This Former Google Director Just Revealed Everything... Self-Improving AI, Acquisitions and Google

What is discussed, and when

1. 0:00 – 10:00: online marketing, clickthrough rate, Interpretability, Joe Tonoski
2. 10:00 – 20:00: Interpretability, Reinforcement learning (RL), creatine supplementation, data centers
3. 20:00 – 30:00: AI alignment, doomers, deniers, and dreamers, AI preparedness teams, Helen Toner
4. 30:00 – 40:00: self-play, Workday, AI agent, proposer
5. 40:00 – 50:00: Alexander Wang, export controls, Huawei, motivated reasoning
6. 50:00 – 1:00:00: DeepSeek, hard drives, motherboards, Absolute Zero (paper)
7. 1:00:00 – 1:10:00: Windsurf, Windsurf, Cursor, Gemini
8. 1:10:00 – 1:20:00: Search, Windows, Windows, data centers
9. 1:20:00 – 1:30:00: Language model, Aerobic exercise, Avatar 4, Brian Johnson
10. 1:30:00 – 1:31:26: AI agent, Concur, Wes Roth, Workday

Notable quotes from "This Former Google Director Just Revealed Everything... Self-Improving AI, Acquisitions and Google"

• Too cuz it seems like the interpretability loses a little bit when we go scale up reinforcement learning.

  about Reinforcement learning (RL) · at 0:44

• Google, eight years in M&A mergers and acquisitions, and later worked at Slack and Salesforce, giving him an understanding of tech, business and strategy, mergers and…

  about Salesforce · at 2:33

• Who cares? All I know is I now can use these things to help me with 15 20% of my work today. So an eight hour workday might go down to a six hour workday so that I can go home…

  about Workday · at 22:52

• 2016 back then like Alpha Go to Alpha Zero. That whole idea of self-evolution, self-play, and then now we seem to be right. So we had superhuman AI that was superhuman at narot…

  about self-play · at 28:50

• And they're not getting that into their training data. And so when they acquired Windsurf, they can actually now say, okay, what developers are using this code that they're…

  about Windsurf · at 35:59

• Of different hype merchants. and I think what me and Joe are trying to talk about is motivate motivated reasoning for a lot of these folks like the people running the AI labs.…

  about motivated reasoning · at 38:07

• Clearly talking your book. It's in your interest for some Chinese lab not to undercut you the way that Deepseek did, for example.

  about DeepSeek · at 44:29

• Would love to invest in it because it's connected to Google I feel like Google's whole thing is you know search ads like one of the best business models ever people clicking on…

  about Search · at 1:13:07

Who and what is mentioned in this episode?

People

• Sam Altman — Person · 6 mentions
• Alexander Wang — Person · 4 mentions · first at 46:38
  Eric Schmidt. You might have heard about him. Oh, God. And so, he wrote that paper with Alexander Wang trying to compare GPU or country AI to nuclear war. That's a slap in the face of JFK.
• Wes — Person · 4 mentions
• Yan LeCun — Person · 4 mentions
• Chamath Palihapitiya — Person · 3 mentions
• Daario Amodei — Person · 3 mentions
• Lee Sedol — Person · 3 mentions
• Lex Fridman — Person · 3 mentions
• Sundar Pichai — Person · 3 mentions
• Wes Roth — Person · 2 mentions · first at 40:26
  Okay. So we are now in part two. we are joined here by the one, the only Wes Roth. Wes, thank you for joining us here on the show. Thank you for having me.
• Eric Schmidt — Person · 2 mentions · first at 46:33
  Why also, yeah, totally unfair. Totally unfair. So, there's a gentleman named Eric Schmidt. You might have heard about him. Oh, God. And so, he wrote that paper with Alexander…
• Balaji Srinivasan — Person · 2 mentions
• Blake Lemoine — Person · 2 mentions
• Demis Hassabis — Person · 2 mentions
• John F. Kennedy (JFK) — Person · 2 mentions
• Larry Page — Person · 2 mentions
• Palmer Luckey — Person · 2 mentions
• Peter Thiel — Person · 2 mentions
• Susan Wojcicki — Person · 2 mentions
• Joe Tonoski — Person · 1 mention · first at 1:32
  You want to skip the origin story of Wes and go straight to AI, start at 10 minutes 25 seconds. Joe Tonoski is an engineer and tech leader. He's a former engineer director at…
• Jordan Tibido — Person · 1 mention · first at 1:45
  Microsoft Splunk Wealthfront and Adobe. Jordan Thibido, am I saying the last name right? Jordan Tibido. It's a terrible name. Team I'm gonna keep it on the air. Tippido.…
• Ray Dalio — Person · 1 mention · first at 2:50
  Spearheaded Google's internal talks program interviewing luminaries like Ray Dalio and Tim Ferrris and so an Wajiki she was the WKEI it's a fun name Wajeki was she the YouTube CEO 23me CEO then your sister Susan ran YouTube and.
• Dr. Monica Gandhi — Person · 1 mention · first at 11:34
  Process, a deeper understanding. like in the sparks of a GI paper, one of them was like write a letter from Gandhi recommending a presidential candidate for the US for US…
• Helen Toner — Person · 1 mention · first at 20:19
  Doomers, deniers, and dreamers. And so you have the doomers who are the AI alignment folks. Helen Toner, Daniel King, whatever his last name was Joe.
• M Philip — Person · 1 mention · first at 53:10
  AI Explained is a channel on YouTube. They he did a Philip I think is the guy behind this. So, he did a great deep dive into the founder of DeepSeek.
• Professor Strover — Person · 1 mention · first at 57:25
  I was in political science, Santa University, and it's professor called Professor Strover. And when you're in school, you're always why am I doing this? And then you get one professor in one class, you're Okay, that was worth the price.
• Jay Cal — Person · 1 mention · first at 1:11:18
  Also people forget that Demis was on this week in Startups and they asked him Jay Cal was Oh, so like when did you go get bought by Google? And Demis said to train more of…
• Wolfgang Amadeus Mozart — Person · 1 mention · first at 1:22:24
  An creative title and it just from the get-go starts hitting you with interesting facts about Moz Mozart and the various writers and the various people like the fact that they…
• Brian Johnson — Person · 1 mention · first at 1:28:19
  Inside joke from last time or whatever that's where it seems to fall apart and most yeah we're still at that Brian Johnson moment where it's like uncanny valley like I want to live forever but then my face looks like a bishop from the alien.
• Arthur Samuel — Person · 1 mention
• Daniel Kokotajlo — Person · 1 mention
• David Sacks — Person · 1 mention
• Eddy Cue — Person · 1 mention
• George H. W. Bush (Bush Senior) — Person · 1 mention
• Isaac Asimov — Person · 1 mention
• Jalani Hussein — Person · 1 mention
• John and Jen — Person · 1 mention
• Jony Ive — Person · 1 mention
• Mark Benioff — Person · 1 mention
• Nikita Khrushchev — Person · 1 mention

Books & papers

• The True Believer — Book · 2 mentions · first at 23:42
  Open AI company and go work on AGI. A lot of them were getting laughed at. there's a book called True Believer. It talks about people who think that the current world that we…
• Absolute Zero (paper) — Book · 2 mentions · first at 32:18
  So I think the absolute zero paper was gang busters. And it's funny because it's a long line of these things, right? There's I don't know when it was like the 60s when Samuels built.
• Sparks of AGI — Book · 1 mention · first at 11:05
  Sparks of AGI paper. Sparks of the AGI paper. Yeah. and that's where it became really obvious to me. It's hard to explain, but basically, sometimes you give GPT 3.5 a certain.
• Star Trek — Book · 1 mention · first at 14:51
  Science fiction books in the past. Isaac Azimov, laws of robotics, Star Trek, etc. All of them describe artificial intelligence as something that humans engineered, something that they've invented and meticulously crafted. So, they describe.
• Alien — Book · 1 mention
• Superforecasters — Book · 1 mention

Companies & orgs

• DeepMind — Company · 8 mentions
• Windsurf — Company · 6 mentions · first at 36:02
  And they're not getting that into their training data. And so when they acquired Windsurf, they can actually now say, okay, what developers are using this code that they're…
• Workday — Company · 5 mentions · first at 22:56
• DeepSeek — Company · 5 mentions · first at 44:35
• Salesforce — Company · 4 mentions · first at 2:38
• Windows — Company · 4 mentions · first at 59:38
  And there'd be another huge wedge that was Microsoft because you had to pay a fat amount of money for Windows and Office. And remember, these are the wedges for profit, right?…
• Huawei — Company · 3 mentions · first at 45:01
  We're going to hold back those researchers because we have DeepSeek now. We have Huawei who needs them. We also have Alib Baba who needs them. The company made Quen 3. And…
• SemiAnalysis — Company · 3 mentions
• Department of Defense — Company · 2 mentions · first at 47:18
  Government believing the garbage he's talking about so they can get more department of defense contracts. Then Alexander Wang who scale AI they've been focusing on labeling…
• Scale AI — Company · 2 mentions · first at 47:20
  Talking about so they can get more department of defense contracts. Then Alexander Wang who scale AI they've been focusing on labeling data for these AI labs. Even Google uses…
• CCP (Chinese Communist Party) — Company · 2 mentions
• Cisco — Company · 2 mentions
• Slack — Company · 2 mentions
• Splunk — Company · 1 mention · first at 1:40
  Engineer director at Google and Facebook and has also held roles at Apple, Microsoft Splunk Wealthfront and Adobe. Jordan Thibido, am I saying the last name right? Jordan…
• Wealthfront — Company · 1 mention · first at 1:40
  Engineer director at Google and Facebook and has also held roles at Apple, Microsoft Splunk Wealthfront and Adobe. Jordan Thibido, am I saying the last name right? Jordan…
• Riverside — Company · 1 mention · first at 2:00
  Gotcha. It's out of frame, so I guess Riverside you can cut it in because they're recording our full There we Yeah. There we go. This is a Fox News alert. There's a minority in your neighborhood.
• Sequoia Capital — Company · 1 mention · first at 29:36
  We're applying it to LLMs. And this is where and actually one of the people at OpenAI they did a talk at the Sequoia Capital AI summit recently saying that the next big wave is…
• All-In Podcast — Company · 1 mention · first at 45:59
  Just that's bad. Now, we also harp about motivated reasoning is because on the All-In podcast, you'll see Chimath and David Sachs circle jerk about export controls and how…
• hedge funds — Company · 1 mention · first at 51:05
  What do you say? Like he's Oh, deep seek. What's really going on is the hedge funds actually put massive short positions on the US market that release this to then capitalize…
• AI Explained — Company · 1 mention · first at 53:06
  AI Explained is a channel on YouTube. They he did a Philip I think is the guy behind this. So, he did a great deep dive into the founder of DeepSeek.
• GitHub — Company · 1 mention · first at 1:05:54
  As I can tell, I don't really understand why Microsoft and especially with GitHub isn't right in the middle of this mix, right? Building a killer version of Visual Studio Code…
• This Week in Startups — Company · 1 mention · first at 1:11:15
  That YouTube worked. Exactly. And then also people forget that Demis was on this week in Startups and they asked him Jay Cal was Oh, so like when did you go get bought by…
• Isomorphic Labs — Company · 1 mention · first at 1:12:48
  I'm very curious because I'm so excited about some of the stuff that's coming out of Google Deep Mind. isomorphic labs I've been kind of following on that's very interesting…
• Healthline.com — Company · 1 mention · first at 1:23:58
  Citations. It's a little bit weak. I mean I get it but I could have went to whatever healthline.com and found that information. Google is No, like here's the And I forget like…
• Sesame AI — Company · 1 mention · first at 1:26:05
  And you mentioned Sesame AI and their voice, the empathetic voices.
• Concur — Company · 1 mention · first at 1:30:42
  Yeah, like dealing with workday or concur and things like that. No one wants to do that. Not even the people who build workday want to touch workday.
• 23andMe — Company · 1 mention
• DOJ (U.S. Department of Justice) — Company · 1 mention
• High-Flyer — Company · 1 mention
• Holiday Inn Express — Company · 1 mention
• SoftBank — Company · 1 mention
• Stability AI — Company · 1 mention

Products & tech

• ChatGPT — Product · 21 mentions
• GPU — Product · 9 mentions
• Grok — Product · 9 mentions
• Windsurf — Product · 6 mentions · first at 36:02
• GPUs — Product · 6 mentions
• Visual Studio Code (VS Code) — Product · 6 mentions
• Windows — Product · 4 mentions · first at 59:38
  And there'd be another huge wedge that was Microsoft because you had to pay a fat amount of money for Windows and Office. And remember, these are the wedges for profit, right?…
• Gemini — Product · 4 mentions · first at 1:01:25
  To control the whole entire hardware stack. Like right now Google's pushing Gemini through my down my throat 247 like Plus back in the day and it's frustrating. But OpenAI…
• Cursor — Product · 4 mentions · first at 1:02:17
  And so it ties to what Joe is saying. Instead of being at behest of Windsurf or behest of cursor, we want to be on the top of that stack and be to have that relationship with…
• Lambda — Product · 2 mentions · first at 26:18
  Mentally. we saw that with Blake Leone when he got freaked out with Lambda. Like everyone laughs now. It's like he saw Lambda back in 2020 and he was also a test engineer and…
• Sesame — Product · 2 mentions · first at 26:57
  Understandable now because it's so much more lifelike, especially the voice mode. I tried the sesame and wow, it feels very realistic, very lifelike. surprise. Yeah, it's hard…
• robots — Product · 2 mentions · first at 30:07
  And then Nvidia is doing the same thing with their Isaac gym training robots on simulation. But I'm yeah, if they if they're able to scale up reinforcement learning on these…
• hard drives — Product · 2 mentions · first at 59:22
  The hardware industry where they would show all the different vendors who made like hard drives and motherboards and cases and operating systems and apps and so on. And when…
• motherboards — Product · 2 mentions · first at 59:22
  The hardware industry where they would show all the different vendors who made like hard drives and motherboards and cases and operating systems and apps and so on. And when…
• Office — Product · 2 mentions · first at 59:41
  That was Microsoft because you had to pay a fat amount of money for Windows and Office. And remember, these are the wedges for profit, right? These are the people who kept a…
• Firebase Studio — Product · 2 mentions · first at 1:02:41
  Valued at 9 billion. Looks like Google is building their own kind of version of it, which is called the Firebase Studio, which I'm very excited about. I've it's not quite…
• Visual Studio — Product · 2 mentions · first at 1:04:36
  Isn't it kind of bizarre to have Microsoft create Visual Studio Code and open source it? And now there's a massive business competition around products that are based on their open source codebase. Yeah, it's it's incredible.
• Apps Script — Product · 2 mentions
• cases — Product · 2 mentions
• PC — Product · 2 mentions
• Qwen 3 — Product · 2 mentions
• Isaac Gym — Product · 1 mention · first at 30:04
  Environments and then Nvidia is doing the same thing with their Isaac gym training robots on simulation. But I'm yeah, if they if they're able to scale up reinforcement learning on these large language models to do coding where.
• supercomputers — Product · 1 mention · first at 45:16
  Harping on how they've failed. In the 90s, we were afraid of China creating their own supercomputers because of CPU use and we blocked them and then China caught up. And then…
• hardware device — Product · 1 mention · first at 1:01:10
  See Open AI right now that partnership with Johnny Ives to create some hardware device which I think is fraught because I'm carrying a cell phone around. Now people have…
• cell phone — Product · 1 mention · first at 1:01:13
  Partnership with Johnny Ives to create some hardware device which I think is fraught because I'm carrying a cell phone around. Now people have watches and then you want me to…
• Gemini 2.5 Pro — Product · 1 mention · first at 1:02:51
  Little bit clunky. It still has issues, but it's got that, it's got the Gemini 2.5 Pro integration and then it has an ability to very easily publish that stuff online. So if…
• Google 2.5 Pro — Product · 1 mention · first at 1:18:00
  More across various things openai when it came out recently been trying doing deep search with Google 2.5 pro and is just incredible I'm blown away by how is it as good as…
• Avatar 4 — Product · 1 mention · first at 1:27:50
  You know hey Jen is doing their avatar 4 or whatever it is. I feel like we're we're so close to just if we're talking about like the fooling the senses. I think we're very close to that in terms of long-term.
• Cisco communications equipment — Product · 1 mention
• CPU — Product · 1 mention
• Google Colab — Product · 1 mention
• Nvidia A100 — Product · 1 mention
• watch — Product · 1 mention

Concepts & topics

• Search — Concept · 8 mentions · first at 1:13:13
• Reinforcement learning (RL) — Concept · 5 mentions · first at 0:49
• AGI (Artificial General Intelligence) — Concept · 5 mentions
• GPT-4 — Concept · 5 mentions
• self-play — Concept · 4 mentions · first at 28:54
• motivated reasoning — Concept · 4 mentions · first at 38:11
• export controls — Concept · 4 mentions · first at 45:10
  Can have your GPUs. We're going to have a research scientist. Another point about export controls. We've been harping on how they've failed. In the 90s, we were afraid of China…
• Interpretability — Concept · 3 mentions · first at 0:46
  Excited about it. A little bit scared too cuz it seems like the interpretability loses a little bit when we go scale up reinforcement learning.
• online marketing — Concept · 3 mentions · first at 3:55
  Decade plus now. So I got started first of all on my own doing just doing some online marketing. Right around that time there was a person that I was following online that was very successful that was he was out in Texas.
• AI alignment — Concept · 3 mentions · first at 20:16
  Doomers, deniers, and dreamers. And so you have the doomers who are the AI alignment folks. Helen Toner, Daniel King, whatever his last name was Joe.
• AI agent — Concept · 3 mentions · first at 37:01
  Don't use Salesforce and I was because you're gonna and what are we going to replace it with? An AI agent just like that emod guy from was it stability?
• AlphaGo — Concept · 3 mentions
• Mergers and acquisitions (M&A) — Concept · 3 mentions
• clickthrough rate — Concept · 2 mentions · first at 9:01
  It has to have a high clickthrough rate.
• supervised fine-tuning — Concept · 2 mentions · first at 17:09
  Happens if we cut out the human data and like Deep Mind they put out a paper recently saying supervised fine-tuning is memorization. Reinforcement learning is generalization.…
• doomers, deniers, and dreamers — Concept · 2 mentions · first at 20:13
  Doomers, deniers, and dreamers. And so you have the doomers who are the AI alignment folks. Helen Toner, Daniel King, whatever his last name was Joe.
• proposer — Concept · 2 mentions · first at 31:05
  Supervised fine-tuning all RL and so what they did was they created two models and one is the proposer and it proposes various questions and problems etc. Then they have the…
• solver — Concept · 2 mentions · first at 31:13
  Proposes various questions and problems etc. Then they have the student or they call it the solver I guess and it solves all the problems.
• takeoff — Concept · 2 mentions · first at 32:03
  If that works within the next couple years, I feel like we're going to see a pretty rapid takeoff in coding. And if we don't, maybe there'll be a bit of a I don't know, AI…
• reward model — Concept · 2 mentions · first at 35:08
  It in the paper, but that's a that's a $64 million question, right? Getting that reward model correct, which ties into the wind surf acquisition for OpenAI. Okay, people are…
• jingoism — Concept · 2 mentions · first at 39:46
  Destroy us. so actually Joe that's actually a good segue. What's going on with this whole jingoism that's going on now that like all of a sudden Silicon Valley was peace and…
• nuclear war — Concept · 2 mentions · first at 46:41
  Him. Oh, God. And so, he wrote that paper with Alexander Wang trying to compare GPU or country AI to nuclear war. That's a slap in the face of JFK.
• Iraq War — Concept · 2 mentions · first at 49:28
  I one thing I talk about Joe is called the Dick Cheney effect. It was m I didn't support the Iraq war but I support the troops but it was master of all Dick Cheney did. He…
• Language model — Concept · 2 mentions · first at 1:27:13
  Give me that person speaking this script. Yeah. You can imagine coupling that with a language model would allow for it to engage in a conversation.
• AlphaZero — Concept · 2 mentions
• Go — Concept · 2 mentions
• JEPA — Concept · 2 mentions
• NDA — Concept · 2 mentions
• P(doom) score — Concept · 2 mentions
• Proto-AGI / proto-AGI claim — Concept · 2 mentions
• viewability — Concept · 1 mention · first at 9:23
  Upload three different images and whichever has the highest combination of clickthrough rate plus viewability, that's the one that's shown. So a lot of people, they'll look at…
• GPT-3.5 — Concept · 1 mention · first at 11:18
  Became really obvious to me. It's hard to explain, but basically, sometimes you give GPT 3.5 a certain prompt and you get that it kind of gets what needs to be said, but it doesn't really thinking about it too deeply.
• proton — Concept · 1 mention · first at 11:45
  Recommending a presidential candidate for the US for US president and that candidate is a subatomic particle proton a subatomic particle, right? So you're like what does that…
• creatine supplementation — Concept · 1 mention · first at 13:19
  He talks about stuff. So in the fitness community, they'll do some deep dive into the study about how creatine supplementation helps muscle contraction, right? So I've done…
• neural nets — Concept · 1 mention · first at 14:30
  This stuff because some of the concepts behind machine learning and neural nets and stuff like that, it just seems so wild and so different like we've discovered something that's fairly so new in life. like for example,.
• memorization — Concept · 1 mention · first at 17:12
  And like Deep Mind they put out a paper recently saying supervised fine-tuning is memorization. Reinforcement learning is generalization. So when we give it human data, it…
• generalization — Concept · 1 mention · first at 17:15
  Recently saying supervised fine-tuning is memorization. Reinforcement learning is generalization. So when we give it human data, it learns to parrot us and it's good. It's like…
• AI preparedness teams — Concept · 1 mention · first at 20:37
  Say, oh, you're saying AI people are doomers, but there's to distinguish it, there's the AI preparedness teams at OpenAI who do real work who make it so that people can't go on…
• red teaming — Concept · 1 mention · first at 20:50
  Catfish kids, groom people or play along with suicidal thoughts and things like that. So that's valid red teaming no one's against. But then you had the AI alignment people who…
• nuke codes — Concept · 1 mention · first at 24:07
  Then they start getting on the wacky sauce and start they're going to get the nuke codes and they're going to destroy everything and it's like hold on time out. Have you talked…

Places & events

• Silicon Valley — Place · 4 mentions · first at 39:48
  Actually a good segue. What's going on with this whole jingoism that's going on now that like all of a sudden Silicon Valley was peace and doves and now we want to go to war? I…
• data centers — Place · 2 mentions · first at 18:58
  Principal engineers, whatever. And you're responsible for getting OpenAI's models to work in your data centers.
• missile silos — Place · 1 mention · first at 24:12
  Everything and it's like hold on time out. Have you talked to anyone who actually worked in missile silos? My uncle actually worked in the silos and a lot of those things were…
• Santa University — Place · 1 mention · first at 57:22
  You think that might be a strategy that's being employed? so I once I was in political science, Santa University, and it's professor called Professor Strover. And when you're in school, you're always why am I doing this? And then you.
• shed — Place · 1 mention

Full transcript with timestamps

0:00 Peter Teal told Sam Alman years ago, um, I am not funding you anymore. I kicked him out of my program and you need to get these EA people out of OpenAI now or you're going to regret it. And Sam was like, uh, well, you know, politics and things like that. Yeah. He gave him a clear warning. He and then also when Joe and I are working at Google, when you're inside these companies, you can't just get on the mic and just start talking about tech because that will violate your NDA or get in trouble. So instead, you get to see Lex Freeman and other people who've never been inside of the corporate operation or been close to how decisions are made just pontificate on crap. Our show is very like let's just let's cut away from the We're replicating that exact thing with large language models now. And it's like that's the next big wave and I'm very excited about it. A little bit scared too cuz it seems like the

0:46 interpretability loses a little bit when we go scale up reinforcement learning. This goes back to where you started with we're not really designing these things. Uh, and the flip side of that is even after they've grown or developed or however you frame it, we can't really understand how they work. And you've got a bunch of engineers and you're like, "Have you taken this thing apart and figured out how it works inside?" And they're like, "We took it apart. We put it back together again. We still don't know how it works." What the hell? Yep.

1:14 Awesome. We're live. As in, we're recording, not live stream. Yes. If I if I screw up, we're cutting it out, right? Yes, of course. See if you can spot the moment Wes realizes that nothing is being cut. If you want to skip the origin story of Wes and go straight to AI, start at 10 minutes 25 seconds. Joe Tonoski is an engineer and tech leader. He's a former engineer director at Google and Facebook and has also held roles at Apple, Microsoft Splunk Wealthfront and Adobe. Jordan Thibido, am I saying the last name right? Jordan Tibido. It's a terrible name. Team I'm gonna keep it on the air. Tippido. Yeahido, which is French Canadian garbage. So, Jordan, that's why I have the American flag over here, just to balance things out.

1:57 Gotcha. It's out of frame, so I guess I guess uh Riverside you can cut it in because they're recording our full There we Yeah. There we go. This is a Fox News alert. There's a minority in your neighborhood. Go. All right. All right. Let me Let me uh let me let me deliver the live without laughing here. All right. Jordan the final we'll make the final cut. I want to end the show if we can go for it. Go. What my Oh, that. Yeah. Yeah. Definitely out takes the out takes, you know. Love it. Love it. Jordan Jordan Tibetto focuses on business and human side of tech. Spent 10 plus years at Google, eight years in M&A mergers and acquisitions, and later worked at Slack and Salesforce, giving him an understanding of tech, business and strategy, mergers and acquisitions, and organizational dynamics. He also spearheaded Google's internal talks program uh interviewing luminaries like Ray Dalio and Tim Ferrris and uh so an

2:54 Wajiki she was the WKEI it's a fun name Wajeki was she the YouTube CEO 23me CEO then your sister Susan ran YouTube and unfortunately Susan passed away uh rest in peace and um yeah Susan was the one that's like hey go buy YouTube which is one of the greatest acquisitions of all Goodbye. Yeah. Okay. Yeah. Very cool. So, I might have to Okay. So, I might have to um do uh like an intro for the for for my side of it later. Um well, we we'll figure it out later, but I just wanted to kind of make sure that sounds good. It's correct. Just use it. It's good. Yeah, it's good. We're we're you can tell we filled things in. I'm recording this with the meth lab, dude.

3:38 And don't leave the comments. Let's get into it. How did you start into like what's your background and what got you into YouTube? Yeah, so I've been um you know doing stuff on the internet for a decade plus now. So I got started first of all on my own doing just doing some online marketing. Right around that time um there was a person that I was following online that was very successful that was he was out in Texas. I was in California and he was launching a new startup in the e-commerce space and he kind of reached out and he was like, "Hey, if you're interested, well, to his audience, so to speak, he's like, we're looking for somebody to come join us doing email marketing." So, this was like 2013. So, this was some time ago.

4:23 Um, and I was like, can I, is this crazy, just pick up and move to some, you know, move across the country from California to Texas and join some startup? But I was like, Let's see. Let me try it. So, I reached out, did a couple of interviews with them, they liked me, joined the company. I was employee number three and they scaled that thing. Yeah. They scaled that thing from zero to at the end of the year one, they were like at 20 million uh plus. They they they end up getting to 33 million um a year at some I I believe. I mean, it was a lot, tens of millions.

4:57 And how many people? So, I was employee number three. Uh, so they had the two owners, then one of the owners hired their brother-in-law, and there was the the director of marketing. So it's like I was the fifth person, but I was employee number three, I guess, technically. Yeah. And yeah, I was just studying online marketing under this guy that just he was in charge of like a million. He was spending like a million a month on ads, right? And I was just sitting there in his room just taking notes. I'm like, "All right." Mhm. And doing email marketing, doing all that stuff, doing split tests, like all the stuff that you do for for that. Um, so I did that for a while. Eventually in 2016, 2015, one of the companies that I was working for, they started it, they had some internal issues, lawsuits, uh, oh yeah. Well, so the the end of that story is the director of marketing one day had a huge problem with the main guy

5:51 into a huge fight. He storms out. He's like, "Screw you guys. I'm going home." And uh and so they kind of turned to me. I've been there for two months at that time. And they're like, "Well, you can do his job, right? I don't even have. Sure. Why not?" I slipped at the Holiday Express. I'm good. Let's go. Yeah. I'm refreshed. I'm ready. Um so I say yes. um which I mean wasn't there yet in terms of like just rationally it should have been that that strong of a yes but I I picked up what he was doing and quickly just learned a lot. So the next few months I just learned a lot about e-commerce online marketing just all the tech behind it how they were building it because at the time there was a lot of tech that went into it with credit cards and hosting and just just everything. So it's not super complicated stuff. It's not rocket science, but they're just like a lot of little things that you

6:46 have to know. Um, start and anyways, fast forward, started my own e-commerce in 2015, 2016. Did pretty well with that. That took me from kind of like having to work a job to maybe being a little bit more, you know, not rich, but you know, at least I don't have to worry about that as much. And then so 20 whenever Chad GPT, so that was the end of 2022, I think, right? So, uh, Elon Musk goes Chad GPT is getting scary good. I'm like, what is this Chad GPT thing? I look at it and that's when it just started coming together. That's I was like, I got to go in all in on this thing. I didn't know what that meant initially. I thought just incorporated into e-commerce because I was like, dude, this is first of all, it automates customer service and I thought all the other stuff, but I didn't know what I was going to do, but I knew I had to be doing something in this arena, in this field. So that's kind of my start. Oh, I

7:41 yeah, put a few YouTube videos out there. They started getting traction and I really just enjoyed it because I was able to learn about it. Like I just sat there and read all the the papers that they the people put out, you know, Microsoft and OpenAI and well OpenAI didn't have as many papers, but you know, all the other companies just sat there reading it, highlighting it, and then posting the results on YouTube. And people liked it. And I was like, this is I'm getting paid to learn this stuff. And that was just like a dream come true. Let's start the genesis of Yeah. Oh boy. Oh yeah. Yeah.

8:18 This is you know I'm so happy that the most controversial thing about my channel is um you know the faces that I make in in in the thumbnails. Um so I think for a lot of people what maybe not is obvious on the surface is if you think about how YouTube is built, it's very different from the other platforms. in that most of the other social platforms the algorithm serves you the content. So on Twitter it shows you the content on you know Instagram it shows you the content on YouTube it doesn't. It shows you whatever the creator chose to wrap it in the thumbnail and the title. Um and so in order for people to see the content that has to be killer.

9:01 It has to have a high clickthrough rate. It has to have well yeah I mean for that it's mainly the clickthrough rate. Then once they click on the content, they have to stick. So people sometimes say like click and stick. So you click on the content, you stick with the or you click on the thumbnail, you stick with the content. Um, and now there's more and more tools. So now we're allowed to upload three different images and whichever has the highest combination of clickthrough rate plus viewability, that's the one that's shown. So a lot of people, they'll look at my video feed, it's always like that face, right? And they're like, "Um, why do you use the same thumbnail in every single one of these images?" It tests. Well, you guys keep clicking on it. Yeah, exactly.

9:40 That's why I always tell I'm like that. It's not true. I always test three different ones and like my goal is to find one that beats it, but all of the collective you, all the people watching it, you always click on that face for some reason. So, you know, don't blame me. You guys click. You know, the game's the game and I have to delete it. never never changed because now if you change people are going to be like what happened to the old thumbnails guy selling us out. Yeah. Like I our video that helped kind of get us going was we did a video on opening has no moat and then I ran it and then I said, "Hey, you know what? I'll clip this into like 10 minutes and I was too lazy to do a thumbnail and it showed my old background of my shed which people think it's messy which it isn't. It's very clean. I think so." And I had my hands up like this and I was like, "Ah, whatever." And that was one of our

10:26 biggest videos of all. Right. Right. No. Yep. Yeah. So, no, I I hear you on that. Um, awesome. Um, I want to know how you thought when you first played with Chad GPT because I'm sure that's uh a lot of people must have had that same experience, but I'm curious what went through your mind. Yeah. Um, I mean, Chad GPT was good and you definitely felt that there was something there. It wasn't quite mind-blowing to me. What really like hammered it in is when GPT4 came out, Microsoft released the proto, you know, it said it was like a protoagi GPT4 was proto AAGI as they called it.

11:05 Um, sparks of AGI paper. Sparks of the AGI paper. Yeah. Um, and that's where it became really obvious to me. It's hard to explain, but basically, you know, sometimes you give GPT 3.5 a certain prompt and you get that it kind of gets what needs to be said, but it doesn't really thinking about it too deeply. GPT4, you can see a much deeper like process, a deeper understanding. like in the sparks of a GI paper, one of them was like write a letter from Gandhi uh recommending a presidential candidate for the US for US president and that candidate is a subatomic particle proton a subatomic particle, right? So you're like what does that even mean? You have to kind of stop and think about what would Gandhi say about that like none of it really makes sense on the surface.

11:57 And so GPT3 is kind of like a, you know, kind of like a kid that didn't do the homework but kind of knows what needs to be said. He's like, "Oh, you know, George Washington was a great president. I'm so proud of him. He was great." Blah, blah, blah. And then GPT4 kind of really breaks it down and really captures these abstract concepts and is able to articulate it and you're like, "Okay, so we just jump from like a kid that knows how to speak English that didn't do his homework, but kind of knows what needs to be said to something that's more advanced." And so I was like, "All right, so just project this forward. This seems like this keeps going, keeps improving. We're going to see some wild stuff." So that was my Chad GPT moment really is just side by side comparing 3.5 and four.

12:38 Interesting. And you were reading all these papers uh I don't know about all all academic background and your experience reading other uh fields of science, but these papers are pretty damn obscure. I mean, there's a lot of whacked out notation and mixtures of statistics and computer science. How did you like what it must have been something about it that was compelling enough for you to slog through all this stuff, you know? I don't know because I haven't had Well, I guess in the past I was interested in like the various workouts.

13:13 So, I follow Dr. Mike and you know how he talks about stuff. So in the in the fitness community, they'll do some deep dive into the study about how creatine supplementation helps muscle contraction, right? So I I've done that a little bit listening to uh people on YouTube uh go over various studies in the fitness community, muscle growth, stuff like that. So something about that I always I say like oh it's like this ADHD thing where you're like uninterested in 99% of the things in the world and then like the one thing that triggers some dopamine response you're just like obsessed with it. So I was I got obsessed with it and I can't do that for anything else. It's only what my brain like really latches on to but that thing I can stick with. So um I I would not be in the past I did it with like various fitness research or like you know health and fitness stuff like that.

14:06 Now it switched to um AI machine learning and but I I could not do that for for anything else that I wasn't like really super focused on. Um and I was just trying to understand these concepts because I feel like it just felt so wild. And I kind of wanted to also, you know, get your guys' take on some of this stuff because some of the concepts behind machine learning and neural nets and stuff like that, it just seems so wild and so different like we've discovered something that's fairly so new in life. Um, like for example, the fact that, you know, in all the science fiction, I'm sure we've all read science fiction books in the past. Isaac Azimov, you know, laws of robotics, you know, Star Trek, etc. All of them describe artificial intelligence as something that humans engineered, something that they've invented and meticulously crafted. So, they describe it, you know what I mean? And now it's

15:04 almost seems like, and let me know if you guys agree with this. It doesn't really seem like we're we're building it as much as we're growing it. Like we're creating an environment like almost like bacteria petri dish. Like if we create the right environment, they multiply. We're not inventing it or whatever. Is does that make sense? Yeah. I mean, well, I would say we're not at the detailed level. We're not designing it. It's not like the way you would design a car, right? You know, you're not making sure all the parts fit together the right way and so on. It's more like you're growing a garden. You sort of set up the environment to be right. Mhm. Put the seeds in and then hope that everything works out. and when you come back it's going to be a good garden, right? You're going to it's going to produce whatever vegetables you planted.

15:51 I feel like that's more of the analogy when I see these experiments. Um, I I also think to your point, it's fascinating how we started out with such tentative results, like so small and weak and it was barely working and then we kind of climbed this this hill where every every paper it just got a little better. And to your point about first seeing GPT3, five, and then four, you're like, okay, each paper it seems like they're making substantial progress. This thing is getting a little more respectable. a little more capable.

16:26 And then they started kind of in my mind kind of cutting out the supports like we're gonna get rid of all these human examples and we're just going to have this thing do selfplay and learn on its own. So now the environment you set up to train became even more important and the amount of effort you put into directly designing it became relatively less. Yeah. And it seems like when they cut out the human data, that's when it really takes off and something just next level begins to happen. And we're seeing that I mean Deep Mind's publishing about it, Deep Seek R10 was an incredible read because that's exactly like okay, what happens if we cut out the human data uh and like Deep Mind they put out a paper recently saying supervised fine-tuning is memorization. Reinforcement learning is generalization. So when we give it human data, it learns to parrot us and it's good. It's like as as as good as

17:23 the best humans like Alpha Go beat Lisa Dole, it was as as good as a world champion. Then the next level when we do, you know, selfplay and all that stuff. Um all of a sudden that one beats the model that beat Lisa Dole like 100% of the time. The one that learned bad habits from all those human games. Yeah. Exactly. So, and and it seems like that's we're replicating that exact thing with large language models now. And it's like that's the next big wave. And I'm very excited about it. A little bit scared too because it seems like we the interpretability loses a little bit when we go scale up reinforcement learning because we don't get what's happening as well. So, I think we're just we're like totally lost on interpretability, you know? we just we have so little idea. This goes back to where you started with we're not really designing these things. Uh and the flip side of that is even after they've grown

18:24 or developed or however you frame it, we can't really understand how they work, you know, even when they're right in front of us. It's like imagine having, I don't know, some complex machine standing right in front of you and you've got a bunch of engineers and you're like, "Have you taken this thing apart and figured out how it works inside?" and they're like, "We took it apart. We put it back together again. We still don't know how it works." Yeah. What the hell? Yeah. Imagine you're working at Microsoft right now. You're like one of the senior, you know, principal engineers, whatever. And you're responsible for getting OpenAI's models to work in your data centers.

19:01 And then the CEO is like, "Okay, do you understand how they work now? Can we start building our own and quit having to pay these guys?" No, we still don't understand how they actually work. The best we can do is train a small one ourselves. Yeah, it it's so new because if any other technology, there's some engineer somewhere in the world, alive or dead, that at some point knew how to make it, how to create it, how it works. Um, nothing was where just no human being ever knew exactly what went into it. And this is kind of different where it's like we don't really know what's happening. And you know Dario Amade recently and anthropically posted a paper where he's saying that they're making some strides in terms of interpretability and understanding what neurons and clusters of neurons do what.

19:48 And they've posted some fascinating research about that. But he's saying like yeah he believes in five years we'll have you know made some great progress into understanding this stuff. But he's like at the rate at which AI progress is going that's going to be way too late. So it's like, yeah, we're making some progress, but not nearly fast enough. So yeah, I also think there's like three classes of folks right now that are getting most of media attention. Doomers, deniers, and dreamers. And so you have the doomers who are the AI AI alignment folks. Helen Toner, Daniel King, whatever his last name was Joe.

20:25 Good. Yeah, thank you. It just flows off the tongue. Then you have Yan Lanky or like Tibido, right? And there's people say, you know, oh, you're saying AI people are doomers, but there's to distinguish it, there's the AI preparedness teams at OpenAI who do real work who make it so that people can't go on there and try to get this model, the catfish uh kids, groom people or play along with suicidal thoughts and things like that. So that's valid red teaming no one's against. But then you had the AI alignment people who are thinking about in 15 20 years now, what will these systems be? And so let's focus on that cluster of all our assumptions and then try to figure out a way to confuse everyone with a modern system that can't do a regular task by itself uh for a decent amount of time while it's self-destructing. So for our show, we've been focusing on is uh how do we be pragmatic here about the AI systems we

21:19 have today? And then also when Joe and I are working at Google, when you're inside these companies, you can't just get on the mic and just start talking about tech because that will violate your NDA or get in trouble. So instead, you get to see Lex Freeman and other people who've never been inside the corporate operation or been close to how decisions are made just pontificate on crap. And so our show is very like let's just let's cut away from the And so you have the doomers here. And then you also have the deniers um who are folks who are just no matter what this tech sucks it's not going anywhere I don't want to use it blah blah blah I remember when chat GPT had its moment and I was over at Salesforce the current VP of AI was like these LLMs are overrated and it sent an internal Slack message these are overrated and you know our methods that we're using are probably better and blah blah blah and

22:07 then with like a few weeks Mark Benov had a talk with him and all of a sudden he magically got the religion and said oh no actually Chad GPT is pretty good and blah blah blah. And then you have like the young lun types who um brilliant, did a lot of great things for black back propagations, super smart dude, but he has a view on where he thinks AI should be going and his jeepa thing and whatnot, but is does not have the political understanding of, hey, right now it's LLM's moments. They're getting attention. This is the first time the AI has been put in the hands of the ley, the peasants who don't know how to code, who can actually get it to do stuff like write copy for us or help us with stuff and let's recognize that and celebrate that. But he's more focused on yes, this is an AGI and people are like, who cares? All I know is I now can use these things to help me with 15 20% of my work today. So an eight hour workday

22:58 might go down to a six hour workday so that I can go home and be with my kids or do things like that. Yeah. Um so you got that camp and then the last one is um the dreamers who are focusing on AGI tomorrow. And so those are a lot of the folks um you had uh people who are in open AI who also are alignment folks who are true believers in AI to the point of they thought they believe that this they have a missionic zeal that what they're doing is going to lead to AGI and a computer that can think consciously and help us evolve and whatnot. And you had to have that belief back in the day if you were going to say, "I'm leaving Google where I'm an AI researcher getting paid millions to go join this open AI company and go work on AGI." A lot of them were getting laughed at. Um, there's a book called True Believer. It talks about people who think that the current world that we have right now is flawed and the only way for us to

23:50 improve upon ourselves is push for a social mission or change that will get us to the promised land. Right? So I call them AI true believers and those are great people. They were right that these LLMs are going to scale up but then they start getting on the wacky sauce and start they're going to get the nuke codes and they're going to destroy everything and it's like hold on time out. Have you talked to anyone who actually worked in missile silos? My my uncle actually worked in the silos and a lot of those things were all mechanical and he actually had a person who was on the other side of the wall with a little little gun sight so that that person didn't pull a truck to fire nukes. He had a pull gun and say, you know, you shoot, you die. So, I never know any governments that were going to say, hey, let's connect these LLMs and elucidate sometimes to launch the nukes. Yeah. So, our show's been trying to be like, hey,

24:35 pro AI, but at the same time, these AGI people are smoking too much crack and these doomers getting a little bit out of hand and these deniers need to kind of get on the train and say, hey, you know, these LM are actually doing something. Yep. So, um, that's kind of my thoughts on things. Wes, Joe, I'd love to hear what your thoughts are. doomers, deniers, and dreamers. Yeah, I mean this everything you said and um I'm 100% on board with with everything. So uh AI doomers you know again a lot of them are very very smart people and a lot of the arguments they present are rooted knowledge of the research into AI and uh so when they when they talk about kind of like concrete examples of problems they see and why it's so difficult they seem very like they know what they're you know they they seem like I agree with a lot of the stuff that they say like yes these are difficult problems yes we have not

25:24 solved these problems and they're like therefore we're all going to die in five years I'm like like badly. I think you lost me. That's where you you lost me. Oh, anytime I hear people say PDUM score, I instantly check out mentally. I just Yeah. Oh, my PUM score is 20 30%. Based on what? And then why was it 15% to 20% like like changes? There's a book called Super Forecasters which goes into the best forecasters in the world and how they actually make those exactly and a lot of time people are just with your fingers in the air and 25% 20 and the reason why I take offense to it is when you're in the corporate setting and you're actually trying to get things done and move things forward, you got these people playing this devil's advocate game of I am just going to put the red flag up and prevent any technology from being launched to the detriment of the org because of my own insecur. ities

26:15 mentally. Um, we saw that with Blake Leone when he got freaked out with Lambda. Like everyone laughs now. It's like he saw Lambda back in 2020 and he was also a test engineer and and everyone at Google explained to you how this thing worked, but he thought it was conscious. There's something here. It's working. We look at trivial now. But in the day, those people can make those arguments that just completely derail your company. But sorry, Wes, I I took over. Go for it. Yeah, that would that model that was like GPT2 level, right? Yeah. roughly. Okay. Okay. Yeah, that's that's interesting. Um, yeah, I I mean I could I guess I could I mean, still some people kind of I think believe that, but it's I guess more more understandable now because it's so much more lifelike, especially the voice mode. I tried the sesame and wow, it it feels very realistic, very lifelike. Um, surprise. Yeah, it's hard to tell what's

27:06 what's real, what's what's not um just based on interacting with it. Um, let's see. So, yeah. So AI doomers again I have some respect for their sort of thought process but the conclusions I just can't get on on board with like that that leap to everybody dies just doesn't make really sense to me. deniers. Um, deniers. Yes. So, a lot of people, it's strange to me how many people are denying the progress. Jan lun or Lan, how would they pronounce that, you know, I watched his um recent talk where he describes Japa. So, his kind of approach to it definitely sounds like it makes a lot of sense. Sounds more like what we do because we're not thinking in terms of words in words out, right? So we have certain encoded abstract concepts and we can learn about them through words or images or sound or whatever. And so we can all think of what a dog is. We might have all kind of a different representation but we can

28:04 all kind of like say what it is described. We can draw it. We can make a sound like whatever. But the point is like it's not like the same tokens in or out. There's some sort of an abstract representation. And I think that's what Jeepa is as far as I can understand. It certainly makes sense to me. But why the hate against LLMs? I don't I don't get that. I think it's it seems to be working. Okay. Like you said, it's the LLM moment. Let's kind of figure out where this thing how far this thing can takes take us. But I guess it's good that he's maybe thinking ahead because 10 years from now, we might be saying, well, good thing somebody was working on that. But yeah. Um, but the people that are saying that it's not going to progress past this is a little bit weird to me because what I'm seeing is we saw massive progress with 2016 back then like Alpha Go to Alpha Zero. That whole idea of self

28:54 self-evolution, self-play, and then now we seem to be right. So we we had superhuman AI that was superhuman at narot tasks. Then we got this LMS that's a little bit more general, right? It's not super human really, but it's generally quote unquote intelligent. It's able to whatever you throw at, it'll attempt to do it. It'll be like, I'll try this, I'll try that. Sometimes it'll do well, sometimes it'll do great. And now I think we're those two sort of different AI uh technology trees or whatever, they're kind of um converging in that we're we're we're taking the stuff that we've learned from Alpha Go, Alpha Zero, and we're applying it to LLMs. And this is where and actually one of the people at OpenAI they did a talk at the Sequoia Capital AI summit recently saying that the next big wave is going to be how we had um you know throwing compute at pre-training then you know so we have test time compute or

29:52 training time train time compute test time compute the next big wave is RL compute so throwing tons of compute at those reinforcement learning environments Um, and then Nvidia is doing the same thing with their Isaac gym training robots on simulation. But I'm like, yeah, if if they if they're able to scale up reinforcement learning on these large language models to do coding where it teaches itself, which they're beginning to do that and it's showing promising things and they're able to scale that up. That seems like at that point I feel like we're going to see some incredible progress really really fast. Um or coding is a is a big use case that'll unlock a bunch of other things. Yeah. Well, so I just uh on my my channel just covered a video. I forget where it's from. I think there's a few Chinese researchers and in collab with US researchers, but what they did is they create exactly that. So they're

30:48 like how do we do it's called absolute reasoner. So how do we train these AI models? You know, we get the base model. How do we train them to do coding without any human data with no supervised fine-tuning all RL and uh so what they did was they created two models and one is the proposer and it proposes various questions and problems etc. Then they have the um the the student or they call it the the solver I guess and it it solves all the problems. And you know I was looking at that paper it looks really smart. they really captured something. How do you take how do you have it self-play? So you have the teacher and the Yeah. So and and it seems like it has good results. So if they're able to take that apply to bigger models, scale up reinforcement learning, it seems like they might get these models to be able to be incredible coders because it seems to like that approach. It generalizes.

31:47 It doesn't just memorize how to do problem X or problem Y. It's sort of like, okay, so I've seen these problems here and I can now solve all of these problems I haven't seen before. So it generalizes to the unseen data. If that works within the next couple years, I feel like we're going to see a a pretty rapid takeoff in coding. And if we don't, maybe maybe there'll be a bit of a I don't know, AI winter or something where it kind of you know what I mean? So I don't know. Happened before a couple times. Yeah. Yeah. Joe Gilfort.

32:18 Uh, so I think the the absolute zero paper was gang busters. And it's funny because it's a long line of these things, right? There's I don't know when it was like the 60s when Samuels built the the checkers player that selfplayed until he could beat anyone in checkers. And then there was Toro and and the back gammon player that did the same thing. Then there was the deep blue team doing chess. And then we have Alph Go doing go because that was considered unsolvable because there's so it's such a large board, so many combinations of moves.

32:52 And then Silver and his team decide, you know what? If we didn't train it on any human data, we make it only learn through self-play. And you get Alpha Zero and Alpha Zero Moo. And the same engine can play chess, go and shogi, I think. Yeah. Um, that's crazy. And it and it becomes superhuman in play at all three games, the same architecture, right? And then these absolute zero guys are like, as you described, we'll we'll do this proposer and this solver, and we'll try and get more and more difficult coding challenges to be proposed, and then the solver will get better and better at solving more and more difficult problems, and then we'll just run another round of RL on those training examples. And now the proposer is smarter, and so he proposes even more difficult problems. Now we're just going to eat our own tail in a loop until this thing breaks out of a glass ceiling, you know, and flies around the city or some Willy

33:46 Wonka right? Anyway, my point is that that algorithm and that self-reinforcing nature, that's gangbusters and it's proven to be gang busters before. The question I had, and probably you thought the same thing when you're looking at the papers, how will you apply this to real world problems? Like if you understand how to do it in a game, you know, a two-player game, total information on the board, rules are really simple. I see how you built a selfplay and I can even probably design one myself if I had to, but how to do that on like math problems or reasoning in a legal case. I have no sense of how to set up a self-play game there until I saw two things in that paper. One is that they figured out how to do it for code. Sort of makes sense because you can evaluate the code and you know know for sure if it ran or not and it gave the right answer or not. But also as they got their system better and

34:43 better at coding, it also became better at solving math problems, right? And they weren't training it on math problems, only on coding. I'm like, wait a minute. Yeah, just getting better at coding makes this thing better at math. I wonder if it's getting better at generic reasoning. And I don't know the answer yet. I don't think they got into it in the paper, but that's a that's a $64 million question, right? Getting that reward model correct, which ties into the wind surf acquisition for OpenAI. Okay, people are generating code and like I was doing. So, I got into chat GPT because at Google I was in HR, which is like the proletariat the company worked in M&A. Anytime I want to get those bougie Joe people to come through and code something, oh, I care busier than me. And so I built uh an appcript to help just automate certain processes and it took me like six months to learn appcript. It was

35:35 just painful. And I was chat GPT and I basically was able to do the same script in about like 30 seconds. And I was just like oh my god. But what I was doing before was I would copy copy code from chat GPT then put it into an IDE play it see if it works and give the error process. And a lot of people were doing it manually. But the issue is for OpenAI is they're not seeing what works for me and they're not getting that into their training data. And so when they acquired Windsurf, they can actually now say, okay, what developers are using this code that they're that we're we're committing which is not working. So they can get a re reward signal from that and then help train their models making it even better. And I think it's always that's a missing aspect. Um, that's why when I see a lot of channels talking about you AGI tomorrow, it's going to take overall project planning and things like that. It's like, okay, how do you

36:24 get a reward model to a uh to train an LLM? Okay, our original goals for this long-term project for X, but then reality changes and we're going to go to Y. Actually, CEO just came in and saying there's different politics, so we're going to Z. Like those things are some things that we haven't we haven't gotten to yet. And so that's why I think I try to narrowly focus on like what can thing what can we get done today. Like for instance um the CLA CEO was coming out a year and a half ago and was saying yeah we don't use workday anymore and we don't use Salesforce and I was because you're gonna and what are we going to replace it with? An AI agent just like um that emod guy from was it stability?

37:08 He like his eyes when you watch him talk it's like the engine on but no one's behind the wheel like what's going on there and someone told him like yeah work day it's gonna get replaced by an AI agent and I'm like time out uh do you think any employees going to be happy when that AI agent hallucinates 70% of the time and doesn't give them the right pay like do you you know how furious people are going to be in that situation and do you also know that workday is just not uh it's a system of record for also Sarbings and oxallay and other socks compliance. So, your companies get sued by the federal government. So, we just had a news update today that Clara CEO um Clarina came out again today and basically was like, "Yeah, actually, we're going to start hiring humans again and we basically overstayed the fact that we didn't just deprecate workday and and Salesforce. We actually just went to different vendors in the first

37:59 place." So I think one issue we deal with too in the space is just the hype gets gets ahead of ourselves and a lot of different hype merchants. Um and I think what me and Joe are trying to talk about is motivate motivated reasoning for a lot of these folks like the people running the AI labs. You want to hype it so that you get um funding for your next round. you get top AI researchers and you want people to make them think that you have secret sauce with we have AGI actually and you're only going to get access to it if you come work with us and so they hype it there so that's their interest and that gets people into a fervor because they now think oh all engineers will be replaced by tomorrow uh all jobs are gone um but then the problem is when they play with that fire too much then they wake up regulators like in California we had California almost pass a bill last year that was going to say hey we need third party

38:51 verification to release any AI models. So a third party vendor would have to come in look at your coding stack and if they raised any red flag, you would have to jump jump through the hoops. Now what if we get some person that believes in the um I forgot what the fallacy is called here. I have my notes here. Sorry. Hold please do the perfect solution fallacy which AI alignment people believe in that before you launch any system there's a way to know every possible externality. what could go wrong. And anyone who's ever had uh either taken care of your parents or raised a kid knows that's impossible.

39:27 You just can't do that. Um so, but what if one of the persons behind the wheel and says, "We're not going to allow you to launch any of these systems anymore." Then we're going to start seeing the AI and mirror will slow down its ability to actually get new techn destroy us. Uh so actually Joe that's actually a good segue. What's going on with this whole jingoism that's going on now that like all of a sudden Silicon Valley was peace and doves and now we want to go to war? I wanted to ask you guys about that. Yeah. What is happening with that? Because everybody's changing.

39:55 Deep mine Daario like they're all now kind of against China. Not against it but they're like be scared of China. We need to be protected. what's happening. Yeah. And also we want to answer that. We also want to talk about this whole talk about Google supposed to crush open AI. Couldn't get into that either. Then we want to talk about benchmark saturation. We also want to get into the guilds and how they get mad. But if you want to learn about that, first you got to do is like and subscribe Wes's video and then click this button for part two.

40:22 Okay. So we are now in part two. Um we are joined here by the one, the only Wes Roth. Wes, thank you for joining us here on the show. Thank you for having me. Appreciate it. Um, if you didn't see part two, part one, go ahead and click here. I'm doing this right now because I hate doing post-production editing. So, Wes had a wonderful question. Like, what's going on with this jingoism in the Silicon Valley? Like, we were doubs five minutes ago and now we're like, we got to mute the Chucky. We're very hawkish all of a sudden. Yeah, the Red Scare. It's kind of Yeah, it's interesting. Uh, I think in some cases, Jordan, it goes back to your description of motivated reasoning. Um, I can imagine if the CEO of Anthropic, uh, Daario, goes to answer questions and says, um, hey, we should we should prevent China from just having unfair advantages and getting ahead of us. That sounds reasonable. Okay, what do you

41:17 have in mind, Dario? Well, for example, we shouldn't ship them the latest chips from Nvidia. That's not an unfair advantage. That's like a level playing field. Why shouldn't Chinese companies get Nvidia chips? Well, because they might get to AGI ahead of us. And you have to understand that whoever gets a slight lead and manages to to achieve AGI or ASI first, then it will be a takeoff, like a runaway where nobody else has a chance to catch up. And even a month or two of being ahead in that in that manner would be just a huge change to the economy, politics, uh military posture. everything would suddenly shift and so we can't allow that to happen where we're behind them.

42:03 So, first of all, that's complete catastrophizing. It's not at all obvious that them getting a little ahead means all these things suddenly happen and it's irreversible. Second, it's not really clear that having China get AGI first is worse for humanity than having the United States get there first. I'm not really 100% sure that I could claim that, but let's say I believe that. Third, why am I punishing Nvidia specifically? They're just one layer in this technology stack. There's a whole bunch of other pieces involved.

42:35 Fourth, uh to to Wes's point earlier about reading all these papers, we were blly publishing all kinds of fascinating uh scientific advances in machine learning right up until just the last couple months when it suddenly became a huge commercial race in the United States. It wasn't until we saw how wildly successful Chad TPT was that suddenly OpenAI is a completely closed shop and Google itself published the Transformer paper to the whole world, right? Not knowing what it meant yet. You know, they had it. The people who worked on it knew that it was something but but the people at Google didn't stop them from publishing. If we really believe that this is a national priority like us versus China, toe-to-toe in nuclear warfare, whatever the crazy Doctor Strange Love thing you want to, you know, describe or frame, we should have never been publishing that research, not just refusing to ship them the

43:38 latest GPUs. So, I'm I'm completely confused by all these things at once. And to to your question, Wes, it's pretty awkward for me to hear someone in charge of a foundation lab say that they don't want China to get uh GPUs, for example, when those foundation labs hire huge numbers of Chinese graduate students who are Chinese nationals and they're producing more than half of all the researchers in the world, that one country like we have if if that's our national priority, why are we taking them into American schools and educating them? Like we have all kinds of conflicts in this, right?

44:24 And if I just focus on anthropic or or open AI, then my reaction is, oh, you're clearly talking your book. It's in your interest for some Chinese lab not to undercut you the way that Deepseek did, for example. I totally agree. And one thing about the awesome Chinese expats we get here who are through all these great AI research labs who are working super hard and they're great. They eventually get their green cards. They stay in America. It's fantastic. If we're going to say, you know, uh, let's hold back GPUs. China can then say, hey, we're going to hold back those researchers because we have DeepSeek now. We have Huawei who needs them. We also have Alib Baba who needs them. The company made Quen 3. And then, hey, you can have your GPUs. We're going to have a research scientist. Another point about export controls. We've been harping on how they they've failed. In the 90s, we were afraid of uh China

45:16 creating their own uh supercomputers because of CPU use and we blocked them and then China caught up. And then now Huawei just came out again and saying we're have coming out with chips are as good as A100. Oh, you skipped the part where we blocked them from buying communications equipment. Go for that joke. From Cisco. Mhm. and they made sure that they funded hundreds of companies and Huawei emerged as a powerhouse building probably better hardware than Cisco and certainly much cheaper. You you know what next time we go barbecue it's it's on me Joe that was a good point. So we create Huawei through our own export controls mistakes and here we are now. Um so that that's just that's bad. Now, we also harp about motivated reasoning is because on the All-In podcast, you'll see Chimath and David Sachs circle jerk about export controls and how great they are. But Chimath doesn't explain what companies

46:07 he invested in. Oh, that's right. Groth. That's right. And they are an Nvidia competitor, right? Right. Knew there was something. I knew there was something going on. Well, what if he went to Chimath and said, "Also, I think there should be export controls on Gro's GP or stop the count, you know." Um, that's why also, yeah, totally unfair. Totally unfair. So, there's a gentleman um named Eric Schmidt. You might have heard about him. Oh, God. And so, he wrote that that paper with Alexander Wang trying to compare um GPU or country AI to nuclear war. That's a slap in the face of JFK.

46:48 You know, JFK had a 13-day missile crisis where he he it was like basically a butterfly dick length a year away from one nuclear war. Was it for JFK, RFK, and Kruev telling the Hawks in their administrations, oh, it's time, let's calm things down a bit. Country annihilated. So, anytime I hear people comparing AI and nuclear war, I'm like, Jesus. So, of course, he does that, but he isn't he fails to mention that. He has multiple startups he's funding right now that would benefit from the government believing the garbage he's talking about so they can get more department of defense contracts. Then Alexander Wang who scale AI they've been focusing on labeling data for these AI labs. Even Google uses them. Well Joe was showing me in Quen 3 and other research papers now they're saying hey LM generate synthetic data and now sometimes they're now they're labeling their own data and now they're also

47:37 cleaning their own data sets. So hm or the absolute zero approach that Wes was describing, right? So hm, do we really need scale AI? So what does scale AI do? Now they're trying to become a a Department of Defense contractor and they've scored a contract with $250 million and so they want to push this AGI, go against the CCP, they're going to destroy us, it's them or us, right? So motivated reasoning is something that we all have to be aware of and do our due diligence on these people because if you go on CNBC anytime an analyst talks about talks about any stock they show a screen saying this person has a position to either buy the stock or sell the stock yet we don't do that for venture capitalists and we don't do this for people who are doing these committees and talking about how bad AI is. So we just raised at least 10 points Wes what do you think of this whole mess? Oh yeah, this huge. And so this is one of

48:27 the things that I do want to learn more about because you guys know tons about like the Bay Area, Silicon Valley, all the shenanigans and what everybody's investing in stuff like that. Thanks, Paul. Yeah. I mean, so obviously, yeah, Shamoff has I think one of the early investors in Grock. Grock obviously is doing very well from from from what I hear. It's an incredible chip. Um, and yeah, so obviously if they can like shut down Nvidia, uh, that's only good for his company. So he he's talking his books. Eric Schmidt, yeah, Alexander Wang was at that Davos meetup where you were saying how Deep has tons of Yeah.

49:04 Nvidia chips that they're not supposed to have. Later, it turns out that was not the case. It was like a a case of broken telephone. Um cuz the guy from semi analysis I guess said one thing then Alexander Wang took it to mean something else then goes on television and go deeper into that because I have a I uh I one thing I talk about Joe is called the Dick Cheney effect. It was m I didn't support the Iraq war but I support the troops but it was master of all Dick Cheney did. He first wrote an op New York Times saying like we know they have yellow cake and blah blah blah blah. Then you want to meet the press and he like the New York Times reported that they had yellow cake and blah blah and so he created his own echo chamber and then people just went with the story and you're seeing these AI labs do it too. Can you break down the whole Alexander Wang telephone issue please again? Yeah, so the guy at semi semi

49:50 analysis posts great stuff very often uh either just post great analysis or leak some information to hold there is no moat memo out of Google back in the days I think they first published it. Um, so definitely and they they know they have a deep view into the industry. So I guess I don't remember the exact details but he said okay yeah Deep Seek must have had their hands on the following chips which was um I think still okay like it didn't necessarily mean that they broke any export controls or anything like that but I guess Alexander Wayne mis either misunderstood that either intentionally or you know or on purpose. Then at the Dalos meetup, he goes on CNBC or whatever it was and goes, "Yes, Deepseek has all these Nvidia chips that they're not supposed to have. That's why they're um able to build this this thing." That that's if I recall correctly, that's what happened.

50:44 And that sort of kind of definitely um And that's the that's the very gentle version compared to like same lucky Palmer, right? Palmer Lucky. I know that Palmer Lucky. Yeah. I mean his his version was much more red meat. What do you say? Like he's like, "Oh, deep seek. What's really going on is the hedge funds actually put massive short positions on the US market that release this to then capitalize on the Magnificent 7 losing money which is just communist trying to destroy capitalism. Scop. Yeah. That's right. Do you think they took out short positions? I mean they do have but but the thing is I think everybody was as they were probably as surprised as everybody else.

51:26 I don't think they were baffled. Yeah. Yeah. Because they released this model for they had it out for a while and no one like really cared. They saw the the research was out for a while, no one really cared. And well, the the negative reaction against the large cap the large cap tech companies was really misapprehending what the what the research meant. You know, it was like an overreaction and kind of unrelated because you still needed to train a very large traditional model and then you could do a little bit of training on top of that RL training to get it to be a reasoning model. And then as a third step, you could take that large reasoning model that you'd now constructed and distill it into a much smaller model. But you still paid the price for the large original model. So it wasn't like all the training just evaporated. which would have been great if they if they discovered that, I'd be

52:18 even happier. But that wasn't the case. And they said so in the paper. They're like, "Well, you still have to train this damn original model to start with the hard way with a lot of compute and a lot of samples." And that's why I got so excited. And I think Wes had the same reaction when I saw the absolute zero paper. I was like, "Okay, now we're getting somewhere. Now we actually see a potential route to not have to train that original massive model as the starting point. Yeah. Um yeah. So yeah, I think the the Deep Seek so they said whatever the the number they said they built it for 6 million or whatever it was that that's what spooked a lot of people. But the amount of money that they had to put into it developing the infrastructure and everything else in order to be able to release that model for 6 million that's what I think everybody missed.

53:06 Um, AI Explained is a channel on YouTube. They he did a Philip I think is the guy behind this. So, he did a great deep dive into the founder of DeepSeek. Seems like an absolutely fascinating person and a lot of the stuff that he talks about like how he approaches benchmarks. He's like, I don't care about the benchmarks. We're not going to game the benchmarks. We're just going to focus on real world use cases and just everything here. Yeah. My man. Amen. And he's not he's not playing the fundraising game. He's not playing he's not playing the shortterm game. he's playing like the really long-term game.

53:37 And so I feel like, you know, we're going to see a lot in we're going to see a lot of incredible things out of that team because I think they're just going to be on a roll creating incredible tech um open sourcing it. And so I was wondering I was going to ask you so in terms of the people that are kind of afraid of China and kind of being a little bit more hawkish about China and stuff like that while at the same time having companies that are sort of in competition with them. How big of a deal is the fact that a lot of these Chinese companies are open sourcing a lot of these models? Is that the big threat? Is that almost like they're afraid that they're going to lose their ability to raise capital and and get paid for the stuff that they produce? If if if Deepc just drops the best models and just like it's open source. Is that correct? Damn Chinese labs are joining Facebook and being open source. Who would have

54:27 thought? Yeah. Do so do you think that's what that's one of the big things that they're afraid of is that that the open source is just going to be undermining open AI and anthropic and such. I mean it forces them to stay at the boundary right they have to stay a step ahead of whatever is being open sourced um because certainly anything that does get open sourced is in a large way commoditized. Um, I think that's true and probably it's a little frustrating for the for the foundation labs. You know, they do all this hard research work and then somebody reverse engineers it in their minds and then open sources it. I think Sam made this point. He was describing how how difficult it is to develop something when you don't yet know if it's possible. Mhm. But then if you see somebody else do it, even if it's proprietary and they don't tell you how they did it now, just the fact that you know it's possible and you have a rough

55:28 idea for what what direction they they used, what approach they used means that it's much easier for you to reproduce it. And he found that kind of annoying that he had a lot of fast followers behind him. Right. But I think to your point uh this goes to the classic commoditizing the compliment strategy that Microsoft was always taking. If somebody opens sources a whole bunch of interesting stuff uh it they they make it no longer be a competitive advantage for anybody else. Very well. And if they have something else that is their competitive advantage, they are relatively advantaged by this other pillar being removed.

56:11 Yeah, and that was the point that Balaji made. Um, I'm not sure what you guys think about, you know. Yeah. So, he's pretty big. He's in the crypto space. I think he's on on Coinbase. One of Wait, wait, wait. BI in crypto? I had no idea. I only watched that 7-hour long brainw Lex Freriedman for Christ's sake. Jesus Christ. Just start your interview. Like, guys, I have a lot of bags in crypto right now. If you just go buy some crypto for me, I appreciate that. I can stop watching. Yeah. But but that was his point kind of I think that this is a way to maybe kind of um undermine some of the US foundational uh like frontier AI labs is by just making everything um all the AI open source because China has an advantage of making physical things. So if they're able to kind of like take the teeth out of the software the AI software etc on the US side they can still produce all the the robots and all

57:01 the trinkets and stuff like that. and he, if I recall correctly, it was almost saying that the the Chinese government might be specifically funding those efforts. What do you think about that? Is just that is just a crazy idea or do you think that might be a strategy that's being employed? Uh, so I I once I I was in political science, Santa University, and it's professor called Professor Strover. And you know, when you're in school, you you're always like, why am I doing this? And then you get one professor in one class, you're like, "Okay, that was worth the price submission." And so he asked us, "Okay, everyone, so why do we go into the Iraq war?" And some person was like, "Oil?"

57:39 And so he put down oil. Um, why? Uh, Saddam might have tried to kill Daddy, tried to kill Bush Senior or something. Um, Israel, US, uh, risk of nuclear weapons. Um, the fact that there was another dictator in the Middle East that we need to overthrow. And so he was like, you know what, you're all right in your own different directions, but actually when you make a decision um in business or in policy, you're kind of in the fog of war where we have all these little different reasons why you're doing something. And then we make a decision, it's easy for someone to go um post hawk and say, "Oh, it was only because of this one element and that's why we got here." It's usually a mix of different things. Like if I was a CCP, they would be crazy right now to say, "You know what? Let's not have an industrial policy that that goes towards us building our own GPUs and AI because there's probably no use for that in the

58:31 future. It's like no, you're going to put resources towards that. But then going back to Highfly and the Deepseek example, that dude before he created his quant fund was super into AI. Like he was a true believer, but eventually he's like, I got to make money somehow. I got new quant fun. My heart heart's not really into it. And then as time went on, hey, I'm making more money than God. I can now crusade like Zuckerberg's well before Zuckerberg I can crusade into AI and start training models. So you have the government aspect of it but then you also had people who are really interested in the in science technology and then you also have tons of Chinese expats who are also into AI are coming to American labs. So you have all these things coming together leading to it.

59:10 But Joe, I'd love to hear what your thoughts are too. Yeah, I like that description. I I think of those old pie charts about the PC software industry or the hardware industry where they would show all the different vendors who made like hard drives and motherboards and cases and operating systems and apps and so on. And when you looked at it, there'd be like a huge wedge that was Intel because the CPU was so expensive. Uh and there'd be another huge wedge that was Microsoft because you had to pay a fat amount of money for Windows and Office. And remember, these are the wedges for profit, right? These are the people who kept a lot of the money out of the sale of every PC. And then you had all the people who made motherboards and cases and hard drives and all these other hard to make components, but their margins were really thin. Even if their revenue was good, they didn't keep much.

1:00:02 Mhm. And I look at that and I think, okay, the same thing is happening to a lot of products in China or like say the modern auto industry where they're looking at, you know, the software they have to pay for in order for their product to work and like, oh, software bastards keep fat margins, right? Like if they sell something for a dollar, they keep, you know, 60 cents of it. Whereas we're over here just beating the hell out of each other, competing to make widgets. And if we sell something for a dollar, we're lucky if we keep a few pennies.

1:00:36 And so if that's your business, you immediately have to be thinking, how do I keep more of the money when I sell this product? And if the software guys are the ones who are collecting all the cash, your target's pretty obvious. That's a really good point. And you know, we saw this too with just Facebook, how they created the Facebook fund, which everyone laughed at. And the big issue was like we're like, I hate that I get squeezed to death by Apple. They control the stack and to play their goddamn games. They decide to put on this privacy thing. Now I got to uh change the way I'm doing my ad targeting. And then you see Open AI right now that that partnership with Johnny Ives to create some hardware device which I think is fraught because I'm carrying a cell phone around. Now people have watches and then you want me to carry a third device like this thing has to be pretty damn good to do that. But the reason why

1:01:21 Open Eye is doing is because they want to control the whole entire hardware stack. Like right now Google's pushing Gemini through my down my throat 247 like Plus back in the day and it's frustrating. But OpenAI wishes they had that so they wouldn't have to deal Google's garbage. Um then you had the Windserve acquisition again like um in in tech especially on Twitter we have a tech ccentric view on how companies are run. We just think oh technology is the most important thing ever and that's why companies are worth a lot of money and and you know Windsurf oh that's a VS fork. I could have created that yesterday. It's like well one didn't do it and then second if you could do it then why aren't you selling it for like two or three billion dollars? Oh that's right you can't. There's a lot of other factors going on into why businesses are successful. And another reason why OpenA did this is because they wanted to

1:02:02 actually be able to have the enterprise relationship the developers are where they're committing code and have access and insight to that so they can get that information to build a better product. Um, and so it ties to what Joe is saying. Instead of being at behest of Windsurf or behest of cursor, we want to be on the top of that stack and be to have that relationship with the end customer and user so that we don't get squeezed in between. Um, so give Wes a crack at what he thinks of that deal. Yeah, go for it, Wes. Yeah, it's uh very interesting. So, they bought Windsor for three billion. There's also cursor valued at 9 billion. Looks like Google is building their own kind of version of it, which is called the Firebase Studio, which I'm very excited about. I've it it's not quite there. It's still a little bit clunky. It still has issues, but it's got that, you know, it's got the Gemini 2.5 Pro integration and then

1:02:55 it has an ability to very easily um publish that stuff online. So if you have an app, you can easily just publish it online. Gives you all the analytics, all that stuff like few clicks. So I think it's making it brain dead simple to do that. So if they are able to get it to the same like level of quality as cursor of or winf I I feel like that's going to be huge for them. And it's the same exact thing, right? So it's VS VS Code, a fork of VS Code plus a chatbot that's sitting on the right side panel. Um, so so if I understand correctly, so is Google trying to build their own cursor to get that that that flywheel going of users using it and building data. That's what's happening there, right? Mhm.

1:03:40 Okay. So they're probably not going to be acquiring a cursor or anything like that. Oh, they probably would still consider it. I mean, they just want to get in the game. This goes to, I think, to Jordan's earlier point about if you're building the models on the back end, you need to see how people are using them and which completions they're choosing and and so on. That means you got to be in the middle of the whole process. You got to sort of own the stack to use the Silicon Valley language. Um, but Wes, is Firebase Studio online only, like in the browser only, or does it come packaged as a standalone app? I haven't seen it. I just seen it online only. I haven't seen any any standalone. Yeah. So, yeah, it's like Google Collab. It It's It's pretty good. Um, I'm I'm enjoying it. It's It's a little clunky, but I feel like if they polished it up, it would be amazing.

1:04:33 Isn't it kind of bizarre to have Microsoft create Visual Studio Code and open source it? And now there's a massive business competition around products that are based on their open source codebase. Yeah, it's it's incredible. I'm I'm I mean I'm I'm glad they did it. Um Yeah. Yeah. What do you guys think about that? I mean, is that going to, you know, is that going to help people create more open-source stuff like this or is now companies are going to be a little bit more careful to put that much value out there? Because I mean, Google's using it, right? If they end up spinning it out into something massive.

1:05:12 Um, obviously Microsoft would want that, would not want that. And Windsurf just got acquired for three billion, something like that. Yeah. And Cursor is worth what? N billion in the last round. Yeah, that's a lot of very valuable companies, at least on the speculation side, all based on a code base and a product that Microsoft sort of built uh promoted, encouraged all those developers and then at some point kind of changed their mind, open sourced it, and kind of backed away from it. As far as I can tell, I don't really understand why Microsoft and especially with with GitHub isn't right in the middle of this mix, right? Building a killer version of Visual Studio Code that's integrated with somebody's language models. I mean, you think that GitHub had the first relationship with OpenAI, like they knew what this this code pile would be and what the possibilities were and then they just couldn't execute on it. Now

1:06:13 Joe, you sent me a research paper called the span of control in businesses and I think this kind of ties to that. Maybe you can talk a little bit like high level about that. No, it's all you man. Okay. So basically there's a theory that uh there's companies that insource things like a core thing that you're doing. Maybe you're a restaurant, you make make you make hamburgers and you make you make you know roasts and soups and things like that. But it gets to a certain point that there's only so many things you can do and eventually outsource certain things. So maybe you might get a pastry chef or you you're going to outsource towards a cleaning company to come in and clean your your your organization your your restaurant.

1:06:48 Now the same thing applies to these tech companies. Um eventually it's like you can be in so many different product areas and you think that maybe you can just keep on hiring people to give you more ability to control different things. The problem is all these decisions end up uh falling on someone's desk at the end of the day of what are we going to pursue, what are we going to improve, and where the trade-offs going to be. And so people have been saying for the longest time, well, you know, Google's going to crush open AI. It's it's going to happen. Well, it's been about what, three years now, and they still are nowhere near closer of crushing open AI. And I think the issue that people don't understand is a lot of these CEOs are so spread out on these big company CEOs are spread out in so many different areas, it's impossible for them to stay on top of all these other different decisions and

1:07:33 industries. And it even happens in startups like uh Sam had a decision of saying hey you know I could pay $3 billion for windsurf or we could try to build something internally but what's the opportunity cost of us going out there and spending my val my engineers valuable time away from building the core chat GPT product that's getting 500 million weekly visits. Um, or I could take this $40 billion that we have and just spend three billion of it and be half cash, half equity and just get 80 people from Windsurf who's who've already died dialed in and can like control the market. And so bring this all back of like where it's going on with Microsoft right now. I just think they have way too much stuff on their plate and they're not able to execute on these things and that's what's hurting Google now too. But I'd love to hear what you two think.

1:08:20 So quick question about what's happening with Google because is it that so there's sort of like Google proper and then there's Google deep mind separate offices separate cultures separate uh management and people and I think Gemini was called Gemini because of the twin kind of companies I don't know if that's the case or not just seemed like to me um how much of the recent Google success would you attribute to the fact that it seems like the development of all these AI models of LM is shifting from sort of the Bayer area Google to the more demisab team and and is that what's happening or no? No, I think it go ties to Joe's point. Remember what Joe what you said about it's so much easier to be a fast follower to someone who's already created the breakthrough compared to creating the breakthrough yourself. So once open AI came out with chat GPT or let's take you a step back. The reason why I maybe Joe also are so cynical

1:09:16 towards these AI doomers and alignment people is that we've dealt with them internally inside corporations and how they have hurt Google who had the opportunity to actually take a lead in this race a very long time ago. Um they could have released Lambda and other things like that. But they had Blake Lemon and all these other AI ethicists say, "Well, what if someone does a completion and it pisses off bees and alphas and dory? What if we can't do this? We're going to hurt people." And so you get stuck in like this endless circle jerk of getting into this whole completeness fallacy of trying to figure out every edge case. And I think that what the the issue is for um Google in this circumstance was that they got caught up in their own internal bureaucracy. But then once check GPT got released, it was like uhoh this is a threat to our core business and it gave Sundor an excuse to douche out of the

1:10:06 system. All the garbage and the trash, it didn't matter anymore. Open eye did the same thing too with with uh with uh the the board drama rid of all the alignment people into anthropic. Peter Teal told Sam Alman years ago um I am not funding you anymore. I kicked him out of my program and you need to get these EA people out of OpenAI now or you're going to regret it. And Sam was like h well politics and things like that. Yeah. He gave him a clear warning. He warned them. Yeah. Right. Exactly. So I see one is this they finally had competition that like woke them up was the ma the main driver. Um but then as far as like was it deep mind and things moving over there it's making things easier. I think it's more of Sundar just saying hey all these all these years I've been trying to make all these different teams happy with AI. I need to consolidate this under some someone.

1:10:57 Now, Deep Mind decision to buy that company in 2014 was Larry's decision and Larry specifically said, "I want to keep this company in an island and I don't want Google proper to f with it." So, let them do what they're doing and don't let our culture get in the way. The way that YouTube worked. Exactly. And then also people forget that Demis was on this week in Startups and they asked him uh Jay Cal was like, "Oh, so like when did you go get bought by Google?" And Demis said to train more of these models we needed more money and it was the only way to do it was to get bought by a bigger company. H reminds you of OpenAI.

1:11:31 Why did they kick a gigantic paycheck for or gigantic investment check Microsoft? Because they need capital to train more models. Soft bank and and now for Soft Bank. Um, so eventually, uh, they were their own thing and for the longest time they tried to get attention for Google proper on seeing if they could work together on things and it was too difficult and it was only until chat GPD came through that eventually it opened the floodgates for them. But Joe, I'd love to hear what your thoughts are on this too. Yeah, I think you described it well. I mean, in the earlier days, you could see papers from these companies that would have researchers from two or even three different companies, including Google and OpenAI, and now you just can't even imagine that.

1:12:12 Yeah. I used to see Google Brain or Google Now, you must see Google Deep Mind as the only one, right? Mhm. Or you'd see a you know a paper that people from Google Brain and people from OpenAI and oh working together on some project because it was a research project. Right. Right. Right. Now everything has an immediate commercial impact and everyone's become much more quiet about what they're doing. Yeah. If like attention is only needed if that paper didn't get published. I wonder where we would all be now. Like how how different everything would have looked. Yeah. Um I'm very curious because I'm so excited about some of the stuff that's coming out of Google Deep Mind. um isomorphic labs I've been kind of following on that's very interesting and just everything they have the the gnome project that's building various uh discovering new material I mean just everything out there is just so

1:13:01 fascinating um at the same time so I feel like if Google deepline mine was its own standalone company I would like would love to invest in it because it's connected to Google um I feel like Google's whole thing is you know search ads like one of the best business models ever people clicking on search ads like whatever billions of dollars they get for that. But doesn't it seem like there's this weird thing whereas AI is taking off, the search is going to start trending down while the revenues from AI is going to be going up? Like what happens to Google? Like how do you even think about where Google's going to be in this environment?

1:13:37 Yeah, that's the the big question, right? Can they or will they make a pivot if it becomes necessary, right? I mean there was a period where Microsoft was sort of uh in the wilderness where you know Windows was trending down and they hadn't Microsoft hadn't really been successful outside of Windows and Office. Um and people kind of left them for dead and then when Satcha came in it was like all of a sudden they started making a bunch of decisions that made sense including the investment in OpenAI. Mhm. and they got a new lease on life. Like suddenly it's a revitalized company. So it's possible to imagine a few things going right for Google and the same kind of uh what what the press calls the redemption arc playing out.

1:14:30 Mhm. Yeah, I wouldn't count them out. I think also they're going to there needs to be a new CEO. Joe and I do a prediction tracker on like what where predictions are for next year and whatnot so we can keep ourselves accountable. I thought Sundar would be gone by end of last year because he's been here for seven or eight years which is like an interity for tech CEOs I think for bet silver trump coins on that. No I I should have damn it I know I missed out on that one. Um that's our usual bet. Exactly. Silver Trump coins novelty factor gets hilarious. Um but I think Google would probably need a new executive to put them into wartime mode where they actually feel okay now we're lo we're losing more market share in search. Um there was a recent uh kurfuffle in the news that during the DOJ uh trial that Google was dealing with for um being being potentially a monopoly. They had an executive from

1:15:21 Apple speak and say yeah Eddie Q there you go. Um once we installed you know a lot of people using perplexity and chatbt search we actually saw a decrease in people using Google search and the whole like stock market was my god then google stairs sold off 8% happening right and then Google came out with the Iraqi information minister and said no no no don't be afraid of this the stock went up but just the sign they could potentially be losing a little market share to proxy chat search is already getting people like on the edge uh but I think in order for the company to switch from like the old search model to pure AI, there would be a rough transitionary process where they wouldn't be able to maybe juice at the same earning per share that Wall Street would want and I don't know if Sigar would be the person to do that because you would have to make some drastic changes internally in the company. That's always tough. Well,

1:16:13 and you made the point about how cheap their money is too, right? I think Perplexity is trying to raise money right now at some crazy valuation like 9 billion or 14 billion or something like that. Yeah, 14 billion. and they're trying to raise 500 million. They're close to closing. Uh so I I would say these different AI companies, if they're perceived as AI companies, have an incredible amount of cash and an easy ability to raise more. So they can get away with losing money on queries that are very expensive to process.

1:16:45 Google can't do that. I mean, if they suddenly put Gemini right in the Google search box for every single query, right, they would immediately lose a god-awful amount of money and probably melt down a bunch of their data centers. So, their transition is much trickier. This is the situation that Microsoft was in where Windows was the cash cow and if you were inside Microsoft and you proposed a project that didn't also promote Windows, it wasn't enough that it was neutral or unrelated. It had to actively promote Windows, right? Or your project would be cancelled, right? And they knew that internally.

1:17:23 They would say it out loud. They'd be like, "Windows is the platform. Your project should be promoting Windows or we see no reason to fund it." And that's, you know, they missed out on a whole bunch of important things that happened in the industry as a as a result. Interesting. Okay. So, maybe Google can take kind of a lesson from that to figure out how to go forward because obviously AI seems like a much bigger thing than s I mean search is incredible obviously they had this cash cow for so long one of the best business models but if it's going I mean I find myself using deep search um more and more across various things openai when it came out uh recently been trying doing deep search with Google 2.5 pro and is just incredible I'm blown away by how is it as good as openai's deep search I think it's better I've been using it in the last week or so and I'm that's an endorsement yeah it It's I I I

1:18:14 love it how deep it goes um and how it really captures exactly what you're going for. And also it really filters cuz I've tried also, you know, on the Grock doing their stuff. Gro seems to be like a kind of um doesn't there's less critical thinking like it doesn't dismiss stuff doesn't that doesn't make sense I mean yeah vain praise would have to do anything with Twitter brain rot affecting Grock that's critical thinking I was told I was told Elon the data set of every other week of someone saying Hitler did nothing wrong is worth so much for Grock for trillions of dollars it's going to improve its ability to think. Go ahead.

1:18:58 Yeah. Well, yeah, I'm curious to see what 3.5 will look like. So, I mean, they're definitely catching up because they're they had a late late start. Um, but so far, I'm Yeah. So, I mean, OpenAI when they launched Deep Research, it was great. I haven't tried up until like a week ago, I haven't tried Google's Deep Research. I've tried the one that's powered by 2.5 Pro and it is just phenomenal. It's absolutely incredible. So, I'll keep testing it. But, um, but I the thing is like back in the days when I used to do these kind of deep dives that I need to do research for, it would be between, you know, like dozens of Google searches, right? Um, going on on Reddit, going down rabbit holes on Reddit or just Google searches and all all the other stuff. And now that's replaced with, you know, as soon as an idea pops to my head, I jot it down into deep research. I hit go and then usually you have to wait five to 10 minutes,

1:19:52 right? So you go do your thing and later you're like, "Oh, here's this massive report that I have to digest." But it's just phenomenal. Yes. Be I love everything about it. Um so I don't see how that doesn't affect search volume because that just eats Yeah. you know 20 to let's say 100 searches that I would have would have done otherwise, right? And there's so many places we could go with that uh as a as different topics. But I'm I'm impressed that you think the Google one is that good. Have you uh you must have read some of the papers about OpenAI's implement uh implementation of open research. Um I have deep research uh versus Google's implementation because they're very different under the hood. Interesting. I will definitely check it out because I've been kind of ignoring I really liked OpenAI's um deep research and that's I've been so I probably I have dozens of more more I probably have 100

1:20:48 plus uh maybe 200 since since they've launched because I have the unlimited plans I just whatever pops into my head I I launch it in there. So definitely probably in the in the hundreds of deep researchers on OpenAI in the last week I started just on a whim doing um Google and so far it's only a handful but those handful have been absolutely mind-blowing. So I I got to do more research but um I will deep research. Are you saying are you saying like 20% better than the open eyes deep research 30% or to the point where like I would never use open eye deep research again because what I have going with me is a switching cost of I'm so used to using open eye for everything and the moment I hear from some oh this new product over here is doing something cool and like okay I'm going to teach myself to use it and then open eyes like oh yeah we just took that feature god damn it no idea come back to where I was. Yeah, I it's

1:21:42 hard to say because I'm still I I haven't made my mind up and it's just from the few interactions. Like one of them I was looking at trying to see what a lot of people that are like are creative types and produce a lot of like creative work. I was curious if there's any sort of overlap in their day-to-day schedule, if there's any stuff that they tend to some commonalities that they tend to do like do they wake up early or later? Do they do meditation or this? So, I kind of like threw that in there and wow, it blew me away cuz first of all, it starts out it had this like great title. I forget what it was, but it was like very like artistic, which usually doesn't do that because this was more like uh creative work. It had like an creative title and it just from the get-go starts hitting you with interesting facts about uh Moz Mozart and the various writers and the various people like the fact that they all

1:22:31 engaged in early morning activities where they would walk on this. Then it went through how many of them used to do TM transcendental meditation and like the quotes about that. So within so I had an idea like okay is there anything I can add to my day to just be able to just kind of have more energy and kind Yeah. And within the first paragraph I was like wow like there's so much here. So well written so well research. Yeah. And the other one I did was um um various things for like stress management and stuff like that. And I was blown away by how deep it went because first it went okay here's what stress is. It's you can think of it as these four biological factors. Here are the research that supports the best things against that and here are the sort of pathways by which you might do that. So like if you go running what does that do? That's the you know it helps with the neurotransmitters

1:23:26 serotonin but also it does this and this and this and this. So as you're sitting there and it links to all the research papers and stuff like that and I was like This is because there's no fluff. There's no BS because like with Gro it's like some people enjoy Yeah. Yeah. Because with Graco I was like well some people enjoy like walking in nature so that might be something to try. And it's like okay I maybe maybe not but like I'm not seeing like the the stuff you're referencing in the in the in the citations. It's a little bit weak. I mean I get it but I could have went to whatever healthline.com and found that information. Google is like, "No, like here's the" And I forget like there's some sort of a method for rating how conf how much confidence you have in a study. It was like Davis's G or I I forget unfortunately, but it ranked them all by how so like uh cognitive behavior therapy versus aerobic exercise versus

1:24:22 yoga versus this and it ranked them and I was like I it seemed like like a lot of thought and effort went into that. Like I don't know how much Okay, I'll try it. Dan, do you have an Do you have an affiliate link I can click so you can get your $10 from Google? Hammering on. How neat. You win. You win. Stop the count. Not in the face. I forgot the original question. I just went up. No, that's like that's my life. I always forget the original question. I'm just like whatever. kind of feature is just such a attractive thing for someone who is interested in getting the complete picture uh and seeing all the links to the research in case they want to dive in. You know, especially if the if the model or the framework depending on how it's implemented has gone and actually looked at all those studies in detail in order to rank them. That's pretty crazy.

1:25:14 Mhm. that's done like to your point, you could have done all those searches yourself and you would have spent hours putting that together. Oh yeah. And you know, depending on how long you spent, done as good of a job or not, you know. Um I'm actually really curious about something you said earlier, all the research and projects you're seeing. And I guess to be kind of kind of um specific, how far away do you think we are from simulated people in meetings that are completely well simulated and you can't really tell?

1:25:51 North Korea is listening right now because they've been trying to do it for a while. Go with Oh, interesting. Um like right now we can't really tell when we go to chat GPT and we pick the most recent model. like it passes the Turing test. And you mentioned Sesame AI and their voice, you know, the empathetic voices. Those are pretty damn good. Like it's hard to tell if it's a real person or or a simulated voice. And you've probably seen a couple services that are already producing uh video in in almost real time, right? Yeah. Yeah. I've tried an open source project. I forget what it was called. Um, but yeah, basically upload any video in real time. Like I it made me look like Elon Musk, Sam Alman. It was pretty good. It was all right. It was copying your movements and your speech real time. And I could shine a a light on myself and it like realistically sort of trace that light and stuff like that. So, um, now so

1:26:47 you're saying simulate people. You're saying so can it simulate somebody like a real person to fool somebody that knows that person or just simulate any random person? Oh, that's a good question. Let's assume it's not someone you know, but you've probably seen examples where you can give it a photograph or a short video clip. Yeah. And then you can say, "Okay, now now give me that person speaking this script." Yeah. You can imagine coupling that with a language model would allow for it to engage in a conversation.

1:27:18 Yeah. No, I feel like we're pretty close in terms of like the fidelity of the sound and the video. I mean, video is still a bit choppy. The sound I feel like is is almost there. The with the sound like if you listen to it long enough you realize it misses the like the inonations because I guess it's not trained on what inonations mean what specific thing. So it fails at those but in terms of just sounding like a real person. It's there video mostly there. Um you know hey Jen is doing their avatar 4 or whatever it is. I feel like we're we're so so close to just if we're talking about like the fooling the senses. I think we're very close to that in terms of long-term fooling people like as if it's the other person that has enough context about the previous conversation to keep go like if you have a long enough conversation can reference stuff from you know some inside joke from last time or whatever

1:28:17 that's where it seems to fall apart and most yeah we're still at that Brian Johnson moment where it's like uncanny valley like I want to live forever but then my face looks like a bishop from the alien you haven't really get there, you know. Yeah. Uh yeah. But I feel like Yeah. And so the the interesting thing to me is like the the whole idea of like the long-term coherence, that's where stuff really falls apart. Whether you're talking to a large language model or the AI agents like OpenAI did a thing where it's like can these AI agents replicate machine learning papers? So if there's some experiment that's proposed, can it replicate the code, run that? And it's like, you know, it's tested against machine learning people with PhDs or like studying for that. And so the AI agents are just so much better out the gate. They're like boom, they start copying the code, they start running, but then they plateau and humans solely

1:29:13 catch up and supersede them. So we win at the long context game. And the same thing with um some of the studies that recently came out is just these AI agents break down over time. And that feels like the biggest stumbling block to really having AI agents. Like if somebody's able to solve that issue, I think it's going to go hard. So I want to we're we're running out of time. I want to get into that because there's been so many companies just banging the door on AI agents 247. Um but then if you work inside those companies, most the employees are like these things don't don't work. They suck. Yeah. Yeah.

1:29:51 And they're convincing everyone like this is the this is the way to go forward. And my real test is I will know that AI agents have arrived when you have managers saying I'm willing to sacrifice headcount in order to get an AI agent. I am willing to not increase incremental headcount growth because I want an AI agent. Yeah, that's a good test. And so far, I don't know anyone worth their salt besides some random person on Twitter who's just smoking crack who says, "Yeah, I'm running everything by AIA agents and blah blah blah." Um, and so maybe that's where we can do part two of this conversation, go into that and what you think about it.

1:30:25 Um, because that's what really got me into this is there's so much uh work out there in corporate that I was like, "Oh, Chibi, you could end this so I can focus on things I want to do." Not the whole I'm going to replace all employees AGI tomorrow, but just more of the the work. offers and jobs. Yeah, like dealing with workday or concur and things like that. No one wants to do that. Not even the people who build workday want to touch workday. So why should I touch it? So So anyways, uh for all of you who are watching, get you can watch part one here. We also create uh a playlist specifically for Wes Roth viewers for some of our videos.

1:31:00 And um thank you Wes for joining us. This has been fantastic. This has been an incredible conversation. I love chatting with you guys. I've learned so much. So, I do appreciate that. We definitely should do it again. Um because this was absolutely incredible. This was very valuable to me. Hopefully, it was valuable to all the watchers and listeners. And I would absolutely yeah would love to do it again one

Questions about this episode

What is discussed in "This Former Google Director Just Revealed Everything... Self-Improving AI, Acquisitions and Google"?

In order: 0:00 online marketing, clickthrough rate, Interpretability, Joe Tonoski; 10:00 Interpretability, Reinforcement learning (RL), creatine supplementation, data centers; 20:00 AI alignment, doomers, deniers, and dreamers, AI preparedness teams, Helen Toner; 30:00 self-play, Workday, AI agent, proposer; 40:00 Alexander Wang, export controls, Huawei, motivated reasoning; 50:00 DeepSeek, hard drives, motherboards, Absolute Zero (paper); 1:00:00 Windsurf, Windsurf, Cursor, Gemini; 1:10:00 Search, Windows, Windows, data centers; 1:20:00 Language model, Aerobic exercise, Avatar 4, Brian Johnson; 1:30:00 AI agent, Concur, Wes Roth, Workday.

What is mentioned in this Wes and Dylan episode?

204 people, works and ideas are mentioned, including ChatGPT, GPU, Grok, Search, DeepMind, Windsurf. Each one on this page links to its own page with every episode that mentions it.

How long is "This Former Google Director Just Revealed Everything... Self-Improving AI, Acquisitions and Google"?

The episode runs 1 h 31 min; the transcript is about 17,776 words in 118 timestamped sections.

More from Wes and Dylan

• All analyzed episodes
• Books mentioned on Wes and Dylan
• People mentioned on Wes and Dylan
• Topics mentioned on Wes and Dylan