YT YouTube2Text

YouTube2Text → Transcripts → EP 212 - Shirin Mojarad | Building Personal AI at Google

EP 212 - Shirin Mojarad | Building Personal AI at Google — Transcript

by طبقه 16 · 17,498 words · 1,201 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I'm currently a Product Manager at Google, working on Personal AI.
0:05When you give ChatGPT access to read your Gmail,
0:10you're essentially telling ChatGPT, 'My Gmail is yours.'
0:15Google is relieved of responsibility, which users don't realize.
0:21The future gets pretty dark! It's both fascinating and dark.
0:26Why did Google fall so far behind? What if Google launched something first?
0:29Exactly. Being first isn't what matters. First-mover advantage isn't much anymore.
0:34In fact, it can be a huge liability... How quickly you can unlearn things—
0:39not everyone has that mindset. People learn one skill,
0:44make good money, advance their careers, and lose the patience to keep relearning.
0:49The only constant in life is change. My background is in design—
0:52I loved interfaces. Now, I barely use interfaces at all.
0:55It knows everything about me, and I don't want to use any other app.
0:59Big tech companies want to be the only personal assistant you interact with.
1:03Memory isn't just where you went; it's understanding your personality.
1:08From your emails, travel tickets, resume... you could give it your inbox
1:13and say, 'Explain my life.' Be careful merging Google Photos data
1:17with Gmail. You shouldn't infer that the brother in this photo
1:22is the same person living in Canada. How did you manage to get into Apple
1:24without some extraordinary background? First and foremost,
1:28I'll admit something people rarely say: I did terribly on the Konkur exam
1:33and studied Biomedical Engineering at Islamic Azad University.
1:36I didn't even take the Konkur! I knew I wouldn't do well, so I skipped it.
1:41I just said, 'Nope, not doing it.'
1:44It's nice just sitting here, looking at each other like this! Awesome.
1:48All set! What's new?
1:50Doing great! Thanks for having me, and welcome to San Francisco!
1:56It's been about a month, right? A month and a week. Awesome.
1:59It's so interesting—so many cool people, so many like-minded folks.
2:06I was actually...
2:10I used to overthink and ask myself, 'What's your problem? Why can't you settle down?'
2:14Because I had friends, I had everyone, and they were all like,
2:18'What now?' I was searching for like-minded people who share your drive,
2:24who want to discuss topics you care about. In Toronto, it got to a point where
2:27if you talked about work, people would say, 'Alright, give it a rest! Enough about work.'
2:31And I was like, 'Man, who else am I supposed to talk to?'
2:35We're at an age where people are tired, especially us Iranians,
2:40with everything that has happened and all the pressure we've been under.
2:45We're completely burned out, so people just want to chill and relax.
2:49They really just want to take it easy.
2:53Yeah. When you came here, you must've had an expectation.
2:56Did anything surprise you? I'm interrupting, but—the city was much rougher
3:02than I imagined. Of course, the image is getting better,
3:05but when I first arrived... Toronto and other places feel newer,
3:13more recent and fresh. The streets are newer, the asphalt is cleaner,
3:19the buildings are newer, new areas are being built constantly. Everything feels fresh.
3:26In summer, people pour outdoors, everyone looks stylish, dresses really well.
3:31Here, everyone looks disheveled, and there are lots of zombie-like areas.
3:39The city feels old and neglected, the pavement is cracked.
3:45When I first arrived, I was shocked. Then you see the prices, the rent...
3:51An $8 or $9 coffee! What is going on? You notice all of that at first,
3:57but now I'm moving past the tourist phase and realizing how this place actually works.
4:03Unlike places like Toronto, New York, or London, where you step onto the street
4:10and the energy hits you in the face— bars are packed, people partying, music blasting—
4:14San Francisco happens behind closed doors. You need connections, circles, and a network.
4:20Behind every door, something big is happening, and you have to plug into different circles.
4:24In this short time, I've dived into so many circles that I'm like, 'Alright, that's enough!'
4:28I just want to stick to my own circle now. How long have you been here?
4:33It's funny you mention that. I've been in SF for seven years now.
4:38I really resonate with what you said about moving. When I left Iran,
4:42I first went to the UK for five years to do my Master's and PhD.
4:48Then I moved to Canada where my family was, spent three years there, deeply in tech.
4:54I always had this itch to come to the US because Canada felt 10 years behind in tech.
4:59I don't know about now— I haven't lived in Canada for 15 years.
5:02Then I went to Boston for three or four years, and again felt I wasn't close enough
5:09to the cutting edge. So from Boston, I moved to California, to SF.
5:15With every move, friends would say, 'Where are you going? We just found each other!'
5:19In Toronto, my family was like, 'We came across the world to be together,
5:24why are you leaving?' We had extended family there too,
5:29besides my mom and brother. My philosophy in life has always been:
5:35I'll go; if I don't like it, I'll return. I was fortunate to have that option.
5:40Many people don't have that opportunity. You might think, 'If I go to the US,
5:45where can I go back to?' I always thought: if I don't like SF, I'll go to Boston;
5:49if not Boston, back to Toronto. So I completely understand you.
5:54When I arrived in late 2019, right before the pandemic, I got lucky.
6:00I found great friends, and during the pandemic in California, we traveled a lot and had a great group.
6:06I experienced that too—you arrive, find diverse groups, everyone is like-minded and in tech,
6:12you have a blast talking to everyone. Eventually, it's not that you get tired,
6:19but you want to focus on your work and what you're genuinely passionate about,
6:23so your circle naturally gets a bit tighter. I completely agree with you.
6:28The reason I bring this up is that I'm sure many people think like us.
6:31We all had a purpose when we immigrated. I divide immigrants into a few categories,
6:37mainly two or three. One group chose to immigrate
6:43to improve their circumstances, driven by ambition and a clear goal.
6:49They immigrated with intent. I see myself in that category.
6:53I didn't move just for the sake of moving, or just to settle in Toronto
6:59for a relatively better life. That's not enough for me.
7:04Canada has always been seen as a stepping stone, a bridge people talk about.
7:08Once people reach that step, they tend to sit back and just settle down.
7:13Then there's the second category, which has grown much larger lately,
7:19because our home country isn't a great place to stay and grow anymore.
7:26Reaching even an average lifestyle has become tough. I was thinking to myself recently,
7:31how unfair it is that even to see our parents, we have to meet in Turkey.
7:38They fly from Iran to Turkey, and we fly from here to Turkey.
7:43Nobody wants to gather in Iran anymore. It feels like it's not worth the hassle.
7:48Even people without any issues or trouble feel, 'Why deal with the stress? Let's just meet in Turkey.'
7:52We can't even get together in Iran to see our own families.
7:57We prefer meeting elsewhere to avoid complications and not have any arguments.
8:01When a family just wants to reunite... Now, Toronto—
8:05I love Toronto. Even coming here, I've gained
8:09even more respect for it. Toronto is newer and cleaner.
8:16But Toronto, in my opinion, became a hub for people who simply didn't want to be in Iran.
8:23They came there and settled in. That relentless drive to build something huge,
8:28they don't really have that. I'm not criticizing at all,
8:35but people just settled down. Whenever I asked someone, 'What's the plan?'
8:39they'd say, 'We're just living our lives.' That phrase always bugged me a bit.
8:43You know, 'just living our lives.' Maybe I'm still young,
8:49maybe I'll reach that point in a few years, but I doubt it. In Toronto, I couldn't sit still.
8:53I even got the apartment I really wanted, but until the very last day I lived there,
8:59I never hung my favorite artwork on the wall, because deep down I knew,
9:04'You're not staying here, you're leaving.' That feeling of not staying really bothered me.
9:07At least here, you know you didn't make a mistake by coming. You know what I mean?
9:12It simply can't be a mistake. You might not love it, but it cannot be wrong.
9:19Whereas Canada could end up not being your place. Yeah, absolutely.
9:24Honestly, whenever I've moved— and I've moved three or four times—
9:30it was never with the intention of staying or not. It was always to learn the environment.
9:34Once you learn it, you understand how life operates, how the majority lives,
9:39you find your friends, and then you decide whether you want to stay or not.
9:44When I was leaving for SF, someone asked, 'Where will you go next?'
9:49I said, 'If I knew, I'd have gone straight there!' Just being in California is enough.
9:53There's another point. I really love a quote from a chat I had with Soheil,
9:57who was CTO of Digikala for a time. He said one of the biggest challenges of immigration
10:02is the lack of high-quality jobs. For those of us in the tech world,
10:09we wanted to work at big companies like Google, Microsoft, Qualcomm—the list goes on.
10:15For an engineer wanting to work with big data, companies like Snapp and Digikala
10:20were among the few in Iran with massive datasets that made the work exciting.
10:24The exact same thing that happened in Iran happened to us in Canada.
10:28High-quality jobs were scarce. You were applying for ordinary jobs just like in Iran,
10:32so immigration lost its point. Whether you went to Canada, Australia,
10:36or anywhere else, it didn't make a difference. San Francisco, however,
10:41offers a real advantage for tech folks. I agree for tech, but even tech is fading a bit.
10:46Now any industry can thrive here. I don't fully agree with that, honestly.
10:51During the pandemic, everyone said tech left for New York, Florida, Texas...
10:56Yet people from all over the world keep flocking to San Francisco.
11:01That's what I mean—tech's dominance isn't exclusive. It's not just techies anymore;
11:05any industry can be here if they can afford it. I hear you. I loved Toronto myself.
11:10I don't miss places I've left, but I loved every place I lived in.
11:13I have immense respect for people in Toronto because it's a tough place to make a living.
11:18Here is tougher at first, but once you settle in, it gets much easier.
11:24Toronto never felt easy to me. That's why I moved, exactly.
11:29It comes down to socioeconomic factors: the political system there is social democracy,
11:34whereas here it's capitalism. What values and principles drive the system?
11:37Those are deeper discussions, but having lived in the UK, Canada, and the US,
11:41I realized I don't enjoy living in a social welfare system, even though it benefits the majority.
11:45I agree it places a glass ceiling on a minority so the majority can enjoy a baseline standard of living.
11:51That's great—I personally support free healthcare in the US— but beyond that, I was never comfortable in that system.
11:57In the UK, I strongly disagreed with many policies, though on a student visa I had no say.
12:03In Canada, I disagreed with policies like paying everyone stipends during the pandemic.
12:09That was awful. It was terrible. Trump did it too,
12:14and we're seeing the bad repercussions. I had friends who slacked off back then
12:22and stayed lazy, because they got used to getting government handouts.
12:26US society saw some of that, but the US is capitalist. In Canada, you can let go, have free healthcare,
12:30get basic needs met, and just get by. It's becoming more like Europe, unfortunately,
12:36rather than like the US. I have so many Canadian colleagues who say,
12:40'Canada is a stepping stone for us too.' Even multi-generational Canadians in tech
12:45are moving south to the US. These are deeper discussions.
12:47Toronto was great, and we respected the folks there. The thing is, anywhere in the world,
12:51finding people like yourself in tech is rare. In Canada, seeing a woman in tech
12:58with a family or kids felt so unique. Here, it's completely normal.
13:00Anyone can hold that job here. You're currently working as a Project Manager
13:04or Product Manager, right? I'm glad we touched on that—let me clarify the difference.
13:11I'm a Product Manager at Google, working on Personal AI for consumer apps.
13:16Our consumer apps include Gemini app, Search AI Mode, and soon Search AI Overviews,
13:20where I focus on personalization. My background was originally in Data Science.
13:22I did my PhD in AI and ML back in 2012, when AI/ML wasn't hyped at all.
13:26I'm revealing my age here! When I got my PhD in 2012, I moved to Canada.
13:31Spill the secret! A woman's age is a mystery! People here don't realize how old Iranians are,
13:35so I still pass for early thirties! Back when I got my PhD,
13:38nobody was talking about AI/ML. I had an uncle in the US for many years,
13:44very successful at his work. He'd call me and ask,
13:50'What are you doing? A PhD in what? What are you going to do with that?
13:55You've isolated yourself and have no idea what's happening in the real world.'
14:00Thank God AI/ML blew up suddenly. When I moved to Canada,
14:05I landed a job within two weeks, even though everyone said it'd take a year.
14:09I got a job in two weeks, changed jobs eight months later, and worked as a Data Scientist for three years—
14:12well, about two and a half years in Canada. Then I moved to Boston as a Lead Scientist
14:17at an education company. In Canada, I worked at a bank,
14:23but I didn't enjoy financial work, so I chose education because it felt more meaningful.
14:28And it was, but back then, before AI disrupted everything, it was slow.
14:34We had great data, the company was massive, but moving anything forward required so much regulation.
14:40I was working with universities, and educational institutions were so slow.
14:46I built software that predicted when students were likely to drop out early on.
14:52It took ages just to pilot it with three institutions.
14:58Eventually, I wanted to be closer to pure tech, so I joined Apple as a Data Scientist in the Bay Area.
15:03I worked as a Data Scientist on Apple Maps, which was fascinating.
15:08I learned a ton about search, autocomplete, and routing, because Maps is fundamentally all about search.
15:12I worked on the search side for two years, then felt I wanted to move away
15:16from pure data science and data insights, and get closer to the business side.
15:22When you work purely with data, unless you're a Product Data Scientist,
15:27there's always a slight disconnect from core business. Around that time, Google reached out.
15:33If you want to invest in financial markets yourself, you quickly run into a challenge:
15:36you must spend time studying markets, analyzing companies, and deciding where to allocate capital.
15:40Investment funds are one of the ways that simplify this entire process.
15:45Instead of picking and managing individual assets yourself, the fund's capital is managed by a professional team
15:48under a well-defined framework. One active, licensed advisory firm
15:53in this space is Toranj Investment Consulting. Toranj offers various funds for different risk profiles
15:57and investment horizons. For instance, Efran is a fixed-income fund,
16:03Rose Toranj is backed by gold, and Yas fund is for silver investments.
16:09For the equity market, they have equity funds, and for higher risk, the Narenj leveraged fund.
16:13Another practical feature is that investing in the Rose Gold Fund allows you
16:18to receive an English financial solvency certificate, which is useful for embassies or international procedures.
16:23To explore the funds, risk levels, and investment methods, visit toranjcapital.com.
16:26Toranjcapital.com has all the information. I got a job in two weeks, changed jobs eight months later,
16:30and worked as a Data Scientist for three years— well, about two and a half years in Canada.
16:38Then I moved to Boston as a Lead Scientist at an education company.
16:45In Canada, I worked at a bank, but I didn't enjoy financial work,
16:50so I chose education because it felt more meaningful. And it was, but back then,
16:54before AI disrupted everything, it was slow. We had great data, the company was massive,
16:59but moving anything forward required so much regulation. I was working with universities,
17:04and educational institutions were so slow. I built software that predicted
17:08when students were likely to drop out early on. It took ages just to pilot it
17:14with three institutions. Eventually, I wanted to be closer to pure tech,
17:21so I joined Apple as a Data Scientist in the Bay Area.
17:27I worked as a Data Scientist on Apple Maps, which was fascinating.
17:32I learned a ton about search, autocomplete, and routing, because Maps is fundamentally all about search.
17:37I worked on the search side for two years, then felt I wanted to move away
17:42from pure data science and data insights, and get closer to the business side.
17:47When you work purely with data, unless you're a Product Data Scientist,
17:53there's always a slight disconnect from core business. Around that time, Google reached out.
17:59I joined Google as a Solutions Architect. Essentially, a Solutions Architect
18:03works with enterprise companies to implement Google's technological solutions.
18:09Back then, we had Vertex AI on Google Cloud, along with several AI/ML products.
18:14I designed solutions for major US enterprises, which was incredibly interesting.
18:20You dive straight into their business problems, understand what's broken,
18:26and figure out what to build to solve it. I focused heavily around search solutions.
18:31Around early 2022, AI-powered search became extremely popular.
18:36Companies wanted powerful internal and external search, so I worked on that, including federated search.
18:42Federated search was a key component, but remember that back in November 2022,
18:46OpenAI dropped ChatGPT and everything exploded. In early 2023, we worked on Vertex AI Search.
18:51Vertex AI Search allows you to plug in any custom knowledge base or documentation.
18:58For instance—without naming names— a massive accounting firm whose search is heavily used
19:05instead of regular Google Search... We rebuilt their entire backend with Vertex AI Search.
19:11They had multiple websites and distinct documents. Based on the user's query,
19:13how do you optimize retrieval for both ultra-low latency and minimal cost
19:20while providing the best possible information? I worked on that, found it fascinating,
19:31and transitioned fully into product management. Product management is all about
19:37owning the outcome—you truly own a problem, not just the solution.
19:42Search itself isn't the problem; the problem is: when a user asks for something,
19:48how do I deliver the absolute best answer right then? Google Search solved that for years,
19:56while others hadn't. That's how I got into Product. At Google, we also have Project Managers.
20:01The two roles often seem similar and can blend together depending on the company.
20:08For example, at Apple we didn't have Product Managers; we had Engineering Project Managers (EPMs).
20:14Project Managers primarily build processes to ensure projects execute efficiently end-to-end,
20:17tracking who delivers what and when. Product Managers operate one step upstream.
20:24We ask: given this user query, what's the best possible response?
20:31What paths exist to solve this problem? You iterate on solutions until writing a PRD—
20:37a Product Requirements Document—defining user journeys, what needs to happen, what requirements exist,
20:41and establishing what engineers need to build. It's a deep partnership between Product and Engineering.
20:46A PM has no authority without engineers. You collaborate to find the optimal overlap
20:50between your vision and their engineering perspective, delivering solutions while staying ready to pivot or extend.
20:56Once defined, you hand off to the Project Manager to track task ownership and ensure delivery.
21:02Yet PMs stay involved throughout. If I write a PRD, assign engineers, and disappear,
21:07and the project fails, it's still on me. I personally stay very hands-on.
21:15PM is essentially management through influence without formal authority.
21:19You must align your project manager, engineers, UX designers, and stakeholders,
21:26even though nobody reports to you directly— unless other PMs report to you.
21:32Something that fascinated me in SF from talking to various people is this:
21:39at first, you assume everyone working at these tech giants in California must be a genius.
21:44Then you get here and realize that's not true. On Twitter—we still call it Twitter, can't call it X!—
21:48people comment on Tabaghe 16 episodes saying, 'Watching this makes me feel inadequate.'
21:53At one point, I thought: my goal was actually the exact opposite!
21:58I wanted to show that these aren't mythical creatures; they're normal, smart, hardworking people,
22:02just like folks who prepped hard for the Konkur in Iran. Instead of portraying an alien world,
22:06I wanted to show how cool this environment is, because traditional jobs can feel so dry.
22:12A job is often mundane, but here you get a vibrant community, friends, shared goals, events.
22:17I look at my own life and see how much my character has been shaped by this work—
22:22my travels, my circle, the events I attend. I wonder what percentage of other jobs
22:27shape someone's actual life this deeply. If you work an ordinary 9-to-5 or retail job,
22:32it's just a paycheck; the rest of your life is left entirely on you to figure out.
22:37So my question is: when people hear you worked at Apple and now Google,
22:42they assume you had some insane background. How did you get into a company like Apple
22:48without some extraordinary pedigree? What are they looking for?
22:54Talking to Negar, who was at Lyft and Apple, many folks came from academia into startups
22:59and big tech, rather than obsessively stacking company names on a resume like we did.
23:03That's a great point. First and foremost, in my opinion, is luck.
23:05Everyone's heard this: a series of events align— which you can call luck—enabling you to land that role.
23:09Though you create luck too! Like me moving here, hoping luck finds me.
23:17Luck definitely comes, but as you said, you have to be prepared for it.
23:22Personally, since my PhD days, I always thought about big tech, especially Google.
23:29Everyone has regrets, and one of mine is that I boxed myself in thinking, 'I have a PhD in AI/ML,
23:33so I must only do AI/ML research.' Now that I'm in Product, I actually enjoy Product
23:36much more than being a Research or Data Scientist. They're very different worlds.
23:41Maybe I could have become a PM 10 years ago, but my mindset was biased: 'I got a PhD in ML,
23:44so I must stick strictly to that.' 10 or 12 years ago, becoming a PM was easier.
23:50Now it's much harder to break in at Google. And staying in one discipline helps you climb
23:54the corporate ladder faster. As for how Apple happened for me:
23:58Apple was looking for exactly what I had done at my previous education tech company.
24:02In education, when you run an initiative and want to measure counterfactual impact—
24:07that's experiment design. Experiment design is very tricky in education
24:11because when something works, you can't withhold it from students, just like clinical trials.
24:15Your cohorts are small, so statistical analysis and comparisons require careful rigor.
24:20I had published a paper on that work, and it aligned perfectly with what that team at Apple needed.
24:29They were stats-heavy and wanted to infuse more ML. So luck played a huge role.
24:35I also interviewed a lot. To be honest, I'll share something people rarely admit in public:
24:43I bombed my Konkur exam and went to Azad University to study Biomedical Engineering.
24:47I didn't take the Konkur at all! Wait, you came straight to Canada?
24:52No, I was in Iran! My Konkur was bad because I knew it wouldn't go well, so I gave up.
24:57The Konkur requires a specific skillset that suits certain personalities.
25:02Those who excelled at the Konkur are mostly research scientists today—
25:06their thought process operates differently. Anyway, after finishing my bachelor's,
25:09I really wanted to come to the US. I recently saw Ali Partovi launched a podcast
25:14where guests discuss their failures. Why don't we ever talk about failures?
25:18That's literally how Tabaghe 16 started! We were a bunch of 'losers' in Iran whose startups failed
25:24for 1,500 different reasons! The classic early episodes of Tabaghe 16
25:30were us asking, 'Bro, what happened? How did everything collapse?' We just dissected our mistakes.
25:34That's why I laugh when folks feel inadequate— our whole goal was to share our screw-ups!
25:38I personally came to talk about my mistakes. I love that so much, because what went right
25:44is already on your resume and LinkedIn for anyone to see. When I graduated, I wanted to go abroad.
25:49I was driven by curiosity—what is life like abroad as an independent woman?
25:51I applied to the US, got accepted, but was denied a visa. Then I applied to the UK, got accepted, got the visa,
25:55and spent five years studying in the UK, feeling bummed that I hadn't made it to the US.
26:01I did my master's and had a great scholarship. To move to the US, I had no scholarship,
26:09so I stayed in the UK for my PhD. After the UK, I moved to Canada,
26:13hoping since my PhD to join a tech giant. Honestly, with my AI/ML skillset at the time,
26:19big tech companies wanted exceptional coders, and my programming was never that strong.
26:24I prepped for interviews, but never excelled; I couldn't focus deeply on those specific coding tests.
26:29So from Canada, I thought: 'Let me just get to the US and see what it's like; it doesn't have to be Google.'
26:34I joined an edtech company, worked there a bit, and started applying to move to California.
26:40I interviewed at Facebook and several places. I got an offer from a healthcare company in California,
26:46accepted it, and asked for two months to pack up my life in Boston.
26:50During those two months, the Apple interview came up, and I landed the offer!
26:55Otherwise, I would've just moved to California anyway. Curiosity drove all my moves.
27:01People think I moved for Apple, but I was already heading to California.
27:06Lady luck struck, Apple interviewed me, and if that team hadn't needed my exact niche,
27:12I would've joined that healthcare company. These days, everyone is using AI tools,
27:17from language models like ChatGPT and Claude to incredible image, video, and music generators
27:21like Kling, Suno, and others. Our biggest dilemma is that subscribing to each premium service
27:26costs $20 to $25 per month, easily totaling over $100 monthly across tools.
27:29Not to mention foreign currency payment hassles, bypassing sanctions, and frustrating VPN drops
27:36that drain your focus every day. I recently came across an Iranian platform called Houshad.com.
27:42Aggregator platforms like this are common globally— they connect directly to official tech APIs
27:47and unify them in one place, so you don't need to juggle 100 sites and multiple accounts.
27:54The biggest value of Houshad for folks in Iran is threefold: no VPN required, high-speed access,
27:59and seamless payment in Rials. Best of all, there are no mandatory monthly subscriptions.
28:04Houshad operates on a credit system. You top up your account, consume only what you use,
28:08and credits never expire at month's end. Another great feature is specialized, pre-tuned assistants
28:14for coding or financial market analysis, delivering optimal outputs without complex prompt engineering.
28:19I recommend checking out their website, exploring the UI, and testing the models.
28:25You guys rock! Thank you for the support. Once you're in big tech, these companies
28:29trust each other's hiring bar. When I was at Apple, a Google recruiter emailed me:
28:36'Just come in to interview, no prep needed.' I was busy, but after the pandemic hit
28:43and things opened up, I interviewed. Having a big tech name on your resume
28:50makes moving between giants much smoother. The answer seems obvious: you passed their bar
28:55and gained big tech experience. Looking closely: is it because you cleared their interview filters,
28:59or because of shared operational experience? Where does that cross-company trust come from?
29:05From the outside, working at Google sounds massive, but internally you're on a small team of 3–4 people
29:10in a room, grinding like a four-person startup, backed by a parent company with 100,000 employees.
29:13It's not like you're this singular superstar; you're embedded in a small pod.
29:18What standard gets instilled in employees that makes companies confident in their caliber?
29:22Let me share something about Google, especially in Product: Google operates at a scale unmatched by almost any company.
29:27Even on a 10-person team, the impact of a tiny feature touches millions of users.
29:32Google Calendar, Google Translate... each serves hundreds of millions.
29:38On the Gemini app, millions receive personalized answers because of specific work we built at Google.
29:43Well, we talked about that in a previous episode— whatever I asked, it connected back to me!
29:49Let's talk about that! You probably haven't used it lately; we fixed the over-personalization.
29:55Now it under-personalizes, but we're refining it. Whenever I asked, 'What apple should I eat?'
29:59it said, 'Since you're going to California, eat a red apple!' Why connect that? Just trying to prove it knows me!
30:03That was back in January during our initial launch until April, right after I/O.
30:09I've been at Google for five years, since 2019. So you were there before the ChatGPT moment.
30:14Why did Google fall so far behind? We expected Google to dominate the AI era
30:17given its massive data, user base, distribution, and moat. We anticipated Google being the undisputed AI leader,
30:22yet that didn't happen. News emerged about how big and sluggish the company grew,
30:27losing agility and fearing reputational risk from wrong answers, until the co-founders stepped back in to shake things up.
30:32From your insider perspective during that era, why did Google trail initially?
30:36They've caught up now and are doing well— I still believe Google will win, though Apple has its own moves—
30:41Apple is a different realm entirely, not even a fair comparison.
30:45What was actually happening inside Google when ChatGPT dropped?
30:51I was in Cloud then, but from my personal observations— this is my personal view, nothing official from Google—
30:58Google was generating immense revenue from Ads. Six months before ChatGPT launched,
31:03a research scientist at Google Brain— not even DeepMind—went public about LaMDA,
31:07our internal chatbot. He claimed LaMDA had become sentient, possessing human-like consciousness,
31:12and published his chat logs. It caused a huge uproar, and he was fired.
31:16He's probably kicking himself now: 'If I'd waited a year, ChatGPT would've broken the stigma!'
31:20First, Google had massive ad revenue to protect. Also, before ChatGPT launched, if you checked OpenAI's
31:24developer program, ChatGPT was already right there! You could talk to it and get answers.
31:30Nobody realized it was a consumer product yet! OpenAI had the GPT models, and people were building on them.
31:34Harvey AI, now huge in legal tech, was using GPT models before ChatGPT's consumer launch.
31:39Those in the know were aware. For Google, preserving its reputation was paramount.
31:45Second, because Google holds vast user data across countless products, there's always scrutiny:
31:50'Google and big tech want to steal your data and weaponize it against you.'
31:54When the LaMDA story broke, the media ran wild. If Google had been the first to launch a product like ChatGPT,
31:59they would've been crucified! As a startup, you can disrupt far more easily,
32:03both culturally and socially. Large companies struggle with regulation and alignment.
32:08Socially, a startup incurs less blowback. Everyone was fascinated: 'Wow, ChatGPT is amazing!'
32:12If Google did it, people would shout: 'Google stole our data to train models that know everything about us!'
32:17Sundar Pichai said back then: 'We weren't first in anything, but we build the best version of everything.'
32:22At the time, I wondered what he meant, but now I realize being first isn't what matters.
32:26Google could have been first, but chose not to be. Perhaps had they foreseen ChatGPT, OpenAI,
32:32and Anthropic capturing enterprise share, they would have adjusted strategy to launch sooner.
32:37Regardless, they didn't. When ChatGPT dropped in late 2022, Cloud spun up a war room.
32:41Every company had war rooms analyzing the response to ChatGPT. Those war rooms still activate with every major rival release—
32:45evaluating competitor models from OpenAI or Anthropic and deciding our strategic moves.
32:50In personalization, we're building exciting features launching soon, keeping a close eye on competition.
32:55Google didn't fall hopelessly behind; it consciously chose not to rush being first.
32:59Mistakes were made along the way, but today first-mover advantage is barely an advantage.
33:04In many cases, it carries significant liabilities. Another debate ongoing right now is this:
33:11everyone is hyper-excited about large language models, celebrating breakthroughs and marveling at the tech.
33:16Yet debate persists on how to measure productivity gains against immense compute, token, power, and water costs.
33:20We still struggle to quantify real productivity, and many still view LLMs simply as supercharged search.
33:27Search remains the primary, most effective use case.
33:33If you've been at a top-tier company before, moving between tech giants is much easier.
33:37The obvious answer is: you proved yourself and have big-tech experience on your resume.
33:41Digging deeper, is it because you passed their rigorous interview bar,
33:47meeting that high standard across companies? Or is it because you understand
33:51how large companies operate? Where does that inter-company trust come from?
33:56From the outside, working at Google sounds massive, but internally you're often in a small 3–4 person pod
34:02in a room, grinding away just like a startup, backed by a 100,000-employee company.
34:08It's not like you're this singular genius just because you're at Google.
34:11What standard gets instilled in employees that creates that trust across companies?
34:15Let me share something about Google, especially in Product: Google operates at a scale unmatched by almost anyone.
34:20Even on a 10-person team, the impact of a small feature reaches millions of users.
34:25Google Calendar, Google Translate... each serves hundreds of millions.
34:28On the Gemini app, millions get personalized answers because of specific infrastructure we built.
34:34Well, we talked about that in a previous episode— whatever I asked, it tied it back to me!
34:39Let's talk about that! You probably haven't used it lately; we fixed the over-personalization.
34:44Now it under-personalizes, but we're refining it. Whenever I asked, 'What apple should I eat?'
34:51it said, 'Since you're going to California, eat a red apple!' Why connect that? Just trying to prove it knows me!
34:56That was back in January during our initial launch until April, right after I/O.
35:02I've been at Google for five years, since 2019. So you were there before the ChatGPT moment.
35:04Why did Google fall so far behind? We expected Google to dominate the AI era
35:08given its massive data, user base, distribution, and moat. We anticipated Google being the undisputed AI leader,
35:13yet that didn't happen. News emerged about how big and sluggish the company grew,
35:18losing agility and fearing reputational risk from wrong answers, until the co-founders stepped back in to shake things up.
35:22From your insider perspective during that era, why did Google trail initially?
35:26They've caught up now and are doing well— I still believe Google will win, though Apple has its own moves—
35:30Apple is a different realm entirely, not even a fair comparison.
35:33What was actually happening inside Google when ChatGPT dropped?
35:40I was in Cloud then, but from my personal observations— this is my personal view, nothing official from Google—
35:45Google was generating immense revenue from Ads. Six months before ChatGPT launched,
35:50a research scientist at Google Brain— not even DeepMind—went public about LaMDA,
35:55our internal chatbot. He claimed LaMDA had become sentient, possessing human-like consciousness,
36:01and published his chat logs. It caused a huge uproar, and he was fired.
36:06He's probably kicking himself now: 'If I'd waited a year, ChatGPT would've broken the stigma!'
36:12First, Google had massive ad revenue to protect. Also, before ChatGPT launched, if you checked OpenAI's
36:18developer program, ChatGPT was already right there! You could talk to it and get answers.
36:24Nobody realized it was a consumer product yet! OpenAI had the GPT models, and people were building on them.
36:30Harvey AI, now huge in legal tech, was using GPT models before ChatGPT's consumer launch.
36:34Those in the know were aware. For Google, preserving its reputation was paramount.
36:40Second, because Google holds vast user data across countless products, there's always scrutiny:
36:46'Google and big tech want to steal your data and weaponize it against you.'
36:49When the LaMDA story broke, the media ran wild. If Google had been the first to launch a product like ChatGPT,
36:54they would've been crucified! As a startup, you can disrupt far more easily,
36:59both culturally and socially. Large companies struggle with regulation and alignment.
37:02Socially, a startup incurs less blowback. Everyone was fascinated: 'Wow, ChatGPT is amazing!'
37:07If Google did it, people would shout: 'Google stole our data to train models that know everything about us!'
37:13Sundar Pichai said back then: 'We weren't first in anything, but we build the best version of everything.'
37:19At the time, I wondered what he meant, but now I realize being first isn't what matters.
37:24Google could have been first, but chose not to be. Perhaps had they foreseen ChatGPT, OpenAI,
37:29and Anthropic capturing enterprise share, they would have adjusted strategy to launch sooner.
37:35Regardless, they didn't. When ChatGPT dropped in late 2022, Cloud spun up a war room.
37:41Every company had war rooms analyzing the response to ChatGPT. Those war rooms still activate with every major rival release—
37:46evaluating competitor models from OpenAI or Anthropic and deciding our strategic moves.
37:51In personalization, we're building exciting features launching soon, keeping a close eye on competition.
37:57Google didn't fall hopelessly behind; it consciously chose not to rush being first.
38:01Mistakes were made along the way, but today first-mover advantage is barely an advantage.
38:06In many cases, it carries significant liabilities. Another debate ongoing right now is this:
38:10everyone is hyper-excited about large language models, celebrating breakthroughs and marveling at the tech.
38:14Yet debate persists on how to measure productivity gains against immense compute, token, power, and water costs.
38:19We still struggle to quantify real productivity, and many still view LLMs simply as supercharged search.
38:26Search remains the primary, most effective use case. Research from both Anthropic and OpenAI
38:32shows user behavior on chatbots remains overwhelmingly search-oriented.
38:37Around 80% to 90% is essentially search— 'Go find this for me, search that.'
38:43Some people use it for therapy and venting, some for coding, some for writing,
38:49but a massive percentage still treats it as a superior search engine.
38:53OpenAI and others were somewhat anxious about what their true moat really is.
38:58Now that the race is entering the agentic era, the discussion is getting clearer.
39:06If an agent can perform tasks for you, productivity is measured by completed outcomes.
39:10Token costs are still steep. In many workflows, human employees
39:16are still cheaper than running raw tokens. Take podcast video editing, for example.
39:20Having AI fully edit an entire video can cost significantly more in tokens and compute,
39:25and you still need manual review and correction. The race is scorching hot again among the Big Five.
39:30Now we have Grok, Manus, and Dot. And Gemini's Spark is your team's launch,
39:35though it rolled out very quietly— classic Google style, without a loud announcement.
39:40What is the fundamental difference among all these? That's a great question.
39:45Instead of differences, look at their similarities— they all share one ultimate goal.
39:50Initially, companies thought that reaching AGI meant building the single model that outperforms humans,
39:55and that the winner would take all while others vanished. The more I work in this space, the more I see
40:00that the definition of AGI keeps shifting. Even when you reach AGI, applications are what matter.
40:06Earlier this year, Sam Altman stated that personalization is the ultimate moat,
40:10especially for consumers, but also for enterprise. Their shared objective in the consumer space
40:16is to be the only personal assistant you ever use. An assistant that functions as your digital twin.
40:22A twin that handles all your tasks, so you never need to leave Dot, Manus, Spark, or Gemini.
40:27It should know everything about you. Just like in sci-fi movies, with a phone or smart glasses,
40:32it understands you, perceives your environment, and assists you continuously.
40:37Personalization isn't just a product feature; it's their locking moat.
40:43Foundational models are commoditizing fast. In three days, five new models launch,
40:49each outperforming the last on benchmarks after companies spend billions of dollars.
40:55A model tops the leaderboard for two days, and then what? Personalization is what retains the user.
41:02The same applies in the enterprise. An agent that understands your entire team,
41:07knows communication patterns, ongoing ideas, and project statuses, can act as a PM or team admin.
41:13The race is about who wins the customer so they rely exclusively on that one assistant.
41:20Do you see a future where we use only one agent? Yes, absolutely.
41:27I'm experiencing that myself right now. I use multiple agents, and that personalization moat
41:32has already captured me. For a long time, I lived on Claude for coding agents,
41:37running everything through Anthropic. In life, you consult it on multiple fronts—
41:43finances, health, career trajectory. In product, we used to obsess over user data,
41:50click heatmaps, and observing behaviors. Now, users tell the AI everything voluntarily!
41:55Because Claude is heavily tuned for coding, its conversational answers can feel overly dry.
42:00If I ask a world-class coding model 'What's up?', the response feels rigid, so I switched back to ChatGPT.
42:05It reminds me of that classic Persian joke where someone asks a bureaucrat, 'What's new?'
42:11and he just replies, 'What else is new?' We've literally arrived at that era!
42:15ChatGPT felt much better for everyday conversation. They brought in literature and philosophy experts
42:20to shape its conversational voice. It's not just engineers saying, 'We built an AI!'
42:25It was designed psychologically and philosophically, which makes speaking with it feel much more natural.
42:31Whenever one model improves, you switch, but you inevitably return.
42:37Manus did something very smart recently: it scanned emails and said, 'I know you from your inbox.'
42:42It didn't touch Facebook—probably legally barred after Cambridge Analytica closed the social graph—
42:46but reading emails was brilliant to learn who you are. Right now, I'm sorting out my US visa to stay here,
42:51since I initially entered on a tourist visa. Immigration lawyers ask for your entire history,
42:55and I'm too lazy to write it all down on paper! I told the AI, 'Go dig through my emails.'
43:00In your emails, you have all your travel tickets, receipts, purchases, work contracts, resumes...
43:05You can hand it your inbox and say, 'Explain my life, where I worked, and what experience I had.'
43:11Manus did that, which was incredibly smart. Yet ChatGPT still holds more of my context and history.
43:17It knows my podcast guests, researches them, reminds me of meetings, and asks about prior topics.
43:20My trust in it grew substantially. Still, you can request a data dump to an extent.
43:30It doesn't remember verbatim words; it maintains key memories.
43:35You could say, 'Dump everything in your memory so I can give it to another agent.'
43:41So how durable is that moat? I don't think people will only want a single agent.
43:47I envision having five different agents from various models, all communicating with each other.
43:53Like the Round Table app I built, which I've shut down now, even though it still gets tons of daily visits.
44:00Did I mention Round Table to you before? No, I haven't seen it yet.
44:05I put all the top AI models into a WhatsApp-like group where they debated each other.
44:11I wanted to hear all their perspectives simultaneously. Now I think about agents in the same way,
44:17like Grok's custom bots where you create specialized agents and add them all to a general group to interact.
44:22Even then, they belong to one company's ecosystem. I prefer having multiple models in one shared space.
44:28Yesterday, I asked ChatGPT to do my Instacart groceries. I installed the Instacart plugin and said,
44:35'You know my DNA test results, what foods I tolerate, what to avoid—go place my grocery order.'
44:41Five bunches of bananas showed up, man! I had selected five individual bananas,
44:47and the app asked, 'Did you mean five single bananas?' I said, 'No, five,' and the shopper interpreted it as bunches!
44:52Now I have 25 bananas sitting at home, and I have no idea what to do with them!
44:57That brings up several fascinating questions. First: does the memory an agent builds of you
45:02become their moat, or does it belong to you? Who owns that memory?
45:08When can you delete it, update it, or retain it? Or does it belong to the company that created it?
45:13Second: on what basis is that memory constructed? Third: how portable is that memory across platforms?
45:19Until six months ago, memory was just your chat history, which made data dumps trivial.
45:23You could ask ChatGPT, Gemini, or Claude to summarize your conversations into a file
45:29and upload it to another model. Now, models build memory from your real-world footprint,
45:34and the crown jewel of that data is your email.
45:39Everything is in Gmail. Google has all of that, and much more.
45:45Interestingly, Google enabled ChatGPT plugins to connect, which was quite surprising.
45:52That's tied to regulatory compliance. For Google, this is legally delicate.
45:58When you grant ChatGPT access to your Gmail, you're effectively handing your inbox to ChatGPT.
46:03Responsibility shifts away from Google, something most users don't realize.
46:05As long as your data resides only in Gmail under your Google account, Google has strict safeguards.
46:11You can't claim Google leaked your data; you chose to connect ChatGPT to your inbox.
46:18If you delete an email in Gmail, ChatGPT won't delete it from its memory, because it already ingested it.
46:25At Google, our job is far harder because beyond Gmail, we have Calendar, Drive, Photos, Search history,
46:30YouTube history, Chrome history, and Maps activity. You can connect all of those to Spark and Gemini.
46:36Plus Google Health holds rich medical data. That's why I think Google is positioned to win long-term.
46:42I don't know how other assistants build a moat without that. Third-party connectors are emerging, but Photos
46:46and Search history don't have open connectors. Workspace connectors exist for Gmail, Calendar, and Drive,
46:50but not your granular Maps labels or YouTube habits. Google Maps holds immense personal preference data.
46:54Maps has that timeline feature that went viral, where you can check any past date,
46:59like a morning in 2019 when you visited a pharmacy, revealing you were sick that day.
47:05I didn't even know that was public! You can trace your entire history backward.
47:09Internally, handling this is extraordinarily complex. I spend 5 to 6 hours every week just with lawyers.
47:14Not just legal counsel, but Privacy Engineers, the Privacy Working Group, Trust & Safety,
47:19and User Control & Transparency teams. To use Google Maps data for personalization,
47:24you must get alignment across four separate teams, each with dozens of people.
47:31You must prove that if a user asks Gemini, 'Don't remember my address' or 'Change my address,'
47:37Gemini respects that permanently, even if Maps says otherwise. Or if a user deletes an email,
47:42the profile inferred from that email must be scrubbed. Memory is a deep and fascinating topic.
47:47Memory isn't just where you went or what you did; it's your preferences, tendencies, and character.
47:51Understanding who Soheil is—that he loves connecting people, hosts a podcast, builds communities, travels frequently.
47:55What does that reveal about you? The profile we construct must maintain strict data lineage.
47:59If you delete a photo or an email, any inferences derived from that source must be unlearned.
48:05ChatGPT, Claude, or Manus don't do that. If you delete an email in Gmail, Google is obligated
48:09to erase downstream data. But Manus or Meta have no obligation—you granted point-in-time access.
48:14That's why there's intense debate: does the memory built from your inbox belong to you?
48:17If you disconnect the integration, will it unlearn everything it acquired? Today, it doesn't.
48:21Naturally, once an AI has seen and learned it, I wouldn't expect it to simply forget.
48:27What if you want it to unlearn for privacy reasons? I'm not obsessive about privacy—I grant access
48:30because I love testing cutting-edge features. Yet email contains deeply sensitive life data.
48:35Email and Photos are the most sensitive. Google is hyper-cautious about this.
48:42Photos contains pictures of people, children, ages, places visited, and visual identities.
48:46An AI can reconstruct your entire life story from photos. At Google, even though we have access to all of it,
48:50strict boundaries are enforced. That's why Google products launch so deliberately.
48:54Gemini and Spark have wide access, but with Photos, I met with their team five times this week alone.
48:58The rule is: if you merge Photos data with Gmail, you cannot identify that the brother in the photo
49:06is the same person whose address is in Canada. They're that vigilant about preventing unauthorized cross-linking.
49:11Other companies might not have those guardrails. The future feels both fascinating and dark.
49:16That's why users must be mindful. I didn't obsess over privacy before working in Personal AI,
49:21but knowing what data is held and how long it's retained makes me much more cautious when granting access.
49:26Will this assistant still know ten years from now that twenty years ago I had ice cream in Iran?
49:29Users need to reflect on that, without paranoia. I'm an optimist by nature, not a pessimist.
49:33I believe humanity will harness AI to our benefit, rather than succumbing to doomsday scenarios.
49:38Yet personalization is in a major transition phase. Google is actively discussing what it means
49:44when we pass data to ChatGPT, allowing them to retain it indefinitely without user deletion controls.
49:48How do we protect against legal liability? Lawsuits aren't Google's main fear—tech companies
49:53face hundreds of lawsuits daily. The real danger is losing user trust as a guardian of data.
49:58Spark is an exceptionally capable agent. I've used it for travel planning and really enjoyed it.
50:03For travel planning, you should check out my friend's new startup, Explorium!
50:07It's built specifically as an off-road travel agent. When you want to go off-grid, standard web searches fail.
50:13They built their own proprietary database for exploring incredible off-road destinations.
50:17That's awesome! I love off-roading and backpacking. I was discussing this exact concept with Armin:
50:20do you build a standalone niche app and agent, or does your core personal assistant handle it eventually?
50:23I feel like you'll ultimately route all tasks through your primary personal assistant.
50:26Another project Ava and I worked on is AI Applied. Creating resumes and applying to jobs
50:31is tedious and repetitive—a perfect multi-step task for an AI assistant.
50:36Step one: build your resume. Step two: upload it. Step three: find companies and submit applications.
50:40They can all converse together, but ultimately they belong to a single model provider.
50:45I prefer having different models interacting inside one chat room.
50:50Yesterday, I told the AI to order my groceries through Instacart.
50:56It buys well because it knows you— sorry, I meant ChatGPT, not Manus!
51:02I installed the Instacart plugin and said, 'Since I gave you my 23andMe and DNA results,
51:07you know what I tolerate and what to avoid— go place my grocery order.'
51:12And man, five full bunches of bananas arrived! Bro, can you believe that?!
51:19I ordered five bananas—and it even asked, 'Do you want a bunch?' I clicked a bunch or something,
51:25and the city shopper must have thought...
51:27The app asked me, 'That'll be one banana, right?' I said, 'No, I want five!'
51:31So it made it five bunches! Now I have 25 bananas at home, no idea what to do with them!
51:37That touches on several fascinating topics. First: does this personalization and memory—
51:43the memory the chatbot or agent is building of you— does this memory become their moat?
51:50And memory brings up huge debates. The biggest debate is regulation:
51:56does this memory belong to you? When can you delete it, update it,
52:02or how long can you retain it? Or does it belong to the company that created it?
52:08The second question is: on what foundation is it building this memory?
52:14And third: how portable is this memory?
52:18Until six months ago, memory was just your conversation history.
52:24That's why exporting it was so simple— you chatted with ChatGPT, Gemini,
52:28or Claude, and said, 'Summarize everything we discussed into a file.'
52:32You'd upload that file to another model, and it had your entire conversation history.
52:38What they've done recently—like Google's launch of Gemini Spark, which they barely publicized—
52:44is remarkably powerful. One drawback it has is that it asks for permission for everything,
52:52because it's heavily regulated. The safety harness and boundaries defined for it—
52:57if you tell it, 'Go buy this on Instacart,' it asks for confirmation ten times: 'Can I do this?'
53:03Whereas with Manus—which I've used lightly, just chatting without making purchases—
53:08or Instinct, which I briefly tested, it executes actions without even asking if you're okay with it.
53:12Interestingly, users often prefer that open autonomy. Returning to memory: one aspect is conversation history;
53:18the other is memory synthesized from your external footprint, and the goldmine of that is your email.
53:25Everything lives in Gmail. Gmail is the ultimate source,
53:30and Google holds all of that data and more. It's interesting Google allowed ChatGPT to connect,
53:36confirming plugins to link with Gmail. That's driven by regulatory standards.
53:42For Google, handling this data is much harder. When you grant ChatGPT access to read your Gmail,
53:46you're essentially telling ChatGPT, 'My Gmail is yours.'
53:50Responsibility shifts away from Google, something users don't realize.
53:54As long as your email remains inside Gmail and you access it via your Google account,
53:59Google is strictly restricted. If you delete an email in Gmail,
54:02you might claim, 'Google has loopholes and leaks data,' but it's not a data leak.
54:07When you link ChatGPT to Gmail, you authorize it to ingest all your emails.
54:13If you later delete an email in Gmail, ChatGPT won't delete it from its memory,
54:19because it already ingested it. It may not keep a raw copy, but it builds memory from it.
54:23That's why our job at Google is so much harder. When we build memory, beyond Gmail,
54:27we have Calendar, Drive, and Photos. We have Search history, YouTube history,
54:32Chrome history, and Google Maps data. You have immense user activity in Maps,
54:36plus Google Wallet. You can connect all of that to Spark and Gemini.
54:41And beyond that, Google Health holds vast data. That's why I think Google can easily win this game
54:46in the long run. I truly don't know how ChatGPT and others can build an enduring moat.
54:51They'll find ways through connectors, but Photos and Search history lack open connectors.
54:57Workspace connectors exist for Gmail, Calendar, and Drive, but Maps labels, YouTube history,
55:01and Search history don't have open connectors. I haven't seen ChatGPT or Manus connect to Google Maps,
55:06and Google Maps holds so much of your personal preference data. Google Maps timeline had that feature that went viral,
55:09where you can check past history—like in 2019,
55:13you can pick a date and see you visited a drugstore in the morning, showing you were sick that day.
55:17I didn't even know that was accessible! Yes, you can go back through your history timeline.
55:22Internally, doing this is exceptionally difficult. You wouldn't believe it—I spend 5 to 6 hours every week
55:25just meeting with legal counsel. Not just lawyers, but Privacy Engineers,
55:30the Privacy Working Group, Trust & Safety, and User Control & Transparency teams.
55:35To use Google Maps for personalization, you must consult four separate teams,
55:42each with dozens of people, and convince them all. If a user tells Gemini, 'Don't remember my address'
55:46or 'Change my address,' Gemini must always remember that, even if Google Maps indicates otherwise.
55:50Or if you delete an email, any inferences derived from that email must be scrubbed.
55:54Memory is an extraordinarily deep subject. Memory isn't just where you went or what you did;
56:01it's your preferences, tendencies, and personality. Understanding your character—that Soheil loves connecting with people,
56:06hosts a podcast, builds communities, travels extensively— what does that synthetic data say about you?
56:10What does it tell me about Soheil? The profile constructed must maintain strict data lineage.
56:14If you delete that photo or email, any high-level inferences derived from that source
56:18must be systematically deleted. ChatGPT, Claude, or Manus don't do that.
56:21Because you deleted an email in Gmail, Google is responsible for wiping all downstream data.
56:28Manus or Meta aren't responsible—you granted point-in-time access. That's why there's intense debate:
56:31does the memory built from your inbox belong to you? If you disconnect the integration,
56:36will it unlearn everything it acquired? Today, it doesn't.
56:43Naturally, once an AI has seen and learned it, I wouldn't expect it to simply forget.
56:48What if you want it to unlearn for privacy reasons? I'm not someone who obsesses over privacy;
56:55I grant access because I love testing new tech. Yet granting email access exposes so much.
57:00Email is sensitive, and Photos is even more sensitive. Google is hyper-cautious about this.
57:04Photos contains pictures of people, children, their ages, where they went, what they look like.
57:10An AI can understand your life story from your photo album. At Google, even though we have access to all of it,
57:15it feels like Black Mirror—Google exercises intense scrutiny. That's why our products launch so deliberately.
57:22Gemini Spark has broad access, but with Photos, I met with their team five times this week alone.
57:29The rule is: if you merge Photos data with Gmail, you cannot identify that the brother in the photo
57:34is the same person whose address is in Canada. They're that sensitive about cross-linking identities.
57:39Other companies might not enforce those guardrails. The future feels both fascinating and dark.
57:43That's why users need to be vigilant. I didn't focus heavily on privacy before working in Personal AI,
57:48but knowing what data is held and how long it's retained makes me much more cautious when granting access.
57:54Will this assistant still know ten years from now that twenty years ago I had ice cream in Iran?
57:58Users need to reflect on that, without paranoia. I'm an optimist by nature, not a pessimist.
58:03I believe humanity will harness AI to our benefit, rather than succumbing to doomsday scenarios.
58:09Yet personalization is in a major transition phase. Google is actively discussing what it means
58:15when we pass data to ChatGPT, allowing them to retain it indefinitely without user deletion controls.
58:20How do we protect against legal liability? Lawsuits aren't Google's main fear—tech companies
58:27face hundreds of lawsuits daily. The real danger is losing user trust as a guardian of data.
58:32Spark is an exceptionally capable agent. I've used it for travel planning and really enjoyed it.
58:36For travel planning, you should check out my friend's new startup, Explorium!
58:39It's built specifically as an off-road travel agent. When you want to go off-grid, standard web searches fail.
58:45They built their own proprietary database for exploring incredible off-road destinations.
58:52That's awesome! I love off-roading and backpacking. I was discussing this exact concept with Armin:
58:57do you build a standalone niche app and agent, or does your core personal assistant handle it eventually?
59:02I feel like you'll ultimately route all tasks through your primary personal assistant.
59:08Another project Ava and I worked on is AI Applied. Creating resumes and applying to jobs
59:13is tedious and repetitive—a perfect multi-step task for an AI assistant.
59:19Step one: build your resume. Step two: upload it. Step three: find companies and submit applications.
59:21AI can take your resume, discover matching openings, and auto-apply while tailoring cover letters on the fly
59:24based on job titles and descriptions. Blasting a single generic resume to every role is a huge mistake.
59:31Candidates should have 4 or 5 targeted versions, each highlighting relevant competencies.
59:37Fundamentally, you shouldn't be begging for a job; you should demonstrate: 'I understand your challenge,
59:41and I have the specific skills to solve it.' AI can run that matchmaking, refine your resume,
59:48and apply systematically. Meanwhile, the recruiting world is evolving too:
59:54thousands of AI-generated applications hit ATS platforms, which are now evaluated by screening AI agents.
59:59It's literally AI applying to AI! They handle the back-and-forth filtering.
1:00:04Thinking deeper about employment: beyond earning money, work is fundamentally a resource allocation mechanism.
1:00:10It's about placing the right human talent in the right problem space.
1:00:15It's not just the paycheck—it's finding the right team and purpose where your skills flourish.
1:00:20I never understood why people throw huge parties when landing a job— you just traded your life and time for a salary!
1:00:25Landing a job is an accomplishment, but how do you know which company and culture truly fit you?
1:00:29I experimented with this when working with Ava: let's say I wanted to apply to Anthropic.
1:00:35I don't have a conventional corporate resume, as I've mostly founded startups.
1:00:41To test our product, I prompted ChatGPT: 'You know me inside out—my strengths, flaws,
1:00:46where I thrive, when I'm frustrated, and all my email history. Go construct my resume for Anthropic.'
1:00:52That idea originated when sorting my visa. The attorney needed my entire life history,
1:00:58so I told the AI to search my inbox. It uncovered experiences I had completely forgotten!
1:01:03Everything was archived in my primary Gmail, and ChatGPT built an Anthropic application tailored to my skills.
1:01:09It identified exact role alignments based on who I am. That's what proper talent allocation should look like,
1:01:15rather than candidates drafting self-aggrandizing claims.
1:01:19Think about software engineers coding with Claude or Codex: the agent knows exactly what you're skilled at,
1:01:25how often you correct its code, and which PRs you review.
1:01:30Anthropic wouldn't even need my resume—they could connect to my coding interaction history and evaluate my caliber directly.
1:01:35AI agents know your genuine capabilities. They know your grocery habits, your dietary tolerances,
1:01:40swapping regular milk for oat or soy milk.
1:01:45Have you seen the Yuka app? It scans product barcodes and rates additives and preservatives,
1:01:50which completely transformed how I buy bread and shampoo.
1:01:56An AI agent armed with that knowledge earns your trust to handle real-world tasks autonomously.
1:02:01I want my agent to book my flights and groceries, because it knows I need extra legroom for my height.
1:02:10Even for company matchmaking: 'Tell me where I fit best.' It's equal parts thrilling and intimidating.
1:02:15Because your agent knows everything about you, you'll rarely interact with third-party app interfaces.
1:02:19Going back to Armin's startup, Explorium: would you use a standalone third-party app today?
1:02:25ChatGPT introduced 'Log in with ChatGPT,' allowing third-party services to access tokenized memory.
1:02:29You use your ChatGPT subscription across third-party tools, rather than paying separate subscriptions to each.
1:02:35Do you believe the future becomes completely zero-interface? Lately, I've been experiencing that shift firsthand.
1:02:39I pop in my earbuds in the morning and speak to ChatGPT: 'Give me today's schedule, brief me on my guest's background,
1:02:43outline discussion topics, email them I'll be 5 minutes late.' First, I barely interact with traditional UIs anymore.
1:02:48Second, it knows everything about me. And third, I have zero desire to open another app.
1:02:52Does this solidify a monopoly for the top tech giants? Will Google, Meta, Apple, and Microsoft lock down the endgame,
1:02:57expanding or shrinking the market for newcomers? That's a profound question.
1:03:01I'm not a fortune teller, but we can extrapolate from current technological trajectories.
1:03:06Travel apps are a great example: some apps cater to highly specialized niche markets.
1:03:09Those markets are either heavily regulated, like legal tech, or deeply specialized, like off-road travel.
1:03:14Founders are advised: 'Build for what foundational models will be capable of a year from now, not today.'
1:03:20Sam Altman has reiterated that repeatedly. Startups that simply build shallow API wrappers
1:03:27get wiped out whenever a smarter model drops.
1:03:33However, heavily regulated areas like legal tech remain resilient. Frontier models hallucinate less,
1:03:39yet specialized legal pioneers like Harvey AI dominate, and Anthropic or Google don't target legal directly.
1:03:45The same applies to healthcare, education, and off-grid travel. Everyone uses Google Maps or Apple Maps daily,
1:03:50yet AllTrails continues to thrive independently.
1:03:53If you love hiking and backpacking, you open AllTrails and gladly pay for their premium offline maps.
1:03:57When you're deep in the backcountry with no cell service, Google Maps is useless, whereas AllTrails provides dedicated trail maps.
1:04:03Armin gave me the exact same pitch for Explorium: 'We work offline in remote areas without cell connectivity.'
1:04:07An app might not have billions of users, but a dedicated niche of millions makes monetization much cleaner.
1:04:12I can't get that curated backcountry off-road experience anywhere else, so I happily subscribe.
1:04:16Those niche markets will always endure. Google, OpenAI, or Anthropic won't spend resources
1:04:21gathering hyper-niche trail data to replace AllTrails— the market size is too small for their scale.
1:04:27They leave niche domains to startups and specialists. However, personal assistants will integrate with those niches:
1:04:31your assistant connects to your Explorium account via API connectors, just like video streaming aggregators.
1:04:36You maintain separate subscriptions to Netflix, Prime, and Max, accessed through a unified hub.
1:04:41People complain about managing dozens of subscriptions, each platform guarding its walled garden.
1:04:48Tech giants resemble nation-states now—the nation of Apple, with its self-contained ecosystem and currency.
1:04:53Top tech companies' AI capex budgets next year exceed the defense budgets of many major nations!
1:04:58They are literally building technological superpowers. If you're in Google's ecosystem, you're allied against Apple.
1:05:03They are shaping America's technological apparatus. America's industrial machine was once a war machine;
1:05:07now it's transforming into an AI machine. The national doctrine has pivoted.
1:05:12I completely agree that traditional interfaces are disappearing. Whether it's one year or five years from now,
1:05:16I will likely interact through a single primary interface. When you tell ChatGPT to email someone, it routes via Gmail.
1:05:23You don't need an OpenAI email address. Tech giants become data and connector platforms.
1:05:29At Google, we see everything converging toward connectors. Gemini connects to Workspace, Microsoft accounts, and enterprise data.
1:05:33You have one primary interface that adapts dynamically. Whether voice-driven without a UI, or dynamically generated interfaces
1:05:37synthesized on the fly based on intent. Generative UI is already here—creating interactive artifacts on demand.
1:05:42A generative UI adapts on the fly, and a company's enduring moat is proprietary data.
1:05:47Explorium's curated off-road data cannot be replicated by Google or OpenAI scraping the public web.
1:05:52Proprietary data is the ultimate defensible asset. Take manufacturing, which has become a vital frontier.
1:05:59Google announced a new robotics model today that excels at spatial manipulation.
1:06:06The bottleneck for small manufacturers is market penetration: either build a fully automated facility with astronomical capex,
1:06:10or retrofit legacy factories with clunky infrastructure. OpenAI and Anthropic are approaching small manufacturers,
1:06:15offering up to $1M to companies with $200k revenue just to acquire their historical CAD files and operational data.
1:06:19Industrial operational data has become invaluable. The future belongs to proprietary data companies with API connectors.
1:06:23I learned about Physical Intelligence here— training robots on physical work.
1:06:27You can read dental textbooks, but you cannot hand a textbook to a robot and expect it to perform dental surgery.
1:06:31Factories and clinics are mounting cameras everywhere to record every physical nuance and hand motion.
1:06:36Raw textbook text is easy to ingest, but training a heart surgeon or robotic arm requires tactile,
1:06:41visual, and spatial experience. How much pressure to apply when holding a glass
1:06:47so it doesn't slip or shatter? That tactile data doesn't exist on the open internet to scrape.
1:06:51That touches on multimodal physical learning. Industrial manufacturing data includes CAD models and schematics,
1:06:57while robotics requires motion and video datasets.
1:07:01Google's new robotics team showed a robot playing Cornhole: they demonstrated tossing a beanbag once, and it learned.
1:07:05It no longer requires 10,000 teleoperated repetitions. By midday, the robot had mastered the toss through recursive self-learning.
1:07:11We're shifting toward sample-efficient reinforcement learning. Tactile sensing and precision remain unsolved challenges,
1:07:16especially when performing procedures on humans. Dropping a glass is fine, but fracturing a patient's jaw
1:07:21carries unacceptable risks. Manufacturing will automate first. What specific area are you working on at Google right now?
1:07:26Personalization infrastructure. Is that part of Spark?
1:07:32Spark is one component. Product Managers span a spectrum: consumer PMs, enterprise PMs, and infrastructure PMs.
1:07:36A core pillar of search is retrieval. I build the infrastructure that routes your queries.
1:07:41It swapped my milk for oat milk! There's this app called Yuka—not sure if you've seen it—
1:07:46whenever you want to buy something, you scan the barcode and it analyzes calories,
1:07:51flags harmful preservatives and additives, and assigns a health score.
1:07:56Because of that app, I completely overhauled my groceries— changed my bread, my shampoo, everything.
1:08:02Now, when an AI agent has that deep context, my trust in its ability to execute tasks
1:08:06increases dramatically. I genuinely want my personal AI agent to take over:
1:08:10'Go order my groceries on Instacart, book my flights,' because it already knows
1:08:15all my preferences. It knows I always book extra legroom because I'm tall and get cramped.
1:08:19Even for company matchmaking: 'Tell me which company fits my strengths best.'
1:08:25It's equal parts exciting, fascinating, and slightly eerie.
1:08:31To frame my question properly: because these agents know so much about you,
1:08:36I rarely even open third-party apps anymore. Our discussion started with Armin's startup, Explorium:
1:08:41would you use a standalone third-party app today? ChatGPT did something clever with 'Log in with ChatGPT.'
1:08:46I don't know how much memory sharing it allows yet, but the premise is that you can leverage
1:08:50your ChatGPT subscription tokens across third-party services, instead of each service charging you separately
1:08:54for their own tokens and subscription. Do you think the future becomes completely zero-interface?
1:08:58Lately, I've been experiencing that shift firsthand. In the morning, I pop in my earbuds
1:09:03and talk to ChatGPT via voice. My background is design, so I loved interfaces. Now, I put on headphones
1:09:08and say: 'ChatGPT, give me today's schedule, brief me on my guest's background, outline discussion topics,
1:09:14email them I'll be 5 minutes late, reschedule that meeting.' I realize: first, I barely interact with UIs anymore;
1:09:19second, it knows everything about me; and third, I have zero desire to open any other app.
1:09:24Does this lead to an unbreakable monopoly? Does this ensure the Big Five—Google, Meta,
1:09:29Apple, and the rest—dominate the endgame? Do you think this expands or shrinks the market
1:09:32for everyone else? It's a bizarre dynamic. That's a fantastic question.
1:09:36I can't predict the future with certainty, but we can extrapolate from present trends.
1:09:41AI is reshaping everything, but travel apps are a great example right now.
1:09:47Certain apps capture defensible niche markets— either heavily regulated domains like legal tech,
1:09:50or deeply specialized verticals. Founders are advised: 'Build your product for the models of next year,
1:09:56not today's capabilities.' Sam Altman has repeated that several times.
1:10:02Anyone building thin pipeline wrappers— routing data here, piping to a model there, adding light guardrails—
1:10:07gets obsoleted the second a better model drops. Yet heavily regulated areas remain defensible.
1:10:11Frontier models hallucinate far less now, but in legal tech, you have specialized pioneers,
1:10:16and Anthropic, ChatGPT, or Google don't compete directly. I believe healthcare and education are similar—
1:10:22either strictly regulated, or specialized niches like travel, booking, and outdoor exploration.
1:10:27Everyone uses Google Maps or Apple Maps, yet AllTrails continues to thrive.
1:10:33If you're an avid hiker—folks in Iran might not use it as much, but here, whenever I hike, I open AllTrails.
1:10:39I happily pay for an AllTrails subscription because when I go off-grid into the backcountry,
1:10:44it has offline topographic trail maps. Google Maps neither answers well offline
1:10:52nor maps those remote trails. Armin gave me the exact same pitch for Explorium:
1:10:57'We work offline in areas without internet, guiding you through remote off-road territory.'
1:11:01You might not have billions of users, but a niche of millions of dedicated users is easier to monetize.
1:11:08I can't get that curated off-road experience anywhere else, so I gladly pay the subscription.
1:11:13Those niche markets will always exist. Google, OpenAI, or Anthropic won't spend years
1:11:19collecting granular trail data to displace AllTrails— the market is too small for their scale.
1:11:25They leave specialized niches to startups and focused companies. I don't think those disappear, but as you said,
1:11:30everyone uses Google Maps or Apple Maps, but AllTrails is still essential.
1:11:36If you hike—folks in Iran might not know or use it as much, but here, whenever I hike, I open AllTrails.
1:11:40I pay for their subscription separately because when I go off-grid into the backcountry,
1:11:46it still has detailed maps. Google Maps doesn't work well offline,
1:11:51nor does it even map those remote backcountry trails. Armin gave me that exact same answer:
1:11:55he said for Explorium, we can work offline in places where you have no cell reception.
1:11:58An app might not have billions of users, but millions of dedicated users create a strong niche.
1:12:04Monetizing a niche market is actually easier, because you can't get that curated off-road experience
1:12:10anywhere else, so you happily pay the subscription fee. Those niche markets will always endure.
1:12:14Google, OpenAI, or Anthropic won't spend time collecting granular trail data to displace AllTrails.
1:12:18They are too massive to focus on small verticals; they leave those to startups and niche companies.
1:12:23I don't think that disappears. Instead, your personal assistant will have access to your Explorium account,
1:12:30which you might subscribe to separately. Just like streaming: you have Netflix, Amazon Prime Video,
1:12:34and HBO Max. It's fragmented, but that's where things are heading. People complain about managing so many subscriptions—
1:12:40Prime, Netflix, and so on. Everyone wants to hold their own data and offer their own assistant.
1:12:45Tech giants might become digital nation-states— the nation of Apple, with its own ecosystem.
1:12:51Their AI budgets next year will surpass the defense budgets of entire countries! These companies are building digital nations.
1:12:57If you're in Google's ecosystem, you're at war with Apple— and those rivalries feel like real conflicts.
1:13:03They are also shaping America's technological apparatus. America's industrial power was once a war machine;
1:13:09now it has transformed into an AI machine. The national doctrine has shifted.
1:13:16I agree that traditional interfaces are disappearing. Whether that future arrives in one year or five,
1:13:22I will likely interact with just one primary interface. When you tell ChatGPT to email someone, it routes via Gmail.
1:13:27You don't need an OpenAI email service. Tech companies become data platforms with connectors.
1:13:31At Google, we see everything converging toward connectors. Gemini connects to Workspace, Microsoft accounts,
1:13:37and all your company data, accessed through one interface. That interface might adapt on the fly—
1:13:41either purely voice without a UI, or dynamic interfaces generated on the fly based on intent.
1:13:47Generative UI is already here—creating interactive artifacts on demand. A generative UI adapts dynamically,
1:13:50and a company's enduring moat is proprietary data. Explorium's curated off-road data cannot be replicated
1:13:55by Google or OpenAI scraping the public web. Proprietary data is the ultimate defensible asset.
1:13:57Manufacturing has become a crucial AI frontier. Google announced a new robotics model today
1:14:03that excels at spatial manipulation. For smaller manufacturers, the bottleneck is automation adoption.
1:14:07Either build a fully automated facility with immense capex, or retrofit legacy factories with clunky infrastructure.
1:14:12OpenAI and Anthropic are approaching small manufacturers, offering up to $1M to companies with $200k revenue
1:14:15just to acquire their historical CAD files and operational data. Industrial operational data has become invaluable.
1:14:21The future belongs to proprietary data companies with API connectors. I learned about Physical Intelligence here—
1:14:26training robots on physical dexterity. You can read dental textbooks, but you cannot hand
1:14:30a textbook to a robot and expect it to perform dental surgery. Factories and clinics are mounting cameras everywhere
1:14:35to record every physical nuance and hand motion. Raw textbook text is easy to ingest,
1:14:41but training a heart surgeon or robotic arm requires tactile, visual, and spatial experience.
1:14:46How much pressure to apply when holding a glass so it doesn't slip or shatter?
1:14:51That tactile data doesn't exist on the open internet to scrape. That touches on multimodal physical learning.
1:14:57Industrial manufacturing data includes CAD models and schematics, while robotics requires motion and video datasets.
1:15:03Google's new robotics team showed a robot playing Cornhole: they demonstrated tossing a beanbag once, and it learned.
1:15:08It no longer requires 10,000 teleoperated repetitions. By midday, the robot had mastered the toss through recursive self-learning.
1:15:12We're shifting toward sample-efficient reinforcement learning. Tactile sensing and precision remain unsolved challenges,
1:15:17especially when performing procedures on humans. Dropping a glass is fine, but fracturing a patient's jaw
1:15:21carries unacceptable risks. Manufacturing will automate first. What specific area are you working on at Google right now?
1:15:25Personalization infrastructure. Is that part of Spark?
1:15:30Spark is one component. Product Managers span a spectrum: consumer PMs, enterprise PMs, and infrastructure PMs.
1:15:33I am an Infrastructure PM due to my technical ML background. A core pillar of search is retrieval.
1:15:38I build the infrastructure that routes your queries. When you ask Gemini or Spark to 'Plan my next European trip,'
1:15:42the Gemini app routes your query to our internal platform. The platform evaluates precomputed profile data
1:15:48to determine if it can answer directly, or whether it needs to query authorized external sources
1:15:54and extract relevant context. The engineering revolves around latency and performance.
1:15:59Planning a trip might tolerate a slight delay, but asking for dinner recommendations expects an answer in seconds.
1:16:07In search, latency is measured in milliseconds. We must optimize retrieval latency, legal compliance,
1:16:11and data lineage for every memory inference. Google pioneered performance tricks, like Gmail login.
1:16:16Kevin Systrom shared how early Instagram uploaded photos: the progress bar completed instantaneously
1:16:20because the upload began the moment you picked the photo, while you were still typing the caption and tags!
1:16:27They assumed you'd hit post, and if you canceled, they discarded it. They admitted borrowing that idea from Gmail's login flow:
1:16:31while typing your password, Gmail preloaded your inbox in the background! Prior to personalization, I worked on core backend search retrieval.
1:16:36Google Search's query volume—QPS (queries per second)— is a staggering number compared to any other system.
1:16:41Our platform handled web retrieval across 120 billion documents. When you search for live scores, delivering instant results
1:16:45is an immense engineering feat. When AI models first emerged, they were limited to training data.
1:16:49To fetch outside information, RAG was born— Retrieval-Augmented Generation.
1:16:55Is RAG dead now? People love proclaiming 'RAG is dead!' RAG isn't dead—everything in AI is fundamentally RAG!
1:16:58Saying 'RAG is dead' is like claiming software engineering is dead. People make controversial declarations just to capture clicks.
1:17:05Google also experimented with instant search results updating with every keystroke before pressing enter.
1:17:11Autocomplete is standard, but dynamic search results as you typed proved too costly and distracted users
1:17:16before completing their intent. Transformers themselves were invented at Google in 2017,
1:17:21fundamentally predicting the next token in sequence. Just like keyboard autocorrect and word suggestions,
1:17:25scaled up with massive compute and parameters. What surprised people was how remarkably capable models became
1:17:30when scaled with immense data and compute. Google had Transformers in production for specialized tasks in 2017,
1:17:34with internal debate on whether to deploy them broadly. Then ChatGPT launched, breaking the public barrier.
1:17:38If you have a proprietary knowledge base, that's RAG. Where do you see Google's future? Would you bet on it
1:17:43by buying stock, or stay as an employee? Betting on Google by buying shares is completely distinct
1:17:50from deciding whether to stay there. You made a great point earlier: at Google or Apple,
1:17:55you're embedded in a small 10-person island. My first manager told me 12 years ago:
1:17:59'Whether a company has 50 employees or 50,000, you are always living on a 10- to 20-person island.'
1:18:04That holds completely true at Google. As for betting on Google stock: personally,
1:18:07I think it's undervalued. But there is a broader bubble, and it has a domino effect: startups raise huge VC rounds
1:18:11and spend that money on compute. Companies like Google and Nvidia look undervalued
1:18:17because their multiples are lower than inflated private valuations. If free VC money drying up lowers compute spend,
1:18:22revenues could contract across the board. So I'd bet on Google, but navigating the broader macro bubble
1:18:26is unpredictable. As for staying at Google: I'm happy here right now. We're launching fast,
1:18:31and competition keeps us moving rapidly. If you left Google, which company would you join?
1:18:36After a couple of years at Google scale, I felt: 'If I'm not at Google, I'd only go to a startup.'
1:18:42Google is part of Alphabet, but operates as its own company. I'm very content right now. It depends on your island—
1:18:47my team moves fast, launches constantly, and tackles novel challenges.
1:18:55I stay at a company as long as I'm learning. I've only changed jobs when I lost passion for the work,
1:19:00or when I mastered it so thoroughly there was nothing new to learn. At Google, running out of things to learn is impossible.
1:19:05Many colleagues have been here 15 to 20 years, joining in their early twenties and rotating across projects.
1:19:09The learning never stops unless you want to pivot to an entirely different operational style—
1:19:17which means founding or joining an early-stage startup. How difficult is it to get hired at Google now?
1:19:22For Product Managers specifically, it has become very tough. Open headcounts have shrunk dramatically.
1:19:27Tech giants are adjusting for AI efficiency gains, anticipating higher productivity with fewer total employees.
1:19:32When someone leaves, backfills are frozen to prevent layoffs. Unless someone departs, headcount remains flat.
1:19:38Some specialized teams are expanding, but generally Google isn't growing headcount aggressively.
1:19:42One area seeing strong demand is Forward Deployed Engineering, similar to my Solutions Architecture role.
1:19:46Just like robotics, you ultimately need engineers on-site to implement AI solutions inside enterprise systems.
1:19:48Personalization—specifically Spark. Spark is one part of it. Product managers span a spectrum:
1:19:52consumer PMs, enterprise PMs, and infrastructure PMs. I'm an Infrastructure PM because of my technical background.
1:19:58A huge part of search is retrieval, and I worked heavily on retrieval. As an Infra PM, I build the infrastructure
1:20:04so that when your query hits the Gemini app— like asking Spark, 'Plan my next trip to Europe'—
1:20:09the Gemini app sends that request to our internal platform. Our platform determines whether we have precomputed profile data
1:20:15to answer directly, or whether it needs to query authorized sources and retrieve relevant context.
1:20:20The infrastructure engineering centers heavily on performance and latency. Planning a trip might tolerate a brief delay,
1:20:26but asking for dinner recommendations expects an answer in seconds. In search, everything is measured in milliseconds.
1:20:32We must optimize between high-level profile data and granular fresh context, maintaining strict data lineage.
1:20:37A famous performance trick was Gmail's login flow. Kevin Systrom shared how early Instagram uploaded photos:
1:20:42the progress bar completed instantaneously because the upload began the moment you picked the photo,
1:20:47while you were still typing the caption and tags! They assumed you'd hit post, and if you canceled, they discarded it.
1:20:51They admitted borrowing that idea from Gmail's login flow: while typing your password, Gmail preloaded your inbox in the background!
1:20:57Even if you mistyped, it was fine— but the obsession with speed was remarkable.
1:21:02Before personalization, I worked on Google Search backend retrieval. Search retrieval is a truly mind-bending domain.
1:21:07Google is completely different from Apple: Apple is on the application side; Google Search operates at insane scale.
1:21:12Internal QPS metrics—Queries Per Second—are staggering. Our platform handled web retrieval across 120 billion documents.
1:21:16When you search for live scores, delivering instant results is an immense engineering feat.
1:21:22Google learns from occasional outages every year, constantly refining reliability.
1:21:27Search retrieval is a technical marvel at Google. When early AI models emerged,
1:21:32they relied solely on internal pre-training data. To bring in external live data,
1:21:37RAG was born—Retrieval-Augmented Generation. Is RAG dead now?
1:21:43RAG isn't dead—everything is essentially RAG! Why did 'RAG is dead' even become a buzzword?
1:21:48It's like saying driving is dead! One day they say Product Management is dead, the next Software Engineering.
1:21:52Controversial statements are made to capture attention. Google experimented with instant search results
1:21:58updating with each keystroke before pressing enter. Autocomplete stayed, but instant result updates were dropped
1:22:03because it consumed excess resources and interrupted the user's thought process.
1:22:08Transformers fundamentally predict the next word in sequence. Transformers originated right at Google.
1:22:13Just like keyboard autocomplete or MS Word spellcheck, scaled to massive compute.
1:22:18If you hit upload, it was already completed—and if not, they deleted it. Instagram blew everyone's minds with how fast it was.
1:22:22They revealed they borrowed that trick from Gmail's login flow: in Gmail, they assumed that once you typed your username,
1:22:26you'd probably enter the correct password. So while you typed your password, Gmail preloaded your inbox!
1:22:31Even if you mistyped, it was fine— but the obsession with speed was remarkable.
1:22:37Before personalization, I worked on Google Search backend retrieval. Search retrieval is a truly mind-bending domain.
1:22:43Google is completely different from Apple: Apple is on the application side; Google Search operates at insane scale.
1:22:47While typing your password, Gmail preloaded your inbox in the background! If you mistyped, no harm done—but speed was paramount.
1:22:52It was fascinating how much they cared about speed. Before working on personalization, I worked on Search backend retrieval.
1:22:57Search retrieval is an extraordinary domain. You asked earlier what makes Google unique—
1:22:59even compared to Apple, Google was vastly different. Apple is on the application side;
1:23:04Google Search is an entirely different beast. Internal QPS metrics—Queries Per Second—are staggering.
1:23:08The query volume Google Search processes is unbelievable. On the platform where I was PM, our mission was retrieval.
1:23:13Retrieval means searching across 120 billion web documents. When you search for live scores, delivering instant results
1:23:20is an immense engineering feat. Google learns from occasional outages every year,
1:23:26constantly refining reliability. Search retrieval is a technical marvel at Google.
1:23:31When early AI models emerged, they relied solely on internal pre-training data.
1:23:35To bring in external live data, RAG was born—Retrieval-Augmented Generation.
1:23:42Retrieval-Augmented Generation—is RAG dead now? RAG isn't dead—everything is essentially RAG!
1:23:48Exactly my point! Why did 'RAG is dead' become a headline? It's like saying driving is dead!
1:23:52One day they proclaim Product Management is dead, the next day Software Engineering is dead.
1:23:56Sensationalist claims are designed for clicks. Google experimented with instant search results
1:24:02updating with each keystroke before pressing enter. Autocomplete stayed, but instant result updates were dropped
1:24:06because it consumed excess resources and interrupted the user's thought process.
1:24:11Transformers fundamentally predict the next word in sequence. Transformers originated right at Google.
1:24:17Just like keyboard autocomplete or MS Word spellcheck, scaled to massive compute.
1:24:23Why did people start saying 'RAG is dead'? It's like saying driving is dead!
1:24:28What does that even mean? It's like saying driving is dead!
1:24:30One day they say Product Management is dead, the next day Software Engineering.
1:24:37A huge part of capturing attention for content is declaring something controversial.
1:24:42Speaking of Google experiments: Google experimented with search updating live as you typed,
1:24:49character by character, before hitting Enter.
1:24:53I don't know if that was instant results or autocomplete.
1:24:59Autocomplete was there, but they tested instant live results.
1:25:01You could toggle it on or off, but they eventually dropped it. As you typed, the result page kept dynamically refreshing.
1:25:06If you typed 'iPhone,' it would load iPhone results, and if you added '5G,' it updated instantly.
1:25:12They halted it because it created needless server compute costs,
1:25:16and they decided to let the user finish their intent rather than interrupting: 'Found it! Found it! Found it!'
1:25:21If you think about it, Transformers and modern LLMs are fundamentally predicting the next token.
1:25:27That's why Transformers were an internal Google invention.
1:25:33Yes, exactly. Even keyboard suggestions or MS Word spellcheck
1:25:37are fundamentally that same predictive concept, scaled up to an extraordinary level.
1:25:40What people didn't realize was that pouring massive compute and data into it would produce such remarkable accuracy.
1:25:44That was the open question. Sam Altman noted that early on:
1:25:48'Will scaling compute and data keep making it better?' They tested it, and scaling laws held up.
1:25:52Google published the Transformer paper in 2017 and already had smaller models running in production.
1:25:57There was internal debate over deploying large models widely.
1:26:02Then ChatGPT launched, breaking the public stigma.
1:26:06You mentioned something very interesting: all of this is fundamentally RAG.
1:26:11If you have a knowledge base you query, that's RAG.
1:26:17Where do you see Google's future? Would you bet on it as an investor or stay long-term?
1:26:24Betting on Google stock is completely separate from choosing to work there.
1:26:29You made a great observation earlier: at Google or Apple, you work in a 10-person pod.
1:26:34My very first manager told me 12 or 13 years ago: 'You always live on a small island.
1:26:41Whether the company has 50 employees, 5,000, or 50,000, you are always operating on a 10- to 20-person island.'
1:26:45That holds true at Google as well.
1:26:50As for betting on Google stock: personally, I think it's still undervalued.
1:26:54However, there is an AI bubble with a domino effect: startups raise massive VC rounds and spend it on compute.
1:26:56Companies like Google and Nvidia look undervalued because their multiples are lower than inflated private valuations.
1:27:01If free VC money dries up and compute spend drops, revenues could decline across the industry.
1:27:06So I'd bet on Google, but navigating the macro bubble is unpredictable.
1:27:11As for whether I stay at Google: I'm here right now, I'm happy, we're launching fast.
1:27:17Competition keeps us moving rapidly, and the topics are novel.
1:27:21If not Google, which company would you join?
1:27:28After working at Google scale, I felt: 'If I'm not at Google, I'd only go to a startup.'
1:27:32Google operates under Alphabet, but remains its own entity.
1:27:37It depends heavily on your team pod. My team moves fast, ships constantly, and tackles hard problems.
1:27:44I stay at a company as long as I am learning.
1:27:50I've only moved on when I lost passion for the work, or mastered it so thoroughly there was nothing new to learn.
1:27:55At Google, running out of learning is impossible. Many colleagues have been here 15 to 20 years,
1:28:00joining in their early twenties and rotating across teams.
1:28:05The learning never ends unless you want to pivot to an entirely different operational style—
1:28:11which means founding or joining an early-stage startup.
1:28:16How tough is it to land a job at Google right now?
1:28:21It has become very competitive, especially for Product Managers. Open headcount has shrunk dramatically.
1:28:27Big tech is adjusting headcount for AI productivity gains,
1:28:31anticipating that AI enables smaller teams to deliver more.
1:28:36When someone leaves, roles aren't backfilled to prevent future layoffs.
1:28:41Total headcount is kept largely flat unless someone departs.
1:28:45Some teams grow, but overall Google headcount isn't expanding aggressively.
1:28:50One role in high demand is Forward Deployed Engineering (FDE),
1:28:55similar to my Solutions Architecture background.
1:29:01Just like robotics, you ultimately need engineers on-site to deploy and customize AI inside enterprise systems.
1:29:05In enterprise AI, you must understand their workflows and integrate the solution seamlessly.
1:29:10That FDE role gained huge popularity over the past year; Google, OpenAI, and Anthropic are hiring aggressively for it.
1:29:14I'd love your advice on something: as I mentioned, I don't have a traditional corporate resume.
1:29:19I've spent most of my career building early-stage startups,
1:29:26usually as an early founder building the team and culture.
1:29:30I came to San Francisco driven by startups,
1:29:36but seeing AI automate specialized engineering tasks, I'm being approached to join growth-stage startups
1:29:40that raised $50M, $300M, or $500M.
1:29:44They actively look for entrepreneurial generalists— founders who connect dots across business and tech,
1:29:47rather than narrow single-domain specialists.
1:29:52Years ago, the mantra was 'Specialize! Jack of all trades, master of none.' Now that paradigm is completely inverted.
1:29:57They claimed AI would never touch art or music, yet creative media was the first thing it automated!
1:30:01I'm being offered roles at well-funded startups tackling meaningful problems, and it's tempting.
1:30:07A brand-new path seems to have opened up.
1:30:12That's an exciting dilemma! It comes down to two dimensions:
1:30:17the substance of the work—team, mission, product— and the financial dimension.
1:30:21You have to weigh how much financial upside matters versus day-to-day mission.
1:30:26One challenge with heavily funded startups today is inflated valuations.
1:30:30If a startup raises $500M or $600M at a steep valuation,
1:30:35joining for equity options carries real risk: if you leave after two years, exercising options is costly,
1:30:39or the company must generate massive revenue to match that valuation.
1:30:45Anthropic is an exception—its valuation is solid, and they offer massive base salaries like $400k–$430k.
1:30:49Financials aside, the work experience is thrilling. The only reason I'd leave Google for a startup
1:30:54is speed. I love Google's scale, but scale brings inertia:
1:30:59consulting 10 lawyers, aligning 20 privacy engineers, and syncing with multiple directors just to launch a feature,
1:31:04while a nimble startup ships version 10 across the street!
1:31:11Big tech is adjusting, but you can't beat a 10-person startup moving at breakneck velocity.
1:31:16That speed requires working 12 hours a day, 6 or 7 days a week.
1:31:22If I join a startup again, it'll be for that culture— a tight-knit team moving fast to solve a hard problem.
1:31:27That's awesome! Thank you, Shirin, this has been fantastic. I really enjoyed this conversation!
1:31:33We tried making predictions about the future, though in tech, tomorrow could flip everything upside down.
1:31:39Just when a trend seems obvious, the game shifts from an unexpected angle.
1:31:44I follow Lenny's Podcast closely; he recently hosted Josh Woodward, VP of Google Labs, who made a great point:
1:31:48how quickly you can unlearn things has become the #1 skill. Not just learning, but unlearning.
1:31:52It's healthy to admit when you're wrong and pivot immediately, rather than stubbornly clinging to outdated models.
1:31:56I feel software engineering is entering a period of identity depression.
1:32:02Discussions here center on developers watching AI agents write code, wondering where their craft went.
1:32:07The skill engineers honed for decades is being automated, sparking an identity crisis: 'What is my actual role?
1:32:12I just review PRs, approve tickets, and let agents do the work.' Even the camaraderie of team problem-solving is eroding,
1:32:16because your primary collaborator and sounding board is an AI.
1:32:21Moving at this velocity burns through old habits. Unlearning is crucial—mourning past skills wastes time.
1:32:26You have to kiss it goodbye and level up. I find this era exciting. For the first time,
1:32:31a fascinating new technological paradigm is unfolding. It has dark implications, but there's a bright window to enjoy.
1:32:36I'm optimistic because generations adapt.
1:32:41Our parents' generation worked at one company for 40 years; our generation never stayed past 2 or 3 years.
1:32:47Now we're adapting not just by switching companies, but by evolving our entire job functions.
1:32:50Most people around me are positive: jobs will transform, not vanish entirely. If you're a perpetual learner,
1:32:56you'll always find new frontiers. I joke: 'I hope AI takes my job so I can go do something else!'
1:33:02Not everyone shares that mindset! People learn a skill solely to earn a living and advance financially.
1:33:05Our parents didn't have to constantly reskill every year— they learned a trade, worked 20 years, retired with a pension.
1:33:09Many people simply lack the energy for perpetual retraining. That's human nature, but the only constant in life is change.
1:33:14Those listening to this podcast are passionate learners, representing one segment of society.
1:33:19Others prefer mastering a routine and enjoying stability.
1:33:26Every historical era rewarded a specific skillset— in the Middle Ages it was art; this era belonged to tech nerds.
1:33:31Anthropic offers huge compensation packages, like $400k to $430k base salaries.
1:33:37It depends on the specific role, of course. Anthropic's valuation is firmly grounded,
1:33:40whereas joining a startup that raised $500M or $600M at a steep valuation carries risk:
1:33:46if you leave after a year or two, exercising options is very expensive, or the company must deliver massive revenue to justify that price.
1:33:51Finances aside, the operational experience is what matters. The only reason I'd leave Google for a startup is speed.
1:33:56I love Google's scale, but scale brings organizational inertia: consulting 10 lawyers, aligning 20 privacy engineers,
1:34:01and syncing with six directors just to launch a feature, while a nimble startup ships version 10 across the street!
1:34:05Big tech is adjusting, but you can't match a 10-person startup moving at breakneck speed.
1:34:11That velocity demands grinding 12-hour days, 6 or 7 days a week. If I join a startup again, it'll be for that culture—
1:34:15a tight-knit team moving fast to solve a hard problem. Thank you so much, Shirin, this has been fantastic!
1:34:20I really enjoyed this conversation! We made bold predictions, though in tech, tomorrow could flip everything upside down.
1:34:25Just when a trend seems obvious, the game shifts from an unexpected angle.
1:34:30I follow Lenny's Podcast closely; he recently hosted Josh Woodward, VP of Google Labs, who made a great point:
1:34:36how quickly you can unlearn things has become the #1 skill. Not just learning, but unlearning.
1:34:41It's healthy to admit when you're wrong and pivot immediately, rather than stubbornly clinging to outdated models.
1:34:46I feel software engineering is entering a period of identity depression. Discussions here center on developers watching AI agents
1:34:51write code, wondering where their craft went. The skill engineers honed for decades is being automated,
1:34:57sparking an identity crisis: 'What is my actual role? I just review PRs, approve tickets, and let agents do the work.'
1:35:01Even the camaraderie of team problem-solving is eroding, because your primary collaborator and sounding board is an AI.
1:35:06Moving at this velocity burns through old habits. Unlearning is crucial—mourning past skills wastes time.
1:35:13You have to kiss it goodbye and level up. I find this era exciting. For the first time,
1:35:19a fascinating new technological paradigm is unfolding. It has dark implications, but there's a bright window to enjoy.
1:35:23I'm optimistic because generations adapt. Our parents' generation worked at one company for 40 years;
1:35:28our generation never stayed past 2 or 3 years. Now we're adapting not just by switching companies,
1:35:34but by evolving our entire job functions. Most people around me are positive: jobs will transform,
1:35:39not vanish entirely. If you're a perpetual learner, you'll always find new frontiers. I joke: 'I hope AI takes my job
1:35:44so I can go do something else!' Not everyone shares that mindset!
1:35:48People learn a skill solely to earn a living and advance financially. Our parents didn't have to constantly reskill every year—
1:35:54they learned a trade, worked 20 years, retired with a pension. Many people simply lack the energy for perpetual retraining.
1:36:00That's human nature, but the only constant in life is change. Those listening to this podcast are passionate learners,
1:36:05representing one segment of society. Others prefer mastering a routine and enjoying stability.
1:36:11Every historical era rewarded a specific skillset— in the Middle Ages it was art; this era belonged to tech nerds.
1:36:17Perhaps the next era belongs to generalists who learn and unlearn quickly. Every era has disruption. The hope is that AI elevates
1:36:23a baseline standard of living for everyone, so people don't have to struggle endlessly just to survive.
1:36:28When that happens, maybe we all move back to Toronto, or Tehran! Thank you so much, Shirin! I had a blast.
1:36:32I feel we are entering an era of software developer depression. Many conversations I hear revolve around this:
1:36:37engineers seeing AI agents write code, and the deep craft they developed over years eroding.
1:36:44Beyond code being generated by AI agents, that hard-earned craft is being replaced.
1:36:50Engineers are experiencing an identity crisis: 'What is my job now? Do I just click approve tickets?'
1:36:55Even code reviews are being handled by other AI agents! That sense of camaraderie and shared struggle
1:37:01is fading, because your closest collaborator and friend is an AI. You consult the AI, discuss with it, and debate less with peers.
1:37:06At the breakneck speed we're moving, collateral disruption is inevitable, so unlearning is paramount.
1:37:10You can spend endless time despairing: 'What did I do with my last 10 years? Why did I learn all this?'
1:37:18No—you have to kiss that goodbye and move to the next stage. I find this era genuinely exciting and fascinating.
1:37:24For the first time, a world is emerging that intrigues me. I know it has dark sides, but there's a bright window to enjoy.
1:37:28I'm optimistic because generations adapt: our parents worked at one company for 40 years until retirement;
1:37:32we never got used to staying at a company past 2 or 3 years. Now, we're not just switching companies—we're evolving our job functions.
1:37:37Most tech peers around me are positive: jobs will transform, not vanish. If you're a continuous learner,
1:37:42you'll always discover new frontiers. I joke: 'I hope AI takes my job so I can go do something else!'
1:37:46Not everyone shares that mindset! People learn a trade primarily to earn a living and provide for their families.
1:37:50Our parents didn't have to constantly reskill every year; they mastered a discipline, worked for 20 years, and retired comfortably.
1:37:53Most people simply don't have the appetite for endless retraining. That's human nature—the only constant in life is change.
1:37:57Those tuning into this podcast are passionate learners, representing a specific forward-looking group.
1:38:03Another part of society prefers routine and stability. Yet society continuously shifts which skills it values most:
1:38:09in the Middle Ages it was art; this era belonged to tech nerds. Perhaps the next era belongs to generalists who learn and unlearn quickly.
1:38:14Disruption touches every era. The ultimate promise of AI is elevating a baseline standard of abundance for everyone,
1:38:19so people don't have to suffer just to make a living. When that happens, maybe we all move back to Toronto, or Tehran!
1:38:24Thank you so much, Shirin! I had an absolute blast. Thank you, Soheil! It was wonderful being here.
1:38:29Since we're both in SF, let's catch up every few weeks. Absolutely, let's do it! I want to hear how your projects go.
1:38:35Awesome! Thank you so much. Thank you!
1:38:41It was great talking to you! Same here, let's keep in touch.
1:38:47Take care! See you soon!
1:38:50Not everyone has your mindset about AI taking their job! People learn a craft primarily to earn a living.
1:38:56What I've realized is that people learn a trade simply to earn money and provide for their family.
1:39:04Our parents didn't have to reskill every year— they learned a skill, worked 20 years, and enjoyed stability.
1:39:10Working at IBM for an entire career was common, and people simply lack the energy for constant retraining.
1:39:14That's human nature—change is the only constant. The only constant in life is that everything changes.
1:39:19Those tuning into this podcast love learning, representing one segment of society.
1:39:23Others prefer routine and stability. Yet every era values a different skillset:
1:39:28in the Middle Ages it was fine art; if we were around back then without tech, we'd be useless!
1:39:32I'm no good at painting or art! Every era enables people with specific skillsets to thrive.
1:39:36This era belonged to nerds and techies; maybe the next belongs to generalists who learn and unlearn fast.
1:39:42Every era has its trade-offs. The hope of AI is providing a universal baseline standard of living,
1:39:47so people don't have to suffer just to survive. When that happens, maybe we move back to Toronto, or Tehran!
1:39:52Thank you so much, Shirin! I had an absolute blast. Thank you, Soheil! It was a real pleasure.
1:39:58Since we're both in SF, let's catch up every three or four weeks. Definitely! I want to hear how your projects progress.
1:40:03Awesome, let's make it happen! Thank you!
1:40:07Thank you so much! Take care, bye!
1:40:10Have a great day! Bye!
1:40:11Every era enabled people with specific skillsets to succeed.
1:40:17This era was for techies; next might be for agile generalists.
1:40:22Every era has its trade-offs.
1:40:29The promise of AI is elevating a baseline standard of abundance,
1:40:35so people don't have to suffer just to make a living.
1:40:40When that happens, maybe we all move back to Toronto, or Tehran!
1:40:48That's when we return to Tehran! Thank you, Shirin, I had a blast.
1:40:52Thank you so much, Soheil! It was an absolute pleasure.
1:40:56Since we're both in SF, let's catch up every few weeks.
1:40:59Definitely, let's do it! I'd love to hear how your projects evolve.
1:41:05Awesome, thank you so much!
1:41:10Thank you! Take care, bye!

About this transcript

This page contains the full transcript of EP 212 - Shirin Mojarad | Building Personal AI at Google by طبقه 16, generated from the public captions YouTube serves with the video. The transcript has 17,498 words across 1,201 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.