask you about Claude Tagg because producer Jasmine, she's very excited about it.
31:15.000–31:16.000
She's trying to get me there.
31:16.000–31:21.000
I don't fully, fully understand the appeal, but she is always right about these things.
31:21.000–31:24.000
I wonder, Aaron, first, I know you've been tweeting about this.
31:24.000–31:26.000
Like, can you break it down?
31:26.000–31:31.000
It's something that always lives in your slack and builds context over time, learns your company.
31:31.000–31:34.000
It feels like lock in to me, but it also seems really important.
31:34.000–31:41.000
Aaron, is this like going to be as big as Co-Work or some of the other kind of land shifting things that Anthropoc has done this year?
31:41.000–31:49.000
I heard to kind of put it in the in the overall kind of graph on that front.
31:49.000–32:00.000
But the reason why it's a pretty big deal is we're very used to AI in kind of single player mode where I have access to Claude or Codex or ChatGPT or Co-Work.
32:00.000–32:02.000
And this is sort of my relationship with AI.
32:02.000–32:08.000
And that AI is effectively acting as me in a variety of systems.
32:08.000–32:15.000
And so it's incredibly powerful because I can then spin up tasks that I would otherwise do a bunch of times across these agents.
32:15.000–32:18.000
But it's again, very single player mode.
32:18.000–32:33.000
And what Claude tag kind of tries to do, and this is, you know, I think building on the recent Zeitgeist of OpenClaw and Hermes and even kind of you're seeing this in agentic coding systems like Factory and Devon, is no, like what if it's a co-worker?
32:33.000–32:40.000
It's not rep, it's not Aaron or Gabe or anybody in the in the organization, it's its own sort of entity.
32:40.000–32:45.000
And it has access to a set of resources that any other kind of entity would have had access to.
32:45.000–32:49.000
And you kind of interacted with it like another user in your system.
32:49.000–32:56.000
And the reason why that's meaningful is that it has shared context across whatever that group that has access to it is doing.
32:56.000–32:59.000
And then you just sort of punt tasks to it and it comes back.
32:59.000–33:06.000
But those are again, kind of group enabled tasks just as again, a colleague would do inside of a Slack channel in this case.
33:06.000–33:14.000
So, so it's, it's, it's probably a very big deal in, in sort of the philosophical direction that it's moving toward.
33:14.000–33:19.000
And, and it has just like a different vector of use cases that, that it enables.
33:19.000–33:26.000
It's less about my personal productivity and more about sort of a shared intelligence for some, from, you know, group of users.
33:26.000–33:29.000
So you're in a Slack channel called the sales team Slack channel.
33:29.000–33:36.000
And, and you're, you're punting off tasks of, Hey, can you generate a deck for this sales presentation?
33:36.000–33:37.000
We're, we're about to go into.
33:37.000–33:44.000
And it needs access to your underlying, you know, marketing assets that are stored in box and your other data, you know, stored in Salesforce.
33:44.000–33:48.000
Just as you would a colleague that was about to go in that sales presentation with you.
33:48.000–33:50.000
So, so that's the kind of benefit of it.
33:50.000–33:57.000
And I think it's another kind of, you know, point in the timeline of like, what is the future of how we're going to work with these agentic systems?
33:57.000–33:58.000
What is the new user experience?
33:58.000–34:00.000
What is the, the, the new interface?
34:00.000–34:04.000
And, you know, from a kind of a lock in point, I think there's an interesting point.
34:04.000–34:10.000
Like on one hand, yes, you're, you're sort of, you know, you're, you're, you're, you're using the cloud particular paradigm.
34:10.000–34:19.000
On the other hand, you can choose to decide how much or how little you want to sort of use the cloud sort of intelligence layer.
34:19.000–34:23.000
Like you can plug in an MCP server of an entirely other agentic system.
34:23.000–34:26.000
So cloud could just be a router to a bunch of other, other tools.
34:26.000–34:29.000
I'm making up the use case, but I'm sure it could be a router to Harvey.
34:29.000–34:39.000
So, so I, I think there's actually like it, you can just design how you want to implement these systems to whatever level of lock in or, or not lock in you want to, you want to drive.
34:39.000–34:40.000
Right.
34:40.000–34:44.000
Feels like it's certainly like the whole industry is heading that way, more of like a buffet.
34:44.000–34:45.000
You can choose what you want.
34:45.000–34:47.000
Gabe, last word to you.
34:47.000–34:51.000
I guess, what should we be looking out for in the next week or so with these live streams?
34:51.000–34:55.000
We really try to be on like what Silicon Valley is talking about this week.
34:55.000–34:59.000
It's very much ZAI's GLM 5.2.
34:59.000–35:02.000
What do you, what should we be watching for the next week ahead?
35:02.000–35:08.000
I feel like it'll be interesting to see what happens kind of with the fable five kind of in the next week.
35:08.000–35:15.000
When that comes back, I think you're going to see more about kind of what Aaron was talking about with Claude Tagg.
35:15.000–35:22.000
Like I think that philosophical shift of just thinking about agents more and more as just employees in your company.
35:22.000–35:29.000
Like you saw Rippling kind of announce they're building kind of a unified like data layer of employees.
35:29.000–35:35.000
I think more and more we're thinking about our product as how do we give agents to a law firm or an in-house department.
35:35.000–35:45.000
And instead of kind of one player like Aaron mentioned, how do you start organizing these kind of like teams of agents and humans to complete these very complex tasks?
35:45.000–35:47.000
So I think you'll kind of keep seeing that trend.
35:47.000–35:50.000
And that's kind of like the trend we're most excited about.
35:50.000–35:54.000
Right. We'll call it multiplayer mode since Aaron's been talking about single player mode.
35:54.000–35:56.000
Thank you both so much, Aaron and Gabe.
35:56.000–35:58.000
Lovely to get you both here.
35:58.000–36:01.000
Thanks for all your insights and hope to talk to you again soon.
36:01.000–36:02.000
Thanks a lot.
36:02.000–36:03.000
All of what we talked.
36:03.000–36:06.000
Distillation open source.
36:06.000–36:09.000
It runs through the same question for investors.
36:09.000–36:11.000
What does it do for the chip trade?
36:11.000–36:14.000
That has been super relevant to Wall Street, of course.
36:14.000–36:15.000
Yesterday we got a big data point.
36:15.000–36:18.000
OpenAI and Broadcom unveiled Jalapeno.
36:18.000–36:22.000
That's OpenAI's first custom inference chip built from scratch in just nine months.
36:22.000–36:25.000
OpenAI used its own AI models to help design it.
36:25.000–36:32.000
Broadcom CEO Hawk Tan says that it cuts inference costs by roughly 50% versus current Nvidia GPUs.
36:32.000–36:35.000
The person I want to break all of this down with is Stacy Rascon.
36:35.000–36:40.000
He has been covering semiconductors at Bernstein for over a decade called the Nvidia trade early.
36:40.000–36:43.000
Stacy, you don't sugarcoat either.
36:43.000–36:44.000
Two decades.
36:44.000–36:45.000
Almost two decades.
36:45.000–36:48.000
Almost 16 years. Is that right?
36:48.000–36:49.000
I should have rounded up.
36:49.000–36:51.000
Oh, then I definitely needed to round it up.
36:51.000–36:54.000
Well, I know that's why we always love talking to you.
36:54.000–36:57.000
You also have all of the technical side, but you can break it down.
36:57.000–36:59.000
Give us your thoughts on Jalapeno.
36:59.000–37:04.000
I mean, to me, it just struck me as a huge, like, develop, like maybe a new era for chip development.
37:04.000–37:08.000
Right? These are some of the most complicated objects ever built by humans.
37:08.000–37:10.000
I guess, like, do you buy it?
37:10.000–37:13.000
Is it this, has it improved this much?
37:13.000–37:17.000
Well, so look, so chips are the most complicated things that humanity has ever built.
37:17.000–37:19.000
And as you know, we've talked before.
37:19.000–37:20.000
I love this space.
37:20.000–37:22.000
And that's one of the reasons why.
37:22.000–37:26.000
So, look, this was not unexpected.
37:26.000–37:27.000
I want to say that.
37:27.000–37:32.000
So we've known for a long time that Broadcom is working on a custom chip with OpenAI.
37:32.000–37:38.000
And in fact, the two companies have a 10 gigawatt deal over the next five years.
37:38.000–37:40.000
And we'll see how much of it actually ships.
37:40.000–37:45.000
But they've got a 10 gigawatt deal for multiple generations of these chips in play already.
37:45.000–37:53.000
On Broadcom's last earnings call they talked about, I think, in 27, they're supposed to ship, I can't remember, 1.2 or 1.3 gigawatts to OpenAI.
37:53.000–37:55.000
So that is presumably this chip.
37:55.000–37:59.000
So the fact that they announced it was not a surprise.
37:59.000–38:03.000
And in fact, one would hope that they would have been announcing it soon.
38:03.000–38:05.000
Because it's supposed to start shipping next year.
38:05.000–38:06.000
They've been talking about it.
38:06.000–38:07.000
Yeah.
38:07.000–38:10.000
Now that being said, look, it's great, right?
38:10.000–38:12.000
And I don't, you know, they put out a lot of stuff.
38:12.000–38:16.000
I guess he had said that, you're right, he said it would lower the cost by 50.
38:16.000–38:22.000
I don't know if that was the cost of inference or the cost of the chip itself relative to the current state of the art, which would probably be an NVIDIA GPU.
38:22.000–38:29.000
You have to be a little careful with those kind of comparisons as well, because you don't exactly know what they're talking about.
38:29.000–38:31.000
Like, people tend to cherry pick stuff.
38:31.000–38:37.000
And at the end of the day, it's not really the cost of the chip all by itself that necessarily matters.
38:37.000–38:42.000
It's the performance per watt, you know, performance per dollar total cost of ownership.
38:42.000–38:47.000
What the press release was talking about was sort of substantial improvements in performance per watt.
38:47.000–38:50.000
I think it said it was still an early test, early phase testing.
38:50.000–38:57.000
They're not in volume production with this yet, but it sounds like the early tests are very, very encouraging along those fronts.
38:57.000–39:00.000
And so, look, I'm sure it'll be a good chip.
39:00.000–39:03.000
And it sounds like they're getting ready to ship a lot of them.
39:03.000–39:06.000
Don't open AI, like I said, well over a gigawatt next year.
39:06.000–39:10.000
And like over the next five years, potentially up to 10 gigawatts.
39:10.000–39:11.000
I'm sure it's not going to be a good chip.
39:11.000–39:15.000
Okay, so I'm glad you stepped back and you said, okay, it comes as a surprise to no one that they have a chip.
39:15.000–39:16.000
They needed to.
39:16.000–39:24.000
If you're like a hyperscaler or an AI lab right now and you're not working on your own chip in some way, like you are doing something wrong here.
39:24.000–39:25.000
Fine.
39:25.000–39:28.000
They're all buying a lot of GPUs too.
39:28.000–39:30.000
So again, to go back to open AI.
39:30.000–39:32.000
So they have this 10 gigawatt deal with Braggam.
39:32.000–39:34.000
They also have a 10 gigawatt deal with Nvidia.
39:34.000–39:38.000
And they have a six gigawatt deal with AMD.
39:38.000–39:40.000
Everyone is just trying to get compute.
39:40.000–39:42.000
They're trying to get their costs down.
39:42.000–39:44.000
They're trying to be less dependent on Nvidia.
39:44.000–39:48.000
So let me tell you the thing that I thought was most interesting about this announcement.
39:48.000–39:52.000
And I don't know, like, that's why I'm curious if you thought it was interesting.
39:52.000–39:56.000
Nine months built in nine months from scratch.
39:56.000–40:00.000
Like that to me is just, that's like mind boggling.
40:00.000–40:03.000
These are the most complicated things to make in humanity, as you said.
40:03.000–40:05.000
So do we believe it?
40:05.000–40:09.000
So that is fast, but Braggam again has talked about exactly that.
40:09.000–40:13.000
And that's part of their differentiation on, on this custom chip.
40:13.000–40:16.000
There's a number of players that can do custom chips.
40:16.000–40:20.000
Braggam has literally said in the past, we can design a chip in nine months.
40:20.000–40:22.000
Like that, that's not something new that they've said.
40:22.000–40:26.000
This is the first time I think that they've actually shown a real example.
40:26.000–40:27.000
Okay.
40:27.000–40:29.000
And called it out that it was developed in nine months.
40:29.000–40:31.000
But again, they have said in the past that they could do that.
40:31.000–40:34.000
And so it's nice to actually see that they can.
40:34.000–40:37.000
It's still that, but okay, here's what I mean.
40:37.000–40:39.000
Does this change the whole chip space?
40:39.000–40:45.000
We've talked about, you know, the cycle getting shorter and Nvidia putting out new models now on a yearly basis.
40:45.000–40:48.000
If you can create it, I know you, I feel like I'm getting skepticism from you.
40:48.000–40:49.000
Is that because?
40:49.000–40:51.000
I think that, no, no, no, no, I'm not skeptical at all.
40:51.000–40:52.000
No, no, no, it looks right.
40:52.000–40:54.000
But I mean the chip space is changing anyways.
40:54.000–40:55.000
Right?
40:55.000–40:56.000
Okay.
40:56.000–40:57.000
So is this an aspect of that?
40:57.000–40:58.000
Absolutely.
40:58.000–41:03.000
Are we seeing a tremendous amount of new advanced silicon designs?
41:03.000–41:06.000
Because now we actually have a reason to do them and someone that's actually willing to pay for them.
41:06.000–41:07.000
Yes.
41:07.000–41:08.000
I'll give you an example.
41:08.000–41:12.000
You know, it's not even this, like we're seeing like chip startups now for a change for, for example.
41:12.000–41:13.000
Yeah.
41:13.000–41:20.000
I remember over 10 years ago, I had started to do a piece or to write a piece on venture capital investments in semiconductors.
41:20.000–41:26.000
I shelved it because there wasn't any, like all of the VC investments back then, it was a corporate VC.
41:26.000–41:27.000
That was it.
41:27.000–41:35.000
There wasn't a lot of, of, of, of actual, like, like a standard VC investment because it costs a lot of money and costs a lot of time.
41:35.000–41:38.000
And you know, the exits, uh, opportunities weren't certain.
41:38.000–41:40.000
And back then they would have just like invested in SAS.
41:40.000–41:41.000
It was a lot easier.
41:41.000–41:42.000
Yeah.
41:42.000–41:52.000
Um, we're actually starting to see a lot of VC investments in, in real meaningful startups that are getting valued both in the private as well as sometimes even in the public markets now, like in the billions or even tens of billions of dollars.
41:52.000–41:56.000
So there's an actual reason to do this now.
41:56.000–42:01.000
Like there, there, there's a, there's a use case and an application where like dollars are flowing in.
42:01.000–42:04.000
And so I think the space is, is absolutely changing.
42:04.000–42:06.000
And I, this is one aspect of it clearly.
42:06.000–42:07.000
Right.
42:07.000–42:08.000
Yeah.
42:08.000–42:09.000
And okay.
42:09.000–42:17.000
So as we've seen more competition, the story has always been like, what does it mean for Nvidia only game in town, but like the whole pie is expanding.
42:17.000–42:18.000
Right.
42:18.000–42:19.000
So that's what.
42:19.000–42:20.000
Yeah.
42:20.000–42:21.000
Tell me.
42:21.000–42:27.000
So as, as, as, as an equity analyst, we get paid to overcomplicate things sometimes.
42:27.000–42:28.000
And there's always.
42:28.000–42:30.000
Worries about competition.
42:30.000–42:32.000
And I would say, first of all, you need to step back.
42:32.000–42:33.000
This is semiconductors.
42:33.000–42:34.000
There was always competition.
42:34.000–42:38.000
There was always somebody like looking to eat your lunch if you will let them.
42:38.000–42:39.000
Right.
42:39.000–42:46.000
So it is imperative on the leaders in this space to make sure that they stay that way to, and to invest in the roadmap and to hire the best people and so on and so forth.
42:46.000–42:47.000
And they're all doing that.
42:47.000–42:54.000
I would also say the opportunity is getting so big that right now it almost doesn't matter.
42:54.000–42:59.000
My general view, for example, you know, you talk about Broadcom versus Nvidia, they talk about GPUs versus TPUs.
42:59.000–43:04.000
And my general view is that I get that, but it's, it's probably the wrong question.
43:04.000–43:08.000
Like the right question probably right now is more, is the opportunity in front of us still bigger?
43:08.000–43:09.000
Is it not?
43:09.000–43:11.000
Because if it's big, you know, everybody will thrive.
43:11.000–43:13.000
And if it's not, everybody's screwed.
43:13.000–43:14.000
Yeah.
43:14.000–43:15.000
And so far it's really big.
43:15.000–43:17.000
The assumption is that it's bigger, right?
43:17.000–43:19.000
The assumption is straight up and to the right.
43:19.000–43:22.000
Well, it's been, it's been straight up to the right so, so far, right?
43:22.000–43:25.000
Is there any reason to think it might be different?
43:25.000–43:29.000
Well, I mean, you're seeing it in G, it's not even just in this space, you're seeing it here.
43:29.000–43:33.000
And you're also seeing it in networking, you see copper versus optical and all kinds of
43:33.000–43:37.000
where there's all, there's always going to be alternatives to this market is, you know, hundreds
43:37.000–43:40.000
of billions or potentially even trillions of dollars.
43:40.000–43:44.000
Somebody is always going to have alternatives, but you can look at the numbers.
43:44.000–43:47.000
So Nvidia, I don't know, they'll do what $500 billion next year, right?
43:47.000–43:50.000
Broadcom got you to a hundred billion, right?
43:50.000–43:51.000
Nope.
43:51.000–43:53.000
And they'll probably do more than that, right?
43:53.000–43:58.000
And I don't know, Qualcomm had an analyst day yesterday and they said they're going to
43:58.000–44:01.000
do $5 billion from zero basically, but still billions.
44:01.000–44:05.000
And you got MediaTek and you got Marvell and AMD and all that.
44:05.000–44:06.000
They're all doing fine.
44:06.000–44:08.000
They're all growing, growing tons.
44:08.000–44:14.000
I mean, so in, in, in the, if you were to calculate the percentages, like our shares
44:14.000–44:18.000
moving up probably a little bit, but is it a really a problem at this point?
44:18.000–44:19.000
I don't think it is.
44:19.000–44:20.000
Right.
44:20.000–44:23.000
And you're saying, even though we have so many of these different players, this is still
44:23.000–44:24.000
a supply problem.
44:24.000–44:27.000
Like we see that especially in memory.
44:27.000–44:30.000
And so I wanted to ask you, I wanted to kind of connect this to the earlier conversation.
44:30.000–44:33.000
We were talking about Chinese open source, right?
44:33.000–44:36.000
Being a solution because it's more abundant.
44:36.000–44:37.000
It's a lot more efficient.
44:37.000–44:38.000
Yeah.
44:38.000–44:39.000
And it's almost at frontier level.
44:39.000–44:44.000
Does it, I feel like as an onlooker and you tell me if this is right or not, the same,
44:44.000–44:46.000
a similar thing is happening in hardware.
44:46.000–44:51.000
Like no China is nowhere close to the leading edge in terms of chips, but in terms of memory,
44:51.000–44:57.000
in terms of, you know, some of the other equipment that you need to build these data centers or
44:57.000–44:58.000
build electronics.
44:58.000–44:59.000
They're getting there.
44:59.000–45:00.000
What does that mean to you?
45:00.000–45:01.000
How is that shift?
45:01.000–45:02.000
How is that shifting?
45:02.000–45:06.000
Well, the Chinese are constrained in some sense, right?
45:06.000–45:08.000
So, you know, I did a piece while back.
45:08.000–45:10.000
It was called the U.S. has chips, but no power.
45:10.000–45:12.000
China has power, but no chips.
45:12.000–45:13.000
Like he's bringing more capacity online.
45:13.000–45:15.000
And the answer was the U.S. is bringing more online.
45:15.000–45:16.000
It's not even close.
45:16.000–45:21.000
The Chinese have been constrained by some of the export controls and the sanctions, particularly
45:21.000–45:26.000
on things like semiconductor manufacturing equipment that forces them to make their local
45:26.000–45:32.000
chips on effectively substandard process technology that impacts their yields and ability to really
45:32.000–45:36.000
produce like local chips at high volume, which is one reason that I think that they have been
45:36.000–45:40.000
forced to innovate along other vectors like like model efficiency and things like that.
45:40.000–45:44.000
They are constrained in terms of the resources that they can deploy.
45:44.000–45:47.000
And so they're forced to do as best as they can with those.
45:47.000–45:50.000
And they're very, very, they've been very, very good at it.
45:50.000–45:51.000
Right.
45:51.000–45:53.000
And this is one thing, you know, engineers are smart.
45:53.000–45:54.000
Right.
45:54.000–45:56.000
You know, if you give them constraints, they'll they'll find their way to make them fit.
45:56.000–45:58.000
And the Chinese are clearly doing that.
45:58.000–46:00.000
There are Chinese memory players.
46:00.000–46:05.000
But I mean, from what we're seeing in memory right now, I mean, you know, I don't cover Micron.
46:05.000–46:08.000
It's a colleague of mine, but I mean, Micron reported last night.
46:08.000–46:13.000
And I mean, it's it's looking like things are going to be tight for a long time, probably.
46:13.000–46:16.000
And it's both a supply issue as well as a demand issue.
46:16.000–46:18.000
And supply will come online over time.
46:18.000–46:23.000
And they have to actually build the buildings first before they have somewhere to put the tools.
46:23.000–46:24.000
Right.
46:24.000–46:25.000
To make the chip.
46:25.000–46:26.000
So it takes time.
46:26.000–46:31.000
But, you know, the question will be once that capacity comes online, like, does the demand rise rise to meet it?
46:31.000–46:33.000
We've actually seen this, by the way, like broadly in semis.
46:33.000–46:34.000
It's really interesting.
46:34.000–46:38.000
We've had this sort of rolling wave of bottlenecks.
46:38.000–46:41.000
Like, AI has gotten so big, it's kind of dragged everything along with it.
46:41.000–46:46.000
And one at a time, the different parts of the industry have sort of been hitting their limits and the stocks have been ripping.
46:46.000–46:51.000
And, you know, we went from the accelerators themselves being the bottleneck a year or two years or whatever.
46:51.000–46:52.000
And then it went to memory.
46:52.000–46:53.000
And then it went to semi-cap.
46:53.000–46:58.000
And then it went to optical and networking and power semis and more recently CPUs.
46:58.000–47:01.000
And you could have almost owned anything in the space.
47:01.000–47:05.000
I think that the SOX index is, you know, it's an index of semi-cap.
47:05.000–47:07.000
It's up 100% year to date.
47:07.000–47:08.000
It's been incredible.
47:08.000–47:09.000
You could have owned anything.
47:09.000–47:10.000
You would have been fine.
47:10.000–47:14.000
Like almost anything you would have been just fine to greater or lesser degree.
47:14.000–47:15.000
In the hardware space.
47:15.000–47:16.000
And it's all being ruined.
47:16.000–47:17.000
What's that?
47:17.000–47:18.000
Yes.
47:18.000–47:25.000
And it's basically you're owning the underlying things, but the app layer or the model layer, it's been a little bit more complicated.
47:25.000–47:26.000
Okay.
47:26.000–47:29.000
So like Stacey, you've been covering this almost two decades.
47:29.000–47:31.000
We're going to say two decades.
47:31.000–47:34.000
This stuff is cyclical, but right now it doesn't feel like cyclical.
47:34.000–47:37.000
And everyone's talking about a super cycle.
47:37.000–47:39.000
Is that still the case right now?
47:39.000–47:42.000
Do you see anything to throw that off?
47:42.000–47:44.000
Well, for now, yes.
47:44.000–47:45.000
Right.
47:45.000–47:51.000
I mean, look, I've been hearing the word super cycle as long as I've ever been doing this job, but this is probably the first real super cycle we've seen.
47:51.000–47:53.000
You know, there's a few different types of cycles, right?
47:53.000–48:00.000
You have inventory cycles, like semis are the back of the supply chain and fluctuations in any demand can propagate backwards.
48:00.000–48:03.000
And they tend to be shorter term typically.
48:03.000–48:05.000
You can get supply cycles.
48:05.000–48:07.000
Supply is tight and pricing goes up.
48:07.000–48:09.000
We're having a big one in memory right now.
48:09.000–48:11.000
You can have product cycles or socket cycles.
48:11.000–48:18.000
I was like, you know, I, I, I win or lose a chip that goes into an iPhone and that can drive my revenue up and down by big amounts.
48:18.000–48:22.000
And then you've got what we have today, which is, I mean, it's a true demand cycle.
48:22.000–48:26.000
And it's not that, you know, we didn't, you know, clearly we didn't have enough supply.
48:26.000–48:28.000
The reason is demand has gotten gotten so big.
48:28.000–48:33.000
It just overpowered and any of the wildest forecasts that were there, like not that long ago.
48:33.000–48:41.000
And if you want to define that as maybe, you know, indicative of what one might call a super cycle, maybe, maybe I would, because it is purely demand driven and it's driving everything.
48:41.000–48:44.000
And so the big question is how long does the demand last?
48:44.000–48:48.000
And I think that's the trillion, maybe it's the quadrillion dollar question.
48:48.000–48:49.000
Like, I don't know.
48:49.000–48:52.000
Right now it's still up and to the right.
48:52.000–48:59.000
And at some point, maybe that won't be the case, but I think all, all signs right now, the only thing we're hearing from anybody is they can't get enough compute.
48:59.000–49:02.000
Right. It's more just sort of the makeup of that, which we discussed before.
49:02.000–49:03.000
Is it going to be the labs?
49:03.000–49:07.000
Is it going to be the open source models, but infrastructure is sort of the system.
49:07.000–49:09.000
For my goals, I don't really care.
49:09.000–49:17.000
I mean, it's computed like Nvidia benefits clearly from both, you know, closed and open source.
49:17.000–49:18.000
Yeah.
49:18.000–49:20.000
Broadcom is doing A6, you know, for the vendors.
49:20.000–49:22.000
I mean, it'll depend on, on building for their models.
49:22.000–49:27.000
Right. But at the end of the day, I think compute is, is demand for compute is good.
49:27.000–49:34.000
It's, it's my guys in some sense, like they'll, by and large, they'll probably do fine as long as compute demand is going up.
49:34.000–49:35.000
Right.
49:35.000–49:37.000
Well, Stacey, it's always great to get your insights.
49:37.000–49:39.000
What's that?
49:39.000–49:42.000
They're selling the picks and the shovels and the gold rush.
49:42.000–49:44.000
So they'll, they'll be fine as long as the gold rush is going on.
49:44.000–49:46.000
Yeah. But have you seen that cartoon?
49:46.000–49:51.000
I feel like I've seen it a lot over the last few months, especially where, you know, someone's like, yeah.
49:51.000–49:52.000
What are you here to do?
49:52.000–49:54.000
No, one's actually mining the gold.
49:54.000–49:57.000
They're just bringing more picks and shovels for like a fraction.
49:57.000–49:59.000
They are, they are mining.
49:59.000–50:05.000
I think we can have genuine questions on, on return in our ROI, because I think that's really where the debate is.
50:05.000–50:11.000
Like what determines whether or not demand continues is, is there a return on the spending or is there not?
50:11.000–50:14.000
And it's still early, but I already think we're seeing evidence.
50:14.000–50:16.000
I mean, we've got you on the rental side.
50:16.000–50:17.000
Yeah.
50:17.000–50:18.000
They're sold out.
50:18.000–50:19.000
There's clearly a return.
50:19.000–50:26.000
I do wonder though, Stacey, how is it that the Chinese are able to do so much on a fraction of the CapEx?
50:26.000–50:32.000
Well, again, you know, it's, they're, they're being forced to be innovative, right?
50:32.000–50:33.000
Yeah.
50:33.000–50:35.000
And by the way, I do not view that as a bad thing.
50:35.000–50:38.000
Like this gets back to the whole deep seek scare from a year and a half ago.
50:38.000–50:39.000
You remember that?
50:39.000–50:40.000
Oh, I do.
50:40.000–50:42.000
I think we might be.
50:42.000–50:44.000
Everybody freaked out.
50:44.000–50:45.000
I know.
50:45.000–50:46.000
Okay.
50:46.000–50:47.000
Last thing I'll say.
50:47.000–50:48.000
Yes.
50:48.000–50:49.000
That was not a blip though.
50:49.000–50:53.000
I mean, we, we may be having another deep seek like moment.
50:53.000–50:54.000
What happened?
50:54.000–50:55.000
So people were worried.
50:55.000–50:59.000
It was like, Oh my God, these guys are so much more efficient.
50:59.000–51:00.000
We won't need as much computers.
51:00.000–51:01.000
We're building.
51:01.000–51:02.000
What happened?
51:02.000–51:05.000
Only thing we've seen since then is skyrocket.
51:05.000–51:07.000
We need costs to come down.
51:07.000–51:08.000
That's how you drive adoption.
51:08.000–51:09.000
That's how you get return.
51:09.000–51:14.000
And everybody throw this company, Devon's paradox, which it's been thrown around to death.
51:14.000–51:18.000
And the idea is when things get cheaper, people use more, but you have to remember I'm a semi
51:18.000–51:19.000
guy.
51:19.000–51:20.000
It's made out like that.
51:20.000–51:21.000
But yeah.
51:21.000–51:25.000
And look, I was, of course I believe in Jevon's paradox and semis cost got cut in half every
51:25.000–51:27.000
two years for six decades.
51:27.000–51:28.000
Was that a bad thing for semi?
51:28.000–51:32.000
No, it was a fantastic thing for semiconductors and for everybody else.
51:32.000–51:34.000
So I think lower, lower cost computers.
51:34.000–51:35.000
Good.
51:35.000–51:37.000
I started by calling you a semi guy.
51:37.000–51:38.000
Now you're calling yourself a semi guy.
51:38.000–51:39.000
I love it.
51:39.000–51:40.000
It was perfect.
51:40.000–51:43.000
You are our semi guy, Stacy, lots of energy, right?
51:43.000–51:45.000
Right place for the last few years.
51:45.000–51:47.000
Thank you so much for coming on the live stream.
51:47.000–51:48.000
We'll talk to you again soon.
51:48.000–51:49.000
I'm sure.
51:49.000–51:50.000
Thanks, Stacy.
51:50.000–51:53.000
Thank you guys for joining another live stream.
51:53.000–52:00.000
Thank you to Jasmine and Janice and Divya, Robert and Evan, and the behind here and Bud
52:00.000–52:01.000
in the control room.
52:01.000–52:02.000
We'll be back next week.
52:02.000–52:06.000
Thanks for watching guys and keep giving in those comments and questions.
52:06.000–52:07.000
Thank you.
52:07.000–52:08.000
Thank you.
52:08.000–52:09.000
Thank you.
52:13.000–52:14.000
Thank you.
52:14.000–52:15.000
Thank you.
52:15.000–52:16.000
Thank you.
52:16.000–52:17.000
Thank you.
52:17.000–52:18.000
Thank you.
52:18.000–52:19.000
Thank you.
0:00.000–0:04.000
（此句尚無繁中翻譯）
0:04.000–0:08.000
（此句尚無繁中翻譯）
0:08.000–0:14.000
（此句尚無繁中翻譯）
0:14.000–0:20.000
（此句尚無繁中翻譯）
0:20.000–0:23.000
（此句尚無繁中翻譯）
0:23.000–0:29.000
（此句尚無繁中翻譯）
0:29.000–0:34.000
（此句尚無繁中翻譯）
0:34.000–0:38.000
（此句尚無繁中翻譯）
0:38.000–0:41.000
（此句尚無繁中翻譯）
0:41.000–0:44.000
（此句尚無繁中翻譯）
0:44.000–0:48.000
（此句尚無繁中翻譯）
0:48.000–0:52.000
（此句尚無繁中翻譯）
0:59.000–1:04.000
（此句尚無繁中翻譯）
1:04.000–1:10.000
（此句尚無繁中翻譯）
1:10.000–1:13.000
（此句尚無繁中翻譯）
1:13.000–1:19.000
（此句尚無繁中翻譯）
1:19.000–1:22.000
（此句尚無繁中翻譯）
1:22.000–1:25.000
（此句尚無繁中翻譯）
1:25.000–1:28.000
（此句尚無繁中翻譯）
1:28.000–1:32.000
（此句尚無繁中翻譯）
1:32.000–1:37.000
（此句尚無繁中翻譯）
1:37.000–1:39.000
（此句尚無繁中翻譯）
1:39.000–1:42.000
（此句尚無繁中翻譯）
1:42.000–1:45.000
（此句尚無繁中翻譯）
1:45.000–1:53.000
（此句尚無繁中翻譯）
1:53.000–1:59.000
（此句尚無繁中翻譯）
1:59.000–2:04.000
（此句尚無繁中翻譯）
2:04.000–2:08.000
（此句尚無繁中翻譯）
2:08.000–2:14.000
（此句尚無繁中翻譯）
2:14.000–2:17.000
（此句尚無繁中翻譯）
2:17.000–2:20.000
（此句尚無繁中翻譯）
2:20.000–2:25.000
（此句尚無繁中翻譯）
2:25.000–2:30.000
（此句尚無繁中翻譯）
2:30.000–2:34.000
（此句尚無繁中翻譯）
2:34.000–2:38.000
（此句尚無繁中翻譯）
2:38.000–2:40.000
（此句尚無繁中翻譯）
2:40.000–2:41.000
（此句尚無繁中翻譯）
2:41.000–2:43.000
（此句尚無繁中翻譯）
2:43.000–2:45.000
（此句尚無繁中翻譯）
2:45.000–2:52.000
（此句尚無繁中翻譯）
2:52.000–2:57.000
（此句尚無繁中翻譯）
2:57.000–3:02.000
（此句尚無繁中翻譯）
3:02.000–3:06.000
（此句尚無繁中翻譯）
3:06.000–3:12.000
（此句尚無繁中翻譯）
3:12.000–3:18.000
（此句尚無繁中翻譯）
3:18.000–3:25.000
（此句尚無繁中翻譯）
3:25.000–3:27.000
（此句尚無繁中翻譯）
3:27.000–3:30.000
（此句尚無繁中翻譯）
3:30.000–3:33.000
（此句尚無繁中翻譯）
3:33.000–3:38.000
（此句尚無繁中翻譯）
3:38.000–3:40.000
（此句尚無繁中翻譯）
3:40.000–3:45.000
（此句尚無繁中翻譯）
3:45.000–3:48.000
（此句尚無繁中翻譯）
3:48.000–3:50.000
（此句尚無繁中翻譯）
3:50.000–3:58.000
（此句尚無繁中翻譯）
3:58.000–4:01.000
（此句尚無繁中翻譯）
4:01.000–4:04.000
（此句尚無繁中翻譯）
4:04.000–4:08.000
（此句尚無繁中翻譯）
4:08.000–4:10.000
（此句尚無繁中翻譯）
4:10.000–4:14.000
（此句尚無繁中翻譯）
4:14.000–4:15.000
（此句尚無繁中翻譯）
4:15.000–4:18.000
（此句尚無繁中翻譯）
4:18.000–4:20.000
（此句尚無繁中翻譯）
4:20.000–4:23.000
（此句尚無繁中翻譯）
4:23.000–4:29.000
（此句尚無繁中翻譯）
4:29.000–4:37.000
（此句尚無繁中翻譯）
4:37.000–4:45.000
（此句尚無繁中翻譯）
4:45.000–4:47.000
（此句尚無繁中翻譯）
4:47.000–4:53.000
（此句尚無繁中翻譯）
4:53.000–5:06.000
（此句尚無繁中翻譯）
5:06.000–5:08.000
（此句尚無繁中翻譯）
5:08.000–5:11.000
（此句尚無繁中翻譯）
5:11.000–5:14.000
（此句尚無繁中翻譯）
5:14.000–5:17.000
（此句尚無繁中翻譯）
5:17.000–5:24.000
（此句尚無繁中翻譯）
5:24.000–5:26.000
（此句尚無繁中翻譯）
5:26.000–5:30.000
（此句尚無繁中翻譯）
5:30.000–5:38.000
（此句尚無繁中翻譯）
5:38.000–5:44.000
（此句尚無繁中翻譯）
5:44.000–5:54.000
（此句尚無繁中翻譯）
5:54.000–6:00.000
（此句尚無繁中翻譯）
6:00.000–6:10.000
（此句尚無繁中翻譯）
6:10.000–6:17.000
（此句尚無繁中翻譯）
6:17.000–6:21.000
（此句尚無繁中翻譯）
6:21.000–6:25.000
（此句尚無繁中翻譯）
6:25.000–6:34.000
（此句尚無繁中翻譯）
6:34.000–6:39.000
（此句尚無繁中翻譯）
6:39.000–6:47.000
（此句尚無繁中翻譯）
6:47.000–6:56.000
（此句尚無繁中翻譯）
6:56.000–7:03.000
（此句尚無繁中翻譯）
7:03.000–7:10.000
（此句尚無繁中翻譯）
7:10.000–7:12.000
（此句尚無繁中翻譯）
7:12.000–7:17.000
（此句尚無繁中翻譯）
7:17.000–7:21.000
（此句尚無繁中翻譯）
7:21.000–7:26.000
（此句尚無繁中翻譯）
7:26.000–7:34.000
（此句尚無繁中翻譯）
7:34.000–7:39.000
（此句尚無繁中翻譯）
7:39.000–7:40.000
（此句尚無繁中翻譯）
7:40.000–7:45.000
（此句尚無繁中翻譯）
7:45.000–7:48.000
（此句尚無繁中翻譯）
7:48.000–7:55.000
（此句尚無繁中翻譯）
7:55.000–7:58.000
（此句尚無繁中翻譯）
7:58.000–8:07.000
（此句尚無繁中翻譯）
8:07.000–8:09.000
（此句尚無繁中翻譯）
8:09.000–8:12.000
（此句尚無繁中翻譯）
8:12.000–8:18.000
（此句尚無繁中翻譯）
8:18.000–8:22.000
（此句尚無繁中翻譯）
8:22.000–8:23.000
（此句尚無繁中翻譯）
8:27.000–8:28.000
（此句尚無繁中翻譯）
8:28.000–8:31.000
（此句尚無繁中翻譯）
8:31.000–8:33.000
（此句尚無繁中翻譯）
8:33.000–8:37.000
（此句尚無繁中翻譯）
8:37.000–8:40.000
（此句尚無繁中翻譯）
8:40.000–8:42.000
（此句尚無繁中翻譯）
8:42.000–8:43.000
（此句尚無繁中翻譯）
8:43.000–8:44.000
（此句尚無繁中翻譯）
8:44.000–8:47.000
（此句尚無繁中翻譯）
8:47.000–8:50.000
（此句尚無繁中翻譯）
8:50.000–8:52.000
（此句尚無繁中翻譯）
8:52.000–8:54.000
（此句尚無繁中翻譯）
8:54.000–8:56.000
（此句尚無繁中翻譯）
8:56.000–8:59.000
（此句尚無繁中翻譯）
8:59.000–9:12.000
（此句尚無繁中翻譯）
9:12.000–9:16.000
（此句尚無繁中翻譯）
9:16.000–9:24.000
（此句尚無繁中翻譯）
9:24.000–9:38.000
（此句尚無繁中翻譯）
9:38.000–9:43.000
（此句尚無繁中翻譯）
9:43.000–9:53.000
（此句尚無繁中翻譯）
9:53.000–10:02.000
（此句尚無繁中翻譯）
10:02.000–10:09.000
（此句尚無繁中翻譯）
10:09.000–10:16.000
（此句尚無繁中翻譯）
10:16.000–10:27.000
（此句尚無繁中翻譯）
10:27.000–10:40.000
（此句尚無繁中翻譯）
10:40.000–10:47.000
（此句尚無繁中翻譯）
10:47.000–10:53.000
（此句尚無繁中翻譯）
10:53.000–10:56.000
（此句尚無繁中翻譯）
10:56.000–11:03.000
（此句尚無繁中翻譯）
11:03.000–11:07.000
（此句尚無繁中翻譯）
11:07.000–11:11.000
（此句尚無繁中翻譯）
11:11.000–11:13.000
（此句尚無繁中翻譯）
11:13.000–11:17.000
（此句尚無繁中翻譯）
11:17.000–11:22.000
（此句尚無繁中翻譯）
11:22.000–11:26.000
（此句尚無繁中翻譯）
11:26.000–11:39.000
（此句尚無繁中翻譯）
11:39.000–11:58.000
（此句尚無繁中翻譯）
11:58.000–12:07.000
（此句尚無繁中翻譯）
12:07.000–12:10.000
（此句尚無繁中翻譯）
12:10.000–12:20.000
（此句尚無繁中翻譯）
12:20.000–12:28.000
（此句尚無繁中翻譯）
12:28.000–12:34.000
（此句尚無繁中翻譯）
12:34.000–12:49.000
（此句尚無繁中翻譯）
12:49.000–13:02.000
（此句尚無繁中翻譯）
13:02.000–13:13.000
（此句尚無繁中翻譯）
13:13.000–13:16.000
（此句尚無繁中翻譯）
13:16.000–13:22.000
（此句尚無繁中翻譯）
13:22.000–13:27.000
（此句尚無繁中翻譯）
13:27.000–13:31.000
（此句尚無繁中翻譯）
13:31.000–13:34.000
（此句尚無繁中翻譯）
13:34.000–13:39.000
（此句尚無繁中翻譯）
13:39.000–13:48.000
（此句尚無繁中翻譯）
13:48.000–13:53.000
（此句尚無繁中翻譯）
13:53.000–14:19.000
（此句尚無繁中翻譯）
14:19.000–14:30.000
（此句尚無繁中翻譯）
14:30.000–14:40.000
（此句尚無繁中翻譯）
14:40.000–14:47.000
（此句尚無繁中翻譯）
14:47.000–14:58.000
（此句尚無繁中翻譯）
14:58.000–15:05.000
（此句尚無繁中翻譯）
15:05.000–15:08.000
（此句尚無繁中翻譯）
15:08.000–15:18.000
（此句尚無繁中翻譯）
15:18.000–15:40.000
（此句尚無繁中翻譯）
15:40.000–15:59.000
（此句尚無繁中翻譯）
15:59.000–16:07.000
（此句尚無繁中翻譯）
16:07.000–16:17.000
（此句尚無繁中翻譯）
16:17.000–16:28.000
（此句尚無繁中翻譯）
16:28.000–16:33.000
（此句尚無繁中翻譯）
16:33.000–16:41.000
（此句尚無繁中翻譯）
16:41.000–16:48.000
（此句尚無繁中翻譯）
16:48.000–16:55.000
（此句尚無繁中翻譯）
16:55.000–17:05.000
（此句尚無繁中翻譯）
17:05.000–17:13.000
（此句尚無繁中翻譯）
17:13.000–17:18.000
（此句尚無繁中翻譯）
17:18.000–17:23.000
（此句尚無繁中翻譯）
17:23.000–17:30.000
（此句尚無繁中翻譯）
17:30.000–17:47.000
（此句尚無繁中翻譯）
17:47.000–17:50.000
（此句尚無繁中翻譯）
17:50.000–17:58.000
（此句尚無繁中翻譯）
17:58.000–18:02.000
（此句尚無繁中翻譯）
18:02.000–18:11.000
（此句尚無繁中翻譯）
18:11.000–18:15.000
（此句尚無繁中翻譯）
18:15.000–18:20.000
（此句尚無繁中翻譯）
18:20.000–18:26.000
（此句尚無繁中翻譯）
18:26.000–18:28.000
（此句尚無繁中翻譯）
18:28.000–18:30.000
（此句尚無繁中翻譯）
18:30.000–18:35.000
（此句尚無繁中翻譯）
18:35.000–18:42.000
（此句尚無繁中翻譯）
18:42.000–18:46.000
（此句尚無繁中翻譯）
18:46.000–18:50.000
（此句尚無繁中翻譯）
18:50.000–19:00.000
（此句尚無繁中翻譯）
19:00.000–19:03.000
（此句尚無繁中翻譯）
19:03.000–19:08.000
（此句尚無繁中翻譯）
19:08.000–19:12.000
（此句尚無繁中翻譯）
19:12.000–19:15.000
（此句尚無繁中翻譯）
19:15.000–19:17.000
（此句尚無繁中翻譯）
19:17.000–19:22.000
（此句尚無繁中翻譯）
19:22.000–19:24.000
（此句尚無繁中翻譯）
19:24.000–19:27.000
（此句尚無繁中翻譯）
19:27.000–19:29.000
（此句尚無繁中翻譯）
19:29.000–19:33.000
（此句尚無繁中翻譯）
19:33.000–19:39.000
（此句尚無繁中翻譯）
19:39.000–19:42.000
（此句尚無繁中翻譯）
19:42.000–19:44.000
（此句尚無繁中翻譯）
19:44.000–19:49.000
（此句尚無繁中翻譯）
19:49.000–19:51.000
（此句尚無繁中翻譯）
19:51.000–20:00.000
（此句尚無繁中翻譯）
20:00.000–20:11.000
（此句尚無繁中翻譯）
20:11.000–20:22.000
（此句尚無繁中翻譯）
20:22.000–20:32.000
（此句尚無繁中翻譯）
20:32.000–20:37.000
（此句尚無繁中翻譯）
20:37.000–20:42.000
（此句尚無繁中翻譯）
20:42.000–20:50.000
（此句尚無繁中翻譯）
20:50.000–20:56.000
（此句尚無繁中翻譯）
20:56.000–21:02.000
（此句尚無繁中翻譯）
21:02.000–21:06.000
（此句尚無繁中翻譯）
21:06.000–21:11.000
（此句尚無繁中翻譯）
21:11.000–21:15.000
（此句尚無繁中翻譯）
21:15.000–21:21.000
（此句尚無繁中翻譯）
21:21.000–21:27.000
（此句尚無繁中翻譯）
21:27.000–21:32.000
（此句尚無繁中翻譯）
21:32.000–21:40.000
（此句尚無繁中翻譯）
21:40.000–21:44.000
（此句尚無繁中翻譯）
21:44.000–21:47.000
（此句尚無繁中翻譯）
21:47.000–21:52.000
（此句尚無繁中翻譯）
21:52.000–21:57.000
（此句尚無繁中翻譯）
21:57.000–22:03.000
（此句尚無繁中翻譯）
22:03.000–22:11.000
（此句尚無繁中翻譯）
22:11.000–22:14.000
（此句尚無繁中翻譯）
22:14.000–22:16.000
（此句尚無繁中翻譯）
22:16.000–22:19.000
（此句尚無繁中翻譯）
22:19.000–22:21.000
（此句尚無繁中翻譯）
22:21.000–22:29.000
（此句尚無繁中翻譯）
22:29.000–22:35.000
（此句尚無繁中翻譯）
22:35.000–22:43.000
（此句尚無繁中翻譯）
22:43.000–22:54.000
（此句尚無繁中翻譯）
22:54.000–23:02.000
（此句尚無繁中翻譯）
23:02.000–23:20.000
（此句尚無繁中翻譯）
23:20.000–23:33.000
（此句尚無繁中翻譯）
23:33.000–23:52.000
（此句尚無繁中翻譯）
23:52.000–24:03.000
（此句尚無繁中翻譯）
24:03.000–24:10.000
（此句尚無繁中翻譯）
24:10.000–24:21.000
（此句尚無繁中翻譯）
24:22.000–24:28.000
（此句尚無繁中翻譯）
24:28.000–24:31.000
（此句尚無繁中翻譯）
24:31.000–24:51.000
（此句尚無繁中翻譯）
24:51.000–24:59.000
（此句尚無繁中翻譯）
24:59.000–25:11.000
（此句尚無繁中翻譯）
25:11.000–25:18.000
（此句尚無繁中翻譯）
25:18.000–25:22.000
（此句尚無繁中翻譯）
25:22.000–25:26.000
（此句尚無繁中翻譯）
25:26.000–25:41.000
（此句尚無繁中翻譯）
25:41.000–25:47.000
（此句尚無繁中翻譯）
25:47.000–25:51.000
（此句尚無繁中翻譯）
25:51.000–25:57.000
（此句尚無繁中翻譯）
25:57.000–26:05.000
（此句尚無繁中翻譯）
26:05.000–26:11.000
（此句尚無繁中翻譯）
26:11.000–26:15.000
（此句尚無繁中翻譯）
26:15.000–26:20.000
（此句尚無繁中翻譯）
26:20.000–26:23.000
（此句尚無繁中翻譯）
26:23.000–26:33.000
（此句尚無繁中翻譯）
26:33.000–26:38.000
（此句尚無繁中翻譯）
26:38.000–26:41.000
（此句尚無繁中翻譯）
26:41.000–26:44.000
（此句尚無繁中翻譯）
26:44.000–26:51.000
（此句尚無繁中翻譯）
26:51.000–26:54.000
（此句尚無繁中翻譯）
26:54.000–26:59.000
（此句尚無繁中翻譯）
26:59.000–27:06.000
（此句尚無繁中翻譯）
27:06.000–27:15.000
（此句尚無繁中翻譯）
27:15.000–27:21.000
（此句尚無繁中翻譯）
27:21.000–27:25.000
（此句尚無繁中翻譯）
27:25.000–27:30.000
（此句尚無繁中翻譯）
27:30.000–27:33.000
（此句尚無繁中翻譯）
27:33.000–27:35.000
（此句尚無繁中翻譯）
27:35.000–27:48.000
（此句尚無繁中翻譯）
27:48.000–27:54.000
（此句尚無繁中翻譯）
27:54.000–27:59.000
（此句尚無繁中翻譯）
27:59.000–28:02.000
（此句尚無繁中翻譯）
28:02.000–28:04.000
（此句尚無繁中翻譯）
28:04.000–28:06.000
（此句尚無繁中翻譯）
28:06.000–28:08.000
（此句尚無繁中翻譯）
28:08.000–28:11.000
（此句尚無繁中翻譯）
28:11.000–28:14.000
（此句尚無繁中翻譯）
28:14.000–28:18.000
（此句尚無繁中翻譯）
28:18.000–28:20.000
（此句尚無繁中翻譯）
28:20.000–28:25.000
（此句尚無繁中翻譯）
28:25.000–28:28.000
（此句尚無繁中翻譯）
28:28.000–28:31.000
（此句尚無繁中翻譯）
28:31.000–28:33.000
（此句尚無繁中翻譯）
28:33.000–28:40.000
（此句尚無繁中翻譯）
28:40.000–28:48.000
（此句尚無繁中翻譯）
28:48.000–28:52.000
（此句尚無繁中翻譯）
28:52.000–28:57.000
（此句尚無繁中翻譯）
28:57.000–29:01.000
（此句尚無繁中翻譯）
29:01.000–29:08.000
（此句尚無繁中翻譯）
29:08.000–29:12.000
（此句尚無繁中翻譯）
29:12.000–29:19.000
（此句尚無繁中翻譯）
29:19.000–29:24.000
（此句尚無繁中翻譯）
29:24.000–29:29.000
（此句尚無繁中翻譯）
29:29.000–29:35.000
（此句尚無繁中翻譯）
29:35.000–29:37.000
（此句尚無繁中翻譯）
29:37.000–29:51.000
（此句尚無繁中翻譯）
29:51.000–30:02.000
（此句尚無繁中翻譯）
30:02.000–30:10.000
（此句尚無繁中翻譯）
30:10.000–30:16.000
（此句尚無繁中翻譯）
30:16.000–30:21.000
（此句尚無繁中翻譯）
30:21.000–30:29.000
（此句尚無繁中翻譯）
30:29.000–30:38.000
（此句尚無繁中翻譯）
30:38.000–30:42.000
（此句尚無繁中翻譯）
30:42.000–30:47.000
（此句尚無繁中翻譯）
30:47.000–30:48.000
（此句尚無繁中翻譯）
30:48.000–30:53.000
（此句尚無繁中翻譯）
30:53.000–30:56.000
（此句尚無繁中翻譯）
30:56.000–30:58.000
（此句尚無繁中翻譯）
30:58.000–31:02.000
（此句尚無繁中翻譯）
31:02.000–31:07.000
（此句尚無繁中翻譯）
31:07.000–31:11.000
（此句尚無繁中翻譯）
31:11.000–31:15.000
（此句尚無繁中翻譯）
31:15.000–31:16.000
（此句尚無繁中翻譯）
31:16.000–31:21.000
（此句尚無繁中翻譯）
31:21.000–31:24.000
（此句尚無繁中翻譯）
31:24.000–31:26.000
（此句尚無繁中翻譯）
31:26.000–31:31.000
（此句尚無繁中翻譯）
31:31.000–31:34.000
（此句尚無繁中翻譯）
31:34.000–31:41.000
（此句尚無繁中翻譯）
31:41.000–31:49.000
（此句尚無繁中翻譯）
31:49.000–32:00.000
（此句尚無繁中翻譯）
32:00.000–32:02.000
（此句尚無繁中翻譯）
32:02.000–32:08.000
（此句尚無繁中翻譯）
32:08.000–32:15.000
（此句尚無繁中翻譯）
32:15.000–32:18.000
（此句尚無繁中翻譯）
32:18.000–32:33.000
（此句尚無繁中翻譯）
32:33.000–32:40.000
（此句尚無繁中翻譯）
32:40.000–32:45.000
（此句尚無繁中翻譯）
32:45.000–32:49.000
（此句尚無繁中翻譯）
32:49.000–32:56.000
（此句尚無繁中翻譯）
32:56.000–32:59.000
（此句尚無繁中翻譯）
32:59.000–33:06.000
（此句尚無繁中翻譯）
33:06.000–33:14.000
（此句尚無繁中翻譯）
33:14.000–33:19.000
（此句尚無繁中翻譯）
33:19.000–33:26.000
（此句尚無繁中翻譯）
33:26.000–33:29.000
（此句尚無繁中翻譯）
33:29.000–33:36.000
（此句尚無繁中翻譯）
33:36.000–33:37.000
（此句尚無繁中翻譯）
33:37.000–33:44.000
（此句尚無繁中翻譯）
33:44.000–33:48.000
（此句尚無繁中翻譯）
33:48.000–33:50.000
（此句尚無繁中翻譯）
33:50.000–33:57.000
（此句尚無繁中翻譯）
33:57.000–33:58.000
（此句尚無繁中翻譯）
33:58.000–34:00.000
（此句尚無繁中翻譯）
34:00.000–34:04.000
（此句尚無繁中翻譯）
34:04.000–34:10.000
（此句尚無繁中翻譯）
34:10.000–34:19.000
（此句尚無繁中翻譯）
34:19.000–34:23.000
（此句尚無繁中翻譯）
34:23.000–34:26.000
（此句尚無繁中翻譯）
34:26.000–34:29.000
（此句尚無繁中翻譯）
34:29.000–34:39.000
（此句尚無繁中翻譯）
34:39.000–34:40.000
（此句尚無繁中翻譯）
34:40.000–34:44.000
（此句尚無繁中翻譯）
34:44.000–34:45.000
（此句尚無繁中翻譯）
34:45.000–34:47.000
（此句尚無繁中翻譯）
34:47.000–34:51.000
（此句尚無繁中翻譯）
34:51.000–34:55.000
（此句尚無繁中翻譯）
34:55.000–34:59.000
（此句尚無繁中翻譯）
34:59.000–35:02.000
（此句尚無繁中翻譯）
35:02.000–35:08.000
（此句尚無繁中翻譯）
35:08.000–35:15.000
（此句尚無繁中翻譯）
35:15.000–35:22.000
（此句尚無繁中翻譯）
35:22.000–35:29.000
（此句尚無繁中翻譯）
35:29.000–35:35.000
（此句尚無繁中翻譯）
35:35.000–35:45.000
（此句尚無繁中翻譯）
35:45.000–35:47.000
（此句尚無繁中翻譯）
35:47.000–35:50.000
（此句尚無繁中翻譯）
35:50.000–35:54.000
（此句尚無繁中翻譯）
35:54.000–35:56.000
（此句尚無繁中翻譯）
35:56.000–35:58.000
（此句尚無繁中翻譯）
35:58.000–36:01.000
（此句尚無繁中翻譯）
36:01.000–36:02.000
（此句尚無繁中翻譯）
36:02.000–36:03.000
（此句尚無繁中翻譯）
36:03.000–36:06.000
（此句尚無繁中翻譯）
36:06.000–36:09.000
（此句尚無繁中翻譯）
36:09.000–36:11.000
（此句尚無繁中翻譯）
36:11.000–36:14.000
（此句尚無繁中翻譯）
36:14.000–36:15.000
（此句尚無繁中翻譯）
36:15.000–36:18.000
（此句尚無繁中翻譯）
36:18.000–36:22.000
（此句尚無繁中翻譯）
36:22.000–36:25.000
（此句尚無繁中翻譯）
36:25.000–36:32.000
（此句尚無繁中翻譯）
36:32.000–36:35.000
（此句尚無繁中翻譯）
36:35.000–36:40.000
（此句尚無繁中翻譯）
36:40.000–36:43.000
（此句尚無繁中翻譯）
36:43.000–36:44.000
（此句尚無繁中翻譯）
36:44.000–36:45.000
（此句尚無繁中翻譯）
36:45.000–36:48.000
（此句尚無繁中翻譯）
36:48.000–36:49.000
（此句尚無繁中翻譯）
36:49.000–36:51.000
（此句尚無繁中翻譯）
36:51.000–36:54.000
（此句尚無繁中翻譯）
36:54.000–36:57.000
（此句尚無繁中翻譯）
36:57.000–36:59.000
（此句尚無繁中翻譯）
36:59.000–37:04.000
（此句尚無繁中翻譯）
37:04.000–37:08.000
（此句尚無繁中翻譯）
37:08.000–37:10.000
（此句尚無繁中翻譯）
37:10.000–37:13.000
（此句尚無繁中翻譯）
37:13.000–37:17.000
（此句尚無繁中翻譯）
37:17.000–37:19.000
（此句尚無繁中翻譯）
37:19.000–37:20.000
（此句尚無繁中翻譯）
37:20.000–37:22.000
（此句尚無繁中翻譯）
37:22.000–37:26.000
（此句尚無繁中翻譯）
37:26.000–37:27.000
（此句尚無繁中翻譯）
37:27.000–37:32.000
（此句尚無繁中翻譯）
37:32.000–37:38.000
（此句尚無繁中翻譯）
37:38.000–37:40.000
（此句尚無繁中翻譯）
37:40.000–37:45.000
（此句尚無繁中翻譯）
37:45.000–37:53.000
（此句尚無繁中翻譯）
37:53.000–37:55.000
（此句尚無繁中翻譯）
37:55.000–37:59.000
（此句尚無繁中翻譯）
37:59.000–38:03.000
（此句尚無繁中翻譯）
38:03.000–38:05.000
（此句尚無繁中翻譯）
38:05.000–38:06.000
（此句尚無繁中翻譯）
38:06.000–38:07.000
（此句尚無繁中翻譯）
38:07.000–38:10.000
（此句尚無繁中翻譯）
38:10.000–38:12.000
（此句尚無繁中翻譯）
38:12.000–38:16.000
（此句尚無繁中翻譯）
38:16.000–38:22.000
（此句尚無繁中翻譯）
38:22.000–38:29.000
（此句尚無繁中翻譯）
38:29.000–38:31.000
（此句尚無繁中翻譯）
38:31.000–38:37.000
（此句尚無繁中翻譯）
38:37.000–38:42.000
（此句尚無繁中翻譯）
38:42.000–38:47.000
（此句尚無繁中翻譯）
38:47.000–38:50.000
（此句尚無繁中翻譯）
38:50.000–38:57.000
（此句尚無繁中翻譯）
38:57.000–39:00.000
（此句尚無繁中翻譯）
39:00.000–39:03.000
（此句尚無繁中翻譯）
39:03.000–39:06.000
（此句尚無繁中翻譯）
39:06.000–39:10.000
（此句尚無繁中翻譯）
39:10.000–39:11.000
（此句尚無繁中翻譯）
39:11.000–39:15.000
（此句尚無繁中翻譯）
39:15.000–39:16.000
（此句尚無繁中翻譯）
39:16.000–39:24.000
（此句尚無繁中翻譯）
39:24.000–39:25.000
（此句尚無繁中翻譯）
39:25.000–39:28.000
（此句尚無繁中翻譯）
39:28.000–39:30.000
（此句尚無繁中翻譯）
39:30.000–39:32.000
（此句尚無繁中翻譯）
39:32.000–39:34.000
（此句尚無繁中翻譯）
39:34.000–39:38.000
（此句尚無繁中翻譯）
39:38.000–39:40.000
（此句尚無繁中翻譯）
39:40.000–39:42.000
（此句尚無繁中翻譯）
39:42.000–39:44.000
（此句尚無繁中翻譯）
39:44.000–39:48.000
（此句尚無繁中翻譯）
39:48.000–39:52.000
（此句尚無繁中翻譯）
39:52.000–39:56.000
（此句尚無繁中翻譯）
39:56.000–40:00.000
（此句尚無繁中翻譯）
40:00.000–40:03.000
（此句尚無繁中翻譯）
40:03.000–40:05.000
（此句尚無繁中翻譯）
40:05.000–40:09.000
（此句尚無繁中翻譯）
40:09.000–40:13.000
（此句尚無繁中翻譯）
40:13.000–40:16.000
（此句尚無繁中翻譯）
40:16.000–40:20.000
（此句尚無繁中翻譯）
40:20.000–40:22.000
（此句尚無繁中翻譯）
40:22.000–40:26.000
（此句尚無繁中翻譯）
40:26.000–40:27.000
（此句尚無繁中翻譯）
40:27.000–40:29.000
（此句尚無繁中翻譯）
40:29.000–40:31.000
（此句尚無繁中翻譯）
40:31.000–40:34.000
（此句尚無繁中翻譯）
40:34.000–40:37.000
（此句尚無繁中翻譯）
40:37.000–40:39.000
（此句尚無繁中翻譯）
40:39.000–40:45.000
（此句尚無繁中翻譯）
40:45.000–40:48.000
（此句尚無繁中翻譯）
40:48.000–40:49.000
（此句尚無繁中翻譯）
40:49.000–40:51.000
（此句尚無繁中翻譯）
40:51.000–40:52.000
（此句尚無繁中翻譯）
40:52.000–40:54.000
（此句尚無繁中翻譯）
40:54.000–40:55.000
（此句尚無繁中翻譯）
40:55.000–40:56.000
（此句尚無繁中翻譯）
40:56.000–40:57.000
（此句尚無繁中翻譯）
40:57.000–40:58.000
（此句尚無繁中翻譯）
40:58.000–41:03.000
（此句尚無繁中翻譯）
41:03.000–41:06.000
（此句尚無繁中翻譯）
41:06.000–41:07.000
（此句尚無繁中翻譯）
41:07.000–41:08.000
（此句尚無繁中翻譯）
41:08.000–41:12.000
（此句尚無繁中翻譯）
41:12.000–41:13.000
（此句尚無繁中翻譯）
41:13.000–41:20.000
（此句尚無繁中翻譯）
41:20.000–41:26.000
（此句尚無繁中翻譯）
41:26.000–41:27.000
（此句尚無繁中翻譯）
41:27.000–41:35.000
（此句尚無繁中翻譯）
41:35.000–41:38.000
（此句尚無繁中翻譯）
41:38.000–41:40.000
（此句尚無繁中翻譯）
41:40.000–41:41.000
（此句尚無繁中翻譯）
41:41.000–41:42.000
（此句尚無繁中翻譯）
41:42.000–41:52.000
（此句尚無繁中翻譯）
41:52.000–41:56.000
（此句尚無繁中翻譯）
41:56.000–42:01.000
（此句尚無繁中翻譯）
42:01.000–42:04.000
（此句尚無繁中翻譯）
42:04.000–42:06.000
（此句尚無繁中翻譯）
42:06.000–42:07.000
（此句尚無繁中翻譯）
42:07.000–42:08.000
（此句尚無繁中翻譯）
42:08.000–42:09.000
（此句尚無繁中翻譯）
42:09.000–42:17.000
（此句尚無繁中翻譯）
42:17.000–42:18.000
（此句尚無繁中翻譯）
42:18.000–42:19.000
（此句尚無繁中翻譯）
42:19.000–42:20.000
（此句尚無繁中翻譯）
42:20.000–42:21.000
（此句尚無繁中翻譯）
42:21.000–42:27.000
（此句尚無繁中翻譯）
42:27.000–42:28.000
（此句尚無繁中翻譯）
42:28.000–42:30.000
（此句尚無繁中翻譯）
42:30.000–42:32.000
（此句尚無繁中翻譯）
42:32.000–42:33.000
（此句尚無繁中翻譯）
42:33.000–42:34.000
（此句尚無繁中翻譯）
42:34.000–42:38.000
（此句尚無繁中翻譯）
42:38.000–42:39.000
（此句尚無繁中翻譯）
42:39.000–42:46.000
（此句尚無繁中翻譯）
42:46.000–42:47.000
（此句尚無繁中翻譯）
42:47.000–42:54.000
（此句尚無繁中翻譯）
42:54.000–42:59.000
（此句尚無繁中翻譯）
42:59.000–43:04.000
（此句尚無繁中翻譯）
43:04.000–43:08.000
（此句尚無繁中翻譯）
43:08.000–43:09.000
（此句尚無繁中翻譯）
43:09.000–43:11.000
（此句尚無繁中翻譯）
43:11.000–43:13.000
（此句尚無繁中翻譯）
43:13.000–43:14.000
（此句尚無繁中翻譯）
43:14.000–43:15.000
（此句尚無繁中翻譯）
43:15.000–43:17.000
（此句尚無繁中翻譯）
43:17.000–43:19.000
（此句尚無繁中翻譯）
43:19.000–43:22.000
（此句尚無繁中翻譯）
43:22.000–43:25.000
（此句尚無繁中翻譯）
43:25.000–43:29.000
（此句尚無繁中翻譯）
43:29.000–43:33.000
（此句尚無繁中翻譯）
43:33.000–43:37.000
（此句尚無繁中翻譯）
43:37.000–43:40.000
（此句尚無繁中翻譯）
43:40.000–43:44.000
（此句尚無繁中翻譯）
43:44.000–43:47.000
（此句尚無繁中翻譯）
43:47.000–43:50.000
（此句尚無繁中翻譯）
43:50.000–43:51.000
（此句尚無繁中翻譯）
43:51.000–43:53.000
（此句尚無繁中翻譯）
43:53.000–43:58.000
（此句尚無繁中翻譯）
43:58.000–44:01.000
（此句尚無繁中翻譯）
44:01.000–44:05.000
（此句尚無繁中翻譯）
44:05.000–44:06.000
（此句尚無繁中翻譯）
44:06.000–44:08.000
（此句尚無繁中翻譯）
44:08.000–44:14.000
（此句尚無繁中翻譯）
44:14.000–44:18.000
（此句尚無繁中翻譯）
44:18.000–44:19.000
（此句尚無繁中翻譯）
44:19.000–44:20.000
（此句尚無繁中翻譯）
44:20.000–44:23.000
（此句尚無繁中翻譯）
44:23.000–44:24.000
（此句尚無繁中翻譯）
44:24.000–44:27.000
（此句尚無繁中翻譯）
44:27.000–44:30.000
（此句尚無繁中翻譯）
44:30.000–44:33.000
（此句尚無繁中翻譯）
44:33.000–44:36.000
（此句尚無繁中翻譯）
44:36.000–44:37.000
（此句尚無繁中翻譯）
44:37.000–44:38.000
（此句尚無繁中翻譯）
44:38.000–44:39.000
（此句尚無繁中翻譯）
44:39.000–44:44.000
（此句尚無繁中翻譯）
44:44.000–44:46.000
（此句尚無繁中翻譯）
44:46.000–44:51.000
（此句尚無繁中翻譯）
44:51.000–44:57.000
（此句尚無繁中翻譯）
44:57.000–44:58.000
（此句尚無繁中翻譯）
44:58.000–44:59.000
（此句尚無繁中翻譯）
44:59.000–45:00.000
（此句尚無繁中翻譯）
45:00.000–45:01.000
（此句尚無繁中翻譯）
45:01.000–45:02.000
（此句尚無繁中翻譯）
45:02.000–45:06.000
（此句尚無繁中翻譯）
45:06.000–45:08.000
（此句尚無繁中翻譯）
45:08.000–45:10.000
（此句尚無繁中翻譯）
45:10.000–45:12.000
（此句尚無繁中翻譯）
45:12.000–45:13.000
（此句尚無繁中翻譯）
45:13.000–45:15.000
（此句尚無繁中翻譯）
45:15.000–45:16.000
（此句尚無繁中翻譯）
45:16.000–45:21.000
（此句尚無繁中翻譯）
45:21.000–45:26.000
（此句尚無繁中翻譯）
45:26.000–45:32.000
（此句尚無繁中翻譯）
45:32.000–45:36.000
（此句尚無繁中翻譯）
45:36.000–45:40.000
（此句尚無繁中翻譯）
45:40.000–45:44.000
（此句尚無繁中翻譯）
45:44.000–45:47.000
（此句尚無繁中翻譯）
45:47.000–45:50.000
（此句尚無繁中翻譯）
45:50.000–45:51.000
（此句尚無繁中翻譯）
45:51.000–45:53.000
（此句尚無繁中翻譯）
45:53.000–45:54.000
（此句尚無繁中翻譯）
45:54.000–45:56.000
（此句尚無繁中翻譯）
45:56.000–45:58.000
（此句尚無繁中翻譯）
45:58.000–46:00.000
（此句尚無繁中翻譯）
46:00.000–46:05.000
（此句尚無繁中翻譯）
46:05.000–46:08.000
（此句尚無繁中翻譯）
46:08.000–46:13.000
（此句尚無繁中翻譯）
46:13.000–46:16.000
（此句尚無繁中翻譯）
46:16.000–46:18.000
（此句尚無繁中翻譯）
46:18.000–46:23.000
（此句尚無繁中翻譯）
46:23.000–46:24.000
（此句尚無繁中翻譯）
46:24.000–46:25.000
（此句尚無繁中翻譯）
46:25.000–46:26.000
（此句尚無繁中翻譯）
46:26.000–46:31.000
（此句尚無繁中翻譯）
46:31.000–46:33.000
（此句尚無繁中翻譯）
46:33.000–46:34.000
（此句尚無繁中翻譯）
46:34.000–46:38.000
（此句尚無繁中翻譯）
46:38.000–46:41.000
（此句尚無繁中翻譯）
46:41.000–46:46.000
（此句尚無繁中翻譯）
46:46.000–46:51.000
（此句尚無繁中翻譯）
46:51.000–46:52.000
（此句尚無繁中翻譯）
46:52.000–46:53.000
（此句尚無繁中翻譯）
46:53.000–46:58.000
（此句尚無繁中翻譯）
46:58.000–47:01.000
（此句尚無繁中翻譯）
47:01.000–47:05.000
（此句尚無繁中翻譯）
47:05.000–47:07.000
（此句尚無繁中翻譯）
47:07.000–47:08.000
（此句尚無繁中翻譯）
47:08.000–47:09.000
（此句尚無繁中翻譯）
47:09.000–47:10.000
（此句尚無繁中翻譯）
47:10.000–47:14.000
（此句尚無繁中翻譯）
47:14.000–47:15.000
（此句尚無繁中翻譯）
47:15.000–47:16.000
（此句尚無繁中翻譯）
47:16.000–47:17.000
（此句尚無繁中翻譯）
47:17.000–47:18.000
（此句尚無繁中翻譯）
47:18.000–47:25.000
（此句尚無繁中翻譯）
47:25.000–47:26.000
（此句尚無繁中翻譯）
47:26.000–47:29.000
（此句尚無繁中翻譯）
47:29.000–47:31.000
（此句尚無繁中翻譯）
47:31.000–47:34.000
（此句尚無繁中翻譯）
47:34.000–47:37.000
（此句尚無繁中翻譯）
47:37.000–47:39.000
（此句尚無繁中翻譯）
47:39.000–47:42.000
（此句尚無繁中翻譯）
47:42.000–47:44.000
（此句尚無繁中翻譯）
47:44.000–47:45.000
（此句尚無繁中翻譯）
47:45.000–47:51.000
（此句尚無繁中翻譯）
47:51.000–47:53.000
（此句尚無繁中翻譯）
47:53.000–48:00.000
（此句尚無繁中翻譯）
48:00.000–48:03.000
（此句尚無繁中翻譯）
48:03.000–48:05.000
（此句尚無繁中翻譯）
48:05.000–48:07.000
（此句尚無繁中翻譯）
48:07.000–48:09.000
（此句尚無繁中翻譯）
48:09.000–48:11.000
（此句尚無繁中翻譯）
48:11.000–48:18.000
（此句尚無繁中翻譯）
48:18.000–48:22.000
（此句尚無繁中翻譯）
48:22.000–48:26.000
（此句尚無繁中翻譯）
48:26.000–48:28.000
（此句尚無繁中翻譯）
48:28.000–48:33.000
（此句尚無繁中翻譯）
48:33.000–48:41.000
（此句尚無繁中翻譯）
48:41.000–48:44.000
（此句尚無繁中翻譯）
48:44.000–48:48.000
（此句尚無繁中翻譯）
48:48.000–48:49.000
（此句尚無繁中翻譯）
48:49.000–48:52.000
（此句尚無繁中翻譯）
48:52.000–48:59.000
（此句尚無繁中翻譯）
48:59.000–49:02.000
（此句尚無繁中翻譯）
49:02.000–49:03.000
（此句尚無繁中翻譯）
49:03.000–49:07.000
（此句尚無繁中翻譯）
49:07.000–49:09.000
（此句尚無繁中翻譯）
49:09.000–49:17.000
（此句尚無繁中翻譯）
49:17.000–49:18.000
（此句尚無繁中翻譯）
49:18.000–49:20.000
（此句尚無繁中翻譯）
49:20.000–49:22.000
（此句尚無繁中翻譯）
49:22.000–49:27.000
（此句尚無繁中翻譯）
49:27.000–49:34.000
（此句尚無繁中翻譯）
49:34.000–49:35.000
（此句尚無繁中翻譯）
49:35.000–49:37.000
（此句尚無繁中翻譯）
49:37.000–49:39.000
（此句尚無繁中翻譯）
49:39.000–49:42.000
（此句尚無繁中翻譯）
49:42.000–49:44.000
（此句尚無繁中翻譯）
49:44.000–49:46.000
（此句尚無繁中翻譯）
49:46.000–49:51.000
（此句尚無繁中翻譯）
49:51.000–49:52.000
（此句尚無繁中翻譯）
49:52.000–49:54.000
（此句尚無繁中翻譯）
49:54.000–49:57.000
（此句尚無繁中翻譯）
49:57.000–49:59.000
（此句尚無繁中翻譯）
49:59.000–50:05.000
（此句尚無繁中翻譯）
50:05.000–50:11.000
（此句尚無繁中翻譯）
50:11.000–50:14.000
（此句尚無繁中翻譯）
50:14.000–50:16.000
（此句尚無繁中翻譯）
50:16.000–50:17.000
（此句尚無繁中翻譯）
50:17.000–50:18.000
（此句尚無繁中翻譯）
50:18.000–50:19.000
（此句尚無繁中翻譯）
50:19.000–50:26.000
（此句尚無繁中翻譯）
50:26.000–50:32.000
（此句尚無繁中翻譯）
50:32.000–50:33.000
（此句尚無繁中翻譯）
50:33.000–50:35.000
（此句尚無繁中翻譯）
50:35.000–50:38.000
（此句尚無繁中翻譯）
50:38.000–50:39.000
（此句尚無繁中翻譯）
50:39.000–50:40.000
（此句尚無繁中翻譯）
50:40.000–50:42.000
（此句尚無繁中翻譯）
50:42.000–50:44.000
（此句尚無繁中翻譯）
50:44.000–50:45.000
（此句尚無繁中翻譯）
50:45.000–50:46.000
（此句尚無繁中翻譯）
50:46.000–50:47.000
（此句尚無繁中翻譯）
50:47.000–50:48.000
（此句尚無繁中翻譯）
50:48.000–50:49.000
（此句尚無繁中翻譯）
50:49.000–50:53.000
（此句尚無繁中翻譯）
50:53.000–50:54.000
（此句尚無繁中翻譯）
50:54.000–50:55.000
（此句尚無繁中翻譯）
50:55.000–50:59.000
（此句尚無繁中翻譯）
50:59.000–51:00.000
（此句尚無繁中翻譯）
51:00.000–51:01.000
（此句尚無繁中翻譯）
51:01.000–51:02.000
（此句尚無繁中翻譯）
51:02.000–51:05.000
（此句尚無繁中翻譯）
51:05.000–51:07.000
（此句尚無繁中翻譯）
51:07.000–51:08.000
（此句尚無繁中翻譯）
51:08.000–51:09.000
（此句尚無繁中翻譯）
51:09.000–51:14.000
（此句尚無繁中翻譯）
51:14.000–51:18.000
（此句尚無繁中翻譯）
51:18.000–51:19.000
（此句尚無繁中翻譯）
51:19.000–51:20.000
（此句尚無繁中翻譯）
51:20.000–51:21.000
（此句尚無繁中翻譯）
51:21.000–51:25.000
（此句尚無繁中翻譯）
51:25.000–51:27.000
（此句尚無繁中翻譯）
51:27.000–51:28.000
（此句尚無繁中翻譯）
51:28.000–51:32.000
（此句尚無繁中翻譯）
51:32.000–51:34.000
（此句尚無繁中翻譯）
51:34.000–51:35.000
（此句尚無繁中翻譯）
51:35.000–51:37.000
（此句尚無繁中翻譯）
51:37.000–51:38.000
（此句尚無繁中翻譯）
51:38.000–51:39.000
（此句尚無繁中翻譯）
51:39.000–51:40.000
（此句尚無繁中翻譯）
51:40.000–51:43.000
（此句尚無繁中翻譯）
51:43.000–51:45.000
（此句尚無繁中翻譯）
51:45.000–51:47.000
（此句尚無繁中翻譯）
51:47.000–51:48.000
（此句尚無繁中翻譯）
51:48.000–51:49.000
（此句尚無繁中翻譯）
51:49.000–51:50.000
（此句尚無繁中翻譯）
51:50.000–51:53.000
（此句尚無繁中翻譯）
51:53.000–52:00.000
（此句尚無繁中翻譯）
52:00.000–52:01.000
（此句尚無繁中翻譯）
52:01.000–52:02.000
（此句尚無繁中翻譯）
52:02.000–52:06.000
（此句尚無繁中翻譯）
52:06.000–52:07.000
（此句尚無繁中翻譯）
52:07.000–52:08.000
（此句尚無繁中翻譯）
52:08.000–52:09.000
（此句尚無繁中翻譯）
52:13.000–52:14.000
（此句尚無繁中翻譯）
52:14.000–52:15.000
（此句尚無繁中翻譯）
52:15.000–52:16.000
（此句尚無繁中翻譯）
52:16.000–52:17.000
（此句尚無繁中翻譯）
52:17.000–52:18.000
（此句尚無繁中翻譯）
52:18.000–52:19.000
（此句尚無繁中翻譯）
0:00.000–0:04.000
DeepSeq showed Wall Street that China could build cheap, powerful AI.
（此句尚無繁中翻譯）
0:04.000–0:08.000
Now, Ji-Poo, the AI, is showing how that advantage could spread.
（此句尚無繁中翻譯）
0:08.000–0:14.000
I've been consistently surprised by how quickly the open source has caught up.
（此句尚無繁中翻譯）
0:14.000–0:20.000
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7
（此句尚無繁中翻譯）
0:20.000–0:23.000
and some of these like very closed source frontier models.
（此句尚無繁中翻譯）
0:23.000–0:29.000
What GLM 5.2 kind of shows is, you know, we might still be in this sort of three to six month territory.
（此句尚無繁中翻譯）
0:29.000–0:34.000
Everything about the economics of AI changes depending on which outcome we're in.
（此句尚無繁中翻譯）
0:34.000–0:38.000
We need costs to come down. That's how you drive adoption. That's how you get returns.
（此句尚無繁中翻譯）
0:38.000–0:41.000
The AI leaderboard is starting to look incomplete.
（此句尚無繁中翻譯）
0:41.000–0:44.000
The next fight may be over intelligence per dollar.
（此句尚無繁中翻譯）
0:44.000–0:48.000
Models that are capable enough for real work and cheap enough to run constantly.
（此句尚無繁中翻譯）
0:48.000–0:52.000
That is where ZAI is forcing a new conversation.
（此句尚無繁中翻譯）
0:59.000–1:04.000
Ji-Poo's latest model, it's called GLM 5.2, comes out of China.
（此句尚無繁中翻譯）
1:04.000–1:10.000
It has landed with a bang in Silicon Valley, blowing past all other open source models,
（此句尚無繁中翻譯）
1:10.000–1:13.000
nearly matching the American frontier for just a fraction of the price.
（此句尚無繁中翻譯）
1:13.000–1:19.000
Developers, they're piling in open router token traffic showing much quicker adoption for GLM 5.2
（此句尚無繁中翻譯）
1:19.000–1:22.000
than DeepSeek's V4 launch back in April.
（此句尚無繁中翻譯）
1:22.000–1:25.000
It was a big deal then, it's a big deal now.
（此句尚無繁中翻譯）
1:25.000–1:28.000
This just carries the story that we've been reporting on even further.
（此句尚無繁中翻譯）
1:28.000–1:32.000
It all matters because Ji-Poo is hitting at a different moment though.
（此句尚無繁中翻譯）
1:32.000–1:37.000
The trillion dollar sell-off post-DeepSeek that was treated as a kind of a one-off shock,
（此句尚無繁中翻譯）
1:37.000–1:39.000
partly because people saw it as a chat bot story.
（此句尚無繁中翻譯）
1:39.000–1:42.000
GLM 5.2, this is different.
（此句尚無繁中翻譯）
1:42.000–1:45.000
It's strong at agentic work and that is key.
（此句尚無繁中翻譯）
1:45.000–1:53.000
On one agentic benchmark, it is just one percentage point away from Opus 4.8 for a fifth of the cost.
（此句尚無繁中翻譯）
1:53.000–1:59.000
Now, Opus 4.8, that is Anthropics' most powerfully available model.
（此句尚無繁中翻譯）
1:59.000–2:04.000
So basically, you're getting very close to the same horsepower for just a fraction of a cost.
（此句尚無繁中翻譯）
2:04.000–2:08.000
And it really comes at a time when expensive AI is already eating into budgets.
（此句尚無繁中翻譯）
2:08.000–2:14.000
And that is a really hard trade-off for enterprises, for companies, for Fortune 500 to ignore.
（此句尚無繁中翻譯）
2:14.000–2:17.000
Agentic AI is only going to intensify those costs.
（此句尚無繁中翻譯）
2:17.000–2:20.000
More complex tasks, more steps, more tokens.
（此句尚無繁中翻譯）
2:20.000–2:25.000
So now that Ji-Poo, also known as Z.AI, is in the picture,
（此句尚無繁中翻譯）
2:25.000–2:30.000
agentic open source, it might be the next big threat out of China.
（此句尚無繁中翻譯）
2:30.000–2:34.000
And here's a part of the story, the AI story, that I think Wall Street is still missing here.
（此句尚無繁中翻譯）
2:34.000–2:38.000
For the last few years, everyone has been obsessed with these AI leaderboards.
（此句尚無繁中翻譯）
2:38.000–2:40.000
Who has the smartest model?
（此句尚無繁中翻譯）
2:40.000–2:41.000
Who's the best at coding?
（此句尚無繁中翻譯）
2:41.000–2:43.000
Which one is ahead on reasoning?
（此句尚無繁中翻譯）
2:43.000–2:45.000
But that is not how companies buy software.
（此句尚無繁中翻譯）
2:45.000–2:52.000
Companies are increasingly asking, what is good enough and what does it cost to run this thing a million times or more?
（此句尚無繁中翻譯）
2:52.000–2:57.000
All of my employees running it a million times over the course of weeks, months, a year.
（此句尚無繁中翻譯）
2:57.000–3:02.000
So the new metric in AI, it is intelligence per dollar.
（此句尚無繁中翻譯）
3:02.000–3:06.000
And one reason Chinese models are pushing so hard on that metric is because of distillation,
（此句尚無繁中翻譯）
3:06.000–3:12.000
which is basically you take a big expensive model, use it to train a smaller, cheaper model to act like it.
（此句尚無繁中翻譯）
3:12.000–3:18.000
The American AI story, it has been built around bigger models, bigger data centers, huge spending.
（此句尚無繁中翻譯）
3:18.000–3:25.000
But Chinese labs, they're putting these cheaper versions front and center, very, very close to the frontier performance without the cost.
（此句尚無繁中翻譯）
3:25.000–3:27.000
That's why GPO is interesting.
（此句尚無繁中翻譯）
3:27.000–3:30.000
Artificial analysis has this chart that looks at both sides of this.
（此句尚無繁中翻譯）
3:30.000–3:33.000
How smart is a model and how much does it cost to run it?
（此句尚無繁中翻譯）
3:33.000–3:38.000
Now, the most attractive quadrant, it's obvious, high performance, low cost.
（此句尚無繁中翻譯）
3:38.000–3:40.000
That's where every company wants to be.
（此句尚無繁中翻譯）
3:40.000–3:45.000
GPO's GLM 5.2, it's getting very close to that sweet spot.
（此句尚無繁中翻譯）
3:45.000–3:48.000
It's not quite at the top in terms of performance.
（此句尚無繁中翻譯）
3:48.000–3:50.000
You can see it here, very close to that green square.
（此句尚無繁中翻譯）
3:50.000–3:58.000
So not at the very top in terms of performance, but it is close enough to the best models from OpenAI and Anthropic to make this price gap hard to ignore.
（此句尚無繁中翻譯）
3:58.000–4:01.000
And on agentic work, that gap matters even more.
（此句尚無繁中翻譯）
4:01.000–4:04.000
These are tasks where models don't just answer one question.
（此句尚無繁中翻譯）
4:04.000–4:08.000
They plan, they code, test, fix mistakes, they loop, they keep going.
（此句尚無繁中翻譯）
4:08.000–4:10.000
That is a lot more expensive.
（此句尚無繁中翻譯）
4:10.000–4:14.000
So if you're a company, you need a great model, cheap enough to use over and over again.
（此句尚無繁中翻譯）
4:14.000–4:15.000
And that is the setup.
（此句尚無繁中翻譯）
4:15.000–4:18.000
I want to get into all of this with two people who are actually living this.
（此句尚無繁中翻譯）
4:18.000–4:20.000
Aaron Levy, he's been running Box for 20 years.
（此句尚無繁中翻譯）
4:20.000–4:23.000
He saw the move to the cloud, the move to mobile, the move to AI.
（此句尚無繁中翻譯）
4:23.000–4:29.000
He is one of the most plugged in enterprise guys in the valley on what this stuff actually looks like when it hits a real company.
（此句尚無繁中翻譯）
4:29.000–4:37.000
We've also got Gabe Pereira, co-founder of Harvey, which is the AI platform that half of the AM Law 100 is now running on.
（此句尚無繁中翻譯）
4:37.000–4:45.000
Now, before that, he was a research scientist at DeepMind and Meta, so he has seen this from both sides, the model layer and the application layer.
（此句尚無繁中翻譯）
4:45.000–4:47.000
Aaron and Gabe, it's great to have you both on.
（此句尚無繁中翻譯）
4:47.000–4:53.000
We got Aaron too?
（此句尚無繁中翻譯）
4:53.000–5:06.000
Okay, we're going to, I can't, let's see, do we have both their sound?
（此句尚無繁中翻譯）
5:06.000–5:08.000
Okay, we're going to try and get this sound fixed.
（此句尚無繁中翻譯）
5:08.000–5:11.000
So, Aaron, you're an observer now.
（此句尚無繁中翻譯）
5:11.000–5:14.000
I'm assuming that you can hear Gabe and I, and then I'll come to you on this.
（此句尚無繁中翻譯）
5:14.000–5:17.000
Gabe, let me start this off with what's happening with, you know, GLM.
（此句尚無繁中翻譯）
5:17.000–5:24.000
It feels like my entire X feed, everyone I'm talking to here just cannot get enough of this new model GLM 5.2 right now.
（此句尚無繁中翻譯）
5:24.000–5:26.000
Kind of feels like DeepSeq all over again.
（此句尚無繁中翻譯）
5:26.000–5:30.000
People also aren't being shy and calling that maybe even bigger.
（此句尚無繁中翻譯）
5:30.000–5:38.000
So, I guess first, how do you explain this to someone outside of tech or like a non-tech CEO who has been all in on American AI,
（此句尚無繁中翻譯）
5:38.000–5:44.000
but sees all of this buzz around a new model that is almost as capable, but a lot cheaper?
（此句尚無繁中翻譯）
5:44.000–5:54.000
Yeah, I would say the big question since kind of ChatGPT came out in these closed source models is just how big is the gap between the closed source and the open source?
（此句尚無繁中翻譯）
5:54.000–6:00.000
And I think I've been consistently surprised by how quickly the open source has caught up.
（此句尚無繁中翻譯）
6:00.000–6:10.000
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7 and some of these like very closed source frontier models.
（此句尚無繁中翻譯）
6:10.000–6:17.000
And exactly to your point, it lets you think about the cost curves, where you need frontier intelligence, where you can use open source intelligence.
（此句尚無繁中翻譯）
6:17.000–6:21.000
And I think this is what we're starting to figure out and a lot of companies are starting to figure out now.
（此句尚無繁中翻譯）
6:21.000–6:25.000
Right. So explain that kind of in like just basic terms.
（此句尚無繁中翻譯）
6:25.000–6:34.000
If you are a company CEO and you have this AI budget, but you've seen, you know, your company blowing through it, when you get something like this, what should they be thinking?
（此句尚無繁中翻譯）
6:34.000–6:39.000
Especially when you say like it almost reaches the frontier, the very best models, but for a fraction of the price.
（此句尚無繁中翻譯）
6:39.000–6:47.000
People you are talking to, Gabe, do you think that they're taking a second look at Chinese and these open source models or should they still be concerned about them?
（此句尚無繁中翻譯）
6:47.000–6:56.000
Yeah, I think every company will use a mix of both, but exactly to what you said, it's starting to think about which tasks do I not need to use frontier intelligence.
（此句尚無繁中翻譯）
6:56.000–7:03.000
And so as these models get smarter, do you need the smartest model reviewing your contracts or doing simple red lines?
（此句尚無繁中翻譯）
7:03.000–7:10.000
And so I think for every company, the same way you organize a large company by figuring out, you know, I need different seniority.
（此句尚無繁中翻譯）
7:10.000–7:12.000
I need to pay different people for different roles.
（此句尚無繁中翻譯）
7:12.000–7:17.000
I think you're going to start seeing this with agents where there'll be some agents that are making critical decisions.
（此句尚無繁中翻譯）
7:17.000–7:21.000
You want frontier intelligence and there'll be some agents that are doing simpler tasks.
（此句尚無繁中翻譯）
7:21.000–7:26.000
And there you can use open source and you can really reduce the cost of of intelligence.
（此句尚無繁中翻譯）
7:26.000–7:34.000
Right. And are you seeing sort of the buzz that I am as well as your feed, the people you're talking to, are they very excited about GLM five, two?
（此句尚無繁中翻譯）
7:34.000–7:39.000
And like, can you explain why this is so interesting, why this is different than deep seek?
（此句尚無繁中翻譯）
7:39.000–7:40.000
I tried to, but I'd love to hear it from you.
（此句尚無繁中翻譯）
7:40.000–7:45.000
Yeah, I think we're seeing the same buzz in terms of our benchmarks.
（此句尚無繁中翻譯）
7:45.000–7:48.000
It was kind of a big open source jump.
（此句尚無繁中翻譯）
7:48.000–7:55.000
I think it's the first model that some of my friends at the labs were like, oh, this is like a very serious open source model.
（此句尚無繁中翻譯）
7:55.000–7:58.000
And so I think it's kind of similar.
（此句尚無繁中翻譯）
7:58.000–8:07.000
I would say to me, the deep seek moment was much more of I think there's always this question of how much money do you need to spend to stay on the frontier?
（此句尚無繁中翻譯）
8:07.000–8:09.000
I don't think that's changing.
（此句尚無繁中翻譯）
8:09.000–8:12.000
I think it's still very expensive to train these frontier models.
（此句尚無繁中翻譯）
8:12.000–8:18.000
I think the question is just how quickly can open source catch up to that frontier as the frontier keeps moving?
（此句尚無繁中翻譯）
8:18.000–8:22.000
Right. And that gap keeps like becoming narrower and narrower.
（此句尚無繁中翻譯）
8:22.000–8:23.000
Aaron, I think we've got you now, right?
（此句尚無繁中翻譯）
8:27.000–8:28.000
Yes. Okay. I can hear you.
（此句尚無繁中翻譯）
8:28.000–8:31.000
You know, it's a live stream, Aaron. I know you're used to TV.
（此句尚無繁中翻譯）
8:31.000–8:33.000
We kind of shoot from the hip here.
（此句尚無繁中翻譯）
8:33.000–8:37.000
I, we, we, we need AGI as soon as we can get it.
（此句尚無繁中翻譯）
8:37.000–8:40.000
So we will have audio figured out.
（此句尚無繁中翻譯）
8:40.000–8:42.000
Yeah. Okay. Agreed.
（此句尚無繁中翻譯）
8:42.000–8:43.000
Okay. Give me your thoughts.
（此句尚無繁中翻譯）
8:43.000–8:44.000
I know you've been tweeting about this.
（此句尚無繁中翻譯）
8:44.000–8:47.000
I know you've been tweeting a lot about open source.
（此句尚無繁中翻譯）
8:47.000–8:50.000
You've been talking about this for years, kind of as I have as well.
（此句尚無繁中翻譯）
8:50.000–8:52.000
What is your take on this moment?
（此句尚無繁中翻譯）
8:52.000–8:54.000
Is it different than anything else we've seen over the last few years?
（此句尚無繁中翻譯）
8:54.000–8:56.000
It certainly feels like that from where I sit.
（此句尚無繁中翻譯）
8:56.000–8:59.000
Yeah. I mean, I totally concur with, with Gabe.
（此句尚無繁中翻譯）
8:59.000–9:12.000
The, the, the, you know, there's been a big question in the Valley, which is how far behind is the open weights, kind of architecture from frontier closed sourced models.
（此句尚無繁中翻譯）
9:12.000–9:16.000
And there's like fundamentally different outcomes in the future of AI.
（此句尚無繁中翻譯）
9:16.000–9:24.000
If we see an, sort of an exponentially kind of increasing gap or a narrowing, or at least sort of sustaining three or six month gap.
（此句尚無繁中翻譯）
9:24.000–9:38.000
And what, what, what GLM five two kind of shows is, you know, we might still be in this sort of three to six month territory, as opposed to the kind of fast takeoff scenario of closed weights models versus open weights.
（此句尚無繁中翻譯）
9:38.000–9:43.000
And, and everything about the economics of AI changes depending on which outcome we're in.
（此句尚無繁中翻譯）
9:43.000–9:53.000
I think to your, to your intro, I actually do think for the, for the next, you know, year or two years, actually frontier level intelligence still is incredibly important.
（此句尚無繁中翻譯）
9:53.000–10:02.000
I think enterprises are only going to kind of peel off sort of workloads that are sort of good enough on, on open weights models.
（此句尚無繁中翻譯）
10:02.000–10:09.000
When they, when they really feel like we've reached a level where you can kind of, you know, bank on, on the level of intelligence we're getting.
（此句尚無繁中翻譯）
10:09.000–10:16.000
I still think that you, you, you need fable level kind of intelligence for a lot of the types of, of use cases that we're seeing in the enterprise.
（此句尚無繁中翻譯）
10:16.000–10:27.000
But what this shows, and you saw the founder of, of ZAI really talk about, you know, the ability to get to fable level intelligence and, you know, within the year with an open weights model.
（此句尚無繁中翻譯）
10:27.000–10:40.000
I think that, that does fundamentally change the architecture of AI because it means that now, whether you're using Harvey or box or factory or cognition or, you know, any, any other kind of neutral, you know, AI layer.
（此句尚無繁中翻譯）
10:40.000–10:47.000
You can kind of route the workloads between frontier intelligence and open weights models as appropriate for, for your tasks.
（此句尚無繁中翻譯）
10:47.000–10:53.000
So it's, it's very meaningful for what the future of the kind of AI stack looks like.
（此句尚無繁中翻譯）
10:53.000–10:56.000
Right. So the big thing right now is model routing, right?
（此句尚無繁中翻譯）
10:56.000–11:03.000
As Erin, as you laid out, it's use the frontier models for the hardest tasks, use the, you know, open source models for the easier ones.
（此句尚無繁中翻譯）
11:03.000–11:07.000
But you did say for the next one to two years, you think that frontier is going to be really important.
（此句尚無繁中翻譯）
11:07.000–11:11.000
And yes, the ZAI see our founder said that it was only a year.
（此句尚無繁中翻譯）
11:11.000–11:13.000
So what happens after that?
（此句尚無繁中翻譯）
11:13.000–11:17.000
Is that gap becoming even shorter?
（此句尚無繁中翻譯）
11:17.000–11:22.000
Like you said, three to six months, does that last or does that eventually go away as a whole?
（此句尚無繁中翻譯）
11:22.000–11:26.000
I think you'd be fine with a three to six month gap.
（此句尚無繁中翻譯）
11:26.000–11:39.000
Because that, that's not enough time for, you know, the, the big issue is, is what is the gap where if I don't use frontier on everything that my firm or my organization loses some kind of meaningful competitive advantage.
（此句尚無繁中翻譯）
11:39.000–11:58.000
And you can totally imagine a plausible scenario where if that gap was two to three years, you would, you would never use the kind of, you know, second best model because you just simply wouldn't be able to produce as effectively as a company in, in the code you're producing or the, the legal, you know, analysis you're doing for your clients or the, the marketing assets that you're generating.
（此句尚無繁中翻譯）
11:58.000–12:07.000
But in a world where there's a three or six month difference, that's, that's totally within the noise of, of the, the general diffusion of AI anyway, within an organization.
（此句尚無繁中翻譯）
12:07.000–12:10.000
So, so that, that's why you kind of see this fundamental shift.
（此句尚無繁中翻譯）
12:10.000–12:20.000
You don't need open weights to be at the exact same, you know, kind of the same week as, as frontier intelligence, but you can't have it be two or three months, two or three years behind.
（此句尚無繁中翻譯）
12:20.000–12:28.000
And it appears that we still are remaining that really kind of important zone where, where you're going to get very, very high quality open weights models.
（此句尚無繁中翻譯）
12:28.000–12:34.000
And then, you know, that gets to the second part, which is now you can post train these models on just your tasks.
（此句尚無繁中翻譯）
12:34.000–12:49.000
You know, everybody kind of looks at it as a, as a pure kind of cost of the token kind of equation, but actually just as big of a deal is the fact that now I can go and post train this model just on the purpose built use cases that my, you know, particular domain requires.
（此句尚無繁中翻譯）
12:49.000–13:02.000
And that has a different set of performance access, axes that matter, which is like, what are the capabilities I need the model to be able to do that actually now can be like built into the model in a way that, that you don't always get from kind of general purpose intelligence.
（此句尚無繁中翻譯）
13:02.000–13:13.000
Right. And that changes the dynamics, like once again, right, Gabe, I was going to come to you because you're actually operating in this app layer and like at the beginning, you're dependent on the frontier labs, open AI, Anthropic.
（此句尚無繁中翻譯）
13:13.000–13:16.000
And that's what we're starting to do for kind of both costs and performance reasons.
（此句尚無繁中翻譯）
13:16.000–13:22.000
And I think to the kind of longer term question of what is going to happen with frontier intelligence, I think the shift we're starting to see from models to agents.
（此句尚無繁中翻譯）
13:22.000–13:27.000
I think when you just use models, it was easy to kind of train a model and then that was separate from the rest of your business.
（此句尚無繁中翻譯）
13:27.000–13:31.000
And you could swap these out. If a model got better, that was kind of easy to replace.
（此句尚無繁中翻譯）
13:31.000–13:34.000
And then you could swap these out. If a model got better, that was kind of easy to replace.
（此句尚無繁中翻譯）
13:34.000–13:39.000
And then you could swap these out. If a model got better, that was kind of easy to replace.
（此句尚無繁中翻譯）
13:39.000–13:48.000
As you start moving to the to these agentic systems, this AI, these agents are getting so embedded in your business that creating the next version of frontier intelligence.
（此句尚無繁中翻譯）
13:48.000–13:53.000
I think a lot of that training data is going to come from the companies themselves.
（此句尚無繁中翻譯）
13:53.000–14:19.000
And so as a simple example, when a law firm works for like a large private equity firm and they help them do M&A, like that training data and creating frontier models can do that type of work like that training data can't go into the general models.
（此句尚無繁中翻譯）
14:19.000–14:30.000
And so I think there's going to be this really interesting question in the future of where is all of that training data eval coming from as these models start doing more and more economically valuable work.
（此句尚無繁中翻譯）
14:30.000–14:40.000
And part of the motivation for us of training these open source models is how do we allow our customers and each of our law firms or kind of in-house teams to build these systems?
（此句尚無繁中翻譯）
14:40.000–14:47.000
Right. And keep them for yourselves, right? You don't want to train someone else's model so that they can go and compete with you.
（此句尚無繁中翻譯）
14:47.000–14:58.000
What I'm hearing from both of you does not sound good for Frontier Labs or a company that has built its entire business model on staying ahead. Am I reading that right, Aaron?
（此句尚無繁中翻譯）
14:58.000–15:05.000
I feel like you're not going to say that, though, because you work with all of them. But like, what is the case for continuing to buy AI?
（此句尚無繁中翻譯）
15:05.000–15:08.000
I feel like I'm trying to ask that. What's that? Sorry?
（此句尚無繁中翻譯）
15:08.000–15:18.000
You know, what's the case for continuing to buy AI from Anthropic or OpenAI when that gap you essentially said is like, we're kind of there. It doesn't matter anymore. We're there.
（此句尚無繁中翻譯）
15:18.000–15:40.000
Well, I actually didn't say that. And this is where it's very nuanced. I think what you're going to see is sort of this barbell dynamic where Frontier Intelligence, again, you know, kind of might be always at two or three, five, 10% better than whatever the open weights equivalent is at any given time.
（此句尚無繁中翻譯）
15:40.000–15:59.000
Like that's the kind of scenario that as just a mental model is sort of, you know, you might want to just sort of, you know, kind of establish. And so what you'll have is that you'll still actually have an incredible demand for Frontier Intelligence, but often as things like the orchestrator or the planner or the reviewer of the work that's happening.
（此句尚無繁中翻譯）
15:59.000–16:07.000
And so interestingly, you might spend the same amount of money on both Frontier Intelligence and Open weights intelligence, but the token volume might be entirely different.
（此句尚無繁中翻譯）
16:07.000–16:17.000
You might have, you might have, you know, a third of or a fifth or a tenth of the tokens going into the Frontier Intelligence. Those will actually cost five or 10 times as much.
（此句尚無繁中翻譯）
16:17.000–16:28.000
But the bulk of sort of the heavy document processing, heavy analysis and combing through every sort of, you know, piece of text in the repository that you're trying to analyze.
（此句尚無繁中翻譯）
16:28.000–16:33.000
All of that might be happening on, again, a cheaper, faster model that is sort of more tuned for that workflow.
（此句尚無繁中翻譯）
16:33.000–16:41.000
So I actually don't know that this spells anything negative for even the Frontier Labs right now, assuming you are bullish on AI.
（此句尚無繁中翻譯）
16:41.000–16:48.000
If you're bullish on AI, what happens is we're going to be using, you know, a thousand times more tokens in 10 years from now than we are today.
（此句尚無繁中翻譯）
16:48.000–16:55.000
And you're still going to want Frontier Intelligence for a lot of your most important sort of work that an agent is doing.
（此句尚無繁中翻譯）
16:55.000–17:05.000
But there's just so many tokens that you now need for these agentic systems that actually you do have to then have something that is sort of cheaper and faster for a large amount of the work.
（此句尚無繁中翻譯）
17:05.000–17:13.000
The other nuance that I would just throw out there is none of this means that a Frontier Lab won't have cheaper and faster models.
（此句尚無繁中翻譯）
17:13.000–17:18.000
And so right now we're seeing that come from these open weights models as an example.
（此句尚無繁中翻譯）
17:18.000–17:23.000
But, you know, OpenAI released an open weights model a year ago as an example.
（此句尚無繁中翻譯）
17:23.000–17:30.000
And there's sort of no reason why if they became a great inference cloud as an example, that you wouldn't just run that model in their environment.
（此句尚無繁中翻譯）
17:30.000–17:47.000
So it's not obvious that this kind of tells you really anything about the market structure at the Frontier Labs other than the market structure of the applied layer is now the very logical place to have a lot of your routing go because it is sort of independent and model agnostic.
（此句尚無繁中翻譯）
17:47.000–17:50.000
And so it's going to just route the workload depending on what the use cases.
（此句尚無繁中翻譯）
17:50.000–17:58.000
Right. So that, you know, three to six month gap that is going to exist is going to be worth paying for, you're saying, for some of the hardest tasks.
（此句尚無繁中翻譯）
17:58.000–18:02.000
But, you know, Aaron, you bring up a good point. Yes.
（此句尚無繁中翻譯）
18:02.000–18:11.000
OpenAI, Google, Gemini have trained open source models, but they don't show up in terms of the adoption data and really on the benchmarks.
（此句尚無繁中翻譯）
18:11.000–18:15.000
Why are they letting the Chinese models run away with this?
（此句尚無繁中翻譯）
18:15.000–18:20.000
And I mean, right now it feels like it's maybe looking for something from reflection.
（此句尚無繁中翻譯）
18:20.000–18:26.000
But where are the Frontier Labs here? Do they turn their attention to it?
（此句尚無繁中翻譯）
18:26.000–18:28.000
It's hard to make money from open source. Is that part of it?
（此句尚無繁中翻譯）
18:28.000–18:30.000
Gabe, maybe you can go first.
（此句尚無繁中翻譯）
18:30.000–18:35.000
Yeah, I think it's kind of depends on where you are in the stack.
（此句尚無繁中翻譯）
18:35.000–18:42.000
So if you're kind of the chip providers, the cloud providers, I think you want there to be a lot of good open source models.
（此句尚無繁中翻譯）
18:42.000–18:46.000
I think if you're a closed source, you'd rather this gap be a bit bigger.
（此句尚無繁中翻譯）
18:46.000–18:50.000
But I think to Aaron's point, there's room for both of these.
（此句尚無繁中翻譯）
18:50.000–19:00.000
And I think like the US versus China have just kind of taken different strategies of I think we are pushing the frontier with closed and their strategy is kind of doing more open.
（此句尚無繁中翻譯）
19:00.000–19:03.000
But I think you've seen companies flip flop. Right.
（此句尚無繁中翻譯）
19:03.000–19:08.000
You've seen companies catch up and then go close source or, you know, release open source.
（此句尚無繁中翻譯）
19:08.000–19:12.000
And so I think it's not clear what the right strategy is yet.
（此句尚無繁中翻譯）
19:12.000–19:15.000
But I think China's like open source models are getting like very good.
（此句尚無繁中翻譯）
19:15.000–19:17.000
And that's something to like pay attention to.
（此句尚無繁中翻譯）
19:17.000–19:22.000
Right. And I wanted to sort of make this connection to I know, Gabe, like Harvey is looking at post training.
（此句尚無繁中翻譯）
19:22.000–19:24.000
I don't know what kind of model you're going to use.
（此句尚無繁中翻譯）
19:24.000–19:27.000
Is it a Chinese open source model, do you think?
（此句尚無繁中翻譯）
19:27.000–19:29.000
We need to support kind of a range of these.
（此句尚無繁中翻譯）
19:29.000–19:33.000
So we think of it like the same way we're agnostic to closed source models.
（此句尚無繁中翻譯）
19:33.000–19:39.000
We need to be agnostic to open source or we have some customers that can't use Chinese models for certain sensitive legal work.
（此句尚無繁中翻譯）
19:39.000–19:42.000
We have European customers that want to use Mistral.
（此句尚無繁中翻譯）
19:42.000–19:44.000
We have customers that want us models.
（此句尚無繁中翻譯）
19:44.000–19:49.000
So I think you'll want kind of this healthy ecosystem of open source models by region, by country.
（此句尚無繁中翻譯）
19:49.000–19:51.000
Yeah, that's right.
（此句尚無繁中翻譯）
19:51.000–20:00.000
Yeah. And I would I would say that that if I if I could kind of like design the perfect strategy of, you know, kind of one of the top few labs and Google is sort of the closest to this with with Gemma.
（此句尚無繁中翻譯）
20:00.000–20:11.000
I actually think you would want to consistently have a just close to frontier open weights model because now you kind of you get the ecosystem effects on both sides.
（此句尚無繁中翻譯）
20:11.000–20:22.000
And and that barbell that I was talking about of you want frontier intelligence from GPD, you know, 5.5 or Fable or, you know, the rumored GPD 5.6.
（此句尚無繁中翻譯）
20:22.000–20:32.000
But there's no reason that the the cheaper, faster version isn't in the same kind of model family from the same from the same sort of underlying weights and provider.
（此句尚無繁中翻譯）
20:32.000–20:37.000
So I actually think from the game theory standpoint, you'd probably want to be heavily supporting open weights.
（此句尚無繁中翻譯）
20:37.000–20:42.000
Now, there's a probably a philosophical reason why Anthropic in particular might not do that strategy.
（此句尚無繁中翻譯）
20:42.000–20:50.000
But for open AI where they already have precedent for doing so for Google, which clearly is doing that with Gemma, which is actually pretty actively used.
（此句尚無繁中翻譯）
20:50.000–20:56.000
It's just not not as obvious that it's at the at the frontier of a GLM or or a Kimmy.
（此句尚無繁中翻譯）
20:56.000–21:02.000
I think actually, you know, from a game theory standpoint, they probably should be pursuing this strategy and historical. Right.
（此句尚無繁中翻譯）
21:02.000–21:06.000
I mean, they sort of led the way on open source and only ecosystem for Android.
（此句尚無繁中翻譯）
21:06.000–21:11.000
Why wouldn't they do that again? It surprises me, too, that we don't hear or see more.
（此句尚無繁中翻譯）
21:11.000–21:15.000
I do think about cursor as well. Right. That was sort of laid the playbook.
（此句尚無繁中翻譯）
21:15.000–21:21.000
Gabe, maybe trained a custom model built on Kimmy's architecture. That's Moonshot, a Chinese company.
（此句尚無繁中翻譯）
21:21.000–21:27.000
It's now one of the hottest AI companies right now. And the whole secret sauce runs on a Chinese base model.
（此句尚無繁中翻譯）
21:27.000–21:32.000
And it feels like nobody actually remembers that or cares that much, especially in the corporate world.
（此句尚無繁中翻譯）
21:32.000–21:40.000
So you can see how you can sort of use these Chinese open source models to get ahead just in terms of especially if you're on the app layer.
（此句尚無繁中翻譯）
21:40.000–21:44.000
Aaron, though, I wonder like because I love that you bring up game theory.
（此句尚無繁中翻譯）
21:44.000–21:47.000
It feels like there's a lot of that going on right now.
（此句尚無繁中翻譯）
21:47.000–21:52.000
And so when I think about Fable and Carson, one of our viewers wrote in with a question on this, too,
（此句尚無繁中翻譯）
21:52.000–21:57.000
it kind of seems like a disaster for Anthropic.
（此句尚無繁中翻譯）
21:57.000–22:03.000
I mean, right after they cut off access to Fable Mythos, their most powerful model,
（此句尚無繁中翻譯）
22:03.000–22:11.000
you had all of these sort of policymakers around the world say, you know, how can you build on a model when your access could be cut off at any time?
（此句尚無繁中翻譯）
22:11.000–22:14.000
And it felt like it just gave so much momentum to open source.
（此句尚無繁中翻譯）
22:14.000–22:16.000
Do you think that that matters?
（此句尚無繁中翻譯）
22:16.000–22:19.000
Do you think that, you know, that was misplayed by Anthropic?
（此句尚無繁中翻譯）
22:19.000–22:21.000
What are the what's the impact of it?
（此句尚無繁中翻譯）
22:21.000–22:29.000
Well, it's it's not obvious that well, it kind of depends on, I guess, what what, you know,
（此句尚無繁中翻譯）
22:29.000–22:35.000
you know, either who you kind of like, you know, maybe blame for the situation we're in or kind of what transpired.
（此句尚無繁中翻譯）
22:35.000–22:43.000
But because it might have been misplayed by the US government for the same for the same kind of economic and national security reasons long term.
（此句尚無繁中翻譯）
22:43.000–22:54.000
If if if if you're kind of like encouraging sovereign AI in these other countries by sort of, you know, being the first to show that will will block access to an AI model through export controls.
（此句尚無繁中翻譯）
22:54.000–23:02.000
So so, you know, I actually think from a game theory standpoint, it might be more on the US government side to to have to kind of consider the implications of.
（此句尚無繁中翻譯）
23:02.000–23:20.000
And but absolutely, like like if you are another country right now, this was a sort of shock to the system that says, oh, actually, like, like, AI is not going to be like most other technologies where you sort of, you know, just kind of assume that software is going to work globally.
（此句尚無繁中翻譯）
23:20.000–23:33.000
There's a few nuances where, you know, there might be some privacy requirements in different regions, but generally it's considered just like a an export that the US is always just going to put out there and and to our kind of, you know, kind of economic benefit.
（此句尚無繁中翻譯）
23:33.000–23:52.000
AI, because the export itself is coming with with, you know, possible risks in at least in the table case, you know, the perceived risk of cybersecurity or other factors out now it's it's showing up as maybe it's something that the government wants a bit more control over and who should have access to it, at least now the precedent has been set for that.
（此句尚無繁中翻譯）
23:52.000–24:03.000
So then, you know, by definition, what another country should at least be doing is saying, oh, actually, I need to be able to protect my my ability to have access to intelligence.
（此句尚無繁中翻譯）
24:03.000–24:10.000
And if you were, you know, the EU and if you were, I mean, you know, China's fine because they're building their own models.
（此句尚無繁中翻譯）
24:10.000–24:21.000
You probably would sort of say, hey, we should actually be, you know, fueling, you know, more of either open source investment or post trained, you know, models and be able to get really, really good at making these things better and better.
（此句尚無繁中翻譯）
24:22.000–24:28.000
At a minimum is a hedge, you know, maybe you still do as much work as you can with US labs.
（此句尚無繁中翻譯）
24:28.000–24:31.000
But now you do need some sort of, you know, backstop.
（此句尚無繁中翻譯）
24:31.000–24:51.000
Because if for some reason, we treated access to AI as another kind of geopolitical or economic kind of weapon, you could see, you know, incredibly interesting consequences as a result of that, you know, imagine, imagine AI being used as a lever in in trade negotiations or trade deals in five years.
（此句尚無繁中翻譯）
24:51.000–24:59.000
You would you would sort of that would that could very much threaten your your kind of economic and national security if that was the case.
（此句尚無繁中翻譯）
24:59.000–25:11.000
So so we're, you know, it's really interesting because what's happening is we have these sort of blunt instruments, you know, export controls is a really interesting blunt instrument that I, you know, again, like very smart people probably predicted this.
（此句尚無繁中翻譯）
25:11.000–25:18.000
But like a year ago, I would not have put that in the in the board of possible chess moves of blocking access to an AI model.
（此句尚無繁中翻譯）
25:18.000–25:22.000
Well, now that that's that now that's a possible move, you sort of open up.
（此句尚無繁中翻譯）
25:22.000–25:26.000
Well, well, wait a second, maybe this is like going to be treated as like nuclear arms or not.
（此句尚無繁中翻譯）
25:26.000–25:41.000
I mean, I mean, not nuclear arms, but, you know, any kind of like weapon, in which case, you know, the set of controls that the government has of who has access to it is is way greater than what we had sort of considered before, which again, then means downstream as another country.
（此句尚無繁中翻譯）
25:41.000–25:47.000
You have to be preparing for that. What you just laid out, though, is kind of like exactly what we saw with chips.
（此句尚無繁中翻譯）
25:47.000–25:51.000
Right. And I'm with you. I never thought that could apply to models, but you saw that.
（此句尚無繁中翻譯）
25:51.000–25:57.000
So this is just telling sort of every government around the world, you need sovereign AI, you need to own something.
（此句尚無繁中翻譯）
25:57.000–26:05.000
And this process of distillation, which chips were actually not even as blunt of an instrument, though chips were even more sort of fine grained.
（此句尚無繁中翻譯）
26:05.000–26:11.000
This was this was this was very, very blunt is like non U.S. citizens is like, wow.
（此句尚無繁中翻譯）
26:11.000–26:15.000
OK, there's like no no real way to kind of execute that at scale.
（此句尚無繁中翻譯）
26:15.000–26:20.000
It's even more blunt, yet the solution is like sitting right there for anyone to take.
（此句尚無繁中翻譯）
26:20.000–26:23.000
Right. And it's just distillation. That is what the Chinese models are doing.
（此句尚無繁中翻譯）
26:23.000–26:33.000
It's still inside baseball, but it's this idea that you take a big, expensive, powerful model and use its outputs to train a smaller, cheaper model to behave like it.
（此句尚無繁中翻譯）
26:33.000–26:38.000
And Gabe, I wonder, like, this is how Chinese AI is spreading around the world.
（此句尚無繁中翻譯）
26:38.000–26:41.000
And what Aaron just laid out is the case for governments.
（此句尚無繁中翻譯）
26:41.000–26:44.000
But, you know, Harvey operates in a very sensitive industry.
（此句尚無繁中翻譯）
26:44.000–26:51.000
Do you see the same thing happening? Have you seen the same blowback happen after what happened with Mythos and Fable?
（此句尚無繁中翻譯）
26:51.000–26:54.000
Yeah, I would say distillation is one way.
（此句尚無繁中翻譯）
26:54.000–26:59.000
But I I think these Chinese labs are actually just training their own models.
（此句尚無繁中翻譯）
26:59.000–27:06.000
Like I think it's a bit like extreme for us to just say the only reason they're catching up is because they're distilling this.
（此句尚無繁中翻譯）
27:06.000–27:15.000
They're U.S. models. I actually think they're doing incredible research and like their open source, even independent of distillation, I think is competitive or ahead of ours.
（此句尚無繁中翻譯）
27:15.000–27:21.000
And then I think what you're seeing the labs do is making it harder to distill their models like that's what you saw with Fable.
（此句尚無繁中翻譯）
27:21.000–27:25.000
And it's like I think that strategy will work as these models get smarter.
（此句尚無繁中翻譯）
27:25.000–27:30.000
You can just tell the model, don't let users distill me when they use them.
（此句尚無繁中翻譯）
27:30.000–27:33.000
And it's as they get smarter, they'll get better at preventing that.
（此句尚無繁中翻譯）
27:33.000–27:35.000
So I think that is defensible.
（此句尚無繁中翻譯）
27:35.000–27:48.000
But then I think the kind of the blowback we saw is kind of exactly what Aaron mentioned of even within the U.S., as these systems start becoming mission critical, like you just need to think about business continuity.
（此句尚無繁中翻譯）
27:48.000–27:54.000
And if one of the labs that you're using runs out of compute, the government prevents you from accessing their models.
（此句尚無繁中翻譯）
27:54.000–27:59.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
27:59.000–28:02.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:02.000–28:04.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:04.000–28:06.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:06.000–28:08.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:08.000–28:11.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:11.000–28:14.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:14.000–28:18.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:18.000–28:20.000
And so I think that's the way that you're using the data to be able to do that.
（此句尚無繁中翻譯）
28:20.000–28:25.000
That's so fascinating. So you see it among sovereigns and then you also see it among corporate America.
（此句尚無繁中翻譯）
28:25.000–28:28.000
And that's why I guess I started by saying it felt like a miscalculation.
（此句尚無繁中翻譯）
28:28.000–28:31.000
You're right, Aaron, probably by the government, less so anthropic.
（此句尚無繁中翻譯）
28:31.000–28:33.000
But the result is sort of the same.
（此句尚無繁中翻譯）
28:33.000–28:40.000
Aaron, I want to give you a word on distillation, too, because, you know, I feel like the smartest people I talk to say that distillation should not be a dirty word.
（此句尚無繁中翻譯）
28:40.000–28:48.000
Right. When you see that anthropic is suing Alibaba, that is non sanctioned distillation, but can actually be a really positive force.
（此句尚無繁中翻譯）
28:48.000–28:52.000
Right. Turning expensive AI into cheaper products that can actually scale.
（此句尚無繁中翻譯）
28:52.000–28:57.000
And as Gabe was saying, like, do you agree that the Chinese are really doing some stuff on the leading edge, innovative edge?
（此句尚無繁中翻譯）
28:57.000–29:01.000
And it's too simplistic to say they're just distilling.
（此句尚無繁中翻譯）
29:01.000–29:08.000
Yeah, I would say I'm not as close to the latest views on what the new method, what the latest methods are.
（此句尚無繁中翻譯）
29:08.000–29:12.000
And I would probably bias toward Gabe's view of that.
（此句尚無繁中翻譯）
29:12.000–29:19.000
And like it also is just like first principles like China has incredible, you know, smart scientists.
（此句尚無繁中翻譯）
29:19.000–29:24.000
Some of them are the ones that came over here and are, you know, leaders at many of these labs.
（此句尚無繁中翻譯）
29:24.000–29:29.000
So it's sort of like a it's a very tractable problem to get very good at at AI and AI training.
（此句尚無繁中翻譯）
29:29.000–29:35.000
And really, it's just been classically like the chip problem of do you have enough compute?
（此句尚無繁中翻譯）
29:35.000–29:37.000
And that's also tractable for the most part.
（此句尚無繁中翻譯）
29:37.000–29:51.000
And then from a kind of philosophical standpoint on distillation, I don't know, I don't I'm not bothered by distillation because, you know, if I were bothered on distillation, I'd probably have to be bothered on the fact that these AI models are just trained on then the public Internet.
（此句尚無繁中翻譯）
29:51.000–30:02.000
It's like we're all benefiting from the collective sort of creation of information and knowledge and everything is sort of riding on on some other set of information that's out there.
（此句尚無繁中翻譯）
30:02.000–30:10.000
Like, you know, I would be bothered by, you know, a model being trained on Wikipedia if I if I was bothered by bothered by distillation.
（此句尚無繁中翻譯）
30:10.000–30:16.000
So so to me, it's sort of all within the realm of, you know, you need access to data.
（此句尚無繁中翻譯）
30:16.000–30:21.000
Obviously, the more intelligent data you can get, the better your model becomes.
（此句尚無繁中翻譯）
30:21.000–30:29.000
And and at the same time, I think that there should be kind of a cat and mouse game because for competitive reasons, you probably want to block that if you're a frontier lab.
（此句尚無繁中翻譯）
30:29.000–30:38.000
But I don't think of it as a as as some kind of, you know, unethical or or, you know, overly kind of, you know, complicated problem.
（此句尚無繁中翻譯）
30:38.000–30:42.000
Yeah, it's a great point. It kind of goes back to the beginning of this, like modern AI era.
（此句尚無繁中翻譯）
30:42.000–30:47.000
All of these models are trained on, you know, the Internet's all of the Internet's knowledge.
（此句尚無繁中翻譯）
30:47.000–30:48.000
That's a form of distillation.
（此句尚無繁中翻譯）
30:48.000–30:53.000
They were they were trained on my my trolling tweets from 15 years ago.
（此句尚無繁中翻譯）
30:53.000–30:56.000
I mean, how mad am I about that? I don't know.
（此句尚無繁中翻譯）
30:56.000–30:58.000
That's a scary, scary thought.
（此句尚無繁中翻譯）
30:58.000–31:02.000
Aaron Levy tweets and all of Reddit.
（此句尚無繁中翻譯）
31:02.000–31:07.000
I don't know, Gabe, I haven't gone too deep into your X feed, but it's trained on that, too, I'm sure.
（此句尚無繁中翻譯）
31:07.000–31:11.000
Last one I wanted to ask you guys about because this is a little bit of a turn.
（此句尚無繁中翻譯）
31:11.000–31:15.000
I want to ask you about Claude Tagg because producer Jasmine, she's very excited about it.
（此句尚無繁中翻譯）
31:15.000–31:16.000
She's trying to get me there.
（此句尚無繁中翻譯）
31:16.000–31:21.000
I don't fully, fully understand the appeal, but she is always right about these things.
（此句尚無繁中翻譯）
31:21.000–31:24.000
I wonder, Aaron, first, I know you've been tweeting about this.
（此句尚無繁中翻譯）
31:24.000–31:26.000
Like, can you break it down?
（此句尚無繁中翻譯）
31:26.000–31:31.000
It's something that always lives in your slack and builds context over time, learns your company.
（此句尚無繁中翻譯）
31:31.000–31:34.000
It feels like lock in to me, but it also seems really important.
（此句尚無繁中翻譯）
31:34.000–31:41.000
Aaron, is this like going to be as big as Co-Work or some of the other kind of land shifting things that Anthropoc has done this year?
（此句尚無繁中翻譯）
31:41.000–31:49.000
I heard to kind of put it in the in the overall kind of graph on that front.
（此句尚無繁中翻譯）
31:49.000–32:00.000
But the reason why it's a pretty big deal is we're very used to AI in kind of single player mode where I have access to Claude or Codex or ChatGPT or Co-Work.
（此句尚無繁中翻譯）
32:00.000–32:02.000
And this is sort of my relationship with AI.
（此句尚無繁中翻譯）
32:02.000–32:08.000
And that AI is effectively acting as me in a variety of systems.
（此句尚無繁中翻譯）
32:08.000–32:15.000
And so it's incredibly powerful because I can then spin up tasks that I would otherwise do a bunch of times across these agents.
（此句尚無繁中翻譯）
32:15.000–32:18.000
But it's again, very single player mode.
（此句尚無繁中翻譯）
32:18.000–32:33.000
And what Claude tag kind of tries to do, and this is, you know, I think building on the recent Zeitgeist of OpenClaw and Hermes and even kind of you're seeing this in agentic coding systems like Factory and Devon, is no, like what if it's a co-worker?
（此句尚無繁中翻譯）
32:33.000–32:40.000
It's not rep, it's not Aaron or Gabe or anybody in the in the organization, it's its own sort of entity.
（此句尚無繁中翻譯）
32:40.000–32:45.000
And it has access to a set of resources that any other kind of entity would have had access to.
（此句尚無繁中翻譯）
32:45.000–32:49.000
And you kind of interacted with it like another user in your system.
（此句尚無繁中翻譯）
32:49.000–32:56.000
And the reason why that's meaningful is that it has shared context across whatever that group that has access to it is doing.
（此句尚無繁中翻譯）
32:56.000–32:59.000
And then you just sort of punt tasks to it and it comes back.
（此句尚無繁中翻譯）
32:59.000–33:06.000
But those are again, kind of group enabled tasks just as again, a colleague would do inside of a Slack channel in this case.
（此句尚無繁中翻譯）
33:06.000–33:14.000
So, so it's, it's, it's probably a very big deal in, in sort of the philosophical direction that it's moving toward.
（此句尚無繁中翻譯）
33:14.000–33:19.000
And, and it has just like a different vector of use cases that, that it enables.
（此句尚無繁中翻譯）
33:19.000–33:26.000
It's less about my personal productivity and more about sort of a shared intelligence for some, from, you know, group of users.
（此句尚無繁中翻譯）
33:26.000–33:29.000
So you're in a Slack channel called the sales team Slack channel.
（此句尚無繁中翻譯）
33:29.000–33:36.000
And, and you're, you're punting off tasks of, Hey, can you generate a deck for this sales presentation?
（此句尚無繁中翻譯）
33:36.000–33:37.000
We're, we're about to go into.
（此句尚無繁中翻譯）
33:37.000–33:44.000
And it needs access to your underlying, you know, marketing assets that are stored in box and your other data, you know, stored in Salesforce.
（此句尚無繁中翻譯）
33:44.000–33:48.000
Just as you would a colleague that was about to go in that sales presentation with you.
（此句尚無繁中翻譯）
33:48.000–33:50.000
So, so that's the kind of benefit of it.
（此句尚無繁中翻譯）
33:50.000–33:57.000
And I think it's another kind of, you know, point in the timeline of like, what is the future of how we're going to work with these agentic systems?
（此句尚無繁中翻譯）
33:57.000–33:58.000
What is the new user experience?
（此句尚無繁中翻譯）
33:58.000–34:00.000
What is the, the, the new interface?
（此句尚無繁中翻譯）
34:00.000–34:04.000
And, you know, from a kind of a lock in point, I think there's an interesting point.
（此句尚無繁中翻譯）
34:04.000–34:10.000
Like on one hand, yes, you're, you're sort of, you know, you're, you're, you're, you're using the cloud particular paradigm.
（此句尚無繁中翻譯）
34:10.000–34:19.000
On the other hand, you can choose to decide how much or how little you want to sort of use the cloud sort of intelligence layer.
（此句尚無繁中翻譯）
34:19.000–34:23.000
Like you can plug in an MCP server of an entirely other agentic system.
（此句尚無繁中翻譯）
34:23.000–34:26.000
So cloud could just be a router to a bunch of other, other tools.
（此句尚無繁中翻譯）
34:26.000–34:29.000
I'm making up the use case, but I'm sure it could be a router to Harvey.
（此句尚無繁中翻譯）
34:29.000–34:39.000
So, so I, I think there's actually like it, you can just design how you want to implement these systems to whatever level of lock in or, or not lock in you want to, you want to drive.
（此句尚無繁中翻譯）
34:39.000–34:40.000
Right.
（此句尚無繁中翻譯）
34:40.000–34:44.000
Feels like it's certainly like the whole industry is heading that way, more of like a buffet.
（此句尚無繁中翻譯）
34:44.000–34:45.000
You can choose what you want.
（此句尚無繁中翻譯）
34:45.000–34:47.000
Gabe, last word to you.
（此句尚無繁中翻譯）
34:47.000–34:51.000
I guess, what should we be looking out for in the next week or so with these live streams?
（此句尚無繁中翻譯）
34:51.000–34:55.000
We really try to be on like what Silicon Valley is talking about this week.
（此句尚無繁中翻譯）
34:55.000–34:59.000
It's very much ZAI's GLM 5.2.
（此句尚無繁中翻譯）
34:59.000–35:02.000
What do you, what should we be watching for the next week ahead?
（此句尚無繁中翻譯）
35:02.000–35:08.000
I feel like it'll be interesting to see what happens kind of with the fable five kind of in the next week.
（此句尚無繁中翻譯）
35:08.000–35:15.000
When that comes back, I think you're going to see more about kind of what Aaron was talking about with Claude Tagg.
（此句尚無繁中翻譯）
35:15.000–35:22.000
Like I think that philosophical shift of just thinking about agents more and more as just employees in your company.
（此句尚無繁中翻譯）
35:22.000–35:29.000
Like you saw Rippling kind of announce they're building kind of a unified like data layer of employees.
（此句尚無繁中翻譯）
35:29.000–35:35.000
I think more and more we're thinking about our product as how do we give agents to a law firm or an in-house department.
（此句尚無繁中翻譯）
35:35.000–35:45.000
And instead of kind of one player like Aaron mentioned, how do you start organizing these kind of like teams of agents and humans to complete these very complex tasks?
（此句尚無繁中翻譯）
35:45.000–35:47.000
So I think you'll kind of keep seeing that trend.
（此句尚無繁中翻譯）
35:47.000–35:50.000
And that's kind of like the trend we're most excited about.
（此句尚無繁中翻譯）
35:50.000–35:54.000
Right. We'll call it multiplayer mode since Aaron's been talking about single player mode.
（此句尚無繁中翻譯）
35:54.000–35:56.000
Thank you both so much, Aaron and Gabe.
（此句尚無繁中翻譯）
35:56.000–35:58.000
Lovely to get you both here.
（此句尚無繁中翻譯）
35:58.000–36:01.000
Thanks for all your insights and hope to talk to you again soon.
（此句尚無繁中翻譯）
36:01.000–36:02.000
Thanks a lot.
（此句尚無繁中翻譯）
36:02.000–36:03.000
All of what we talked.
（此句尚無繁中翻譯）
36:03.000–36:06.000
Distillation open source.
（此句尚無繁中翻譯）
36:06.000–36:09.000
It runs through the same question for investors.
（此句尚無繁中翻譯）
36:09.000–36:11.000
What does it do for the chip trade?
（此句尚無繁中翻譯）
36:11.000–36:14.000
That has been super relevant to Wall Street, of course.
（此句尚無繁中翻譯）
36:14.000–36:15.000
Yesterday we got a big data point.
（此句尚無繁中翻譯）
36:15.000–36:18.000
OpenAI and Broadcom unveiled Jalapeno.
（此句尚無繁中翻譯）
36:18.000–36:22.000
That's OpenAI's first custom inference chip built from scratch in just nine months.
（此句尚無繁中翻譯）
36:22.000–36:25.000
OpenAI used its own AI models to help design it.
（此句尚無繁中翻譯）
36:25.000–36:32.000
Broadcom CEO Hawk Tan says that it cuts inference costs by roughly 50% versus current Nvidia GPUs.
（此句尚無繁中翻譯）
36:32.000–36:35.000
The person I want to break all of this down with is Stacy Rascon.
（此句尚無繁中翻譯）
36:35.000–36:40.000
He has been covering semiconductors at Bernstein for over a decade called the Nvidia trade early.
（此句尚無繁中翻譯）
36:40.000–36:43.000
Stacy, you don't sugarcoat either.
（此句尚無繁中翻譯）
36:43.000–36:44.000
Two decades.
（此句尚無繁中翻譯）
36:44.000–36:45.000
Almost two decades.
（此句尚無繁中翻譯）
36:45.000–36:48.000
Almost 16 years. Is that right?
（此句尚無繁中翻譯）
36:48.000–36:49.000
I should have rounded up.
（此句尚無繁中翻譯）
36:49.000–36:51.000
Oh, then I definitely needed to round it up.
（此句尚無繁中翻譯）
36:51.000–36:54.000
Well, I know that's why we always love talking to you.
（此句尚無繁中翻譯）
36:54.000–36:57.000
You also have all of the technical side, but you can break it down.
（此句尚無繁中翻譯）
36:57.000–36:59.000
Give us your thoughts on Jalapeno.
（此句尚無繁中翻譯）
36:59.000–37:04.000
I mean, to me, it just struck me as a huge, like, develop, like maybe a new era for chip development.
（此句尚無繁中翻譯）
37:04.000–37:08.000
Right? These are some of the most complicated objects ever built by humans.
（此句尚無繁中翻譯）
37:08.000–37:10.000
I guess, like, do you buy it?
（此句尚無繁中翻譯）
37:10.000–37:13.000
Is it this, has it improved this much?
（此句尚無繁中翻譯）
37:13.000–37:17.000
Well, so look, so chips are the most complicated things that humanity has ever built.
（此句尚無繁中翻譯）
37:17.000–37:19.000
And as you know, we've talked before.
（此句尚無繁中翻譯）
37:19.000–37:20.000
I love this space.
（此句尚無繁中翻譯）
37:20.000–37:22.000
And that's one of the reasons why.
（此句尚無繁中翻譯）
37:22.000–37:26.000
So, look, this was not unexpected.
（此句尚無繁中翻譯）
37:26.000–37:27.000
I want to say that.
（此句尚無繁中翻譯）
37:27.000–37:32.000
So we've known for a long time that Broadcom is working on a custom chip with OpenAI.
（此句尚無繁中翻譯）
37:32.000–37:38.000
And in fact, the two companies have a 10 gigawatt deal over the next five years.
（此句尚無繁中翻譯）
37:38.000–37:40.000
And we'll see how much of it actually ships.
（此句尚無繁中翻譯）
37:40.000–37:45.000
But they've got a 10 gigawatt deal for multiple generations of these chips in play already.
（此句尚無繁中翻譯）
37:45.000–37:53.000
On Broadcom's last earnings call they talked about, I think, in 27, they're supposed to ship, I can't remember, 1.2 or 1.3 gigawatts to OpenAI.
（此句尚無繁中翻譯）
37:53.000–37:55.000
So that is presumably this chip.
（此句尚無繁中翻譯）
37:55.000–37:59.000
So the fact that they announced it was not a surprise.
（此句尚無繁中翻譯）
37:59.000–38:03.000
And in fact, one would hope that they would have been announcing it soon.
（此句尚無繁中翻譯）
38:03.000–38:05.000
Because it's supposed to start shipping next year.
（此句尚無繁中翻譯）
38:05.000–38:06.000
They've been talking about it.
（此句尚無繁中翻譯）
38:06.000–38:07.000
Yeah.
（此句尚無繁中翻譯）
38:07.000–38:10.000
Now that being said, look, it's great, right?
（此句尚無繁中翻譯）
38:10.000–38:12.000
And I don't, you know, they put out a lot of stuff.
（此句尚無繁中翻譯）
38:12.000–38:16.000
I guess he had said that, you're right, he said it would lower the cost by 50.
（此句尚無繁中翻譯）
38:16.000–38:22.000
I don't know if that was the cost of inference or the cost of the chip itself relative to the current state of the art, which would probably be an NVIDIA GPU.
（此句尚無繁中翻譯）
38:22.000–38:29.000
You have to be a little careful with those kind of comparisons as well, because you don't exactly know what they're talking about.
（此句尚無繁中翻譯）
38:29.000–38:31.000
Like, people tend to cherry pick stuff.
（此句尚無繁中翻譯）
38:31.000–38:37.000
And at the end of the day, it's not really the cost of the chip all by itself that necessarily matters.
（此句尚無繁中翻譯）
38:37.000–38:42.000
It's the performance per watt, you know, performance per dollar total cost of ownership.
（此句尚無繁中翻譯）
38:42.000–38:47.000
What the press release was talking about was sort of substantial improvements in performance per watt.
（此句尚無繁中翻譯）
38:47.000–38:50.000
I think it said it was still an early test, early phase testing.
（此句尚無繁中翻譯）
38:50.000–38:57.000
They're not in volume production with this yet, but it sounds like the early tests are very, very encouraging along those fronts.
（此句尚無繁中翻譯）
38:57.000–39:00.000
And so, look, I'm sure it'll be a good chip.
（此句尚無繁中翻譯）
39:00.000–39:03.000
And it sounds like they're getting ready to ship a lot of them.
（此句尚無繁中翻譯）
39:03.000–39:06.000
Don't open AI, like I said, well over a gigawatt next year.
（此句尚無繁中翻譯）
39:06.000–39:10.000
And like over the next five years, potentially up to 10 gigawatts.
（此句尚無繁中翻譯）
39:10.000–39:11.000
I'm sure it's not going to be a good chip.
（此句尚無繁中翻譯）
39:11.000–39:15.000
Okay, so I'm glad you stepped back and you said, okay, it comes as a surprise to no one that they have a chip.
（此句尚無繁中翻譯）
39:15.000–39:16.000
They needed to.
（此句尚無繁中翻譯）
39:16.000–39:24.000
If you're like a hyperscaler or an AI lab right now and you're not working on your own chip in some way, like you are doing something wrong here.
（此句尚無繁中翻譯）
39:24.000–39:25.000
Fine.
（此句尚無繁中翻譯）
39:25.000–39:28.000
They're all buying a lot of GPUs too.
（此句尚無繁中翻譯）
39:28.000–39:30.000
So again, to go back to open AI.
（此句尚無繁中翻譯）
39:30.000–39:32.000
So they have this 10 gigawatt deal with Braggam.
（此句尚無繁中翻譯）
39:32.000–39:34.000
They also have a 10 gigawatt deal with Nvidia.
（此句尚無繁中翻譯）
39:34.000–39:38.000
And they have a six gigawatt deal with AMD.
（此句尚無繁中翻譯）
39:38.000–39:40.000
Everyone is just trying to get compute.
（此句尚無繁中翻譯）
39:40.000–39:42.000
They're trying to get their costs down.
（此句尚無繁中翻譯）
39:42.000–39:44.000
They're trying to be less dependent on Nvidia.
（此句尚無繁中翻譯）
39:44.000–39:48.000
So let me tell you the thing that I thought was most interesting about this announcement.
（此句尚無繁中翻譯）
39:48.000–39:52.000
And I don't know, like, that's why I'm curious if you thought it was interesting.
（此句尚無繁中翻譯）
39:52.000–39:56.000
Nine months built in nine months from scratch.
（此句尚無繁中翻譯）
39:56.000–40:00.000
Like that to me is just, that's like mind boggling.
（此句尚無繁中翻譯）
40:00.000–40:03.000
These are the most complicated things to make in humanity, as you said.
（此句尚無繁中翻譯）
40:03.000–40:05.000
So do we believe it?
（此句尚無繁中翻譯）
40:05.000–40:09.000
So that is fast, but Braggam again has talked about exactly that.
（此句尚無繁中翻譯）
40:09.000–40:13.000
And that's part of their differentiation on, on this custom chip.
（此句尚無繁中翻譯）
40:13.000–40:16.000
There's a number of players that can do custom chips.
（此句尚無繁中翻譯）
40:16.000–40:20.000
Braggam has literally said in the past, we can design a chip in nine months.
（此句尚無繁中翻譯）
40:20.000–40:22.000
Like that, that's not something new that they've said.
（此句尚無繁中翻譯）
40:22.000–40:26.000
This is the first time I think that they've actually shown a real example.
（此句尚無繁中翻譯）
40:26.000–40:27.000
Okay.
（此句尚無繁中翻譯）
40:27.000–40:29.000
And called it out that it was developed in nine months.
（此句尚無繁中翻譯）
40:29.000–40:31.000
But again, they have said in the past that they could do that.
（此句尚無繁中翻譯）
40:31.000–40:34.000
And so it's nice to actually see that they can.
（此句尚無繁中翻譯）
40:34.000–40:37.000
It's still that, but okay, here's what I mean.
（此句尚無繁中翻譯）
40:37.000–40:39.000
Does this change the whole chip space?
（此句尚無繁中翻譯）
40:39.000–40:45.000
We've talked about, you know, the cycle getting shorter and Nvidia putting out new models now on a yearly basis.
（此句尚無繁中翻譯）
40:45.000–40:48.000
If you can create it, I know you, I feel like I'm getting skepticism from you.
（此句尚無繁中翻譯）
40:48.000–40:49.000
Is that because?
（此句尚無繁中翻譯）
40:49.000–40:51.000
I think that, no, no, no, no, I'm not skeptical at all.
（此句尚無繁中翻譯）
40:51.000–40:52.000
No, no, no, it looks right.
（此句尚無繁中翻譯）
40:52.000–40:54.000
But I mean the chip space is changing anyways.
（此句尚無繁中翻譯）
40:54.000–40:55.000
Right?
（此句尚無繁中翻譯）
40:55.000–40:56.000
Okay.
（此句尚無繁中翻譯）
40:56.000–40:57.000
So is this an aspect of that?
（此句尚無繁中翻譯）
40:57.000–40:58.000
Absolutely.
（此句尚無繁中翻譯）
40:58.000–41:03.000
Are we seeing a tremendous amount of new advanced silicon designs?
（此句尚無繁中翻譯）
41:03.000–41:06.000
Because now we actually have a reason to do them and someone that's actually willing to pay for them.
（此句尚無繁中翻譯）
41:06.000–41:07.000
Yes.
（此句尚無繁中翻譯）
41:07.000–41:08.000
I'll give you an example.
（此句尚無繁中翻譯）
41:08.000–41:12.000
You know, it's not even this, like we're seeing like chip startups now for a change for, for example.
（此句尚無繁中翻譯）
41:12.000–41:13.000
Yeah.
（此句尚無繁中翻譯）
41:13.000–41:20.000
I remember over 10 years ago, I had started to do a piece or to write a piece on venture capital investments in semiconductors.
（此句尚無繁中翻譯）
41:20.000–41:26.000
I shelved it because there wasn't any, like all of the VC investments back then, it was a corporate VC.
（此句尚無繁中翻譯）
41:26.000–41:27.000
That was it.
（此句尚無繁中翻譯）
41:27.000–41:35.000
There wasn't a lot of, of, of, of actual, like, like a standard VC investment because it costs a lot of money and costs a lot of time.
（此句尚無繁中翻譯）
41:35.000–41:38.000
And you know, the exits, uh, opportunities weren't certain.
（此句尚無繁中翻譯）
41:38.000–41:40.000
And back then they would have just like invested in SAS.
（此句尚無繁中翻譯）
41:40.000–41:41.000
It was a lot easier.
（此句尚無繁中翻譯）
41:41.000–41:42.000
Yeah.
（此句尚無繁中翻譯）
41:42.000–41:52.000
Um, we're actually starting to see a lot of VC investments in, in real meaningful startups that are getting valued both in the private as well as sometimes even in the public markets now, like in the billions or even tens of billions of dollars.
（此句尚無繁中翻譯）
41:52.000–41:56.000
So there's an actual reason to do this now.
（此句尚無繁中翻譯）
41:56.000–42:01.000
Like there, there, there's a, there's a use case and an application where like dollars are flowing in.
（此句尚無繁中翻譯）
42:01.000–42:04.000
And so I think the space is, is absolutely changing.
（此句尚無繁中翻譯）
42:04.000–42:06.000
And I, this is one aspect of it clearly.
（此句尚無繁中翻譯）
42:06.000–42:07.000
Right.
（此句尚無繁中翻譯）
42:07.000–42:08.000
Yeah.
（此句尚無繁中翻譯）
42:08.000–42:09.000
And okay.
（此句尚無繁中翻譯）
42:09.000–42:17.000
So as we've seen more competition, the story has always been like, what does it mean for Nvidia only game in town, but like the whole pie is expanding.
（此句尚無繁中翻譯）
42:17.000–42:18.000
Right.
（此句尚無繁中翻譯）
42:18.000–42:19.000
So that's what.
（此句尚無繁中翻譯）
42:19.000–42:20.000
Yeah.
（此句尚無繁中翻譯）
42:20.000–42:21.000
Tell me.
（此句尚無繁中翻譯）
42:21.000–42:27.000
So as, as, as, as an equity analyst, we get paid to overcomplicate things sometimes.
（此句尚無繁中翻譯）
42:27.000–42:28.000
And there's always.
（此句尚無繁中翻譯）
42:28.000–42:30.000
Worries about competition.
（此句尚無繁中翻譯）
42:30.000–42:32.000
And I would say, first of all, you need to step back.
（此句尚無繁中翻譯）
42:32.000–42:33.000
This is semiconductors.
（此句尚無繁中翻譯）
42:33.000–42:34.000
There was always competition.
（此句尚無繁中翻譯）
42:34.000–42:38.000
There was always somebody like looking to eat your lunch if you will let them.
（此句尚無繁中翻譯）
42:38.000–42:39.000
Right.
（此句尚無繁中翻譯）
42:39.000–42:46.000
So it is imperative on the leaders in this space to make sure that they stay that way to, and to invest in the roadmap and to hire the best people and so on and so forth.
（此句尚無繁中翻譯）
42:46.000–42:47.000
And they're all doing that.
（此句尚無繁中翻譯）
42:47.000–42:54.000
I would also say the opportunity is getting so big that right now it almost doesn't matter.
（此句尚無繁中翻譯）
42:54.000–42:59.000
My general view, for example, you know, you talk about Broadcom versus Nvidia, they talk about GPUs versus TPUs.
（此句尚無繁中翻譯）
42:59.000–43:04.000
And my general view is that I get that, but it's, it's probably the wrong question.
（此句尚無繁中翻譯）
43:04.000–43:08.000
Like the right question probably right now is more, is the opportunity in front of us still bigger?
（此句尚無繁中翻譯）
43:08.000–43:09.000
Is it not?
（此句尚無繁中翻譯）
43:09.000–43:11.000
Because if it's big, you know, everybody will thrive.
（此句尚無繁中翻譯）
43:11.000–43:13.000
And if it's not, everybody's screwed.
（此句尚無繁中翻譯）
43:13.000–43:14.000
Yeah.
（此句尚無繁中翻譯）
43:14.000–43:15.000
And so far it's really big.
（此句尚無繁中翻譯）
43:15.000–43:17.000
The assumption is that it's bigger, right?
（此句尚無繁中翻譯）
43:17.000–43:19.000
The assumption is straight up and to the right.
（此句尚無繁中翻譯）
43:19.000–43:22.000
Well, it's been, it's been straight up to the right so, so far, right?
（此句尚無繁中翻譯）
43:22.000–43:25.000
Is there any reason to think it might be different?
（此句尚無繁中翻譯）
43:25.000–43:29.000
Well, I mean, you're seeing it in G, it's not even just in this space, you're seeing it here.
（此句尚無繁中翻譯）
43:29.000–43:33.000
And you're also seeing it in networking, you see copper versus optical and all kinds of
（此句尚無繁中翻譯）
43:33.000–43:37.000
where there's all, there's always going to be alternatives to this market is, you know, hundreds
（此句尚無繁中翻譯）
43:37.000–43:40.000
of billions or potentially even trillions of dollars.
（此句尚無繁中翻譯）
43:40.000–43:44.000
Somebody is always going to have alternatives, but you can look at the numbers.
（此句尚無繁中翻譯）
43:44.000–43:47.000
So Nvidia, I don't know, they'll do what $500 billion next year, right?
（此句尚無繁中翻譯）
43:47.000–43:50.000
Broadcom got you to a hundred billion, right?
（此句尚無繁中翻譯）
43:50.000–43:51.000
Nope.
（此句尚無繁中翻譯）
43:51.000–43:53.000
And they'll probably do more than that, right?
（此句尚無繁中翻譯）
43:53.000–43:58.000
And I don't know, Qualcomm had an analyst day yesterday and they said they're going to
（此句尚無繁中翻譯）
43:58.000–44:01.000
do $5 billion from zero basically, but still billions.
（此句尚無繁中翻譯）
44:01.000–44:05.000
And you got MediaTek and you got Marvell and AMD and all that.
（此句尚無繁中翻譯）
44:05.000–44:06.000
They're all doing fine.
（此句尚無繁中翻譯）
44:06.000–44:08.000
They're all growing, growing tons.
（此句尚無繁中翻譯）
44:08.000–44:14.000
I mean, so in, in, in the, if you were to calculate the percentages, like our shares
（此句尚無繁中翻譯）
44:14.000–44:18.000
moving up probably a little bit, but is it a really a problem at this point?
（此句尚無繁中翻譯）
44:18.000–44:19.000
I don't think it is.
（此句尚無繁中翻譯）
44:19.000–44:20.000
Right.
（此句尚無繁中翻譯）
44:20.000–44:23.000
And you're saying, even though we have so many of these different players, this is still
（此句尚無繁中翻譯）
44:23.000–44:24.000
a supply problem.
（此句尚無繁中翻譯）
44:24.000–44:27.000
Like we see that especially in memory.
（此句尚無繁中翻譯）
44:27.000–44:30.000
And so I wanted to ask you, I wanted to kind of connect this to the earlier conversation.
（此句尚無繁中翻譯）
44:30.000–44:33.000
We were talking about Chinese open source, right?
（此句尚無繁中翻譯）
44:33.000–44:36.000
Being a solution because it's more abundant.
（此句尚無繁中翻譯）
44:36.000–44:37.000
It's a lot more efficient.
（此句尚無繁中翻譯）
44:37.000–44:38.000
Yeah.
（此句尚無繁中翻譯）
44:38.000–44:39.000
And it's almost at frontier level.
（此句尚無繁中翻譯）
44:39.000–44:44.000
Does it, I feel like as an onlooker and you tell me if this is right or not, the same,
（此句尚無繁中翻譯）
44:44.000–44:46.000
a similar thing is happening in hardware.
（此句尚無繁中翻譯）
44:46.000–44:51.000
Like no China is nowhere close to the leading edge in terms of chips, but in terms of memory,
（此句尚無繁中翻譯）
44:51.000–44:57.000
in terms of, you know, some of the other equipment that you need to build these data centers or
（此句尚無繁中翻譯）
44:57.000–44:58.000
build electronics.
（此句尚無繁中翻譯）
44:58.000–44:59.000
They're getting there.
（此句尚無繁中翻譯）
44:59.000–45:00.000
What does that mean to you?
（此句尚無繁中翻譯）
45:00.000–45:01.000
How is that shift?
（此句尚無繁中翻譯）
45:01.000–45:02.000
How is that shifting?
（此句尚無繁中翻譯）
45:02.000–45:06.000
Well, the Chinese are constrained in some sense, right?
（此句尚無繁中翻譯）
45:06.000–45:08.000
So, you know, I did a piece while back.
（此句尚無繁中翻譯）
45:08.000–45:10.000
It was called the U.S. has chips, but no power.
（此句尚無繁中翻譯）
45:10.000–45:12.000
China has power, but no chips.
（此句尚無繁中翻譯）
45:12.000–45:13.000
Like he's bringing more capacity online.
（此句尚無繁中翻譯）
45:13.000–45:15.000
And the answer was the U.S. is bringing more online.
（此句尚無繁中翻譯）
45:15.000–45:16.000
It's not even close.
（此句尚無繁中翻譯）
45:16.000–45:21.000
The Chinese have been constrained by some of the export controls and the sanctions, particularly
（此句尚無繁中翻譯）
45:21.000–45:26.000
on things like semiconductor manufacturing equipment that forces them to make their local
（此句尚無繁中翻譯）
45:26.000–45:32.000
chips on effectively substandard process technology that impacts their yields and ability to really
（此句尚無繁中翻譯）
45:32.000–45:36.000
produce like local chips at high volume, which is one reason that I think that they have been
（此句尚無繁中翻譯）
45:36.000–45:40.000
forced to innovate along other vectors like like model efficiency and things like that.
（此句尚無繁中翻譯）
45:40.000–45:44.000
They are constrained in terms of the resources that they can deploy.
（此句尚無繁中翻譯）
45:44.000–45:47.000
And so they're forced to do as best as they can with those.
（此句尚無繁中翻譯）
45:47.000–45:50.000
And they're very, very, they've been very, very good at it.
（此句尚無繁中翻譯）
45:50.000–45:51.000
Right.
（此句尚無繁中翻譯）
45:51.000–45:53.000
And this is one thing, you know, engineers are smart.
（此句尚無繁中翻譯）
45:53.000–45:54.000
Right.
（此句尚無繁中翻譯）
45:54.000–45:56.000
You know, if you give them constraints, they'll they'll find their way to make them fit.
（此句尚無繁中翻譯）
45:56.000–45:58.000
And the Chinese are clearly doing that.
（此句尚無繁中翻譯）
45:58.000–46:00.000
There are Chinese memory players.
（此句尚無繁中翻譯）
46:00.000–46:05.000
But I mean, from what we're seeing in memory right now, I mean, you know, I don't cover Micron.
（此句尚無繁中翻譯）
46:05.000–46:08.000
It's a colleague of mine, but I mean, Micron reported last night.
（此句尚無繁中翻譯）
46:08.000–46:13.000
And I mean, it's it's looking like things are going to be tight for a long time, probably.
（此句尚無繁中翻譯）
46:13.000–46:16.000
And it's both a supply issue as well as a demand issue.
（此句尚無繁中翻譯）
46:16.000–46:18.000
And supply will come online over time.
（此句尚無繁中翻譯）
46:18.000–46:23.000
And they have to actually build the buildings first before they have somewhere to put the tools.
（此句尚無繁中翻譯）
46:23.000–46:24.000
Right.
（此句尚無繁中翻譯）
46:24.000–46:25.000
To make the chip.
（此句尚無繁中翻譯）
46:25.000–46:26.000
So it takes time.
（此句尚無繁中翻譯）
46:26.000–46:31.000
But, you know, the question will be once that capacity comes online, like, does the demand rise rise to meet it?
（此句尚無繁中翻譯）
46:31.000–46:33.000
We've actually seen this, by the way, like broadly in semis.
（此句尚無繁中翻譯）
46:33.000–46:34.000
It's really interesting.
（此句尚無繁中翻譯）
46:34.000–46:38.000
We've had this sort of rolling wave of bottlenecks.
（此句尚無繁中翻譯）
46:38.000–46:41.000
Like, AI has gotten so big, it's kind of dragged everything along with it.
（此句尚無繁中翻譯）
46:41.000–46:46.000
And one at a time, the different parts of the industry have sort of been hitting their limits and the stocks have been ripping.
（此句尚無繁中翻譯）
46:46.000–46:51.000
And, you know, we went from the accelerators themselves being the bottleneck a year or two years or whatever.
（此句尚無繁中翻譯）
46:51.000–46:52.000
And then it went to memory.
（此句尚無繁中翻譯）
46:52.000–46:53.000
And then it went to semi-cap.
（此句尚無繁中翻譯）
46:53.000–46:58.000
And then it went to optical and networking and power semis and more recently CPUs.
（此句尚無繁中翻譯）
46:58.000–47:01.000
And you could have almost owned anything in the space.
（此句尚無繁中翻譯）
47:01.000–47:05.000
I think that the SOX index is, you know, it's an index of semi-cap.
（此句尚無繁中翻譯）
47:05.000–47:07.000
It's up 100% year to date.
（此句尚無繁中翻譯）
47:07.000–47:08.000
It's been incredible.
（此句尚無繁中翻譯）
47:08.000–47:09.000
You could have owned anything.
（此句尚無繁中翻譯）
47:09.000–47:10.000
You would have been fine.
（此句尚無繁中翻譯）
47:10.000–47:14.000
Like almost anything you would have been just fine to greater or lesser degree.
（此句尚無繁中翻譯）
47:14.000–47:15.000
In the hardware space.
（此句尚無繁中翻譯）
47:15.000–47:16.000
And it's all being ruined.
（此句尚無繁中翻譯）
47:16.000–47:17.000
What's that?
（此句尚無繁中翻譯）
47:17.000–47:18.000
Yes.
（此句尚無繁中翻譯）
47:18.000–47:25.000
And it's basically you're owning the underlying things, but the app layer or the model layer, it's been a little bit more complicated.
（此句尚無繁中翻譯）
47:25.000–47:26.000
Okay.
（此句尚無繁中翻譯）
47:26.000–47:29.000
So like Stacey, you've been covering this almost two decades.
（此句尚無繁中翻譯）
47:29.000–47:31.000
We're going to say two decades.
（此句尚無繁中翻譯）
47:31.000–47:34.000
This stuff is cyclical, but right now it doesn't feel like cyclical.
（此句尚無繁中翻譯）
47:34.000–47:37.000
And everyone's talking about a super cycle.
（此句尚無繁中翻譯）
47:37.000–47:39.000
Is that still the case right now?
（此句尚無繁中翻譯）
47:39.000–47:42.000
Do you see anything to throw that off?
（此句尚無繁中翻譯）
47:42.000–47:44.000
Well, for now, yes.
（此句尚無繁中翻譯）
47:44.000–47:45.000
Right.
（此句尚無繁中翻譯）
47:45.000–47:51.000
I mean, look, I've been hearing the word super cycle as long as I've ever been doing this job, but this is probably the first real super cycle we've seen.
（此句尚無繁中翻譯）
47:51.000–47:53.000
You know, there's a few different types of cycles, right?
（此句尚無繁中翻譯）
47:53.000–48:00.000
You have inventory cycles, like semis are the back of the supply chain and fluctuations in any demand can propagate backwards.
（此句尚無繁中翻譯）
48:00.000–48:03.000
And they tend to be shorter term typically.
（此句尚無繁中翻譯）
48:03.000–48:05.000
You can get supply cycles.
（此句尚無繁中翻譯）
48:05.000–48:07.000
Supply is tight and pricing goes up.
（此句尚無繁中翻譯）
48:07.000–48:09.000
We're having a big one in memory right now.
（此句尚無繁中翻譯）
48:09.000–48:11.000
You can have product cycles or socket cycles.
（此句尚無繁中翻譯）
48:11.000–48:18.000
I was like, you know, I, I, I win or lose a chip that goes into an iPhone and that can drive my revenue up and down by big amounts.
（此句尚無繁中翻譯）
48:18.000–48:22.000
And then you've got what we have today, which is, I mean, it's a true demand cycle.
（此句尚無繁中翻譯）
48:22.000–48:26.000
And it's not that, you know, we didn't, you know, clearly we didn't have enough supply.
（此句尚無繁中翻譯）
48:26.000–48:28.000
The reason is demand has gotten gotten so big.
（此句尚無繁中翻譯）
48:28.000–48:33.000
It just overpowered and any of the wildest forecasts that were there, like not that long ago.
（此句尚無繁中翻譯）
48:33.000–48:41.000
And if you want to define that as maybe, you know, indicative of what one might call a super cycle, maybe, maybe I would, because it is purely demand driven and it's driving everything.
（此句尚無繁中翻譯）
48:41.000–48:44.000
And so the big question is how long does the demand last?
（此句尚無繁中翻譯）
48:44.000–48:48.000
And I think that's the trillion, maybe it's the quadrillion dollar question.
（此句尚無繁中翻譯）
48:48.000–48:49.000
Like, I don't know.
（此句尚無繁中翻譯）
48:49.000–48:52.000
Right now it's still up and to the right.
（此句尚無繁中翻譯）
48:52.000–48:59.000
And at some point, maybe that won't be the case, but I think all, all signs right now, the only thing we're hearing from anybody is they can't get enough compute.
（此句尚無繁中翻譯）
48:59.000–49:02.000
Right. It's more just sort of the makeup of that, which we discussed before.
（此句尚無繁中翻譯）
49:02.000–49:03.000
Is it going to be the labs?
（此句尚無繁中翻譯）
49:03.000–49:07.000
Is it going to be the open source models, but infrastructure is sort of the system.
（此句尚無繁中翻譯）
49:07.000–49:09.000
For my goals, I don't really care.
（此句尚無繁中翻譯）
49:09.000–49:17.000
I mean, it's computed like Nvidia benefits clearly from both, you know, closed and open source.
（此句尚無繁中翻譯）
49:17.000–49:18.000
Yeah.
（此句尚無繁中翻譯）
49:18.000–49:20.000
Broadcom is doing A6, you know, for the vendors.
（此句尚無繁中翻譯）
49:20.000–49:22.000
I mean, it'll depend on, on building for their models.
（此句尚無繁中翻譯）
49:22.000–49:27.000
Right. But at the end of the day, I think compute is, is demand for compute is good.
（此句尚無繁中翻譯）
49:27.000–49:34.000
It's, it's my guys in some sense, like they'll, by and large, they'll probably do fine as long as compute demand is going up.
（此句尚無繁中翻譯）
49:34.000–49:35.000
Right.
（此句尚無繁中翻譯）
49:35.000–49:37.000
Well, Stacey, it's always great to get your insights.
（此句尚無繁中翻譯）
49:37.000–49:39.000
What's that?
（此句尚無繁中翻譯）
49:39.000–49:42.000
They're selling the picks and the shovels and the gold rush.
（此句尚無繁中翻譯）
49:42.000–49:44.000
So they'll, they'll be fine as long as the gold rush is going on.
（此句尚無繁中翻譯）
49:44.000–49:46.000
Yeah. But have you seen that cartoon?
（此句尚無繁中翻譯）
49:46.000–49:51.000
I feel like I've seen it a lot over the last few months, especially where, you know, someone's like, yeah.
（此句尚無繁中翻譯）
49:51.000–49:52.000
What are you here to do?
（此句尚無繁中翻譯）
49:52.000–49:54.000
No, one's actually mining the gold.
（此句尚無繁中翻譯）
49:54.000–49:57.000
They're just bringing more picks and shovels for like a fraction.
（此句尚無繁中翻譯）
49:57.000–49:59.000
They are, they are mining.
（此句尚無繁中翻譯）
49:59.000–50:05.000
I think we can have genuine questions on, on return in our ROI, because I think that's really where the debate is.
（此句尚無繁中翻譯）
50:05.000–50:11.000
Like what determines whether or not demand continues is, is there a return on the spending or is there not?
（此句尚無繁中翻譯）
50:11.000–50:14.000
And it's still early, but I already think we're seeing evidence.
（此句尚無繁中翻譯）
50:14.000–50:16.000
I mean, we've got you on the rental side.
（此句尚無繁中翻譯）
50:16.000–50:17.000
Yeah.
（此句尚無繁中翻譯）
50:17.000–50:18.000
They're sold out.
（此句尚無繁中翻譯）
50:18.000–50:19.000
There's clearly a return.
（此句尚無繁中翻譯）
50:19.000–50:26.000
I do wonder though, Stacey, how is it that the Chinese are able to do so much on a fraction of the CapEx?
（此句尚無繁中翻譯）
50:26.000–50:32.000
Well, again, you know, it's, they're, they're being forced to be innovative, right?
（此句尚無繁中翻譯）
50:32.000–50:33.000
Yeah.
（此句尚無繁中翻譯）
50:33.000–50:35.000
And by the way, I do not view that as a bad thing.
（此句尚無繁中翻譯）
50:35.000–50:38.000
Like this gets back to the whole deep seek scare from a year and a half ago.
（此句尚無繁中翻譯）
50:38.000–50:39.000
You remember that?
（此句尚無繁中翻譯）
50:39.000–50:40.000
Oh, I do.
（此句尚無繁中翻譯）
50:40.000–50:42.000
I think we might be.
（此句尚無繁中翻譯）
50:42.000–50:44.000
Everybody freaked out.
（此句尚無繁中翻譯）
50:44.000–50:45.000
I know.
（此句尚無繁中翻譯）
50:45.000–50:46.000
Okay.
（此句尚無繁中翻譯）
50:46.000–50:47.000
Last thing I'll say.
（此句尚無繁中翻譯）
50:47.000–50:48.000
Yes.
（此句尚無繁中翻譯）
50:48.000–50:49.000
That was not a blip though.
（此句尚無繁中翻譯）
50:49.000–50:53.000
I mean, we, we may be having another deep seek like moment.
（此句尚無繁中翻譯）
50:53.000–50:54.000
What happened?
（此句尚無繁中翻譯）
50:54.000–50:55.000
So people were worried.
（此句尚無繁中翻譯）
50:55.000–50:59.000
It was like, Oh my God, these guys are so much more efficient.
（此句尚無繁中翻譯）
50:59.000–51:00.000
We won't need as much computers.
（此句尚無繁中翻譯）
51:00.000–51:01.000
We're building.
（此句尚無繁中翻譯）
51:01.000–51:02.000
What happened?
（此句尚無繁中翻譯）
51:02.000–51:05.000
Only thing we've seen since then is skyrocket.
（此句尚無繁中翻譯）
51:05.000–51:07.000
We need costs to come down.
（此句尚無繁中翻譯）
51:07.000–51:08.000
That's how you drive adoption.
（此句尚無繁中翻譯）
51:08.000–51:09.000
That's how you get return.
（此句尚無繁中翻譯）
51:09.000–51:14.000
And everybody throw this company, Devon's paradox, which it's been thrown around to death.
（此句尚無繁中翻譯）
51:14.000–51:18.000
And the idea is when things get cheaper, people use more, but you have to remember I'm a semi
（此句尚無繁中翻譯）
51:18.000–51:19.000
guy.
（此句尚無繁中翻譯）
51:19.000–51:20.000
It's made out like that.
（此句尚無繁中翻譯）
51:20.000–51:21.000
But yeah.
（此句尚無繁中翻譯）
51:21.000–51:25.000
And look, I was, of course I believe in Jevon's paradox and semis cost got cut in half every
（此句尚無繁中翻譯）
51:25.000–51:27.000
two years for six decades.
（此句尚無繁中翻譯）
51:27.000–51:28.000
Was that a bad thing for semi?
（此句尚無繁中翻譯）
51:28.000–51:32.000
No, it was a fantastic thing for semiconductors and for everybody else.
（此句尚無繁中翻譯）
51:32.000–51:34.000
So I think lower, lower cost computers.
（此句尚無繁中翻譯）
51:34.000–51:35.000
Good.
（此句尚無繁中翻譯）
51:35.000–51:37.000
I started by calling you a semi guy.
（此句尚無繁中翻譯）
51:37.000–51:38.000
Now you're calling yourself a semi guy.
（此句尚無繁中翻譯）
51:38.000–51:39.000
I love it.
（此句尚無繁中翻譯）
51:39.000–51:40.000
It was perfect.
（此句尚無繁中翻譯）
51:40.000–51:43.000
You are our semi guy, Stacy, lots of energy, right?
（此句尚無繁中翻譯）
51:43.000–51:45.000
Right place for the last few years.
（此句尚無繁中翻譯）
51:45.000–51:47.000
Thank you so much for coming on the live stream.
（此句尚無繁中翻譯）
51:47.000–51:48.000
We'll talk to you again soon.
（此句尚無繁中翻譯）
51:48.000–51:49.000
I'm sure.
（此句尚無繁中翻譯）
51:49.000–51:50.000
Thanks, Stacy.
（此句尚無繁中翻譯）
51:50.000–51:53.000
Thank you guys for joining another live stream.
（此句尚無繁中翻譯）
51:53.000–52:00.000
Thank you to Jasmine and Janice and Divya, Robert and Evan, and the behind here and Bud
（此句尚無繁中翻譯）
52:00.000–52:01.000
in the control room.
（此句尚無繁中翻譯）
52:01.000–52:02.000
We'll be back next week.
（此句尚無繁中翻譯）
52:02.000–52:06.000
Thanks for watching guys and keep giving in those comments and questions.
（此句尚無繁中翻譯）
52:06.000–52:07.000
Thank you.
（此句尚無繁中翻譯）
52:07.000–52:08.000
Thank you.
（此句尚無繁中翻譯）
52:08.000–52:09.000
Thank you.
（此句尚無繁中翻譯）
52:13.000–52:14.000
Thank you.
（此句尚無繁中翻譯）
52:14.000–52:15.000
Thank you.
（此句尚無繁中翻譯）
52:15.000–52:16.000
Thank you.
（此句尚無繁中翻譯）
52:16.000–52:17.000
Thank you.
（此句尚無繁中翻譯）
52:17.000–52:18.000
Thank you.
（此句尚無繁中翻譯）
52:18.000–52:19.000
Thank you.
（此句尚無繁中翻譯）
影片筆記：Z.AI And The Chinese Open Source Moment
一句話總結
中國開源模型（如 GLM 5.2）在「代理工作」與成本效益上迅速追趕美國前沿模型，推動企業採用混合架構與模型路由策略；同時，地緣政治出口管制促使各國發展主權 AI，而硬體供應鏈正經歷由真實需求驅動的「超級週期」，OpenAI 與 Broadcom 合作推出自研晶片 Jalapeno 標誌著硬體生態系的擴張與競爭加劇。
核心重點
- 開源與前沿模型的差距縮小：
- 中國模型（如 GLM 5.2）在代理工作（agentic work）基準測試中表現優異，僅以極低成本（約 Opus 4.8 的五分之一）達到接近前沿模型的表現。
- 開源與前沿模型的差距縮短至 3 到 6 個月，這在組織內部的 AI 擴散噪音範圍內，使得企業可以接受開源模型，無需等待兩三年。
- 企業轉向「混合架構」：關鍵決策使用高成本的前沿智能（Frontier Intelligence），簡單任務使用低成本開源模型，類似於根據員工職級分配薪資。
- AI 經濟學與指標轉變：
- 關注點從單純追求「最聰明模型」轉向「每美元智力」（intelligence per dollar）。
- 企業更關注運行成本，能否以足夠低的成本反覆運行百萬次。
- 蒸餾技術（Distillation）讓中國實驗室能夠以較低成本逼近前沿性能，但也引發了關於授權與道德的爭議。
- 從模型到代理（Agents）的系統性轉變：
- AI 從單純的模型轉向嵌入業務流程的代理系統，訓練數據將更多來自企業自身（如律所的 M&A 數據）。
- 出現「多人遊戲模式」（Multiplayer mode）：AI 不再是個人生產力工具，而是組織內的「協作同事」，具備共享上下文與資源，類似於團隊成員在 Slack 頻道中互動。
- 應用層（App Layer）成為路由工作負載的邏輯位置，可靈活選擇底層模型。
- 地緣政治與主權 AI（Sovereign AI）：
- 美國透過出口管制限制 AI 模型存取，打破了「軟體全球通用」的假設。
- 各國（如歐盟、中國）必須建立自己的 AI 能力或後盾，以確保國家安全與經濟利益。
- 蒸餾技術被視為應對策略，但也引發了美國實驗室（如 Anthropic）對「非授權蒸餾」的訴訟。
- 硬體供應鏈與晶片創新：
- OpenAI 與 Broadcom 合作推出首款自研推論晶片 Jalapeno，耗時九個月設計完成，旨在降低推論成本約 50%。
- 半導體產業結構變化：VC 投資進入，市場規模擴大，競爭加劇但整體餅圖擴大。
- 供應鏈出現「滾動式瓶頸」，從加速器轉移至記憶體、半導體設備、光學與網路、電源半導體，甚至 CPU。
- 當前被視為「真正的需求週期」（True demand cycle）或「超級週期」，由極度龐大的需求驅動。
詳細大綱
1. GLM 5.2 的市場衝擊與經濟學轉變
- GLM 5.2 的發布：由 Ji-Poo (Z.AI) 開發，在矽谷引起轟動，在 Open Router 上的採用速度快於 DeepSeek V4。
- 代理工作表現：GLM 5.2 在代理工作基準測試中表現優異，僅以 Opus 4.8 五分之一成本達到接近其表現（相差僅 1%）。
- 每美元智力：華爾街開始關注「每美元智力」，企業關注運行成本，能否以足夠低的成本反覆運行百萬次。
- 蒸餾技術：-distillation（蒸餾技術）讓中國實驗室能夠以較低成本逼近前沿性能。
2. 企業採用策略：混合架構與模型路由
- 混合架構：企業不再只買單一模型，而是根據任務複雜度進行分配。關鍵決策使用前沿智能，簡單任務使用開源智能。
- 模型路由：這種做法類似於企業根據員工職級分配薪資。只要差距不大於讓組織失去競爭優勢的閾值，3-6 個月的差距是可以接受的。
- 嘉賓觀點：
- Gabe Pereira：開源正在迅速追趕，GLM 5.2 是第一個真正與封閉源前沿模型競爭的模型。
- Aaron Levy：目前可能仍處於開源與前沿模型差距為 3 到 6 個月的狀態，而非快速取代之前的場景。未來 1-2 年，前沿級智能對企業仍至關重要。
3. 開源權重與前線智能的關係
- 時間差距縮小：若差距為兩到三年，企業不會使用次優模型；若差距為三到六個月，企業可以接受開源模型。
- 後訓練（Post-training）的價值：企業可以針對特定領域任務對開源模型進行後訓練，這比單純比較 token 成本更重要。
- 從模型到代理的演變：隨著 AI 從單純的模型轉向嵌入業務流程的代理系統，訓練數據將更多來自企業自身。
- 啞鈴型動態（Barbell Dynamic）：前線智能仍將存在，主要用於高價值、複雜的規劃或審查任務；而大量繁重的文檔處理和分析工作將由更便宜、更快的開源模型承擔。
4. 前線實驗室的開源策略與市場結構
- 現狀觀察：OpenAI, Google (Gemini) 已訓練開源模型，但在採用數據和基準測試中未顯著體現。中國開源模型在採用率上領先。
- 遊戲理論觀點：理想策略是前線實驗室應持續提供接近前線水平的開源權重模型，以獲取兩側的生態系效應。
- 應用層的靈活性：應用層獨立於模型，成為路由工作負載的邏輯位置。例如 Cursor 使用基於中國公司 Moonshot 的 Kimmy 架構訓練自訂模型。
5. Anthropic 政策與地緣政治影響
- 出口管制：美國透過出口管制阻斷對 AI 模型的存取，打破了「軟體全球通用」的假設。
- 主權 AI：各國需保護其獲取智慧的能力，避免經濟與國家安全受威脅。歐盟需投資開源或後訓練模型作為美國模型的後盾。
- 蒸餾技術的爭議：蒸餾被視為一種應對策略，但也引發了美國實驗室對「非授權蒸餾」的訴訟。
- Claude Tagg：Anthropic 的新產品概念，將 AI 從個人生產力工具轉變為組織內的「協作同事」，具備共享上下文與資源。
6. AI Agent 的未來工作模式與架構
- 多人遊戲模式：未來工作模式將涉及多個 Agent 與人類組成團隊，處理複雜任務。
- 技術架構與鎖定效應：雲端可作為路由器，連接不同的 MCP 伺服器或其他 Agent 系統。使用者可選擇使用多少雲端智慧層，設計實施方式以控制鎖定程度。
7. OpenAI 與 Broadcom 合作晶片分析 (Jalapeno)
- 產品發布細節：OpenAI 與 Broadcom 聯合發布 Jalapeno，這是 OpenAI 首款完全從零開始設計的自研推論晶片，耗時九個月。
- 商業合約：OpenAI 與 Broadcom 簽署總值 100 億美元（原文提及 gigawatt deal，疑點見下文）、為期五年的合約。
- 半導體產業結構變化：VC 投資進入，市場規模擴大，競爭加劇但整體餅圖擴大。
8. 硬體供應鏈瓶頸與投資回報
- 瓶頸的輪動現象：從加速器轉移至記憶體、半導體設備、光學與網路、電源半導體，甚至 CPU。
- 中國硬體現狀：受出口管制影響，被迫使用次標準製程技術，但在模型效率上進行創新。
- 超級週期：當前被視為「真正的需求週期」，由極度龐大的需求驅動。
- 傑文斯悖論：成本下降會推動採用率上升，從而增加總體需求。半導體成本每兩年减半，長期來看對半導體產業是好事。
工具 / 模型 / 名詞整理
- GLM 5.2：Ji-Poo / Z.AI 發布的最新模型，在代理工作基準測試中表現優異。
- Ji-Poo：開發 GLM 5.2 的實體/公司。
- Z.AI：Ji-Poo 的別稱。
- DeepSeek / DeepSeek V4：DeepSeek 在四月發布的版本，此前引發市場震動。
- Opus 4.7 / Opus 4.8：Anthropics 最強大的可用模型，GLM 5.2 在其代理基準測試中僅相差 1%。
- Anthropic：Opus 系列模型的開發者。
- Open Router：顯示 GLM 5.2 採用速度較快的平台。
- Box：Aaron Levy 經營的公司。
- Harvey：Gabe Pereira 聯合創始的 AI 平台，被半數 AM Law 100 公司使用。
- AM Law 100：美國律師事務所排名前 100 的列表。
- DeepMind：Gabe Pereira 曾工作的研究機構。
- Meta：Gabe Pereira 曾工作的研究機構。
- Factory：Aaron Levy 提到的 AI 層之一。
- Cognition：Aaron Levy 提到的 AI 層之一。
- Artificial Analysis：提供性能與成本對比圖表的機構。
- 前線實驗室/公司 (Frontier Labs)：OpenAI, Anthropic, Google。
- Gemini / Gemma：Google 的模型系列及開源模型。
- Fable Mythos / Fable：Anthropic 的模型，文中提及被切斷訪問。
- GPD 5.5 / GPD 5.6：疑似口誤，可能指 GPT-5.5 或類似版本。
- Kimmy：中國模型，用於 Moonshot/Cursor 的底層架構。
- Mistral：歐洲客戶使用的模型。
- Cursor：應用層公司，提及使用中國底層模型。
- Moonshot：中國公司，與 Cursor 相關。
- Claude Tagg：Anthropic 的新功能或產品，具備協作與上下文學習能力。
- OpenClaw：協作編碼系統之一。
- Hermes：協作編碼系統之一。
- Devon：協作編碼系統之一。
- Slack：提及作為協作平台。
- Jalapeno：OpenAI 與 Broadcom 合作推出的自研推論晶片。
- Nvidia GPUs / TPUs：現有的 GPU 與張量處理單元。
- SOX index：半導體設備指數。
- Micron：美光，記憶體廠商。
- Rippling：提及建立員工數據層。
- MCP server：Model Context Protocol server。
- Distillation (蒸餾技術)：使用大型模型輸出訓練較小模型。
- Post-training (後訓練)：針對特定領域任務對模型進行訓練。
- Agentic systems / Agents (代理系統)：嵌入業務流程的 AI 系統。
- Orchestrator (協調器)：前線智能的角色之一。
- Planner (規劃者)：前線智能的角色之一。
- Reviewer (審查者)：前線智能的角色之一。
- Single player mode (單人模式)：目前的 AI 主要處於此模式。
- Multiplayer mode (多人遊戲模式)：未來工作模式，多個 Agent 與人類協作。
- Shared intelligence (共享智慧)：轉向共享智慧而非僅個人生產力。
- Sovereign AI (主權 AI)：各國建立自己的 AI 能力。
- Export Controls (出口管制)：美國限制 AI 模型存取的措施。
- Backstop/Hedge (後盾/對沖)：歐盟需投資開源或後訓練模型作為美國模型的後盾。
- Non sanctioned distillation (非授權蒸餾)：美國實驗室訴訟針對的對象。
- Blunt instruments (鈍器/粗獷工具)：形容出口管制。
- Fine grained (細粒度)：形容晶片管制。
- Accelerators (加速器)：硬體供應鏈瓶頸之一。
- Memory (記憶體)：硬體供應鏈瓶頸之一。
- Semi-cap (半導體設備)：硬體供應鏈瓶頸之一。
- Optical and networking (光學與網路)：硬體供應鏈瓶頸之一。
- Power semis (電源半導體)：硬體供應鏈瓶頸之一。
- CPUs (中央處理器)：近期瓶頸之一。
- Picks and shovels (鏟子與鑿子)：比喻基礎設施供應商。
- Super cycle (超級週期)：當前被視為真正的需求週期。
- True demand cycle (真正的需求週期)：由需求規模驅動。
- Inventory cycles (庫存週期)：供應鏈後端，需求波動向後傳播。
- Supply cycles (供應週期)：供應緊張導致價格上漲。
- Product/Socket cycles (產品/插座週期)：單一晶片進入特定設備對營收影響巨大。
操作流程整理
- 評估任務複雜度：企業根據任務性質（關鍵決策 vs. 簡單任務）決定使用前沿智能還是開源智能。
- 選擇模型與架構：
- 對於關鍵任務，選擇高成本的前沿模型（如 Opus 4.8）。
- 對於簡單任務，選擇低成本開源模型（如 GLM 5.2）。
- 考慮使用蒸餾技術或後訓練來優化開源模型在特定領域的表現。
- 實施模型路由：在應用層（App Layer）建立路由邏輯，根據任務適配使用 frontier 或 open weights 模型。
- 部署代理系統：
- 從單人模式轉向多人遊戲模式，讓 AI 成為組織內的協作同事。
- 利用共享上下文與資源，通過類似 Slack 頻道的方式與 AI 互動。
- 監控供應鏈與成本：
- 關注硬體供應鏈的瓶頸轉移（從加速器到記憶體、光學等）。
- 評估推論成本（如使用 Jalapeno 晶片降低 50% 成本）。
- 監控需求週期，確保投資回報（ROI）。
值得注意的限制或風險
- 出口管制與地緣政治風險：美國透過出口管制限制 AI 模型存取，可能影響全球供應鏈與各國主權 AI 發展。
- 蒸餾技術的爭議：蒸餾技術可能引發法律訴訟（如 Anthropic 對 Alibaba 的訴訟），影響商業連續性。
- 模型鎖定效應：隨著 AI 深度嵌入業務流程，替換模型變得困難，可能產生鎖定效應。
- 供應鏈瓶頸：硬體供應鏈出現滾動式瓶頸，可能影響 AI 發展速度。
- 中國硬體限制：中國受出口管制影響，無法使用先進製程，導致良率與產能受限，需依賴模型效率創新。
- 前沿模型差距縮小：開源與前沿模型差距縮小，可能影響前沿實驗室的商業模式，需通過生態系效應維持競爭力。
逐字稿辨識疑點
- DeepSeq：逐字稿中出現 "DeepSeq showed Wall Street..."，疑為 DeepSeek 的聽寫錯誤或口誤。
- Ji-Poo：逐字稿中多次出現 "Ji-Poo" 作為開發 GLM 5.2 的公司名稱，疑為 01.AI（零一萬物）或其他中國 AI 公司的音譯錯誤或特定稱呼，需查證。
- GPO：逐字稿中出現 "That's why GPO is interesting"，疑為 Z.AI 或 GLM 的聽寫錯誤，需查證。
- fable level：Aaron Levy 提到 "fable level kind of intelligence"，疑為 frontier level（前沿級）的聽寫錯誤。
- Opus 4.7 / Opus 4.8：逐字稿中交替出現 Opus 4.7 和 Opus 4.8，需確認具體指代的模型版本編號是否一致。
- ZAI / Z.AI：逐字稿中交替使用 "ZAI" 和 "Z.AI"，需確認官方名稱寫法。
- agentic work：多次出現，指代代理型 AI 工作負載，此為專業術語，保留原樣。
- open weights：Aaron Levy 使用 "open weights" 指代開源模型，與 "open source" 混用，保留原樣。
- AM Law 100：指代美國
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