---
title: 'Do This to Get Hired at Google'
source: 'https://youtube.com/watch?v=NVTDicU4UX4'
video_id: 'NVTDicU4UX4'
date: 2026-07-31
duration_sec: 1469
---

# Do This to Get Hired at Google

> Source: [Do This to Get Hired at Google](https://youtube.com/watch?v=NVTDicU4UX4)

## Summary

In this podcast, Omar, a lead at Google DeepMind and former Hugging Face platform head, discusses how frontier models like Gemini are built and evaluated, and what it takes to get hired on his team. He emphasizes high agency, building in the open, and strong technical fundamentals over pedigree. The conversation also covers AI's impact on engineering careers, the evolution of technical interviews, and the trade-offs between open and closed models.

### Key Points

- **Omar's background** [01:00] — Omar leads developer experience at Google DeepMind, overseeing model launches like Gemini and Gemma, and the Gemini API. He previously led platform and community at Hugging Face, working on technical integrations for open-source libraries.
- **How frontier models are built** [02:24] — The core building blocks are compute, data, and training recipes. Google has innovated across all layers, including custom TPUs and continuously incorporating community feedback into the modeling roadmap.
- **Model evaluation** [03:05] — Gemini is evaluated using internal and community benchmarks, as well as industry signals like Artificial Analysis and LMArena to gauge performance and user preference.
- **Build in the open** [05:09] — Openly sharing your work — even if not fully open source — shows initiative. It can mean building a platform, sharing technical details, or contributing to a community, all within your constraints.
- **What DeepMind looks for in hires** [06:00] — The team wants strong technical skills, a habit of building in the open, and high agency. Degrees from prestigious universities are not required; many team members come from all over the world.
- **AI replacing engineers** [08:30] — The obstacles to AI replacing humans are cross-domain integration and agentic capabilities. With the right tools and harnesses, models can already do 90% of desired tasks, so engineers must rethink the value they add.
- **The combination that matters** [11:00] — Combining deep domain expertise, high agency, and years of software experience — plus AI nativeness — makes you extremely impactful. Non-technical founders are now shipping products too.
- **AI nativeness gap** [12:31] — Many senior engineers resist AI while junior high-agency people adopt it and become more productive. Pairing senior experience with AI adoption is ideal.
- **Advice for students** [16:44] — Keep building and experimenting with AI tools, but don't ignore fundamentals: data structures, algorithms, and system design still matter. LeetCode-style interviews are becoming less relevant.
- **Interviews with AI tools** [18:06] — Omar's team allows AI coding tools like AI Studio or Antigravity during interviews. What they test is whether you can introspect and validate the model's output, not blindly copy it.
- **Master's or PhD?** [20:53] — A PhD is mainly needed for research scientist roles. Many models are built by research engineers without PhDs who handle post-training recipes and infrastructure.
- **Open vs closed models** [22:44] — Open models are ideal for privacy-sensitive use cases and niche customization (e.g., medical data). Frontier intelligence like Gemini remains server-side, but small models can run on devices like Pixel phones.

### Conclusion

Omar's insights show that success in the AI era comes from staying action-driven, building visibly, and pairing domain expertise with technical fundamentals. Rather than fearing AI, engineers should become AI-native and focus on what models can't yet do: deep judgment, creative problem-solving, and integration across domains.

## Transcript

people don't come from Stanford or MIT. I don't think lead code style interviews make any sense anymore. Though our team does allow people to use AI studio high agency, one of the things that we really, really, really look for. I got
an opportunity to talk to someone who is involved in building Google's Gemini involved in building Google's Gemini models. Omar Senseiro is this person. He works for Google's Deep Mind team. Prior to Google, he was a head of platform and
community at hugging face. In this podcast, we have discussed how Google trains and builds these Gemini models. Omar has also given insights on how he interviews people in his team. So, if you want to apply in Google's Gemini
team, you're going to get some tips on the interview. We also discussed some general career tips to thrive in this AI era. Without further ado, let's get started with the podcast. Omar in brief can you talk about your experience at
both Google and hugging face what are the kind of projects you work on and >> at the moment I'm leading developer experience at deep mind uh but yeah to give you to share with you my high level overview I was at Google many years ago
I was working in ondevice models as a software engineer in Google assistant h there for three and a half years leading all of the developer community efforts those partnerships with different companies and research labs that were
releasing open models and for the last year and a half I've been leading the developer experience organization within Google deep mind. So all of the model launches such as Gemini in an banana vo GMA and so on all of the API ergonomics
Gemini API and AI studio. >> Nice. So you know all the internals all the [snorts] internals of how magic happens right? So without revealing confidential details, can you walk us through the
details, can you walk us through the life cycle of how a frontier model like Gemini is built right from data collection to training? Because when I use Gemini personally or GPT [snorts] whatever, it just sounds like magic.
It's like a digital god that can answer all your questions. But you guys are the creator of that digital god. >> Tell us like how do you do this magic? foundations and the different pieces are like usually the same like they're usual
suspects, right? So you have the compute, you have the data and you have the the training recipes. uh what Google has is that Google has innovated across all the layers of the stack for many many years right so on the hardware side
of things we have TPUs uh on the data side of things something that we're working a lot over the last few years is really talking with the community much feedback as we can from the community and try to incorporate all of
modeling road map which are the capabilities that we want to introduce feedback >> okay when you train this model How do you evaluate it? Because evolution seems so hard. It's native AI and it can
perform wide range of tasks. So how do you do that? community benchmarks that that we also use to do evaluation. We have our own internal benchmarks as well. We use also benchmarks in the industry such as
artificial analysis or LA Marina to get a signal or a proxy of how much these models are liked or how the models perform against each other. well? >> Yeah. Okay, cool. All right. Now, u many
of the people who are watching this video right now, they are building career in AI. They want to know how what kind of skills they want to learn going ahead. You have worked at Google. You have worked at hugging face. By the way,
thanks for all the work at hugging face because many of our learners use hugging face models and they use that open-source ecosystem. So, awesome work. >> Thank you. uh and you you were a community and platform head in hugging
>> Yeah. Yeah. So pretty much I was leading a team uh so I began my journey at hugging face by doing technical integrations between the hoging face hub and different open-source libraries. So enabling the people that are using
sentence transformers which is a library for embeddings for example to be able to share their models through hoging face. So that's how my role started at hoging head of platform and community role which is a very horizontal role. So all
of the technical partnerships, developer relations, ondevice machine learning, the moonshot teams as well. So we were working at the intersection of machine learning and other disciplines, right? So ML and healthcare, ML for biology, ML
for different industries. So that was something that we were working quite a something that we were working quite a bit. Erh and yeah pretty much uh bringing a strong communityentric mindset and that means building the
building in the open operating a bit more as a startup. Uh so yeah >> extending the conversation along the same line how much is the importance of opensource contribution for a let's say college student or a working
professional who want to build a career in AI how important is open source contribution for them? Yeah, definitely building in the open helps a lot. I industry in which you need to be able to show the work that you're doing, right?
you work in a company where you cannot share anything, right? But building the open and building in the open again does not mean necessarily you need to do everything open source but it can mean hey I build this platform share it with
a community share maybe some of the technical details about it >> share what you can of course like always within the constraints of of your building in the open is important like get them the product or the things that
you've built in front of others make sure that people are aware about what >> okay so when you hire people in your own degree that much right like >> no
>> we want to see a couple of things we want to see strong technical uh skill set of course we want to see that the people want to build in the open so that's another part and really like the mindset like high agency
>> I think it's one of the things that we really really really look for which is very difficult right because there's not like a lead code style of interview for something that >> h you need to show you need to to be you
>> Yeah. >> Okay. So if there is a person from a unknown university or let's say dropout but they have contributed in opensource they have built something really good >> then you would you would like to
>> Yeah yeah yeah for sure like really we don't care about degrees. So in the team we are a very distributed team. So I'm I come from a university in Mexico. I'm from Peru. We have team members basically in in Japan, in Australia, in
basically in in Japan, in Australia, in Nuremberg, in Berlin, in London, in Paris and in in San Francisco and yeah, Austin and Brazil and so on. But most of the people don't come from Stanford or MIT or any of these big name
is people that >> are ready to to jump to do to experiment that have the hunger to build the latest things and get those things that they build out there in front of others as well. We reciprocate this feeling in our
own community in our boot camps and in our videos that we do. Uh we uh put a very high emphasis on building in open like in our boot camps we have a special
initiative of building in open where we teach how to work on a pull request and how do you effectively contribute to open source and all of that. uh so that way they uh learn the true skills not like going to college and just
>> Yeah. Yeah. >> Yeah. So uh now what is the biggest >> Yeah. So uh now what is the biggest obstacle for AI in order to replace human engineers? Because this is the question uh that we get a lot
>> and this question I I got from my YT community. So just yesterday I posted that I'm going to do interview with Omar. What kind of question you have for Omar? And few people ask this question because people have this fear that if AI
is doing coding and everything then what are we going to do? So what is that one thing which is an obstacle for an AI in order to replace humans completely? I think there are different aspects to this question. The the first one is how
do we enable models to do things in different domains. So previously I think there's a big potential a big opportunity for people and startups and companies that come come from a very specific field. So let's say legal
domain for this use case right people with very strong niche expertise domain expertise >> how can those people adopt AI into their own specific industries. Uh so that that's one area. This connects to the
second one which is how do we enable agents to perform actions in different things and services right and that's a place where we need to talk about skills agentic harnesses integrations product integrations right so how do we enable
these agents or these models to perform actions that we can do as humans but as models they struggle because the right levels of integrations are not are not the more fundamental like just coding on the day-to-day we need to accept that
software development is happening like that has been happening already for the last few years. Uh the jump from 1.5 years ago to now has been quite point in finally the models can just do
think the models today with the right tooling, the right integration, the right identic harnesses can do 90% of the things we want them to do which for me is very exciting. Uh it also means that we need to reevaluate what's the
value we bring as engineers, right? like the way we build systems, the the ways >> Who who is reading documentation nowadays like I think most people are nowadays like I think most people are copy pasting it into like their model of
the reference or maybe they are just using MCP or some level of integration documentation anymore. I think most of the time what we see is people are using agents, APIs, MCPS, browser control to go and read the documentation and I
think that's something very interesting. It doesn't require us to rethink how we It doesn't require us to rethink how we build products in a way that will be friendly towards agendic experiences. >> So having a deep domain expertise can be
>> Having a very strong domain expertise can be helpful. The second part uh which I already touch a bit on the things we look for is the high agency. >> High agency. I think like if you're someone that is able to become AI native
and uh you have high agency, I think you can be extremely impactful to the current ecosystem, right? And of course, if you're able to combine the the domain expertise and the high agency with strong experience,
you have 10 years of experience or five years of experience, you have been building with >> software engineering for many years and you're able to become AI native and adopt AI in a productive efficient way
in your day-to-day and you have high agency. I think the limit is is is huge, right? There's so much that you can be doing. That said, you don't need to be building. And I think that's the second part which is pretty exciting which is
nowadays with >> right like demo. >> Yeah. Yeah. Exactly. Exactly. Like there are partners, startups, people with the reaching out like telling us hey I
He's building a game. week he's sharing a video of the new feature he had to his game right. We have the the gaming app for Mac, right?
a very small team. They launch 100 features in 100 days. I think Josh shared yesterday and we're releasing a word in which we we see many tech adjacent people, people from business schools, design schools, people that
have the hunger to build but don't come from a traditional technical background >> and being able to build even in that context is very exciting. I think again like we need to rethink what it means to be a developer, what it means to be a
software engineer. >> Uh but definitely the the industry is to accept. So >> uh one of the things I see many very senior people they struggle with is AI nativeness. I'm seeing many software
engineers that have lots of experience that don't want to adopt AI or are struggling to adopt AI quickly, right? while maybe like some very very junior people very high agency they are learning with AI and they are using AI
in their day-to-day and they are hence usually not always much more productive because they can yeah move faster they can iterate faster of course if you're able to pair the high agency and the expertise the software engineering
skills for many years h I think you can even be much more productive but you do need to have that openness to experiment understand the capabilities of the models see what the models can do because the jump that we have seen in
like in a year year and a half is massive. Yeah, it's massive. >> And this is what I have observed that the people who have experience 5 years, 10 years experience, they become kind of rigid. They don't want to learn these
>> They are under some kind of denial. >> Yeah. >> Yeah. >> And they keep on working the old ways. >> Whereas the fresh, you know, fresh person out of college, their mind is
very open. They know how to use cursor or anti-gravity and they can code faster. But the senior engineers have this benefit that they have the >> Mhm. >> Okay. So if I have to summarize folks
>> Okay. So if I have to summarize folks like as Omar suggested a strong domain expertise is important whether you're in healthcare or finance or whatever just build a strong expertise. There was a person who came to me for an an advice.
He he was getting an admission of his son into a college and they have a program which is uh finance plus technology. I'm like go for it. >> Because you will become finance expert. You will build the domain expertise
>> and then you will use AI tool to build something amazing. >> Right. So domain expertise is important. When Omar suggested high agency, what it means is you are very agile. You can learn things quickly and you also have
original thought. It's not like you are doing something you're just copy pasting. You you have an original idea, creativity. >> It's like a mix of having like this vias to action being creative. Uh be ready to
experiment when there's something new. >> Just think about maybe how that new thing can be helpful for your day-to-day or for your own use cases, right? because this experiment heavy kind of mindset is quite powerful.
>> So high agency means experiment heavy mindset that is one of the key >> Yeah. Actually for me action driven is like the core of it. So having this bias to action to try and around that of course you will want to experiment you
to test which are the latest capabilities. >> All right amazing next question is partially answered but I'll still ask you. So if you are getting a new person in your team at
are the skills you look at? If you have already answered it, you can skip that. But anything else you would like to add for our team again like the high agency aspect, the technical expertise is very
being connected with the community, right? Uh and that goes back to this point of building in the open, but really like talking with the startups, having this developer empathy. And developer empathy means when you're
building a developer tool, understand what will be painful for developers, right? Understanding what will be a great experience. What is a SDK or an great experience. What is a SDK or an API that makes sense versus an API that
will be confusing? How do we build those great experiences from all the levels, right? From documentation to API design to SDK. So that developer empathy is learn. I think it's something that you learn by building, by testing and by
talking with people and that's something that we really look quite a bit into >> developer empathy. Wow, that's that's a great point. What advice would you give to student who has just started their college, they in computer science, they
are in AI and data science, AI is moving fast. Now they're confused that by the time they get out of the college after 3 years world will be different. So what should they do? What kind of skills they should focus on?
Yeah, it's a difficult question because the ecosystem is moving very fast, will be in four years from now. I have my own guesses and opinions, but I think it's very difficult to predict where we will stand in three, four years from
build, build build, build build like keep experimenting, keep trying new tools just as you should like even before AI like the people that were the those people that during college were not just going to classes but also
that were building their own website, their own game, their own side projects, they should be doing the same just with AI, right? I mean, not necessarily AI, but it's one of the things that I think they should be doing. So, building as
think is also important is the fundamentals are still there, right? It's not like oh suddenly big notation or like complexity does not matter at course, like someone can buy code and build a platform very easily. That does
of system design >> learn data structure and algorithms. >> Okay. >> That's I mean that's that's my opinion. I don't think lead code style interviews make any sense anymore.
>> So when when you guys interview at Google, do you ask people to write code coding tools? >> That depends interview. >> That depends from team to team. Okay. Our team does allow people to use AI
studio for example or anti-gravity uh during the process >> but things that we look at during those kind of interviews uh do you accept what the model tells you directly >> or do you maybe like like we want to see
that level of introspection of thinking and realizing oh maybe this path that >> right uh >> so don't do a blank copy >> exactly exactly so when you're coding with anti-gravity for example example,
everything the model gives you is correct. Right? I'm also not saying, oh, you need to review every single line of code in an interview, but you should be you're building, which are the errors, which is high level overview of the
system, the architectural design, and those are the things that we really care about. I think lead code style interviews were really designed for a world in which you are hired a lot for people out of university very young
people and there it's very difficult to get signal at that level right like can get and lead code can give you like a very strong signal of whether a person can do something with data structures or yeah the traditional algorithms
>> I like some of the lead code questions because they test your understanding of So you know that this person has strong >> Yeah. >> Yeah. The way I see is we need to we
will head towards a direction in which we do more and more applied kind of >> Uh whether you can use AI or not during the interview. I think that will like be see like both things. But rather than an interview about
designing a a queue that can do something in under certain constraints, problementric uh interviews and maybe you will just
this as part of the >> of the project that you that you build or the dancer of the interview but I think there will be less and less h focus on it as the core of what is being evaluated.
>> Okay. Okay. So if you are going to appear at an interview at Google, uh they will allow you to use VIP coding. >> Depend depends on the team. Like that's with our team. Our team does allow to use like a anti-gravity but in other
teams they still do like the all the hand code the whiteboard. >> Okay. So if you're applying for inst maybe they will allow but they will not allow you to blindly copy. They will ask you okay what what is this doing?
you understand what you're building. you understand what and for that reason fundamentals is super important. Your computer architecture, how memory works, compute works, algorithms and and all of that. All right. So folks, that is still
a value in learning those fundamentals. How useful is it to pursue master or PhD nowadays? It depends on what you want to do. Uh if you want to be a research scientist, PhD is usually the easiest way to go there, right? It provides you
way to go there, right? It provides you the the fundamental critical thinking and research skills that will enable you to go in a more research path. That significantly to the machine learning ecosystem even like from a modeling
infrastructure tooling architectural level as a software engineer. I work on open models. Many of the people that work you work on Gemma. >> Yeah. Yeah. So I was at Hogfest. I was doing open models before, right? So I
have a bit of a bias towards open models. And on the GMA side, many of the folks that are building the model are not research scientists. They are research engineers that are uh building recipes for post training, building the
infrastructure, right? I mean these people don't have a PhD. They come from a master's come from a PhD. So definitely it's not a requirement to be able to build great models. >> That's it. If your goal as a person is
to be a research scientist, most likely you will want to PhD. >> Got it. Do you think every company will eventually run their own local or private models? See I have a consultancy company called ATL technologies and when
we are building projects for our clients and these are usually small to medium scale enterprises they have the first question in the very first meeting they ask is okay how secure is my data
>> uh we were dealing with some clients in Middle East and they don't want their data to go out of their boundaries of their office premises what are your thoughts on this >> I think there's a place for open models.
Uh there are a couple of places. So one is this use cases with high privacy data to leave your servers or your computers at all. Right? Uh that's
definitely one. Another one is like so AI use cases where again like the data cannot leave at all. Use cases where you may want to modify the model to be even better at your use case. Right? So we talk for medical use cases for example.
Maybe you want a model that is specialized for your specific test or specialized for your specific test or multimodel data uh for your specific medical domain. So I do think there will be quite a bit of that there are quite a
bit of a use cases where as an individual or as a startup I would want to use an open model. Again, it depends. If you want frontier intelligence, like the most intelligent uh raw intelligence, you usually go and use
like Gemini. And there are different services within cloud, Google cloud that could enable you to run even like a airup solution uh not doing any calls to to a Google cloud server. Uh but if what you want is just a capable enough model
that can run in I have here my Pixel phone from two years ago and I can run yama for the two billion parameter which is on the smaller side. I can run the model here. So there are many use use cases where you as a company may be
totally fine using a local model but there are also many where you will want to use a serverside model. >> All right. All right. Thank you very much for your insights folks. If you have any questions please post a comment
in the comment box below. Thank you for spending the time with us today. >> Thank you. Bye. >> Thank you.