跟讀練習: What's next for mainframes and AI? - 通過影片學習英語口語
選擇練習模式
每種模式都會開啟全屏練習室。
104 句
如果句子過短或過長，請點選 Edit 進行調整。
1
A trillion web transactions daily.
2
More than 300 billion inference requests in a single day.
3
A million plus transactions in a second.
4
The modern mainframe can handle all that and more with ease.
5
Welcome to AI Academy.
6
I'm Lou Folia, a writer and journalist.
7
And I'm here with Christian Jacoby, IBM Fellow and CTL of Systems Development at IBM.
8
On average, Visa processes around 8 ,500 global transactions every second.
9
A mainframe which can handle one million transactions per second would be able to process more than 100 times Visa's volume.
10
So, Christian, to you, what makes the mainframe remarkable?
11
And what makes them key to AI's next big breakthrough?
12
Is it the hardware?
13
Is it the software?
14
Is it somewhere in between?
15
It's really the hardware and the software together.
16
When we design the next generation of the mainframe, we're enhancing the performance, the scalability, availability features.
17
But we're doing this in conjunction as a hardware team.
18
We're working closely with the firmware team and software team to make sure
19
that we end up with a stack optimized end -to -end solution that clients can use to run their workloads.
20
IBM has been thinking about AI on mainframes for a long time now, right?
21
This isn't something we've recently come to.
22
Ten years at least.
23
And the IBM Telum processor includes an on -chip AI accelerator.
24
Why is that important, having the accelerator on chip?
25
And how does it enhance AI performance?
26
Yeah, as you're running workloads on the processor, of course, a lot of the data resides within the chip.
27
And within the chip, we can communicate between different components of the chip, literally in nanoseconds, right?
28
Nanoseconds is a tiny sliver of time, of course.
29
Versus if you go off -chip and go to a separate card, you're talking microseconds.
30
And if you go to a different system, you might be talking milliseconds, right?
31
And so by putting AI acceleration directly onto the processor chip, we could achieve the lowest possible latency for some of those traditional models.
32
And then now with the Spire acceleration card that goes into the IO subsystem,
33
we can run generative AI also with very low latency without needing to cross system boundaries and go into a separate system.
34
It matters how big that slice of the second is.
35
It matters big time.
36
Absolutely.
37
When you're running millions of transactions every second
38
and the latency requirements from a business perspective is that transactions need to finish within a handful of milliseconds,
39
then you really don't want to waste milliseconds and traveling the data between here and some other system for AI inferencing.
40
Let's drill in on that a little bit.
41
So what makes mainframes so pivotal for AI?
42
Their hardware capabilities, their software capabilities, What is it that sort of makes it perfectly suited for AI?
43
Yeah, so clients are running mission -critical central workloads for their corporations.
44
They run the central ledger of a bank or they run insurance claims processing on mainframes.
45
That's where a lot of the data resides.
46
That's where the transactions get processed.
47
And of course, you can use AI to increase the sort of business value
48
of each transaction by checking for fraud on an insurance claim
49
or on a banking transaction and being able to integrate that directly into the workload, detect fraud, for example,
50
in the instant of time where the transaction is getting processed is of huge value to clients compared to,
51
for example, doing it in a cloud a few minutes or an hour after the transaction has already completed.
52
By doing it directly in the mainframe box itself, you cut out all that latency and that's how you get to the milliseconds of latency for, for example, fraud detection.
53
You already have all the data, all the transaction processing happening in the mainframe system by doing the AI directly in the processor for traditional models
54
or in the I .O subsystem on a Spire accelerator for generative models.
55
You can maintain low latency, get the good throughput and performance that the mainframe is known for,
56
but also preserve the security capabilities by not needing to send the data to a separate system.
57
When you're sending the data out into the world, it's slowing things down and it's creating security risk.
58
Correct.
59
Exactly.
60
You're sending it somewhere else that inherently takes time and you're moving the data out of the highly secure mainframe environment.
61
So now it gets exposed to different attack surfaces that hackers could exploit.
62
What should organizations consider when choosing AI -ready infrastructure?
63
And how do mainframes address those needs better than other platforms?
64
When clients embark on AI journeys in the context of the mission -critical workloads, We often talk to them about the readiness of the data,
65
the readiness of the models.
66
Do they have the capabilities in the data science departments to kind of train their own data into the models
67
that they will be using?
68
And then also look at, you know, what latency requirements do you have?
69
What throughput requirements?
70
How many inferences per second do you need to run in that workload?
71
And by working with clients through those considerations early on,
72
we can make sure that clients are actually successful in deploying AI solutions, not just developing them, and then kind of getting surprised about how hard it is to scale them to actual deployment.
73
So we're in a room that has seen a lot of mainframe history.
74
You spent the bulk of your career working on mainframes.
75
What excites you about the future of mainframes?
76
We've been already developing the next generation and the generation after that.
77
We're always working on three mainframe generations at the same time.
78
So it's your job to think about what the future's going to look like and build a machine that serves it.
79
And that's exciting and a lot of fun.
80
Could you talk a little bit about some of the use cases that may have surprised you?
81
Yeah, I'll give you an example.
82
There's one client who's actually using mainframes to analyze imagery of seagrass, satellite images.
83
Like seagrass, like on the beach seagrass?
84
Yes, yeah.
85
And they see beach grass dying in certain areas, and they use that information
86
that they obtained from the AI models to then do ecological measures to protect more of the seagrass.
87
So when you start designing something, and then you see where it goes, like protecting seagrass, like how cool is that for you to see the expanse?
88
That is super exciting.
89
Like we have, when we design the hardware and the software, we have use cases in mind that we're optimizing for.
90
And then once it's in the hands of our customers, their innovation starts and how they can use all these capabilities for use cases we would have never imagined.
91
And that is one of the most fascinating parts to see how this sometimes evolves.
92
Yeah, totally.
93
So that's the future.
94
But even today, just think about it.
95
Every time you pay using a credit card, withdraw cash from an ATM, make reservations with any airline,
96
file your taxes or access your medical records or basically do a thousand other things, it's likely you are being served by a mainframe.
97
Now, imagine a day without mainframes.
98
Mainframes are without a doubt integral to enterprise IT.
99
And today, with new integrations and enhancements, they are redefining their role in modern IT.
100
So, to answer the question we started with, what makes the mainframe unique?
101
It's the hardware and the software and everything in between.
102
It's how the end -to -end stack is optimized from the Silicon App for mission -critical workloads.
103
And I think that we could both agree that that's what makes the mainframe so remarkable.
104
And that's what makes it great for AI.
同頻道
FOR YOU
推薦影片
C118
104 句
How Smart Organizations Will Use AI: Jevons Paradox & Workforce Impact
Business EnglishUniversity StudentsWorking Adults
C11
68 句
System Design: Why is Kafka fast?
Business EnglishUniversity StudentsWorking Adults
C12
146 句
How a CPU Works in 100 Seconds // Apple Silicon M1 vs Intel i9
American AccentBusiness EnglishUniversity StudentsWorking Adults
C13
108 句
Jensen Huang & Satya Nadella on unmetered intelligence | Microsoft Build 2026
Business EnglishUniversity StudentsWorking Adults
C15
222 句
The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, Factory
Business EnglishUniversity StudentsWorking Adults
C11
462 句
Expensive RTX 5090 for LLMs? NO. Use This Instead. (Tesla V100 SXM2 + Z8 G4, #RACERRRZ)
Daily ConversationUniversity StudentsWorking Adults
C13
43 句
P1 机器学习的应用【2024公认最好的 | 吴恩达机器学习 | 教程 | Machine Learning Specialization(超爽中英!)】
Business EnglishUniversity StudentsWorking Adults
C11
394 句
Chip War 2.0: The Global Battle for Semiconductor Supremacy│Chris Miller (The Author of “Chip War")
IELTS PrepBusiness EnglishUniversity Students
C148
8 句
Neurosymbolic AI Explained
Business EnglishUniversity StudentsWorking Adults
C11
402 句
FULL DISCUSSION: Google's Demis Hassabis, Anthropic's Dario Amodei Debate the World After AGI | AI1G
IELTS PrepBusiness EnglishUniversity Students
C11
12 句
Rapid Engineering With AI
Business EnglishUniversity StudentsWorking Adults
C111
98 句
What are AI Agents?
Business EnglishUniversity StudentsWorking Adults
對話背景與場景
這段影片是關於大型機與人工智慧未來發展的訪談，主持人與IBM專家探討了大型機的效能、AI加速技術及實際應用。對話中包含專業術語與技術講解，語速適中，適合提升科技類英語聽力與口語能力。通過學習這段內容，不僅能瞭解前沿技術，還能掌握商務訪談中的表達邏輯。
日常交流高頻短語5則
- end-to-end solution：端到端解決方案。例：我們致力於提供最佳化的端到端解決方案。
- mission-critical workloads：關鍵任務工作負載。例：企業需確保關鍵任務工作負載的穩定執行。
- latency requirements：延遲要求。例：金融交易對延遲要求極高。
- attack surfaces：攻擊面。例：資料傳輸會增加系統的攻擊面。
- scale to deployment：擴充套件至部署。例：AI模型需能擴充套件至實際部署場景。
影子跟讀分步指南（針對本影片）
本影片涉及科技術語，跟讀時需注意專業詞彙的發音與連讀。以下是 proven 技巧：
第一步：拆分片段 將影片按話題分段（如大型機效能、AI加速、應用案例），每段30秒左右。使用“shadowspeak”技巧，逐句跟讀，重點模仿重音與語調。
第二步：攻克術語 預先熟悉“mainframe”（大型機）、“inference”（推理）等詞彙的發音，確保跟讀時流暢度。可結合“看影片學英語”的方式，邊看畫面邊理解術語含義。
第三步：提升反應速度 影片中對話節奏較快，可採用“shadow speak”訓練法，滯後1-2秒跟讀，鍛鍊聽力與口語的同步反應。反覆練習直至能跟上原速。
第四步：模仿語氣 注意訪談中專家的自信語氣與主持人的提問節奏，學習商務溝通中的表達技巧。這對“英語口語練習”至關重要。
通過以上步驟，結合專業的“shadowing site”資源，可有效提升科技英語的聽說能力。堅持練習，你會發現應對複雜對話的能力顯著增強。
影片中的語法
說話人最常用的結構，並附上影片中的原話：
| 結構 | 影片中的用法 |
|---|---|
| 現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事 | has already completed · has seen · We've been |
| 被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的 | is known · being served · is optimized |
什麼是跟讀法？
跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。