Two Pivots: A Googler and a PhD Founder

Conversations with a TPM at Google in New York and a US AI founder with a PhD.

Whenever I have time, I keep doing coffee chats with people I know in the US. This time it was an older friend who works as a TPM (Technical Program Manager) at Google's New York office, and an acquaintance who finished a PhD in machine learning in the US and was about to join a lab building AI foundation models, then suddenly started a company instead. Both have lived in the US for more than 15 years.

Honestly, before every coffee chat I have no idea where the conversation will go or how it will wrap up. So while we talk, I race to draw out better answers, and at the same time I concentrate hard so I can ask exactly what I'm curious about. The two of them have very different backgrounds. One graduated from an Ivy League college and has worked in finance and big tech ever since. The other graduated from a top US engineering school and seemed set to stay in academia, then suddenly started a company. But when I chewed over the conversations to write this post, I realized they have a lot in common. Both feel a real sense of crisis about the AI wave, and both are scrambling, pivoting and adapting to ride it while actively sketching out their futures. Neither has found a clear solution yet (I doubt anyone has right now), but it was obvious that both are fighting hard to turn this into a second chance, whether in their careers or in life.

My older friend, for example, was in a relatively stable support organization on the operations side at Google. To get more directly into AI, they asked for an internal transfer to the AI Google Search team, one of the core teams. Their reason: "to survive five more years." Nobody knows what the world will look like five years from now, but they figured you have to swim against the wave one more time to have a shot at a more hopeful future after that. Big tech is fiercely competitive inside too, and in an environment where about 70% of people are engineers, it's even tougher for non-engineers. My founder acquaintance, on the other hand, majored in machine learning itself at a top US engineering school. More precisely, they studied the mathematical theory of neural networks and wrote quite a few meaningful AI papers, and I heard they got a lot of offers from foundation model labs. They'd always been deeply interested in that kind of research, so that seemed like the natural path. But a close friend from college talked them into (or lured them into, maybe) co-founding a startup. That co-founder is a former Google engineer in Silicon Valley, worked at several US startups before that, and back in college made Dogecoin keychains that completely sold out. They'd always had a knack for business and a strong interest in startups, so my acquaintance said they were very open to the idea too.

Of course, as my Google friend put it, "In two or three years I'll probably have to find something else to survive again." And my founder acquaintance said the AI frontier companies are moving so much faster than expected that, after focusing on a B2B productivity tool until now, they're considering a hard turn to B2C, toward something that triggers people's dopamine at a primal level.

Listening to what they're wrestling with right now gave me a lot of insight, so I wrote it down.

1. Big tech is thriving, but individuals are fighting to survive

My Google friend used to be on an operations team for Google services. The work was stable and important, but they worried about how much it would help their own career in the AI era. So they asked to move to Google's AI Search team. That's the feature where AI gives you a summary right away when you search on Google.

They said they were really surprised after the move. "I've been at Google for years, but this is new to me. This team in particular moves with real urgency, like a startup." For example, one PM went on a business trip for a week, and in the meantime two of the three projects they'd been running got handed to other people, and a whole new scope got added to their work, shifting its direction. Step away for a moment and huge changes happen while you're gone. Miss two or three days of Slack and you simply can't keep up with where a project is heading. My friend is so busy that they can't keep up with all of Slack either, so they always hook Claude up to it to pull out real-time summaries and insights, at least, so they can follow up fast.

"I had no choice but to move teams to survive another five years or so. But in two or three years I'll probably have to find something else to survive again." Even inside the big tech companies driving the AI wave, survival is fierce, so instead of settling in one place, people keep moving to wherever the next wave is breaking. "I'm not normally the type to jump on trends fast. I usually watch for a while and then follow slowly. But right now I really feel that this isn't a time when you can just watch."

2. Google stopped for a week: AI Builders Week

In an earlier post, about a conversation with a friend in finance in New York, I wrote that Stripe shut down the whole company for a week and had every employee focus only on adapting to AI, so it could get its AI transition right. Google did the same, and it seems to be running that kind of week fairly often. The AI wave keeps crashing in, and something you learned just a month ago is already history.

So internally they call it AI Builders Week. They block out an entire week, and of course everyone gets AI training, but people also form study groups and study together like college students, and even run mini hackathons to build things. "It's fine if work slows down a little. Use AI, no matter what." My friend's only complaint was that the week is too short.

What was interesting was how the non-engineers reacted. "Engineers were already doing this little by little. It's their job, so it's natural for them. But once non-engineers got this dedicated time, a lot of them got excited and started trying all kinds of things." They knew they should, but hadn't been doing it, and now they had a reason. Slack channels are blazing until midnight, and people who are far apart even gather in the same city to study harder and bond.

"The team I just joined is really motivated." When I asked why, my friend said it's because of how big the project and the mission are. Google Search is Google's main product, and this team is putting AI on top of it, so a mission statement like "we want to be the one that people go and interact and get information" is baked deep into everyone's DNA. True Googlers.

3. Execs don't ask me for updates anymore: using agents for real

According to my Google friend, a coworker said this: "In 10 years of working, this is the first time leadership hasn't asked me for a status update in two weeks." Now that executives can go into an AI tool themselves, run a query and understand the situation right away, the PM's role of organizing information in the middle and passing it along has shrunk.

This is a change that shakes the very reason the PM role exists. A PM basically sits between engineers and leadership, figuring out how a program is running, organizing it and communicating it. AI has started taking over that "organize and communicate" layer.

So as a PM, my friend keeps looking for ways to use AI to make communication between leadership and engineers more efficient. For example, the tools engineers build are so huge and complex that in the past my friend couldn't touch them directly. They'd ask the engineers for documentation, try to understand it, and wrestle with the engineers to polish it into something leadership could follow. Leadership, after all, would know far less about the context of the project.

Now, if there's a tool an engineer built, my friend just pulls the data, takes out only what matters from leadership's point of view, builds a custom tool or web app for that one use case, and puts it in the leadership report. There are two big advantages. First, there's no unnecessary back and forth with engineers. The engineers don't have to worry as much about documentation for leadership reports, and the PM can understand the context and customize things more easily and quickly. Second, even a well-summarized document is hard for leadership to follow as text alone, but if they look at the custom tool first and then read the document, they understand it much faster and more accurately. You could say it's a move from a written summary to a visual summary.

In the end, my friend has redefined their role around "reducing mental load" for everyone.

4. The day an academic theorist became a startup builder

The founder I met the same day had a completely different background. They graduated from a foreign language high school in Korea (an elite specialized school), then went to the US and studied statistical theory grounded in neural networks the whole way through. They did nothing but research for more than 10 years. I don't fully understand their exact work, but it involves assuming a certain data distribution and mathematically proving how well a model performs on new samples depending on model size, algorithm and amount of data. They also worked on dimension reduction theory. For example, theory on the information loss that happens when you reduce a million-dimensional vector to 100 dimensions. They apply this to neural networks to study generalization bounds.

As a side note, during college they interned on the AI team at a tech company in Korea. That was back when Korean companies were first saying they'd build their own AI, and the initial team was six or seven people. Since it was so early, the team kept bouncing between domains, from commerce to games, and maybe because of that, things were a bit makeshift. Everyone came from a service background, and my acquaintance wanted a more principled approach, developing models grounded in theory. That's why they went on to do a PhD.

After finishing the PhD, at the crossroads between continuing research and starting a company, they chose the company. AI played a big role in that. They're not an engineer by training. Their background is in math and statistical theory. In grad school they barely coded at all and mostly used statistical tools like R. But to run a business you have to build a product, and to build a product you have to code. Luckily, by the time they started the company, AI was already there. "I could build what I was thinking about really fast, without any complex about coding," they said.

5. US academic publishing is changing too

Papers in statistics and math have traditionally centered on theory and proof. You set up an idea, prove it mathematically, then attach a simple numerical experiment. In the past, computation was limited, so those experiments were run on small synthetic datasets. They were "nice to have, but it didn't really matter either way." Now things are different. There's far more data than before, and the models you can compute with are bigger. "There used to be the excuse that we couldn't do more because it wasn't possible, but that excuse is gone now," they said. If you have a theory, the room to show computationally where it actually works well has grown a lot.

We also talked about the peer review system papers have to get through. Reviewers for AI/ML conference papers get about three weeks, and in that time it's practically impossible to verify whether a paper's mathematical proofs are correct. The odds that a reviewer studies exactly the same area as the paper are low, since three reviewers are assigned at random. On top of that, most reviewers are PhD students. So "even if you write it well, it might not get in. Luck plays a big part," they said.

6. The reality of competing with frontier models

My founder acquaintance's startup builds an AI productivity tool. A user drops in source files, the AI drafts a narrative, suggests a fitting layout, and once the user confirms, it fills it in. What set it apart was that you could click on any piece of content and immediately see which part of the source files it references.

At first they targeted B2B, things like portfolio reports for real estate funds. Managers each send in material in different formats, Excel, Word, notes and so on, and someone has to pull it all together into investor decks. That's repetitive work. But security was too big an issue. The bigger the enterprise customer, the higher the wall against sending data outside. So they pivoted to B2C, which is easier to get into, and they're preparing to relaunch now.

There's a reason this hit home for me. My company is wrestling with something similar. We're using AI to build a service for US government contracting, and these days the foundation models are getting really good even in niche business domains. New updates keep coming out every few weeks, and they're so powerful that there's always uncertainty about how long we can last.

The second worry is the gap between early adopters and everyone else. There are way more people not using AI than you'd think. Most people only use ChatGPT, and when you ask why, they say "security" or "no time," but honestly they just haven't had a trigger. Some of my friends are in the AI field and some aren't, and no matter how much I talk to the ones who aren't, it ends with "That works? Wow, cool. But we can't do that at my company." I always find myself thinking: if even Google can't convince these people, how am I supposed to?

7. You have to set off people's dopamine at a primal level

There's a post that's gone pretty viral on LinkedIn lately: "Cold cakes instead of cold emails." Exactly what it sounds like. Since everyone ignores cold emails these days, you design a cake that will catch the other person's attention and send them an actual cake. The recipient finds it surprising and funny, takes a photo and posts it on social media, and that goes hugely viral too. It's a Norwegian startup making more than 165 million won a month (a bit over $100,000) from this alone, and it's really taking off on the US West Coast right now.

Actually, I got a DM from that founder a few days ago too, asking if I wanted to do a cold cake. At first I wasn't sure whether they meant an actual cake or whether it was some kind of metaphor, but it was a real cake. Now even the biggest US VCs like a16z and Sequoia order as customers all the time, and people take photos in their offices or in front of their car trunks, and those keep going viral.

The funny part is that my founder acquaintance actually wanted to do B2B sales in several US cities, mainly to real estate companies. Cold emails and cold calls weren't getting through, so they just showed up at offices with donuts. If they were lucky, they'd get connected. If not, they got stopped in the lobby.

They went through a ton of grind that way and often wondered what on earth they were doing. Seeing this cake post, they said they actually thought, "Wow, people go this far?"

Anyway, we're now in an era where you have to grab people's attention to do business at all. Anyone can build a decent-looking product now, so competition is fiercer, social media is oversaturated, and if my content and branding look like everyone else's, it won't land. It'll just be more noise.

In the end, what I think about most is how to trigger people's primal dopamine, keep them hooked and turn their attention into business.

Think of the US startup I mentioned in an earlier post, which raised about 20 billion won (roughly $14 million) to launch satellites that reflect sunlight down to the ground with mirrors, a subscription service for sunlight (for things like nighttime search and rescue or film shoots). And a few days ago I came across a service in New Zealand called a virtual fence: you put collars on cows and manage where they can go with electric stimulation instead of a physical fence.

https://www.halterhq.com/

It's really clever. Whether or not it actually works that well, things are heading toward a structure where you have to deliver this kind of shock for our dopamine to fire and for us to pay more attention.

8. Surviving as a human in the AI era

Finally, I asked my Google friend: AI keeps getting smarter, so how can humans keep growing? AI getting smarter and us actually growing while understanding the context are two different things.

My friend started with their kids. They said they think about what their children should do when they grow up. When we were in school, it was enough to memorize well and solve problems well, but that skill is going to become useless. Instead, kids will need to know more about the humanities, understand people better and be better at building relationships.

The way my friend summed it up stuck with me: "Be more human in your study and personality while continuously practicing AI." Using AI is a given. You have to keep practicing it. But you should develop more of what we can do better, the parts that draw out people's originality.

My friend was doing exactly that. They leave the execution to AI and focus on building relationships. "Rather than having no time to meet people because I'm busy executing, I'll leave execution to agents and focus on people." The center of gravity of the PM role is shifting from organizing documents to building relationships.

My friend also sees physical AI as the next frontier. Right now everyone only talks about LLMs, but honestly, in the US, LLMs are already old history. Physical AI is going to be really big down the road, they said. Then again, that's old news too, and physical AI is already right in front of us, we just don't see it at mass scale yet. They said it makes them think hard about what we should actually do as humans. Just as they moved teams with the next two or three years in mind, what mattered was getting a foot in the door on this team now, so they can jump into that kind of field later.

Two people I met on the same day. Different backgrounds, different positions, but both were doing the same thing: pivoting and adapting, over and over. There's no clear answer yet, but they're moving. Fast and nimbly.

Just as I was writing this, an American friend reached out. They run businesses in Silicon Valley and Japan, and recently came to Seoul after touring the AI scenes in SF, Shanghai and Bangkok. What they said stuck with me.

What I realized revisiting Silicon Valley recently is that it's not about the big AI events, it's about the near daily conversations that everyone is having in person about their latest bleeding edge learnings.

What matters in the end isn't the big events. It's the face-to-face conversations that happen almost every day. The frequency and density of people sharing the latest tech with each other is what makes Silicon Valley Silicon Valley.

I think the same way. That's why I'm preparing a community event in Seoul this April with an official Anthropic (Claude) Ambassador, on the theme of Claude Blue and Bloom. I don't want it to be just a place to trade productivity tips. I want to create a space where we can talk, in depth but without pretense, about how to live as humans now that the AI era is here. I'll announce the details on LinkedIn first.

Originally written in Korean on Brunch, April 3, 2026.