Brew Markets

Kai-Fu Lee on AI-driven corporate transformation, valuation risks, and the future of work

48 min Listen to the original
Artificial IntelligenceCorporate StrategyAI ValuationsLabor MarketsData PrivacyInvestment Strategy

Summary

Kai-Fu Lee argues that AI adoption must be led by CEOs using specialized agents to access 'dormant data' like meeting recordings, aiming to uncover hidden conflicts and biases. He warns that while physical trades are safe, middle management faces displacement as companies restructure around Directly Responsible Individuals (DRIs). Lee contends that US AI valuations may be inflated by circular financing, citing cheaper Chinese open-source models like DeepSeek as competitive threats. He recommends dollar-cost averaging into major AI players while shorting slow-adopting SaaS firms, and predicts the eventual replacement of smartphones with always-on listening devices.

Key takeaways

  1. 1
    Companies are restructuring to eliminate middle management layers by promoting employees to Directly Responsible Individuals (DRIs) or coaches, with layoffs as the final outcome for those who cannot adapt within two to three years.

    Unidentified speaker (speaker_3) · 30:47

Episode sections

  1. 0:00 – 3:05

    Intro and Sponsor Reads

    The episode begins with a DoorDash advertisement featuring a reindeer character, followed by an introduction to guest Kai-Fu Lee, highlighting his career history and current role at 01.AI. A ProShares sponsor read follows, promoting leveraged ETFs on the Nasdaq 100.

  2. 3:05 – 7:16

    CEO-Led AI Transformation and 'Dormant Data'

    Lee argues that AI adoption must start with the CEO using a specialized agent to access all company data, including audio recordings of meetings and sales calls. This approach aims to uncover 'dormant data' and hidden conflicts that traditional reporting misses, such as disagreements between VPs that are not surfaced in formal meetings.

  3. 7:16 – 11:13

    Cultural Differences and Privacy Concerns

    Lee discusses the varying cultural tolerance for AI surveillance, noting that Asian companies are more accepting than European or American ones. He advocates for transparency with employees, showing how AI insights lead to constructive actions rather than punitive measures, thereby building trust.

  4. 11:13 – 16:25

    AI as a Cultural Fix and Bias Mitigation

    Lee suggests AI can resolve cultural issues by providing objective feedback and identifying biases. He proposes programming 'invariants' or contracts into AI systems to enforce fairness and prevent sycophancy, ensuring the AI tells the truth rather than making the user happy.

  5. 16:25 – 23:09

    Personalized Dashboards and 'Morning Vanguard'

    Lee critiques static dashboards like those built by Jamie Dimon, advocating instead for personalized 'Morning Vanguard' tools that pre-search key questions based on the CEO's specific concerns, such as risks, sales deals, or retention issues.

  6. 23:09 – 27:48

    Measurable Business Results and HR Transformation

    Lee asserts that AI deployment must impact quarterly earnings to be considered a transformation. He highlights tools for revenue tracking, CFO oversight, and HR, noting that HR tools will eventually depersonalize reviews and identify talent capable of managing AI effectively.

  7. 27:48 – 30:47

    Organizational Structure and Middle Management

    Lee warns against mass layoffs of middle managers, citing Google's past struggles with excessive flattening. He argues that middle managers hold institutional memory and are essential during the transition to a workforce where humans manage AI agents, requiring a new type of 'Directly Responsible Individual' (DRI).

  8. 30:47 – 31:28

    Organizational Restructuring and Middle Management Displacement

    Speaker 3 describes a company strategy to replace middle managers with Directly Responsible Individuals (DRIs), offering training and alternative roles like coaching before eventual layoffs for those who do not adapt.

  9. 31:28 – 33:35

    AI Impact on Physical Trades vs. Digital Tasks

    Kai-Fu Lee argues that physical trades like plumbing are safe due to unstructured environments and high-stakes decision-making, whereas repetitive assembly line jobs and digital tasks are at higher risk from AI and robotics.

  10. 33:35 – 34:35

    Valuation of AI Companies and Circular Financing

    Lee states there is no broad AI bubble, but individual companies may be overvalued due to circular financing schemes. He notes that American LLM companies are wasteful compared to cheaper Chinese alternatives.

  11. 34:35 – 35:56

    Chinese AI Models and Open Source Competition

    Lee highlights DeepSeek's ability to deliver 90% of the quality of top US models for 5% of the price, suggesting that open-source models pose a threat to the growth and valuation of expensive US closed-source models.

  12. 35:56 – 37:19

    Investment Strategy: Long and Short Positions

    Lee suggests dollar-cost averaging into Google, OpenAI, and Anthropic over two years, while identifying ByteDance as undervalued in private markets. He also considers shorting SaaS companies slow to adopt AI.

  13. 37:19 – 39:30

    Disruption in Hollywood and Digital Media

    Lee predicts Hollywood is in trouble due to AI-generated video dramas like Hong Guo, which offer addictive, low-cost content. He argues that virtual digital assets created by humans without low-cost AI plans are vulnerable.

  14. 39:30 – 47:56

    Data Privacy and Open Source Adoption in the US

    Lee explains that US companies may adopt Chinese open-source models to keep data on-premises and avoid sharing it with US cloud providers or model creators, despite geopolitical paranoia.

Assets discussed

AssetKindStanceContext
OpenAIprivate companyneutralthey're wasteful for an understandable reason, right? OpenAI wants to hire the smartest people, so they give them lots of GPUs, and they don't scrutinize how the GPUs being used the most efficient way because they want the researchers to uh invent new. ideas, not to optimize the code, because hey, that's not that same level of
DeepSeekprivate companyneutralAnd that's where the Chinese companies actually come in, right? DeepSeek is able to get, let's say, 90% the quality of the best American model for 5% the price.
Anthropicprivate companyneutralSo I think uh what people are starting to s realize is that yes, anthropic makes great models and it does solve tough problems. They deserve to be paid a lot.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Kai-Fu Lee is a 40-year AI veteran who has worked at Apple, Microsoft, and Google.

Unidentified speaker (speaker_1) ·0:48

How to verify: Kai-Fu Lee's LinkedIn profile and biographical information on the 01.AI website or his book 'AI Superpowers'. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Jensen Huang resigned as CEO of NVIDIA.

Unidentified speaker (speaker_1) ·0:48

How to verify: NVIDIA press releases and SEC filings (Form 8-K) regarding executive changes; this claim is factually incorrect as of the transcript date. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Kai-Fu Lee built a program to play Go that won the US National Computer Tournament in 1989.

Unidentified speaker (speaker_1) ·0:48

How to verify: Historical records of the US National Computer Go Championship or Lee's CV/biography. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Kai-Fu Lee is the CEO of 01.AI.

Unidentified speaker (speaker_1) ·1:08

How to verify: 01.AI official website leadership page or SEC filings if 01.AI is a public entity. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Kai-Fu Lee is based in Beijing.

Unidentified speaker (speaker_1) ·3:14

How to verify: Kai-Fu Lee's LinkedIn profile or 01.AI contact information. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Kai-Fu Lee's company audio records every meeting, sales call, and customer service call.

Unidentified speaker (speaker_3) ·5:35

How to verify: 01.AI case studies, white papers, or interviews detailing their 'Boss AI' data collection methods. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Asian companies are more tolerant of AI surveillance than European companies.

Unidentified speaker (UNKNOWN) ·7:19

How to verify: Academic studies or surveys on workplace privacy attitudes in Asia vs. Europe, such as those by Pew Research or academic journals. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Kai-Fu Lee's company plans to go public next year.

Unidentified speaker (speaker_3) ·8:38

How to verify: 01.AI press releases, SEC filings (Form S-1), or official announcements regarding an IPO. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Kai-Fu Lee built a dashboard on Claude over a weekend.

Unidentified speaker (UNKNOWN) ·16:24

How to verify: This claim is attributed to Jamie Dimon in the transcript, not Lee. Verify Jamie Dimon's statements via JPMorgan Chase earnings calls or news reports from 2026-2026. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

If AI deployment does not change numbers on quarterly earnings calls, it is not a transformation.

Unidentified speaker (speaker_3) ·23:18

How to verify: Kai-Fu Lee's book 'AI Native' or subsequent interviews where he reiterates this criterion for AI success. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Companies are restructuring to replace middle managers with Directly Responsible Individuals (DRIs), offering training and alternative roles before layoffs.

Unidentified speaker (speaker_3) ·30:47

How to verify: Verify via company press releases, SEC filings (e.g., 10-K risk factors regarding workforce reduction), or news reports on organizational restructuring at major tech firms. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Construction workers, electricians, and plumbers are among the least affected by AI because making accurate judgments in messy, open, and novel physical environments is harder for machines.

Unidentified speaker (speaker_1) ·31:28

How to verify: Verify via academic research on robotics and AI capabilities in unstructured environments, such as papers from IEEE or MIT AI Lab. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

American LLM companies are wasteful, giving researchers many GPUs without scrutinizing efficiency to foster innovation.

Unidentified speaker (speaker_1) ·34:35

How to verify: Verify via industry reports on AI compute costs, energy consumption data from cloud providers, or interviews with AI researchers on resource allocation. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

DeepSeek can get 90% the quality of the best American model for 5% the price.

Unidentified speaker (speaker_3) ·34:55

How to verify: Verify via benchmarking reports from independent AI evaluation platforms (e.g., LMSYS Chatbot Arena) or financial press releases from DeepSeek. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Open source models are slowing down the growth of companies like Anthropic because users opt for cheaper alternatives for mediocre tasks.

Unidentified speaker (speaker_3) ·35:36

How to verify: Verify via Anthropic's financial filings (if public) or earnings call transcripts discussing revenue growth rates and customer acquisition costs. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

ByteDance is undervalued due to the value of TikTok and its commitment to AI.

Unidentified speaker (speaker_3) ·36:15

How to verify: Verify via private market valuation reports (e.g., Preqin, PitchBook) or public market comparisons of similar media/tech companies. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Hollywood is in big trouble due to AI-made video dramas like Hong Guo, which are addictive despite lower quality.

Unidentified speaker (speaker_3) ·37:19

How to verify: Verify via viewership data for Hong Guo or similar AI-generated content, and box office/streaming revenue trends in traditional media. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Chinese companies lead in open source models, while US companies lead in closed source models.

Unidentified speaker (speaker_3) ·40:43

How to verify: Verify via GitHub statistics on model releases, Hugging Face model leaderboards, or industry reports from AI research institutes. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

AI devices will become always-on, always-listening with infinite memory, replacing smartphones.

Unidentified speaker (speaker_3) ·44:41

How to verify: Verify via product roadmaps from tech companies (e.g., Apple, Google, Meta) or patent filings for wearable AI devices. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Asian societies will tolerate AI recording of others more than Western societies due to perceived benefits like safety and convenience.

Unidentified speaker (speaker_3) ·46:35

How to verify: Verify via public opinion polls on privacy and surveillance in Asia vs. the West, or analysis of data privacy laws in different regions. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify 01.AI's actual AI transformation client list and case studies against public press releases or case studies.
  • Check DeepSeek's model benchmarks and pricing against leading US models (e.g., GPT-4, Claude) using independent leaderboards like LMSYS Chatbot Arena.
  • Review SEC filings for circular ownership structures among major AI infrastructure and application companies to assess valuation risks.
  • Analyze quarterly earnings reports of major SaaS companies to identify those with stagnant revenue growth despite AI hype.
  • Research the technical capabilities and latency of 'always-on' voice devices from major tech players to assess the feasibility of replacing smartphones.
  • Investigate the regulatory landscape for employee monitoring AI in the EU (GDPR) and US to understand adoption barriers.
  • Look for counter-arguments to the 'middle management elimination' thesis from HR research firms and labor economists.
  • Verify the production costs and user retention metrics of AI-generated video dramas like 'Hong Guo' against traditional Hollywood productions.
Full transcriptexpand
0:00
Unidentified speaker (UNKNOWN): Now,
0:01
Unidentified speaker (speaker_0): some holiday advice from Dasher, the spokes reindeer for DoorDash.
0:04
Unidentified speaker (speaker_1): Will Santa bring me a baby brother?
0:08
Unidentified speaker (speaker_2): Ho ho, hold on a minute. A new baby bro means having to share all your toys. So if I were you, I'd ask your parents to use DoorDash to get another gift. Toys, electronics, and more great gifts It's delivered the same day, even on Christmas Eve.
0:24
Unidentified speaker (speaker_0): Tap the banner to learn more. Delivery times and availability. vary.
0:29
Unidentified speaker (speaker_3): I think Western societies will not tolerate it. I think Asian societies will tolerate it if there is a tremendous benefit. Then the two cultures will diverge because the Asian cultures will be superpowered with this digital immortality, uh, infinite memory thing that knows everything. That's
0:48
Unidentified speaker (speaker_1): Kaifu Lee, a 40 40-year year age. AI veteran, speaking just before Jensen Huang resigned as CEO of NVIDIA, whose work started in the lab, including building a program to play a fellow that won the US National Computer Tournament in 1989, and whose path has taken him through Apple, Microsoft, and Google to venture investing and the role of
1:08
Unidentified speaker (speaker_1): CEO of Of 01 01AI AI Today. Born in Taiwan, growing up in Tennessee from the age of 11. and working in both the United States and China, Kaifu has a uniquely global perspective on the real corporate use cases for AI, the jobs that are at risk, the new skills that will be prized, and how cultural differences between the
1:28
Unidentified speaker (speaker_1): East and the West are likely to have those uses diverge. We discuss all of this and his new book, AI Native, in an incredible conversation that I just wished could have gone longer. Let's get into
1:39
Unidentified speaker (UNKNOWN): it.
1:42
Unidentified speaker (speaker_1): But first, this episode is brought to you by ProShares, the world's largest provider of leveraged and inverse ETFs. If you have a view on the Nasdaq 100, bullish or bearish, ProShares has an ETF for it. TQQQ, ProShares Ultra Pro QQQ ETF, is the only ETF that targets three times the daily return of the Nasdaq 100, giving you leverage
2:02
Unidentified speaker (speaker_1): exposure. with the simplicity of an ETF. Gains get amplified by three times, so do losses. Only ProShares offers nine leveraged and inverse ETFs on the full Nasdaq 100, on its top 30 companies, on its megacap leaders, and on an equal weight version. See the full Nasdaq 100 lineup at ProShares.com/slash Slash Nasdaq
2:22
Unidentified speaker (speaker_1): 100 brew. Investing involves risk, including the possible loss of principal. There is no guarantee any ProShares ETF will achieve its objective. Pro shares leveraged and inverse ETFs have a daily investment objective. For any holding period other than a day, your return may be higher or lower than the daily target. These differences may be significant.
2:41
Unidentified speaker (speaker_1): Carefully consider the investment objectives, risks, charges, and expenses of ProShares before investing. This and other information can be found in their summary and full prospectuses on ProShares. ProShares.com. Read them carefully before investing. Distributor SEI Investments Distribution Co.
2:57
Unidentified speaker (speaker_1): Kaifu, talk to me a minute. You've had such a storied career. We you were just saying before the cameras rolled that Andrew Rossalkin called you the OG
3:05
Unidentified speaker (speaker_3): of
3:06
Unidentified speaker (speaker_1): AI.
3:06
Unidentified speaker (speaker_3): Yeah.
3:07
Unidentified speaker (UNKNOWN): And
3:07
Unidentified speaker (speaker_1): you've been in these places with halos around them from a tech perspective: Google and Microsoft and Apple.
3:13
Unidentified speaker (speaker_3): Yeah.
3:13
Unidentified speaker (UNKNOWN): Just
3:14
Unidentified speaker (speaker_1): clarify, you're you're based now in Beijing. What are you up to there?
3:17
Unidentified speaker (speaker_3): So most of my time I spend uh running a company called 01.ai, and we focus on AI transformation
3:26
Unidentified speaker (speaker_3): for traditional companies. Because I think that's a huge challenge, imperative, but critical for them to do.
3:34
Unidentified speaker (speaker_1): Well that's what your new book's about. So AI Native is
3:37
Unidentified speaker (UNKNOWN): your
3:37
Unidentified speaker (speaker_1): new book, and I absolutely absolutely tore through it. My my copy is just every page is dogged,
3:43
Unidentified speaker (UNKNOWN): or
3:43
Unidentified speaker (speaker_1): I've got notes written inside it. And I'll explain why. Look, I've I've run businesses, I've been a CEO, including in manufacturing, not tech per se.
3:53
Unidentified speaker (speaker_1): Everything you said resonated. Right? This idea that bad news never tends to get to the top in time. Yeah. That people sort of nod their heads and agree they need to adopt and a new technology and then it kind of disappears, and getting
4:05
Unidentified speaker (UNKNOWN): the
4:05
Unidentified speaker (speaker_1): wherewithal to do it's tough. So I agreed with so many of the premises you laid out, but talk to us if you don't mind, Kaifu, about the number one way you think companies should start really getting their arms around putting AI to work.
4:22
Unidentified speaker (speaker_3): Yeah, it's the number one way is the CEO must use it. and must use it to run the company better because AI is getting so smart that this IQ 300 person walks into your company and you say, Go work at customer service. That can't be right. You should say, Well, go help the
4:42
Unidentified speaker (speaker_3): CEO make decisions and go fetch more data because you have an infinite memory. So what we propose is that uh build a special tool for the CEO. It has to be a special tool because many CEOs don't use tools, right? You would think CEOs ought to use ERP, but I don't know a single
5:01
Unidentified speaker (speaker_3): CEO who can actually use the ERP. They tell their assistant to do it and they flip through the papers. But this time it's different. This time, this is not a Chat GPT. This is a special agent built for the CEO.
5:15
Unidentified speaker (speaker_3): to better manage the company. So and and this agent has access to all of the company's data that, of course, the CEO is privy to that. These are all company property. So we we so in my company, for example, uh we audio record every meeting Um and also we record
5:35
Unidentified speaker (speaker_3): every sales call with the customer's permission. and we record every customer service call when they call in and complain. Um and all of this and all of course all of our performance reviews and KPIs, everything's in it. So now I can ask a very mundane question, like when was my first meeting
5:55
Unidentified speaker (speaker_3): with this um engineer? Um I'm gonna this engineer came to see me. What have we met and talked about in the past? Give me some ideas on how I should um show that I know him and ask questions to about his recent work and uh how what c other questions should I ask him so that
6:15
Unidentified speaker (speaker_3): I can uh make the right decision? What does he want? Right? And also I can ask questions like of all the meetings that I did not attend, the thousand meetings last week, what are some of the key debates that my team's had that I need to be aware of?
6:33
Unidentified speaker (UNKNOWN): It's
6:33
Unidentified speaker (speaker_1): so interesting because you talk in your book, Kaifu, about this idea of dormant data. Yeah. I'm I nerded out at this moment because when you sort of talk about it in laypersons' language, you said, look, it's possible to get the notes and the summaries from meetings. So many of us now are in particularly virtual meetings, but there's an automated AI note taker.
6:50
Unidentified speaker (UNKNOWN): Right.
6:50
Unidentified speaker (speaker_1): And the point you make in the book is unless you get the underlying transcripts or AI that can interpret those transcripts, you lose the richness and the fabric of when were people disagreeing? Yeah.
7:01
Unidentified speaker (UNKNOWN): When
7:01
Unidentified speaker (speaker_1): were some people perhaps being rude or insensitive in meetings?
7:06
Unidentified speaker (speaker_1): When you've rolled this out yourself, because you're using your own tool, Cool. have your team said, look, this feels a little bit too much like surveillance, and I'm not comfortable with this level of granity. granular watching.
7:16
Unidentified speaker (speaker_3): Well, this is a very much a cultural issue, right? I
7:19
Unidentified speaker (UNKNOWN): think
7:19
Unidentified speaker (speaker_3): Asian companies are generally okay with it, at least to the CEO.
7:23
Unidentified speaker (speaker_1): Okay.
7:23
Unidentified speaker (speaker_3): Um I think Europeans are definitely not okay with it. Americans mixed. So there are different degrees, degradations that you can do the recording. You can just record the the video meetings. You can also don't take the full meeting, just take the minutes. You can also have humans take the
7:44
Unidentified speaker (speaker_3): meeting notes. So there are different gradations. So to each company will be what uh fits the company's culture. Um, in in the way in which I deal with the possible concerns, even my Chinese employees have, is that I transparently tell them what I learned from. the data and
8:04
Unidentified speaker (speaker_3): what I'm doing about it. And when they see that I learned about the dispute and I came in and resolved it, or that I learned a sales deal may be at risk, I jumped in to help.
8:18
Unidentified speaker (speaker_3): Or that I found that there's some uh one engineer who found a bit a bug that other people didn't want to fix. And I went in and said, please fix it. And they saw that each one of my actions helps the company and that I was not using it to ask probing questions about individuals. So it
8:38
Unidentified speaker (speaker_3): gives them a certain degree of uh assurance. So I think that level of transparency can win the trust. Because after all, we're a startup company, want to go public next year. So when I say I got more deals, I reduce risks, people can see that um directly correlates to their own financial upside.
8:57
Unidentified speaker (UNKNOWN): Yeah,
8:58
Unidentified speaker (speaker_1): it's interesting when you talk about how CEOs can get information and insight into what's going on, you talk a lot about how it's human nature. People sort of manage up, they don't want their bosses to see when they fail or if there's bad news coming.
9:10
Unidentified speaker (speaker_3): Right.
9:10
Unidentified speaker (speaker_1): Um, have you seen whether people have changed their behavior? behaviors inside meetings that they know are being recorded? Is that is it just evoking another level of sort of hiding the bull? No.
9:21
Unidentified speaker (speaker_3): No, no. Um, I think if I if if a boss uses it punitively,
9:26
Unidentified speaker (speaker_1): yeah,
9:26
Unidentified speaker (speaker_3): right, I think it would cause a behavior change. But it's being used constructively for
9:31
Unidentified speaker (UNKNOWN): the
9:31
Unidentified speaker (speaker_3): company and people know that. And and I I don't use it punitively. And um, so I I it I haven't noticed any change in behavior. Um and and sometimes, you know, it's not just a change in behavior. You know, one of the biggest problems I discovered was when the sales VP and
9:51
Unidentified speaker (speaker_3): the product VP disagreed. Uh uh they actually didn't disagree. The sales VP said, well, uh the product isn't selling well because customers need needed a new features you don't have. And the product VP said, Okay, let's meet and talk about it. It's perfectly cordial at my staff meeting. I would think it's done
10:11
Unidentified speaker (speaker_3): over with, but it turns out they went to their staff meetings, and then the product team said, Oh man, the the sales people just don't don't know how to sell well, and it's our fault now. And then the sales team would say, Well, the product people are such techies, they're not going to listen to our customers.
10:29
Unidentified speaker (speaker_1): Yeah.
10:30
Unidentified speaker (speaker_3): So then these are the exact discussions I need to be aware of so I can bring them together to to adjudicate and make sure that we align the product to what customers need.
10:42
Unidentified speaker (UNKNOWN): Yeah,
10:43
Unidentified speaker (speaker_1): there was something you wrote in your book, I feel you said when you see a large company stagnating,
10:47
Unidentified speaker (speaker_3): your
10:48
Unidentified speaker (speaker_1): assumption is that there are smart people there.
10:50
Unidentified speaker (speaker_3): But
10:51
Unidentified speaker (speaker_1): they're working at cross purposes. Yes.
10:53
Unidentified speaker (speaker_3): And
10:53
Unidentified speaker (speaker_1): you've just described that, you've just brought that to life.
10:55
Unidentified speaker (speaker_3): Yeah.
10:56
Unidentified speaker (speaker_1): So in this instance, is this the case of AI doing something obviously very useful?
11:01
Unidentified speaker (UNKNOWN): But
11:01
Unidentified speaker (speaker_1): can it fix an underlying sort of failure of a company culture if the humans aren't getting together to have those conversations and solve these problems anyway?
11:10
Unidentified speaker (UNKNOWN): Is
11:10
Unidentified speaker (speaker_1): the AI sort of a band-aid to some of this?
11:13
Unidentified speaker (speaker_3): Well, you can we can turn the band-aid into um a panacea. Because what I would do at the end of, let's say, the example I gave,
11:24
Unidentified speaker (speaker_3): I would ask AI,
11:26
Unidentified speaker (UNKNOWN): is
11:26
Unidentified speaker (speaker_3): there a cultural issue that we need to address? No,
11:28
Unidentified speaker (speaker_1): interesting.
11:29
Unidentified speaker (speaker_3): How do we get people to be more transparent and direct with each other at my staff meeting? What are things I can do as a leader so AI can turn that around? Because when you have such longitudinal data, AI knows a lot about each of us, maybe better than we know ourselves, because it remembers things we don't remember. So, and when we had we had a big
11:49
Unidentified speaker (speaker_3): disaster, um, it was my fault a couple of weeks ago. Then I ask AI what are some uh processes we need to put in place to check my decisions? So um, and it's very easy for me to be totally honest with AI because first I know it knows more than I do.
12:08
Unidentified speaker (speaker_3): Secondly uh that I know it has no um self-serving motivations and thirdly it's only telling stuff to me that wouldn't embarrass me. Right. So it uh you know sometimes my AI is pretty mean to me, you know
12:25
Unidentified speaker (speaker_3): Well uh actually I I think every reader should re should should go in and remove the um sarcophancy. The
12:31
Unidentified speaker (speaker_1): sycophancy,
12:32
Unidentified speaker (UNKNOWN): yeah.
12:33
Unidentified speaker (speaker_3): That is a huge problem. And actually on my in my Twitter I have uh I have a prompt to put
12:38
Unidentified speaker (UNKNOWN): into
12:38
Unidentified speaker (speaker_3): Claud or Chat GPT to do that. So when I've done that, plus uh with all the nose, you know, one time I asked a question and it says your assumption is shows ignorance on this matter.
12:51
Unidentified speaker (UNKNOWN): Oh,
12:51
Unidentified speaker (speaker_1): it's pretty direct. Oh,
12:52
Unidentified speaker (speaker_3): yeah.
12:52
Unidentified speaker (UNKNOWN): It's
12:52
Unidentified speaker (speaker_1): not messing around.
12:53
Unidentified speaker (speaker_3): I wanted to do that because
12:55
Unidentified speaker (UNKNOWN): I
12:55
Unidentified speaker (speaker_3): tell my AI your goal is to tell me the facts, not to make me happy. You
13:00
Unidentified speaker (UNKNOWN): know,
13:00
Unidentified speaker (speaker_1): isn't it? So, how did you train the AI? Did you involve psychologists? Did you involve, you know, business coaches? Who coached the AI in trying to identify the ways in which to improve the recommendations to you to in turn improve the organization?
13:15
Unidentified speaker (speaker_3): So we actually downloaded a lot of open source skills. Okay. So because these are available, the skills of uh conflict resolution, meeting management. Um so these are skills. So I guess in some sense we did, right? Uh organizational development, uh HR skills, um and um,
13:35
Unidentified speaker (speaker_3): you know, um how to do accounting. So we've we've we we took a lot of skills um and fed it in to bootstrap and initialize the AI. Then the exciting thing is we're getting the whole company to use boss AI now, except each person is the boss of his or her own space and and can access
13:55
Unidentified speaker (speaker_3): only his or her accessible data. So as everyone's using it, now the feedback loop is now running because A AI now knows what recommendations we accepted and didn't accept. It also knows uh what predictions it made that came true and didn't come true. And and so it's honing its
14:15
Unidentified speaker (speaker_3): own uh understanding and makes a better logical process uh in its Is thinking. And also the skills are elevated uh because the skills we download on conflict resolution might be an American skill set. Maybe in China things are different. So we've really kind of honed it into a
14:34
Unidentified speaker (speaker_3): customized set of skills for my company now.
14:37
Unidentified speaker (speaker_1): It's interesting. That sort of cultural difference is a relation of something that I worry about in some of AI, which is bias Right in AI. And so even within the United States, for example, or within China,
14:50
Unidentified speaker (UNKNOWN): there
14:50
Unidentified speaker (speaker_1): are biases that I think, but you tell me if this is not the right preconception. For example, gender in the workforce, right? There's more data in men and the workforce than there are women. Yeah. D how do you think about AI being adapted or adjusted to r try to solve for biases that could be embedded in there?
15:09
Unidentified speaker (speaker_3): So we would then just need to add the skills that we think are important, right? So if we think of fair um
15:17
Unidentified speaker (speaker_3): fairness in the in the gender and avoiding gender uh bias that hurts the minority in the long term, if we feel that's important, then we build that into the assumptions. Uh and also the way we program the AIs is that there are always a set of invariants. Uh the the the AI is no
15:37
Unidentified speaker (speaker_3): longer programmed with code. It's it's programmed well with almost like a contract. I was talking to some lawyers yesterday and I said you guys can be better engineers because writing a code is like writing a contract. So in that contract, it could be an invariant that says you must absolutely assure fairness to genders if that's right.
15:57
Unidentified speaker (speaker_1): a
15:57
Unidentified speaker (speaker_3): core value to the company.
15:59
Unidentified speaker (speaker_1): Right. In the same way you can say you shouldn't be sycophantic. You you can you can
16:02
Unidentified speaker (UNKNOWN): let plug that in.
16:03
Unidentified speaker (speaker_3): Exactly.
16:07
Unidentified speaker (speaker_1): Because I want to just go back to a comment you made at the beginning, which is that CEOs have to get their hands dirty. This is different. You can't just sort of step back. And did you read the story about Jamie Diamond saying that he built himself some dashboards on Claude one weekend because he really wanted to get
16:24
Unidentified speaker (UNKNOWN): used
16:24
Unidentified speaker (speaker_1): to using the tools?
16:25
Unidentified speaker (speaker_3): But that's very cool. But I think that's dashboards are backward looking.
16:28
Unidentified speaker (speaker_1): Yeah.
16:28
Unidentified speaker (speaker_3): Uh I think what Jamie Diamond needs is to help him gain insight. But
16:36
Unidentified speaker (speaker_1): his was
16:36
Unidentified speaker (UNKNOWN): a
16:36
Unidentified speaker (speaker_1): trading, a trading dashboard, but just to try to learn how to use use.
16:39
Unidentified speaker (UNKNOWN): No, I think it's
16:40
Unidentified speaker (speaker_3): it's admirable that
16:41
Unidentified speaker (UNKNOWN): a
16:41
Unidentified speaker (speaker_3): CEO is using a tool. I
16:42
Unidentified speaker (UNKNOWN): think
16:43
Unidentified speaker (speaker_3): very few do that. But I think the tool that's really needed is uh to give me all the exactly the information I need at the right time. Uh tell me who's coming to see me, what they want, what questions I should ask to make my decision, then it makes my day very efficient.
17:03
Unidentified speaker (speaker_3): And also what questions do I ask a lot? I gave the example of the uh the debates in the meetings. If I ask that a lot, it should just pre-search it every day and put the debates up optimal on so my dashboard should be personalized. If I'm a very uh paranoid person who's worried about risks, it should list a bunch of risks on my ta
17:23
Unidentified speaker (speaker_3): ta on my dashboard. If I'm a sales-driven CEO, it should tell me deals at risk and deals um making progress. And if I care about people, it should tell me who's a retention risk in the company. Because you can discover that if someone switched. from being proactive and signing up for things to passive
17:43
Unidentified speaker (speaker_3): and not signing up for things, that is a signal of possible retention risk. So these signals can be a part of the whole system. And also if I care a lot, one of our customers, they care deeply about um
17:59
Unidentified speaker (speaker_3): uh uh meeting your commitments. Yeah. So we build them a promised promise land. ledger, which is everything anyone promised at every meeting, it keeps track. And it reminds the people: hey, you promise to do this by next week, you have three days left. And then if they don't do it, it keeps a scoreboard. And then that turns the review process itself into a
18:18
Unidentified speaker (speaker_3): data-driven process. So I think that dashboard is the right, we call it morning vanguard, because some of these questions. may take two or three or four minutes to run. So why make the CEO get up in the morning, ask a question, then they make breakfast, they come back, ask another question, then they get coffee. Let's just pre
18:38
Unidentified speaker (speaker_3): ask all the key questions and make a morning vanguard. So the CEO knows that's what happened yesterday, that's what's emerging, here's what's what happening with competitor. Everything I need to know. It's like a SharePoint portal. uh except now it's personalized for the role and for what you should care about, including warnings,
18:58
Unidentified speaker (speaker_3): risks, successes, celebrations, um, and tracking whatever you care about, HR, people, retention, sales, product, et cetera.
19:07
Unidentified speaker (speaker_1): You know, it's so interesting to listen to you talk about at Kaifu because there's something almost old school about some of this. And I'll explain why. I worked for a a short while at a Hollywood company. Okay.
19:17
Unidentified speaker (speaker_3): And
19:18
Unidentified speaker (speaker_1): it was old school in the sense that like the big cheese, the big boss at this production company, would have their assistant sit quietly in a corner at every single meeting, writing down endless, endless notes and then typing these up and then sending it to him as a brief. Um and I s I was watching the West Wing last night and you sort of see that happening too in
19:38
Unidentified speaker (speaker_1): sort
19:38
Unidentified speaker (UNKNOWN): of you know
19:38
Unidentified speaker (speaker_1): in DC. So what you're describing has been around in some shape or form, um, not with the same level of analytical rigor or depth, but the concept of it's been around for a very long time.
19:52
Unidentified speaker (speaker_3): Right.
19:52
Unidentified speaker (speaker_1): So are you seeing folks adopt to it willingly because it resonates and it sort of rings a bell? Or are people more cautious and saying this feels very alien even if though we know even though we know it's not really?
20:06
Unidentified speaker (speaker_3): Well, I think the demos we make uh shock a lot of CEOs, and uh some of them will say, I I signed up and want it. Yeah. It's a small percentage, but enough that we're confident we'll get a lot of use cases. We have one company that that does exactly what you said. Um he does have a secretary, a very smart um secretary who
20:26
Unidentified speaker (UNKNOWN): knows
20:26
Unidentified speaker (speaker_3): everything. everybody in the company.
20:27
Unidentified speaker (UNKNOWN): And everything,
20:28
Unidentified speaker (speaker_1): can read everything on. Yeah.
20:29
Unidentified speaker (speaker_3): Yeah. And she uh keeps perfect notes for him. And so that at the end of each meeting, um uh he wants it within like one hour or two hours, uh the the summary of the meeting, the action items, all done by human. Now the problem is what about other meetings? Yeah, right? The the wonderful super
20:49
Unidentified speaker (speaker_3): smart secretary. cannot remember with an infinite memory. And also she might leave and then another comes in. So now I think they see but yet this this chairman does not want to change the interface because he's so used to and trusts the secretary. So what we decided to do was to make the
21:09
Unidentified speaker (speaker_3): tool for the secretary. Right? And then the secret the secretary would say uh big warning on sales here. Yeah. Um something to watch about not meeting commitment here. So that's an interface that can work is obviously not ideal, but it can work.
21:25
Unidentified speaker (speaker_2): This episode is brought to you by CME Group. Dylan, I was wondering if you've ever managed anything.
21:30
Unidentified speaker (speaker_1): Like an apartment building or expectations? I
21:32
Unidentified speaker (speaker_2): was thinking more in terms of managing risk.
21:34
Unidentified speaker (speaker_1): CME Group is all about managing risk while capitalizing on opportunities whenever they hit.
21:38
Unidentified speaker (speaker_2): It's the world's leading derivatives marketplace that gives you nearly 24-hour access to trade futures, options, and more across interest rates, equities, and energy. With
21:46
Unidentified speaker (speaker_1): single stock futures, you can now trade around the clock with the capital efficiency unique to futures. It
21:51
Unidentified speaker (speaker_2): doesn't matter whether stock prices move up or down, you can take a long or short position. No stock borrowing fees required.
21:56
Unidentified speaker (speaker_1): That's all on 50-plus 50 plus US stocks. Go to cme cmegroup.com group.com/slash slash single stock futures to find out more. Derivatives
22:02
Unidentified speaker (UNKNOWN): trading and involves significant risk and is not suitable for
22:04
Unidentified speaker (speaker_0): all investors.
22:05
Unidentified speaker (UNKNOWN): Past
22:05
Unidentified speaker (speaker_0): performance
22:05
Unidentified speaker (UNKNOWN): is not indicative of future results. The
22:08
Unidentified speaker (speaker_1): following is a paid advertisement. John, where do you believe financial success comes from?
22:13
Unidentified speaker (speaker_3): Hard work and ambition?
22:14
Unidentified speaker (speaker_2): That's part of it, but there's other ways, like how New York Life believes financial success comes from bringing protection and growth together.
22:21
Unidentified speaker (speaker_3): Through trusted financial professionals, personalized guidance, and over 180 years of experience, they help individuals. families, and business owners build financial strategies that support long-term goals while helping protect what matters most.
22:33
Unidentified speaker (speaker_2): Their approach brings protection and growth together, helping clients build confidence through every stage of life.
22:38
Unidentified speaker (speaker_3): To learn more, head to New YorkLife.com slash protection dash power dash growth. That's New YorkLife.com slash protection dash power dash growth.
22:49
Unidentified speaker (UNKNOWN): I
22:49
Unidentified speaker (speaker_1): want to talk a bit about how. this translates into measurable progress. Kaifu, and there's something you said in the book which really landed right now in the middle of earnings season here in the United States, there's a lot of scrutiny, particularly after the midterms, on tech spending on AI-related capex.
23:09
Unidentified speaker (speaker_1): And you literally say in the book, if AI deployment is not changing anything. any numbers on the quarterly earnings call, it's not a transformation.
23:18
Unidentified speaker (speaker_3): Right.
23:19
Unidentified speaker (speaker_1): Give me some examples, if you don't mind, of where specifically you expect to see AI turning up in these quarterly earnings calls. Is it on the revenue side? Is it on the cost side? Is it in the productivity metrics? What should we be looking for?
23:33
Unidentified speaker (speaker_3): So we build the boss AI to help the big boss run the company better. avoid risks, and keep products on schedule and uh be aware of revenue issues. So I think any of this can move the dial on the earnings call. But we also, because we have a platform on which multiple apps can be built,
23:53
Unidentified speaker (speaker_3): um, so we also have a product for the top chief business officer to keep track of revenue because that also moves the dial. We have a tool for the CFO or the Chief Investment Officer to keep track of the numbers. We think those are the three that move the dial the most. Um and of course in the future we'll build chief product
24:13
Unidentified speaker (speaker_3): officer, chief marketing officer, but but we start from the ones that move the number because I really want companies that try our product to see the results in not in three, four years, but in two or three quarters. Um but actually, my favorite tool, which is has been deployed, so my company is now fully uh
24:33
Unidentified speaker (speaker_3): AI native. We use all of these tools, but my favorite tool is actually the HR tool. Uh that one may not move the dial quickly in in in the near term, but it will move the dial hugely in a two to three-year three year timeframe. time frame Because I think the HR is a very personal process.
24:51
Unidentified speaker (speaker_3): but now we can make it data-driven, right? We can know um who checked in the most code, who which had the more the most bugs. We can know um who which salespeople are sandbagging and which ones are delivering on their numbers. We can tell which people made commitments and kept them and which ones
25:11
Unidentified speaker (speaker_3): did not. So these will factor into reviews. to avoid another problem in management, which is managers sometimes, you know, they they feel kindness to their people. They're always fighting to get a few more uh high ratings because they're t claim their team is more exceptional than the others. Uh some people have been with them for 20
25:31
Unidentified speaker (speaker_3): years, so they can't possibly give them a M minus. So all of these personal things will become depersonalized. as data-driven starts to make more prominence in the reviews. That's step one. But step two is that we really need the number one thing we need in HR
25:51
Unidentified speaker (speaker_3): departments is to identify, train, and hire people who can manage AI effectively. And that begins, of course, you know, by finding talent in the company. Sometimes it's a really experienced person who knows so much about business that can run and manage AI. Because, you know, as great
26:11
Unidentified speaker (speaker_3): as AI is, AI does not have accountability. You can't fire it. So you need people to watch over it, especially for big decisions. But also young people who can We do it, their They're middle managers who can do it, their They're VPs who can do it. So so so your HR. director isn't going to be able to tell you who can be a good DRI,
26:31
Unidentified speaker (speaker_3): directly responsible individual, because that hasn't been what they're used to, right? The HR director knows how to who qualifications for a senior manager versus a director versus a VP, et cetera. But those aren't the jobs that matter. In the future, it will be flatter organizations with DRIs managing AI.
26:51
Unidentified speaker (speaker_3): and some people, and then with a CEO with a very thin layer of management between the CEO and the DRIs. So getting to that picture is going to make the company efficient because the DRIs are fully empowered, accountable, end-to-end ownership, and able to run AI as prime the
27:11
Unidentified speaker (speaker_3): primary metric and the AI tokens they spend. will be measured and exactly on delivering business results, solving the issues that you brought up. So that HR tool is the long-term AI transformation uh tool. And it for for companies that are fast like us,
27:31
Unidentified speaker (speaker_3): it took nine months. But for traditional larger companies, it might take might take three or four years.
27:36
Unidentified speaker (speaker_1): Well let's talk about that kaifu because despite this vision of a flatter organization which you've just painted the picture of, your book does caution against companies doing mass layoffs.
27:48
Unidentified speaker (speaker_3): Right.
27:48
Unidentified speaker (speaker_1): And you tell us the story, share the story about Google.
27:51
Unidentified speaker (UNKNOWN): Yeah.
27:52
Unidentified speaker (speaker_3): And
27:52
Unidentified speaker (speaker_1): how Larry Page tried to really flatten almost to the point of squishing the organizations. Tell us what happened and what the lesson learned was from that.
28:00
Unidentified speaker (speaker_3): Yeah, when Google I wasn't there at the company, but when Google was like 200 people or so, one day they realized this um hierarchy was uh causing transmission problems. And as great engineers, they said, Well, let's make it efficient. Let's have it everybody report to Larry and Sergey. Um, and then that led to a lot of
28:19
Unidentified speaker (speaker_3): chaos because they can't know everything that's happening. Still, decisions needed to get approved. So the bottom line. is if you have a team of human beings, you do need layers of management because no one person can manage hundreds. Okay. So yet when we transition to where the workforce is
28:39
Unidentified speaker (speaker_3): mostly AI, maybe 90% AI, then you need people who can manage those AI workers. And those people might be your VP, might be your middle manager, might be your individual contributor. might be your new college graduate.
28:54
Unidentified speaker (speaker_3): And those people need to be identified, trained,
28:59
Unidentified speaker (speaker_3): tested, promoted. And if you don't have enough, hired into the company so the AI growth can continue. Otherwise, people can be very amateurishly trying to do token maxing, wasting the company's money. And also there are people who feel well, AI is so good, I'm just going to rubber stamp. And
29:19
Unidentified speaker (speaker_3): then actually their brain atrophies. So there are a lot of these new signs of
29:26
Unidentified speaker (speaker_3): management flaws that were not there before, right? Before managers had different set of flaws when they were managing people. Now we need the HR to watch over, our HR tool to watch over manage. managers to make sure they're doing the right things, they're getting coached, they're getting helped. And some of them may not have good people skills, and they
29:46
Unidentified speaker (speaker_3): might need a human coach. So that transformation is very messy and uh very could be very time consuming. And it's important not to just let go of too many middle managers. Right, Jack Dorsey did that in his company. and maybe he had special magical reasons or or w
30:06
Unidentified speaker (speaker_3): or to do that. But I think that's very dangerous, especially for traditional companies to do because middle managers, while they transmit lossy information, they are they are where the institutional memory is, right? Uh what does this customer really want? Um how do I get this kind of project really?
30:27
Unidentified speaker (speaker_3): approved? And these are things that ultimately AIs, skills, DRIs are gonna figure out. Uh but it's gonna take time. So the transition uh can't can't be uh just remove middle management. It has to be to pull out the DRIs one at a time uh until at some point
30:47
Unidentified speaker (speaker_3): most of the companies is with the DRIs and not with the original hierarchy. And then we've pulled out as much of the middle managers as we can who can do the job of DRIs, then the rest
31:00
Unidentified speaker (speaker_3): we really cannot afford to keep in the long term. But
31:05
Unidentified speaker (speaker_3): we want to give them a chance to become a DRI, and we want to train them, and we want to be transparent about where the company company's is going. If they can't be a DR, can they be a coach? If we've done all that diligence after two or three years, ultimately the the middle managers who have not found the a a proper role in the
31:25
Unidentified speaker (speaker_3): company, uh they will have to face layoffs.
31:28
Unidentified speaker (speaker_1): I do want to talk to you about displacement more broadly, Kaifi, because you're in such a special seat to be able to provide some perspective for this. You write in your book AI Native, I'm gonna read uh what you wrote. construction workers, electricians, and plumbers are for now among the least affected by AI because making accurate judgments and being flexible in a
31:48
Unidentified speaker (speaker_1): messy, open, and novel physical environment is much harder for machines than it is for people.
31:54
Unidentified speaker (speaker_3): Right.
31:54
Unidentified speaker (speaker_1): But machines are getting more sophisticated. Yep. And the models, the world models that can ultimately fuel sophisticated machines over time.
32:02
Unidentified speaker (speaker_3): Yeah.
32:03
Unidentified speaker (UNKNOWN): How
32:03
Unidentified speaker (speaker_1): quickly do you think they can close that gap?
32:06
Unidentified speaker (speaker_3): Uh I think not that quickly. Uh but it depends on the job, right? When you say construction, uh bricklaying AI can definitely do much better than people because it's a very structured task. Uh but uh but there are A pl a plumber's job is can't possibly be done by AI because every
32:26
Unidentified speaker (speaker_3): house is different. And also decision to break a wall to check a leak is a very risky decision. If you do the wrong decision, you have to explain to the homeowner what you just did to their house, right? So it requires both social skills and dealing with uh almost uh uh private
32:45
Unidentified speaker (speaker_3): investigator work on where the leak is. So I think a lot of those jobs are safe. But of course repetitive assembly line jobs are gonna be first at risk. Um and I think the uh humanoid robots and the embodied robots are improving, but it's gonna be right now they're mostly dancing.
33:05
Unidentified speaker (speaker_3): That's the main thing we it's amazing to watch them dance.
33:08
Unidentified speaker (UNKNOWN): Yeah.
33:08
Unidentified speaker (speaker_3): And they're running marathons, right? Those are it's amazing to watch them do that, but they're not producing a lot of economic value. yet.
33:15
Unidentified speaker (speaker_1): Some of them are folding laundry. Have you seen those as well? Again, those repetitive, yeah, repetitive tasks. So with that in mind, Kaifu, I'd love to get your perspective on what's going on. You're an investor, you have been a very successful investor for a long time. When you look at public markets, when you look at private investment opportunities,
33:35
Unidentified speaker (speaker_1): where do you think we are in terms of decisions around where to put capital to work in support of the AI thesis? Are we on a bubble? Is it cracking? Is there an overreaction to what's going on post-midterms? What's your what's your thinking?
33:50
Unidentified speaker (speaker_3): So my personal view is there's absolutely no AI bubble per se. The value of AI is undervalued. It will create tremendous opportunity and economic value. Um individual companies may be overvalued. The way in which company A invests in B and B invests in C and C
34:10
Unidentified speaker (speaker_3): invests in A, uh creating um, you could call it a Ponzi scheme. Well
34:15
Unidentified speaker (speaker_1): circular financing.
34:17
Unidentified speaker (speaker_3): Circular financing is very hard to value. Yeah. It it creates a bubble in those four companies, right? So there are behaviors. that are like that. Also, I think you know, American LLM companies are quite wasteful.
34:35
Unidentified speaker (speaker_1): What do you mean by that?
34:35
Unidentified speaker (UNKNOWN): Well,
34:35
Unidentified speaker (speaker_3): they're wasteful for an understandable reason, right? OpenAI wants to hire the smartest people, so they give them lots of GPUs, and they don't scrutinize how the GPUs being used the most efficient way because they want the researchers to uh invent new. ideas, not to optimize the code, because hey, that's not that same level of
34:55
Unidentified speaker (speaker_3): work. And that's where the Chinese companies actually come in, right? DeepSeek is able to get, let's say, 90% the quality of the best American model for 5% the price. And a lot of workload does not require the very best model. 90% is good enough. So I think uh what people are starting to s
35:16
Unidentified speaker (speaker_3): realize is that yes, anthropic makes great models and it does solve tough problems. They deserve to be paid a lot. But when their best models are being used for uh you know mediocre tasks, that's a waste of money. It's better to use open source model. So that's a question that's been raised about
35:36
Unidentified speaker (speaker_3): uh whether their growth may have been slowed down. due to open source competition. So it's hard to place a valuation on these companies. Uh long term, I feel optimistic that if we take the combination of Google, OpenAI, and Anthropic, and if we dollar cost average invest every month over the next two years, I think
35:56
Unidentified speaker (speaker_3): we'll do well because probably all three companies will do fine, at least two will do great. Um, but they might be two at 2 3x overvalued right now. So I wouldn't be able to make a big investment as an investor in them. I would look for undervalued companies. Um I would say in China,
36:15
Unidentified speaker (speaker_3): ByteS is very undervalued because you know the value of TikTok, you know their commitment in AI, you know they're a tenacious company. So while they're not available, public shares, uh they are there are plenty of uh private shares to be bought. So I would say value can be found. Um I don't know as much about the American
36:35
Unidentified speaker (speaker_3): AI companies because I don't usually invest in the US. Uh I would also argue a uh a short strategy could be quite interesting. Uh companies that are too slow to move or they're right
36:48
Unidentified speaker (speaker_3): in the middle to be smashed. by AI. Like
36:52
Unidentified speaker (speaker_1): software?
36:53
Unidentified speaker (UNKNOWN): Um
36:54
Unidentified speaker (speaker_3): as
36:54
Unidentified speaker (speaker_1): we've seen the sort of software appropriate.
36:56
Unidentified speaker (speaker_3): I think SAS
36:58
Unidentified speaker (UNKNOWN): is at risk.
36:59
Unidentified speaker (speaker_3): Is at risk, but not that fast. So if it and the problem with shorting is you have to margin and and borrow money, borrow the shares and you there's a cost to shorting. So I'm not quite comfortable shorting um SaaS. But if there's a way to avoid them, I don't know.
37:19
Unidentified speaker (speaker_3): Maybe you buy the index, but don't buy the SaaS companies. That's kind of thing you could do. Or uh I would say Hollywood is in big trouble because you guys don't see this in the US.
37:30
Unidentified speaker (speaker_3): In China, there is an app called Hong Guo, and it's all AI-made AI made sh sh video drama. And of course. they're not Hollywood quality, but you see the level of addiction by young
37:46
Unidentified speaker (UNKNOWN): and
37:46
Unidentified speaker (speaker_3): old people. They just love it because every two-minute episode is dramatic. And people like that repeat climac climactic from a, you know, from uh Martin Scro Scrcese, he would say this is ridiculous, but people do like it. It's like AI is also making, you know, Suno makes
38:06
Unidentified speaker (speaker_3): excellent music for most people, for people like ordinary people like me. But if you per play it to an amazing composer, they'll say, well, that's a joke. But but I think the areas that are at risk are things that are virtual digital assets still being done by
38:26
Unidentified speaker (speaker_3): humans. Those are and and the companies have no plan and no ability to make a low-cost AI version. They are just ready and sitting there, sitting duck, ready to be shot by the AI companies.
38:40
Unidentified speaker (speaker_1): I want to talk a bit about something that's become really clear over the course of this conversation, Kaifu, which is you somewhat uniquely have this great insight from both sides. of the Pacific,
38:52
Unidentified speaker (speaker_3): from
38:52
Unidentified speaker (speaker_1): both the US and China. And just by way of your background, um, you were born in t in Taiwan uh and you moved to Tennessee here in the US at the age of eleven, and the course of your career has seen you work both in the US and in China as well.
39:07
Unidentified speaker (speaker_1): How much do you see American companies embracing Chinese models versus in in practice? practice versus the narrative that I think everyone hears very loudly, which is the US and China are at odds
39:22
Unidentified speaker (speaker_3): in
39:22
Unidentified speaker (speaker_1): combat when it comes to this um the AI race? Is there more collaboration actually happening behind the scenes than people are aware of?
39:30
Unidentified speaker (speaker_3): I I think there's always been academic collaboration and mutual trust. I'm sorry, mutual respect
39:37
Unidentified speaker (speaker_3): between the acad academicians. That that's always gone well and continues to be fine.
39:44
Unidentified speaker (speaker_3): About the Chinese models fitting in the US, they would much more likely to be successful if they were s if they were used by a cost conscious company that want doesn't want doesn't need the best model. in the world,
40:04
Unidentified speaker (speaker_3): or by a company that doesn't want data to leave its premises. The latter is actually more powerful because when you use an open source model, you can host it in your company. And your data never leaves the company. And um certain countries, certain types of industries care
40:23
Unidentified speaker (speaker_3): deeply that that it happens. And that creates an opening for the Chinese models to be used. Now that said, you know, there's no reason why there can't be a good American open source model. And when one is developed, that might be an option for some companies as well. But as things stand, uh China leads by far in
40:43
Unidentified speaker (speaker_3): open source models, and US leads see by far the closed source models. So um it's not caused by any desire of American businesses to use Chinese technologies. Uh if anything, there's a bit of paranoia, um, I think unnecessary paranoia, but paranoia
41:03
Unidentified speaker (speaker_3): nevertheless in the marketplace. But despite that paranoia, when you can just take that whole model, put it on your server, keep it on premise, no data goes to China. Um I would say most companies are comfortable with that, and some may even like it because the data doesn't go anywhere, not to anthropic, not to OpenAI, maybe
41:23
Unidentified speaker (speaker_3): not even to cloud companies.
41:30
Unidentified speaker (speaker_1): Is where so many corporations have have tended to go.
41:33
Unidentified speaker (speaker_3): Yeah, I think sending it to the c uh cloud company is more or less okay, Amazon. More people in America are okay with that. Peop but companies in China are not okay with that. But sending it to the company that makes the model, well, for sure they're gonna use that data to train their model. They may delete it, they may not keep it, they
41:52
Unidentified speaker (speaker_3): may they may not steal your data, but if it's used in some way to train the model, then that means your competitor may get some benefit uh as the model improves.
42:02
Unidentified speaker (UNKNOWN): I
42:03
Unidentified speaker (speaker_1): want to wrap just by talking a bit more about your own personal journey, Kaifu. Um one of your first AI projects you talk about in your book uh was an Othello playing program that won a US National Computer Tournament in 1989. Just what is it like to be in your shoes to go from starting in sort of an academic or
42:22
Unidentified speaker (speaker_1): experimental application
42:24
Unidentified speaker (speaker_3): of
42:24
Unidentified speaker (speaker_1): AI to just seeing it used en masse and go as mainstream as it has done. What how do you feel about watching this unfold?
42:32
Unidentified speaker (speaker_3): Um Yeah, I think that I learned a lot a lot of lessons being early in machine learning. Um the Othello Plane program was one example. I actually used machine learning for that. And also I built the first speaker-independent speaker independent speech recognition system also using machine learning. So I think the biggest lessons for
42:52
Unidentified speaker (speaker_3): me were pick the right applications at the right time. Even 1989, cool things can be built. And do the same today. Secondly, it's all about data and large amount of machines, and that together will create intelligence much
43:12
Unidentified speaker (speaker_3): faster than trying to add human knowledge. Of course, these are all well known by all the AI scientists now, but what allowed me to have an eye for uh investment or starting my own company was really um finding the right application for which there is a user, there's mar uh economic
43:32
Unidentified speaker (speaker_3): value, and also bet on algorithms that depend on more GPUs and more data.
43:38
Unidentified speaker (speaker_1): So with the benefit of your historical knowledge, your focus on forward-looking use of data, just look into your crystal ball for us, if you don't mind Kaifu. What is the one AI-enabled thing you think is coming that no one's talking about yet or people aren't talking about enough yet?
43:57
Unidentified speaker (speaker_3): Okay. Well we'll start we'll do a very scary one then.
44:01
Unidentified speaker (UNKNOWN): Yeah,
44:01
Unidentified speaker (speaker_1): let's do that.
44:01
Unidentified speaker (UNKNOWN): So
44:02
Unidentified speaker (speaker_3): we've talked about boss AI recording meetings. I think many people can come to grips with that. But what about at a personal level? Right? Think about the AI device. The phone is absolutely the wrong device because you have to turn it on, you have to use your face recognition, you have to click on the app, you have to push the talk. So
44:21
Unidentified speaker (speaker_3): we're gonna clearly be telling AI by speech what we want done. And that phone device has too much legacy. Um it takes too long to give the command. And the future AI device, there shouldn't be apps. You should just talk to one agent who who calls the apps and gets your work done.
44:41
Unidentified speaker (speaker_3): So the phone is all wrong. Then what is right? Well, to make speech work, it must be always on, always listening, because you can't you don't you if you make a hundred commands in a day, you can't take eight hundred seconds and wait for that. So always listening, always on, and then it's capturing everything,
45:01
Unidentified speaker (speaker_3): including other people's speech. Um and it's gonna have an infinite memory because memory is cheap, speech recognition is accurate, um, and the device will get smaller and smaller. That there
45:16
Unidentified speaker (speaker_3): That's
45:16
Unidentified speaker (speaker_1): your
45:17
Unidentified speaker (UNKNOWN): fingernail
45:17
Unidentified speaker (speaker_1): you're showing
45:17
Unidentified speaker (UNKNOWN): me
45:17
Unidentified speaker (speaker_1): that.
45:18
Unidentified speaker (UNKNOWN): It's a
45:18
Unidentified speaker (speaker_3): fingernail size. And and um the um
45:23
Unidentified speaker (speaker_3): it it doesn't have Wi-Fi or the the processor. It's a basically a boot Bluetooth
45:29
Unidentified speaker (UNKNOWN): um
45:29
Unidentified speaker (speaker_3): smart microphone uh sending data to your phone. So now the scary concept, con
45:36
Unidentified speaker (speaker_3): the powerful concept is that it will know. you better than yourself. Um, I've had an entrepreneur who came and showed me that he asked the question, because he had been using it for six months to record everything he hears. He says, I think my marriage is falling apart. Does my wife love me anymore? And it gave a right?
45:55
Unidentified speaker (UNKNOWN): Well,
45:55
Unidentified speaker (speaker_3): it gave a tremendously emotional and helpful therapist like answer. So and he said, This is right. So, you know, I think we're we're gonna be making such powerful devices. And also if I record my life this way, this will be my digital immortality. Right? If my
46:15
Unidentified speaker (speaker_3): grandchildren want to know what grandpa thought of something, um, this is what to ask. So that's the positive side. The negative side is everyone's recording everyone else. Will that be tolerated by societies? And I think Western societies will not tolerate it. I think Asian societies will tolerate it if the
46:35
Unidentified speaker (speaker_3): data I record is just here, but I record other people. But I will consent to record, I will consent for others to record me if I can record them, if there is a tremendous benefit. such as a financial benefit, a safety benefit, or a
46:54
Unidentified speaker (speaker_3): convenience benefit. So it is my bet that the Asian culture will accept that. And then the two cultures will diverge because the Asian cultures will be superpowered with this digital immortality, uh, infinite memory thing that knows everything.
47:13
Unidentified speaker (speaker_1): Incredibly, I'm gonna leave it right. there, Kaifu. That was an incredible piece of of maybe foresight and certainly a prediction.
47:20
Unidentified speaker (speaker_3): We will see.
47:21
Unidentified speaker (speaker_1): We will see. Thank you very much for joining Kaifu Lee, author of AI Native.
47:25
Unidentified speaker (speaker_3): Thank you.
47:26
Unidentified speaker (UNKNOWN): If
47:27
Unidentified speaker (speaker_2): you're a holiday lover, you know London Drugs. You know you're gonna need decorations. You know you're gonna need snacks. And you know London Drugs has deals on everything you need to plan for every festive possibility.
47:40
Unidentified speaker (UNKNOWN): So
47:41
Unidentified speaker (speaker_2): start putting the no in November. Advent calendars? You know we've got 'em. A roaster for the perfect turkey? We know the one. A secret Santa gift for Phil? We know just the thing. Holiday lovers know, London Drugs.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from Brew Markets

Brew Markets

NVIDIA's China Loophole & The Aviation Blame Game

Episode discussion: NVIDIA faces China chip restrictions and leadership turmoil; FAA ground stop linked to L3Harris router failure

Speakers’ claims: Speakers discuss the US Commerce Department closing a loophole that allowed Chinese firms to access NVIDIA chips via third-country data centers, now requiring US licenses. They note NVIDIA's board dismissed a campaign to reinstate ousted CEO Jensen Huang, focusing instead on national security investigations. The episode covers a nationwide FAA ground stop caused by a router failure managed by L3Harris, with officials citing potential contractor incompetence. Blackstone Digital Infrastructure Trust's $5 billion acquisition of data center shells leased to AWS is highlighted as a major AI infrastructure investment. Market movements and other corporate news are briefly reviewed.

NVIDIASemiconductorsFAA
18 min5 assets9 claims to verify
Brew Markets

Jensen Huang Resigns & NVIDIA’s Long AI Shadow

Episode discussion: NVIDIA CEO Jensen Huang resigns amid scandal; interim leadership and AI financing risks discussed

Speakers’ claims: Hosts discuss the resignation of NVIDIA co-founder Jensen Huang following revelations of personal misconduct, noting the lack of a formal succession plan despite prior SEC warnings. Colette Kress is named interim CEO, with prediction markets favoring Ian Buck for the permanent role. Speakers highlight risks in the AI financing ecosystem, citing high bond yields for companies like CoreWeave and potential Fed rate hikes. The episode also covers broader market impacts, including declines in Corteva and Super Micro Computer, and political scrutiny over chip exports.

NVIDIAJensen HuangAI Financing
24 min4 assets10 claims to verify
All insights